When I started investing in YC, I read every Request for Startups as an indication of where the next wave of founders was heading. I read it very differently now.

We’ve checked every RFS against the batches that followed it.

Across 83 requests, the typical theme turned up in the batch at almost exactly the same rate it had before the list came out.

RFS might change who applies, but it doesn’t affect who YC ends up funding.

So, then, what role is RFS playing in YC? And why do investors continue to pay attention to it?

What the RFS data says

Counting W27, YC has now published 107 requests across 9 batches, from Summer 2024 to Winter 2027. There are 7 complete batches, totaling 83 requests. Our research compares those 83 requests against 6,341 public YC directory records, measuring how often each request’s theme appears in the batch it named, compared with the three batches before it.

The typical request’s theme shows up at 1.04× its prior rate. Across 9 ways of running the test, the median stays between 1.01× and 1.10×.

We also ran a shuffle test to see if the timing mattered. We randomly mixed which list each request belonged to 1,000 times, as if YC published its ideas at random moments. If the real timing impacted the outcomes, then the first results would be significantly different from the shuffle test. In the end, the results looked very similar.

Then we tested if there was a lag time on requests by following the S24 and W25 releases through 6 subsequent batches. Across each batch, the numbers move up and down. For the Summer 2024 request, the six batches run between 0.83× and 1.22×. For Winter 2025, they span from 0.77× to 1.20×.

The peaks shown are no more significant than the swings before each list was published. They also shrink or disappear when we look only at a company’s one-line descriptions.

By the sixth batch after the request, both groups are sitting lower than when they started.

So the RFS doesn’t shape the next batch in any meaningful way, nor is there a strong indication of a delayed influence on later batches.

The list steers founders

YC describes the RFS as a list of “ideas that should be made real”, following on from conversations among partners about where founders are doing well, and where the gaps are. It’s intended as an inspiration for founders who haven’t started building, or encouragement for those who might already be building in these areas, to apply to YC.

In that sense, the RFS does serve its purpose, because one thing it does do is change who applies. In February 2024, YC ran early interviews for the summer batch after many people had said the new RFS encouraged them to apply.

In 2014, Ginkgo Bioworks applied after reading the biotech RFS, as an established company with years of technology behind it. Similarly, Pachama and Prometheus applied to and joined W19 after YC called for more carbon-removal startups in October 2018. So for companies outside the mainstream software remit, RFS acts as a permission to apply, knowing there is interest and funding.

The list follows founders

While YC writes the list to steer founders, it also follows previous successful ones. Both these things can be true at the same time, because the list is a reflection of what YC founders have already shown to work.

The S26 hardware request named two existing companies, Hlabs and Prototyping.io, as examples. In the most recent W27 request, Gustaf Alströmer specifically mentioned the companies they had backed selling AI to the government, some currently showing staggering growth, and asking for more.

There’s also a call for startups building cheap access to space, which could be follow-on from the hardware cluster of S26, and the success of Ethos Space Resources and Hop Aero. It won’t necessarily always be such a linear cycle, but it is happening.

Similarly, the RFS can sharpen some themes already present, and the sharpening tells you where the bottleneck has moved to. Diana Hu names a specific bottleneck in W27, asking for domestic rare-earth magnets after YC backed Mass Magnetics, which makes them from recycled material.

This is YC building on its own advantage: learning from its best founders, sharing it openly, and recruiting more of them. It’s a compounding feedback loop.

Of course, if YC is asking for more of what it’s already funding, then there’s a limit to how much the RFS might push these themes up. But naming them shows there’s an interest, which could keep the prevalence more or less consistent.

Everyone has the same list

But that’s where it becomes problematic, because once the RFS is published, it’s public knowledge. Specter routinely releases a list of emerging startups that map to each new RFS release. Regardless of whether those companies go on to get funded, the RFS and the trickle-down effect shape where the industry’s attention goes.

A list that everyone has gives no one an edge and has everyone looking in the same places. The better approach is to read the list as a compass, indicating a direction.

Ultimately, YC will always select the best founders, in whichever sector they sit. More founders from a specific sector might apply, which could generate a small uptick: more applications increase the likelihood of those sector applicants being successful, but it doesn’t guarantee it.

The RFS is also better at point than at timing. The only themes that lined up more strongly after their first request were defense, robotics, and US manufacturing. Those sectors are the slowest, most physical categories, and have always played out on longer timescales, so it makes sense.

Read the list, back the founders

So the list shifts the attention of founders and investors alike, but naming a trend doesn’t create it. At the end of the day, the fundamentals of venture and company-building still prevail: who the best founders are, where they sit, and what they’re building.

Now when I read each RFS, I look at it as a prompt to look wider, past the fashionable sectors, and where success has already been proven, with more room to build. So that means that Lobster Capital’s monitoring system, Antenna, is tracking growth, hiring, and commercial traction across every company in every sector in the batch. When it finds an outlier company, I give it the attention it deserves, and adjust the checklist for the sector.

For software, a company has to clear a threshold, and then we price the growth slope. For hardware and deep tech, where revenue doesn’t always arrive on the YC clock, we look for other signals: a completed proof event, a credible counterparty, an agreed contract.

But let’s remember, by the time a theme makes it to RFS, the best founders have probably already started building it the cohort before. That’s when you want to have met them, and it’s where we’re doing the work.

Visit www.lobstercap.com for more.

Is AI fatigue a thing? If you’ve read as many pitch decks this year as I have, you already know the answer.

So we’re taking a brief detour in the world of GLP-1s.

They aren’t new. One in eight American adults is already on one, and the market is now bigger than AI, at least in revenue:

But more recent regulatory changes mean the old model that telehealth apps profited from no longer works, so a new model is needed. Enter RonanRX. Lobster Capital’s Fund II invested in RonanRX in August.

The FDA broke the model

GLP-1 companies surfaced in 2022 when branded semaglutide and tirzepatide ran short due to mass demand. Pharmacies were allowed to produce copies of the drugs, and telehealth platforms built out a business model on top: sourcing patients, arranging the prescription, selling the drugs at their price.

It’s the trusted model that has generated steady revenue across businesses for years: own the supply chain that feeds the demand. Across one dataset of 52 private obesity companies, delivery businesses have raised around $2.6B, against $1.7B for companies developing the new drugs.

The FDA declared the shortage over by February 2025, yet many companies kept selling their drug copies anyway. In 2026, that no longer works. The FDA sent at least 55 warning letters to telehealth companies this year and proposed barring large outsourcing facilities from producing the drugs in bulk. Without cheap copies, a platform has to sell the branded drug, which it buys from the manufacturer at their price.

Manufacturers are also now competing for patients. Lilly sells Zepbound direct from $299 a month and its new pill, Foundayo, from $149. Compare that to the government’s price of $245 a month with a $50 copay under Medicare for obesity.

So the middle ground has lost its value. The Hims & Hers outcome shows by how much: a gross margin fall from 76% to 64% following Novo Nordisk suing the company for patent infringement of its Ozempic drug. Hims agreed to sell Novo’s branded drugs at Novo’s prices instead of its own versions, keeping its patients but losing the price.

Enter RonanRX

Clearly, the original model isn’t going to work anymore, so the revenue lever needs to be found elsewhere. We believe RonanRX has found it.

Lloyd Armburst is a two-time YC founder who previously built a U.S. medical-manufacturing business to $50M in revenue. When he chose what to build next, he picked pharmaceuticals partly because a frontier model couldn’t replicate it. His co-founder, Emily Colman, spent a decade in pharma at Eli Lilly and AbbVie. Their background in the field is more important than it might have been if they were building something else because RonanRX has gone straight to the source and is building the drug, rather than a storefront.

RonanRX co-manufactures the product and is bringing production in-house. They then sell directly to patients, which cuts the costs between each layer and means they can offer patients a better price while still keeping a healthy margin themselves. Where Hims is reselling a drug that someone else prices, RonanRX is pricing its own. Lloyd says that patients are typically paying three to ten times less than they would elsewhere.

Personalization

GLP-1s come in a handful of fixed doses, and patients don’t respond uniformly. RonanRX has made the treatment personal, with labs, wearable data, and targeted dosing plans. A physician can adjust the treatment plan according to how a patient responds, and the pharmacist releases a prescription for that specific person’s needs.

Plenty of companies now claim “personalization” as a means to stay on the right side of the FDA. RonanRX ensures it is remaining legally compliant by tying every dose to the individual patient’s recorded data and the physician’s decision.

Owning the loop

The intake, prescribing, fulfillment, and refill processes all run through one system, and RonanRX owns the system. If AI has taught us anything, it’s the importance of proprietary data.

Every dose change and patient outcome is documented and stays with the company. As the patient base grows, it has become an extensive record of how real people are responding to non-standard doses. RonanRX can build this database because it is controlling the full journey, from initial dosage, to adjustment, monitoring and outcome. Resellers are unable to do this because they don’t set the doses. As the GLP-1 market becomes ever larger, this is becoming a very valuable database with insights that both the private and the public market may one day want.

Secondly, in having proprietary ownership of every step, RonanRX is well-placed to respond as the regulation continues to be honed out. The rules on GLP-1s are still in flux: 55 warning letters this year, a new Medicare price, proposed bans on bulk production… As a company that controls everything, they can quickly adapt if new regulations demand it, be that in dosage, documentation, or manufacturing standards. The price of not owning those steps has already been shown with Hims.

Trust

In April, the FDA sent warning letters to seven online sellers marketing GLP-1 copies as “research use only”. The customers those sellers were serving haven’t disappeared, and now they’re looking for a legal supplier they can trust.

RonanRX meets the need: made in the USA, prescribed by a doctor, and with transparent manufacturing.

The Growth is the Signal

Five days before Demo Day, and eight weeks after taking its first patients, the founder reported $500K in ARR, growing roughly 50% week over week. Every new customer brings in 1.7 more, and that’s before the company has spent anything on paid acquisition.

At that rate, a business more than doubles every two weeks. The growth was the signal, and it was even stronger because of where the growth came from. The demand was spreading without any paid help, through word of mouth alone. In the GLP-1 market, that is much harder than you might think.

People talk about drugs and healthcare a lot. Around 30% of American who know about GLP-1s heard about them from a family member, a friend, or other acquaintance. Yet users are private about their treatment. A 2026 survey showed that 69% of current and formers users said they were unlikely to tell anyone.

So people talk about the drug, but they’re not saying what they’re taking or who they’re getting it from. A patient who names RonanRX outright to a friend is therefore one of the strongest trust signals you can get in the drug market right now. It’s trust, earned on price, the personal approach of the treatment, and how the patient felt.

Organic growth also has a cost advantage. Hims spent $262M on marketing last quarter, 34% of its revenue, and $229.5M of that went on acquiring customers. For RonanRX, one patient brings in the next, which carries no cost and makes the savings that make its lower price possible.

We watched the growth start and escalate, completely naturally, and we priced for it accordingly. Now we’re watching to see if personalization stays substantive as volume grows, and whether manufacturing stays in-house, and clinical guardrails are upheld.

GLP-1s aren’t bigger than AI yet, but with one in eight American adults now using them, and companies like RonanRX coming up, they could be. So watch this space. That’s certainly what I’m doing.

Visit lobstercap.com for more

Two new AI agents want to run your errands. One is a startup valued at $10 billion, and the other sits inside WhatsApp. Both run into the same math problem when users expect the work to be free.

On September 28, Instinct announced a $1 billion Series C at a $10 billion valuation from Sequoia Capital, Benchmark, and Coatue. Just 33 days earlier, on August 26, its Series B priced the company at $2.5 billion. That is a massive pile of venture capital behind an assistant that still charges users nothing.

Three weeks before Instinct’s raise, on September 8, Meta launched Muse. Within two weeks, it became the most downloaded free app on the US App Store. It is free for everyday tasks, with $20 and $100 monthly plans in testing for heavy users, and it works directly inside WhatsApp.

Instinct is still in invite-only early access while Muse is already live across the US, yet they are fighting for the exact same job. You text the agent a task, and it handles the websites, phone calls, forms, and follow-ups you want to avoid.

Everyone wants that. Giving it away for free is an expensive trap.

The product is the finished task

Instinct’s interface is plain on purpose. You text, call, or email it, and it uses its own virtual phone and computer to order groceries, book flights, or cancel subscriptions. As founder Noah Shinn put it on X when opening the beta:

When I tested it on booking a nice reservation in SF for a Friday night dinner in SF, the appeal was obvious. You ask for a finished errand instead of bouncing between five tabs yourself. In early September, Instinct even gave each user’s agent its own email address and 1Password integration so it can create accounts and manage booking confirmations on its own.

Meta pitches Muse the same way. Powered by Meta’s Muse Spark model, it browses the web, fills out forms, books travel, and keeps running in the background after you close the app, pausing to ask permission before sending an email or charging a single-use virtual card via Stripe Link.

Chatbots spit out an itinerary you still have to book yourself. Agents only matter if they finish the boring parts without making a mistake.

A clean demo proves very little. I want to see what happens when a flight gets canceled, a restaurant refuses online reservations, or a merchant blocks the bot entirely. On September 20, Amazon began blocking Meta’s Muse, displaying an error that an unauthorized AI agent was violating its conditions of use because Muse failed to identify itself and allegedly stored customer credentials.

Twenty-four hours later, Shopify CEO Tobi Lütke moved in the exact opposite direction:

Shopify shares jumped 7.3% on the news, and Meta Chief AI Officer Alexandr Wang joked on X that new users just needed to “catch up on the multi-year history of shopify vs amazon.” Still, for a regular user, if you have to remember which storefronts block your agent and babysit every checkout failure, doing the errand yourself is faster.

Free gets people through the door

Every completed errand burns serious compute. The agent has to plan steps, spin up an isolated cloud browser, place voice calls, and loop when a checkout page fails. Unlike traditional software, where an extra click costs almost nothing, your most loyal power user is also the one bleeding your gross margins dry.

On Patrick O’Shaughnessy’s Invest Like the Best podcast on September 28, Shinn admitted that Instinct’s five-invite cap per user comes down to hardware constraints rather than marketing hype. With usage growing 10% to 11% day-over-day on zero marketing spend and beta invites reselling on eBay for around $300, adding 10x more users overnight would overwhelm their compute capacity.

Instinct raised $1 billion to buy time for product and distribution without disclosing a dollar of revenue. Raising a billion dollars buys runway, but it doesn’t solve unit economics.

Meta is playing a different game. It dropped Muse straight into WhatsApp, an app billions of people already open every day, and paired the free tier with two paid plans listed in Meta’s Help Center: Power at $20 a month (500 million Muse tokens per week) and Max at $100 a month (3 billion Muse tokens per week). Meta still burns compute on every task, but its advertising engine lets it subsidize those losses far longer than a standalone startup can.

For users, free is an easy choice while you are just testing both tools. Once an agent holds your calendar, inbox, seat preferences, and saved cards, the switching costs lock in. Moving to a competitor means teaching a blank system how you live all over again.

Trust is harder than access

WhatsApp gives Meta a massive distribution head start, whereas Instinct has to convince every user to try a new service and link their accounts from scratch.

Distribution inside WhatsApp gets users to try Muse once, but it doesn’t automatically make them trust Meta with their credit card. An agent that books travel and replies to messages requires deep permissions across your personal life. Meta built isolated cloud virtual machines and a secondary “Sentinel” oversight agent to approve data leaving the sandbox, though Wired noted that user interactions still train Meta’s models by default unless you opt out.

When an agent has real-world permissions, small glitches turn into real-world problems fast:

• The Marketplace Address Leak: In late September, YouTube tech reviewer Matt J. Robb let Muse handle a Facebook Marketplace listing for a computer and selected “Allow Always.” As The Guardian reported on September 28, Muse handed out Robb’s home address to a buyer and scheduled a pickup without his knowledge, sending a stranger to his front door.

• The Hallucinated Breach: A week earlier, an Instinct user (@prit4k) posted screenshots on X after the assistant apologized that another person’s photo had “crossed into” their chat. Shinn had to explain publicly that no data had actually leaked; Instinct had fabricated a proper noun and convinced itself a privacy breach had happened.

The real test is whether a user trusts the agent with a second task after it botches the first one. One wrong email to a boss or one stranger showing up on your porch, and most people will revoke access immediately.

Over time, memory becomes the real moat. An agent that remembers you hate loud restaurants, fly aisle seats on United, and keep emails under three sentences is hard to replace. Shinn noted on Invest Like the Best that once a user trusts Instinct enough to link a single payment method, retention hits roughly 80%:

What the $10 billion bet assumes

Pricing Instinct at $10 billion assumes personal agents will become the default interface for everyday commerce. With Shinn stating that over $1 billion in annualized volume is already flowing through Instinct, the market is clearly huge if it works.

Still, Instinct’s revenue model has to catch up to its cap table. If Instinct charges a subscription, it has to beat a free alternative sitting inside WhatsApp. If it takes affiliate cuts or merchant commissions to stay free, users will wonder whether the agent picked the best product or the one paying the highest fee. In early October, three Instinct users reported receiving unrequested product recommendations drawn from earlier chats, highlighting how delicate a merchant-funded model will be to pull off after Shinn himself warned on Invest Like the Best against having an “alternative incentive stream” guide an agent’s behavior.

Sequoia, Benchmark, and Coatue just priced Instinct at $10 billion before anyone knows how free personal agents make money. Right now, I trust both agents to call a pharmacy or cancel a gym membership. Handing either of them my inbox and an open credit card is a bet I’m not ready to make.

Visit lobstercap.com for more.

In every batch, there are always a few companies built on something the model couldn’t quite do yet. Then a new model would ship, and the gap gets filled.

I’ve always passed on those companies. But in S26, those companies just weren’t there.

Founders have stopped racing the model. Now, they assume it wins and build around it. But just as the founders caught up, the customer started to change.

The big platforms are now building AI agents that research, book, buy, and cancel on a person’s behalf. The person doesn’t even have to open the product themselves anymore.

Which means that the user on the other end of a startup’s product is increasingly an agent.

So once again we have to change the question: if the next model can do the product, and the customer’s own agent can route around the interface, what is an investor paying for at seed?

What happened to the wrappers?

Only a few months ago I wrote about the different positions AI founders were building in. Some build on the model layer, where AI makes the product better. Others build against it, where a capable model eventually renders the product unnecessary.

AI wrappers are in the second group. Sometimes a wrapper has made it into YC because YC backs founders over ideas. In 2024, many of the strongest founders were working on gaps in the models, and the logic was simple: if OpenAI ships your feature, you pivot. A wrapper with proprietary data or deep workflow integration could still be a good business.

Backing founders before the market settles is the whole point of YC, but in a period of fast change some of those early bets can look premature (especially in hindsight). The wrapper era was one of those periods.

It seems with S26 and beyond the path is much clearer, so what’s changed?

The shrinking gap

The gap where founders were building for now exists for less than the cycle of a batch. This year, every batch has coincided with flagship releases from both OpenAI and Anthropic. In the run-up to S26 Demo Day, OpenAI shipped GPT-5.6, then GPT-6, and Anthropic shipped Claude Opus 5. So a founder might start the batch with one generation of model and then be pitching in a world using the next.

Founders learned a trick

We all caught on to the AI wrapper problem fairly quickly, and founders took note. After watching wrappers get absorbed with each new release, the best founders skate to where the puck is going and are building ahead of model capability, treating the underlying model provider as a competitor.

YC is correcting

I don’t think we’ve strayed into over-correcting yet. We’re correcting in line with the pace of the AI market, and YC partners are selecting for founders who build with awareness of the model’s future runway. So now, the model risk is effectively screened out before companies go to pitch.

Where did risk move to?

No more wrappers is great news. But while founders adjusted to the model, the platforms started building something else: an agent that acts for the customer. OpenAI has ChatGPT Pulse, Meta launched Muse in September, and Instinct, a personal agent still in beta, raised at a $10B valuation, which was 4x its markup a month earlier.

These agents sit between the product and the person, and it almost eliminates the workflow and switching cost defense from our May framework.

The switch costs grow as a product becomes more integrated into a team’s workflow. But if the human using the product is replaced by an agent, then switching costs are minimal because an agent will use any interface that does the job. If a better one appears, it also handles the switch, and the customer never feels the cost.

The workflow defensibility is gone. So how do you mitigate that risk? Increasingly, it will be the businesses that are dealing with the agent directly and where they still own the products and the terms that the agent can’t supplant.

The Change in S26

It seems that the S26 founders saw what was happening early and adapted. Across the batch, only one or two companies launched a personal assistant, while around 20 were selling directly to agents.

The founders I now want to meet will be building in three places:

1. The Agent’s Destination

This would be the licensed or regulated business at the end of the agent’s request. Radley runs an AI-native radiology practice, and RonanRX is a licensed peptide pharmacy that owns manufacturing and distribution. An AI agent can book the appointment, but it doesn’t hold the license to issue the medication or perform the procedure.

2. The Parts the Agent Needs

Graphify became the first YC company to pass 100,000 GitHub stars during its batch. It describes itself as “shared memory of your codebase, for humans and agents.” Memory is a risky layer (don’t forget Meta and Limitless), but Graphyify is open source, runs locally, and works across over 20 coding assistants, so it doesn’t rely on a single platform.

3. The Physical World

Robotics, chips, and infrastructure. Agents can run the command, but they don’t carry out the work. Whatever the agent asks for still has to be built or fixed by a physical machine, and S26 founders are betting on that gap with more hardware and deep tech companies than usual.

For LPs, the seed opportunity now lives in the companies that every agent needs and will continue to need.

What Lobster Capital is pricing now

Traction and growth rate still set the price, but we’re adding another question to the list: can a general system or the customer’s agent route around you?

There are three signs that it can’t:

• You’re the licensed or regulated endpoint

• You own data or memory the agent needs (and are not tied to one platform)

• You’re built to be called by agents in a way the platforms would rather use than copy.

What we are looking for now is the companies the customer’s agent still has to come to, or simply can’t yet touch.

Visit lobstercap.com for more.

Paul Graham recently shared a VC’s finding that YC startups are raising at valuations 79% higher than non-YC companies, despite having lower ARR.

It resparked the “YC premium” debate: if the industry is paying more for companies showing an ARR 84% lower, then are you just paying for the YC label?

But the argument comes from the assumption that pricing is based on revenue. ARR tells an investor what the customer is willing to pay. Growth rate tells us what the company is becoming, which is the outcome we are paying for.

Those are two different metrics, not to be confused. If you do, then the math starts to fall apart.

Where the Maths breaks down

The full post from investor Nichole Wischoff showing a comparison of her firm’s meetings with YC’s Summer 2026 batch against non-YC companies raising in the same window.

If revenue priced these rounds, YC companies should be worth less, not more. When revenue is used as the value input, however, the math doesn’t add up regardless of what accelerator a company is in.

Public SaaS companies trade at roughly 6x ARR, and Jason Lemkin showed that if you apply that to an early round, a startup will end up with a negative pre-money valuation.

“Pre-seed and Seed rounds can’t really be done at 6x ARR typically. A startup doing $300k in ARR is rarely valued at $1.8m post-money these days. If they raised $2m, that would be a –$200k pre-money valuation.”

A seed valuation is a bet on the next five to seven years, and the revenue a company has right now is a poor guide for what that timeline could produce.

It extends beyond the seed stage too. There are two companies in our portfolio that both do around $100M in revenue. One of them was at zero a year ago. The other has been at $100M for four years, on a healthy 30% margin, and its last round valued it at less than 1x. The revenue is the same for both, so if we priced off that, they would have the same value. Yet one is clearly moving a lot faster than the other, and this is where a different metric fits in.

How We Price: Threshold then Slope

Lobster Capital’s process has two stages, and revenue only decides the first.

Threshold
Investor intuition won’t always be reliable, including my own, which is why we rely on traction data. We built a system to track signals across every YC batch and catch the numbers that we either miss completely or misread.

I passed on RealRoots at the first look, but then our data flagged the traction and I took the meeting. Lumeria went the other way. I was impressed, but the company only had $100K in pre-orders and so we passed. At $1M, we probably would have invested.

In both cases, the threshold did its job, overruling my first instinct when the numbers were there, and holding me in check when they weren’t.

The Slope
Once a company clears the bar, price follows the growth rate. We invested in RonanRx at a $50M valuation ten days before Demo Day, where it ended up raising at $100M. The company’s revenue didn’t double in ten days, but the demand for the round pushed up the valuation because everyone could see how fast it was moving.

The reverse also holds, so if a company has cheap revenue it’s worth looking at the growth. Some companies could recover after years of going nowhere, but most don’t.

Growth is a skill

A company’s growth tends to come from two places. One is the sales machine. Crustdata recently announced it had reached $10M ARR, adding $2M in the previous 30 days. The product isn’t particularly glamorous and isn’t something people would recommend to their friends. But the sales team is relentlessly experimenting with new features for existing customers to try, which they then sell.

The second is product pull. Every RonanRx customer brings in another 1.7 customers, so the company is spending its time trying to keep up with demand rather than create it. Whichever stream is driving the growth, in both instances there is a team that knows how to produce it behind it.

Batch averages tell you little

Where the YC premium debate goes wrong is in using averages for comparison and as a substitute for a single company. The premise runs against the power law of venture.

A YC batch is a barbell, with the breakout companies at one end raising in oversubscribed rounds, and the bottom 20% at the other that never get close to even a $2-3M round. An average blends both the extremes together as a population statistic and eliminates the individual evidence for one founder or one company. Trying to underwrite a unique company based on an average is never going to be accurate, and for an LP, your returns depend on which end of the barbell you are buying from.

So no, we are not going to tell you whether the YC 79% premium is “earned” because it is a question that assumes that revenue at seed is doing something it isn’t fit to do.

What matters is whether a company clears the bar, and then whether it’s growing fast enough to justify the price. This is how Lobster Capital prices, and it’s why we’ll pay $50M for a company moving at RonanRx’s pace, and nothing for a $100M company that stopped moving.

Visit lobstercap.com for more.

I’ve sat through more YC Demo Days than I can count, and every cycle the room falls in love with a new word.

Right now it’s AI-native.

Take a look at the S26 directory, and you’ll find it on a property manager, a defense supplier, an industrial controls company, and a new chain of GP clinics. When a word is equally applied to all of those, it doesn’t really tell you anything.

One label, four businesses with nothing in common.

An AI-native bad business is still a bad business.

At Lobster Capital, we have gone deep into what to look for amongst the ‘services is the new software’ thesis, and the company that taught me what to look for is Harper.

They came into W25 building software for insurance brokers. Smart team, crowded idea. Then they tore it up, became the brokerage, and put AI in charge of the operation. We wrote the check pre-Demo Day, and were later followed Antler at Seed, then a $47M Series A led by Emergence.

In four months, Harper’s revenue per work hour went from $6 to $71.

The number that decided the check.

That number is crucial, and what I’m paying for. Did the AI move the margin, and can they do it again and again for the next hundred customers? If a founder can’t show me both, I don’t care about AI-native.

Margin: how good is your math

AI-native aside, margin decides whether business growth compounds into profit, which is why it’s one of the more important signals for investors. If the numbers are not pointing in the right direction, then a business is heading for loss. It’s even more important for AI-native companies, because for these the margin math looks a little different.

On the surface, AI companies appear to slash the cost of a human doing the work for the cost of AI doing it, and faster. But the trade only improves the margin if the AI cost is less than the labor it is replacing. Whether or not that equation pays off depends on the sector and how AI fits into the workflow, which is different for every company and product.

In insurance, applying AI for underwriting has proven to be a profitable cost-value trade. McKinsey research estimates 30-40% of underwriters’ time is spent on administrative work, such as rekeying data and manually running analyses. This is a large labor cost that is well suited to AI automation. On the value side, BCG’s recent research showed that underwriting captures around 36% of all the value AI is creating across insurance.

Convergence on one side, a false signal on the other.

So the high labor cost is being actively solved by AI. The convergence makes the margin move, and it is why the current wave of AI-native insurance startups is economically coherent as a category. Harper found the same admin burden on the broker side, which is exactly where its margins moved.

Across other sectors, the numbers are not quite as clear, which is partly due to falling token prices creating a false sense of margin growth. Per-token prices have reduced significantly, and inference costs are forecast to fall by more than 90% by 2030. Yet a single-token price isn’t a useful unit of math for margin calculations, because every company’s workflow consumes them differently.

Agentic, multi-step workflows consume between 5-30x more tokens per task than a single chatbot exchange. The token price may have dropped, but an AI agent running code carries a higher token cost than a chatbot, so a generic human-for-token swap doesn’t suffice. To accurately understand the margin, the math has to be run against the individual company, matching the token consumption for each specific task against the human cost it replaces.

Repeatability: test it once, test it twice

We were always taught in science that getting the right answer once doesn’t prove a hypothesis. Repeatability works in a similar way: the same core product needs to solve the same problem, using mostly the same process, for multiple different customers.

If the product transfers easily across varied customers, it’s repeatable. If it requires a redesign every time, then a company will face friction when it tries to scale. In venture, this means repeatability is the variable that determines if a company can grow big enough and fast enough to be one of the outliers that make the fund’s return math work. A business could be profitable but still fail the repeatability test, and an AI-native pitch can hide it because AI can look identical across customers even when the work underneath is different.

Admin repeats. Trust and local rulebooks do not.

Some sectors are more likely to run into this, and for these, going AI-native probably won’t help. Accounting is one of them, for a reason most investors are familiar with. In venture, closing a deal comes down to trust and the relationship, and accounting is the same. There are around 40,000 accounting firms in the US, but the Big Four hold 74% of the global market, secured through trust in the brand. Trust isn’t a repeatable formula an AI model can churn out. AI can make the paperwork cheaper, but it can’t make the relationship repeat, and it can’t build the relationship in the first place. We still need humans for that, at least.

Real estate has the same problem in that every city runs on different regulations and ownership norms. The model built for one city’s rulebook won’t work in another, and no amount of automation will make two regulatory regimes the same product.

That’s not to say we should be skipping over these sectors, but it is a reason to be skeptical of AI-native as a route to scaling them. AI doesn’t fix those particular bottlenecks.

Peeling back the label

When you’ve sat through as many YC batches and Demo Days as Lobster Capital has, the pattern of new tech, new hype, new market obsession becomes quickly evident. Every time a new buzzword starts circulating, it becomes less valuable and more noisy. AI-native is headed that way, if not already there.

AI-native could mean that a company has built something that will make it an outlier, or it could mean that a company is applying AI where it hasn’t added any real value. It is not an investment thesis that automatically signals returns. Under the label, the same two variables that mattered before the phrase existed determine what the business is worth.

Lobster Capital doesn’t price the AI-native label, and we also don’t discount for it. The traction filter still governs how we think about revenue, and margin and repeatability shape how we assess the company once it clears that bar. More than three-quarters of venture capital deployed this year has been concentrated on a label that the same people writing the check have publicly acknowledged is now detached from business quality.

While it’s not a reason to avoid the category, it is a reminder to any LP that marketing is designed to grab your attention, but due diligence demands that you take a moment to peel back the label and check what is going on underneath.

Visit lobstercap.com for more.

Venture is a game of risk vs returns and the solo founder used to tip the scales too far the wrong way. AI changed that.

The traditional objections were:

• Skills gap: no one person can be equally strong across product, sales, and operations.

• Concentration risk: one person carries the whole company and the pressure.

• Low bandwidth: a single founder, however talented, still only has limited hours in a day.

• Weak governance: a solo founder doesn’t have pushback or a sounding board the way co-founders do.

Then came AI, and it has effectively broken the ceiling of what a solo founder can do, which means that these risks are either no longer relevant, or look very different.

As Garry Tan commented on a recent a16z podcast, with agentic tooling “any person could be 400 of that person”, replicated across their own army of AI agents, performing any task with speed. Sam Altman also confessed to being part of a betting pool among tech CEOs for when a solo founder will reach the billion-dollar company mark. The industry seems comfortable with the fact that the solo founder is now a viable choice, and on the rise.

We are less comfortable with how to assess the risk and what that might look like at the early stage.

The skills gap is diluted

The traditional objection that a founder cannot possess all the skills required to build and execute a company is now somewhat redundant. One founder with the right agentic tooling can build a thousand markdown files to carry out just about any task, across every layer of a company. Garry has done this with GBrain, which he uses to manage the demands of running YC, working autonomously across the equivalent of tens of thousands of people’s work. Of course, there is a degree of knowledge required to understand what those agents need to do, but it can still be done by one person in front of a computer.

Medvi is the most visible example of solo founder capability with AI. Medvi reached $401M in revenue in 2025, its first year, and is projected to reach $1.8B in sales this year. The company comprises one founder, Matthew Gallagher, one other (human) employee, and a comprehensive AI stack. With AI tooling, Gallagher closed the execution-skills gap completely and scaled successfully without a team.

Medvi scaled without a team, and without a check.

However, there is a caveat. Medvi received an FDA warning over misleading marketing, allegations of AI-generated fake doctors, and a lawsuit against its fulfillment partner. Nothing in Medvi’s AI stack caught or prevented any of those issues. AI can cover the execution gap, but it doesn’t cover the judgment gap, and Medvi is now facing the consequences.

Every founder is now their own product manager, which means that while AI shrinks the technical-execution gap, judgment becomes concentrated instead of outsourced.

Garry made the same point. Judgment is simply stretched across more domains, orchestrated by a single person, which creates less coverage and more risk.

The concentration risk changed shape

The concentration risk is a concern over who is carrying the operational load of the company. The right AI stack removed the bandwidth limitation, but changed where the concentration risk now sits.

With a solo founder and an AI-native company, the entire operating memory of a company is captured in a single person’s prompts and skill files, instead of distributed across a team. As Garry noted, this new operating model requires provenance management, and “someone to keep sweeping that and making sure that things are accurate”. Garry frames this as an engineering chore, monitoring files for updates and drift, but for the investor it is the new surface layer for risk.

The old ceiling of human mental capacity forced founders to hire, and those additional team members built in a level of redundancy because if someone left, there was someone else who still knew how the business worked. AI took away the ceiling without replacing the redundancy, which means a company can now scale further while becoming more fragile with a single point of failure. The concentration risk hasn’t gone away. It now looks different, but is still reliant on one person’s judgment, running continuously.

Speed without governance

Bandwidth used to cut against solo founders because a team can move faster than one person. Again, AI solved that one. More companies than ever before are now hitting $10M ARR in three months, with rapid scaling that was unheard of a couple of years ago. However, speed without governance at an equal pace is not the golden goose.

A team provides a built-in debate and a shared decision-making framework, which works as a proxy for governance quality for investors. A solo founder doesn’t have someone at the table with an equal stake to push back and question where that scaling is headed. AI agents don’t measure up as a substitute for this, and the evidence suggests it’s more likely the opposite.

A Stanford study found that AI models affirm user behavior around 50% more often than human respondents do, even in scenarios where the user was very clearly wrong. This is a structural feature resulting from how the models are trained, which uses reinforcement learning and rewards agreement as a better response. A co-founder will question a decision because of their own stake in the outcome; the AI agent is more likely to agree, and often in the moment where a disagreement matters the most. So the structural check that used to catch a founder’s blind spot has been taken away, and replaced with something that more often than not will tell the founder they are right. What’s more, the absence of pushback is unlikely to have a negative early-stage impact, but it will show up later on, when it is more expensive and possibly too late to navigate out of.

The underwriting position

The industry is already adapting to understand early-stage traction in the era of AI, understanding which signals remain intact, and which are now either unreliable or need a different approach. The number of solo founders in YC batches is almost double that of a few years ago, and it will likely continue to grow. Any investor looking for the top 2% needs to pay equal attention to solo founders and adapt the traditional playbook accordingly, just as we have been doing for the rest of the ecosystem.

Of the four traditional solo founder risks, AI has dissolved one, diluted one, and reshaped the other two. The bandwidth problem is no longer a limitation, the skills gap is significantly diluted, and concentration risk now wears a different costume. The concentration risk is a continuity problem that is the cost of the model, and therefore should be underwritten as one.

The fourth risk, governance, is different, and carries a more answerable assessment. The greatest role of the co-founder is as an independent source of disagreement and reasoning. Agents don’t provide that, and so we need to consider whether or not the solo founder has replaced that check with a functional substitute from elsewhere: their external network.

The check moves outside the company, and has to be earned.

YC is naturally structured to test this. The group partners, office hours, and peer founders are the institutional version of the co-founder check inside the batch, and how a founder engages with these resources becomes the signal. Do they actively pursue feedback and expert insights? Do they listen to these and act on them or adapt accordingly? The engagement factor is particularly important because it distinguishes performative input from functional input. A name on a deck is performative. The ability to change direction because of the advice of someone outside their own head (including the AI stack they created) is evidence that a founder will override their agent stack when it matters. The strength of a founder’s outside relationships therefore becomes a legible proxy for the disagreement a co-founder used to provide.

There isn’t a clear number for this signal the way there is for traction, and it won’t be visible in one pitch meeting. This is a human characteristic demonstrated through how the founder behaves inside the ecosystem, not what they say about themselves in the deck, and an investor can only truly understand the strength of the signal with proximity. The same relationships that get Lobster Capital into a deal before Demo Day are what let us watch a founder take advice and adjust, in real time. As AI keeps expanding the possibilities of what one person can build, the proximity is what lets us see the next solo billion-dollar-company founder before the rest of the market.

Visit lobstercap.com for more.

Lobster Capital’s traction filter is based on the simple thesis: that revenue is the closest thing to proof for early-stage companies. The filter has worked, batch after batch, and it’s the reason we continue to secure deals with the top 2%.

The YC Summer 26 batch might be changing things.

S26 is clustered with hardware and deep-tech companies: space infrastructure, defense robotics, custom silicon, autonomous logistics. Most already have government contracts or live field deployments, but on paper they also have close to zero revenue.

Deep tech and hardware don’t produce long-term revenue on a YC clock, which means that in this category, our traction filter has to shift.

When there’s no dollar to filter on, what do we underwrite instead?

Hardware became possible in YC

Hardware traditionally suffers in the compressed timeline of a YC batch. Three months is a tight squeeze if you’re designing, building, testing, and iterating a physical prototype, which all comes before scaling the manufacturing and the revenue.

That said, the pressure of the batch timeline does have its benefits. It forces founders to streamline testing and product iteration, and pushes them to find a contract or a committed customer almost immediately. So the constraint filters out the founders who can operate at speed and know when to move forward even if something isn’t quite perfect yet.

But for this batch to look the way it does, something bigger must have changed.

The Tooling collapsed

A hardware development loop is lengthy and repetitive: design, build, test, adapt. Simulation, foundation models, and cheaper components for building are now compressing the cycle from years to months. There are three companies clearly demonstrating this in S26.

Baud

Baud, a company building a custom AI training chip, validated its design through FPGA emulation before committing to a physical tape-out. What would have been a month-long and expensive fabrication run became one of the last steps in the process instead of the first, and so everything else sped up.

Proprio Robotics

Proprio Robotics is building autonomous maintenance robots for data centers. They skipped the wait for access to a live site and built a working AI-managed data center inside their own office, with one of their robots maintaining it around the clock. Their exact product on a microscale.

Cosmic Robotics

Cosmic Robotics is deploying the same generalized Cosmic-1 design across varied solar construction sites. The generalization is the capability, eliminating the need for bespoke rebuilds for every new environment.

Credit: Cosmic Robotics

In each case, a step of the runway that used to be slow, sequential, and expensive has become compressed or even skipped entirely.

The Capital Structure collapsed

Faster building doesn’t necessarily translate to faster adoption. More often, lengthy procurement cycles kill a three-month hardware pilot before it even starts. The difference with this batch is that hardware companies are structured as an outcome or a service, effectively skipping over the step requiring the CapEx decision that stalls customer adoption.

All three of the above companies are structured this way. YC describes Baud as building both the chip and a developer platform for training and deploying models. Baud is giving access to a capability rather than a piece of hardware the customer has to buy and own. Manifold, another robotics company, put the logic into their own pitch: a pay-per-pick model that costs less than manual labor and requires no upfront capital or infrastructure changes. Essentially, why buy a robot you can rent for a fraction of the cost, immediately?

Proof without Revenue

Once you’ve caught onto the pattern, it’s visible across a bigger slice of the batch, and the range of companies it covers is impressive.

Hop Aero is building an autonomous cargo rocket that delivers 250kg payloads up to 750km in 15 minutes, launched from a standard shipping container. They have already completed a hot-fire test of their proprietary engine and been awarded a $1.25M contract from the US Air Force.

Credits: Hop Aero’s launch video YC S26

Ethos Space Resources is planning to extract silicon from the Moon to manufacture solar panels and compute infrastructure in space. NASA put a contract on the table, and there’s a mission target in the calendar for 2027.

Edgerun builds a sub-10lb powered military exoskeleton, cutting the metabolic effect of moving under combat load by 30-45%. It’s already been field-tested by Ukrainian infantry units.

Vernius Systems builds autonomous radar guidance that turns any drone into an interceptor. A YC partner noted that the founders spent the batch far from SF, building and deploying alongside their users.

Put those alongside Cosmic, Proprio, and Baud, and there are companies across space, defense, energy infrastructure, logistics, and semiconductors. They have all been verified in some way. It just has not been through ARR.

Speed of shipping

In hardware and deep tech, where the physical product is paramount, speed of shipping is a key signal. The idea is there, and we are underwriting the team’s ability to execute it at speed and scale. Any company successfully shipping several iterations and landing a contract is demonstrating that at the early stage.

A few companies in this batch have tapped into AI tooling to make this happen. Standard Machines is training AI models to design chips end to end, effectively compressing a design loop that normally takes specialist engineers months. The framework is a step beyond the simulation and foundation models compressing the loop for human teams. AI is running the loop itself.

Adapting our filter

So if it’s not revenue, what do we underwrite instead?

Other signals exist, and if we are precise about what they look like, then the filter will still hold.

Looking at the above companies, there are three distinct thresholds we can apply:

1. A credible institutional counterparty. A government contract or a defense manufacturer is a fair stand-in for a paid customer pilot. It’s a demonstration of market demand.

2. A completed physical proof event. A field deployment or hot-fire test can be exchanged for a working product’s usage data.

3. The deal structure. A deal that doesn’t require the company to prove revenue before it proves deployment.

The last one ties back to the capital-structure point, and it is a necessary adaptation as hardware founders increasingly adopt the outcome/service model.

None of these are a hard number, but they don’t need to be. Each is still a high threshold filter that only a company building a product with market demand, proven product, and potential could secure. All of which would be the engine behind ARR in companies outside this sector.

We already made similar adaptations to the filter for AI-native companies in terms of ARR, compressed runways, and inflated valuations. If there’s anything to learn here, it’s that the market is rapidly evolving. YC S26 is a batch that is proving hardware and deep tech can move at YC’s pace, where previously it fell away. If you’re not watching and evolving the strategy along with it, then you will soon be left behind as well.

Visit lobstercap.com for more.

A recent M&A analysis noted the traditional playbook of buying 100% of a target’s equity is no longer the only way that acquisitions are happening, especially with AI companies.

While this is not new, it is a rapidly accelerating trend.

The $60B Cursor acquisition is the biggest example so far. Instead of a payout, Cursor’s investors gained shares in a company that had gone public four days earlier, at a valuation topping $2T. Thrive Capital, which held positions in both Cursor and SpaceX going in, now has an estimated combined stake of over $10B, while a16z’s 10% of Cursor converted into an estimated $6B hold.

Overnight, a startup’s cap table turned into a slice of a trillion-dollar company.

So what happens when the same mechanism shows up at YC scale? It’s already happening, and it raises questions about how we evaluate those companies for the potential return profile of our fund.

The New Deal Structure at YC Scale

In July, MoonPay bought Glide, a four-person YC crypto startup processing $100M in annualized volume. Then Harvey, the $11B legal AI company, acquired Benchmark, a YC decision-infrastructure platform used by firms managing over $2T in assets.

The Glide deal was the same all-equity stock swap as Cursor/SpaceX. The Benchmark financial details are undisclosed, but with the team staying on it is likely to also be equity-based. These are just two of 34 YC companies acquired in H1. From inside the ecosystem, the pattern is clear. It’s a deal structure I am now calling “side equity”, and I expect it will keep showing up.

The Side Equity Structure

“Side equity”, as I am using it, refers to what a fund ends up holding when a portfolio company is acquired for stock rather than cash. The investor gains an ongoing position in whatever the acquirer’s stock does next, rather than a fixed value payout.Holding side equity is passive, the same as with public shares. The fund has no influence over the acquirer’s strategy or the future of a company, and so upside is a result of how the company - the “stock” - performs.

There are two different ways a fund can gain side equity, and they produce opposite outcomes for the fund’s return profile.

Pattern A: the roll-up

The company is rolled into the acquirer and keeps operating inside it. A share in the acquired company becomes live equity, and shareholders are exposed to wherever the acquiring company goes from there. The MoonPay/Glide acquisition is a prime example, and we would assume that the Harvey/Benchmark deal is too. Cursor/SpaceX is the potential.

Pattern B: the talent acquihire

The team and the product are absorbed into the acquirer, and investors get some capital back in a return event, but no ongoing upside exposure. Anthropic’s acquisition of Mendral falls under this category. The Mendral team has been absorbed into Anthropic to work on the Claude platform, not Mendral, and the company has wound down its own product.

Pattern B sub-category: the license-and-hire

The license-and-hire deals seen with Microsoft/Inflection, Amazon/Adept, and Google/Character.AI, are also Pattern B, although legally not classified as an “acquisition”, which is deliberate.

The bigger company “licenses the technology” and “hires the team”, which gets the same outcome (control of the tech, absorption of the talent) without filing for merger clearance or attracting the same level of FTC/DOJ regulatory scrutiny. The original company is left as a residual entity, wound down on purpose, and whatever the acquirer pays for the technology is used to buy out investors, no ongoing share.

The FTC has opened inquiries into these deals, but taken no firm action yet, so for big tech companies growing their AI teams, it’s the most attractive deal structure available.

A “startup acquired” headline misses the detail that matters. Is the fund walking away with a check, or are they maintaining an ongoing stake and the possibility of exponential returns to come?

Does this Matter for Lobster Capital’s Fund?

Side equity raises a question for Lobster Capital. Should a fund think twice before passing over the companies seen as too small to produce high returns? Glide is a small company, but it follows Pattern A, so the deal numbers aren’t indicative of what the return value could become. For all we know, Glide could become a Cursor re-run.

Of course, that kind of re-rating is rare, so I don’t think we need to start underwriting for Cursor-scale outcomes across the portfolio. But, even a small Pattern A exit carries optionality a cash acquisition or a Series B markup doesn’t because the position keeps moving after the deal closes.

There are also other considerations for LPs.

Fund return timeline

LPs like to know when the returns will arrive before they invest, but with side equity it’s harder to answer. When a deal closes, the position is marked at the share price, which lifts TVPI immediately. DPI doesn’t move until the fund sells the stock and distributes the cash, and that decision doesn’t have a date and the fund still has to choose when to convert that into a distribution. The public, sellable position is more attractive than a private, illiquid one, but an LP may also have to become comfortable with the uncertainty of the exact return date.

Access-thesis connection

Spotting the Pattern A deals before the market notices relies on the same strategy that underpins Lobster Capital: position and access. If you’re on the outside, you likely won’t be able to differentiate between which small acquisition is a roll-up, and which is a quiet wind-down. It is the Index vs Picker dynamic we spoke about previously.

Pro rata connection

Pro rata matters here for the same position-building reason as above. When a fund holds real pro rata and above-average ownership at entry, it leverages more side equity than a fund holding a stake bought later and diluted early. Lobster Capital insists on pro-rata rights, even at the early stage, because early access translates into sustained ownership.

The future of M&A

The Cursor/SpaceX acquisition is the most visible example of what an all-stock, no cash rollups acquisition could produce for a fund. What I don’t think the industry has yet caught onto is that this structure is already a pattern, playing out on a smaller scale across early-stage companies.

How this continues to evolve across companies that are currently much smaller than Cursor, but with just as much potential, is where a much more interesting dynamic is waiting to unfold.

Who will be the next Cursor? I can’t say for sure, but if you’re sitting inside YC, it’s likely we will spot it before the rest of the industry does.

Visit lobstercap.com for more.

Being on both sides of the table and sat inside the world’s biggest startup engine, provide two things making him one of the few people in the industry with the knowledge, experience, and understanding to attempt what he and the Standard Capital co-founders Paul Buchheit and Bryan Berg are now doing…

Building a YC at Series A.

I sat down with Dalton to talk about this on our recent episode:

What YC built, and how

The Y Combinator of today is a product of multiple experiments that tested conventional investment theory. Paul Graham did two things. The first was to introduce an application and an interview, followed by an investment decision within a few days. The second was to take a step back, skipping the board seat and encouraging founders to engage with each other for peer-to-peer learning. The result was one of the strongest founder networks in the ecosystem.

The network is the engine under YC. The founder relationships continue beyond Demo Day, and when a founder needs advice, they return to the peers that helped them get there, or the partners that mentored them. The dynamic extends to later-stage investment too. As Dalton commented, part of his job as a YC partner was to point founders towards Series A firms, and he openly states that he would recommend those he already knew, inside the YC network.

It’s a traceable pattern. YC companies raise a Series A at around 45%, compared to 33% for venture-backed seed companies generally, which is a meaningfully higher conversion rate, but the most telling insight is the investors writing those checks. Crunchbase’s active-investor data shows that the same small handful of firms topped the rankings for post-seed capital almost every month, largely from follow-on investment into YC companies.

This is the dynamic that is now channeling founders towards Standard Capital. The current YC partners and mentors know and trust Dalton and Paul, so when a founder asks who to approach at Series A, Standard Capital will be the first name on the table.

Building the post-YC case

Dalton watched YC prove that unconventional investment theory works. Alongside that, he also heard firsthand exactly what founders want, or don’t want, when they leave the batch. With years spent referring founders to YC-adjacent firms, Dalton, Paul, and Bryan decided instead to build their own, using the insight of YC’s model and founder needs, and their position inside the ecosystem to shape Standard Capital.

A standardized application

Anyone can apply on a quarterly cycle, with no warm-intro required. It’s the YC instinct, applied one stage later: minimize time spent fundraising and networking, and concentrate on speaking to customers and building a business.

The format also does useful filtering of an investor’s inbox that is increasingly cluttered by AI-generated decks. In an environment where pitch decks are becoming less trustworthy, a structured application requiring founders to explain their numbers in depth is now a better signal.

The peer group dynamic

Paul Buchheit’s personal contribution, which he pioneered at YC, is peer-to-peer founder learning, mirroring office hours, and no Standard Capital board seat. As Dalton explains, an investor with a board seat sits in other companies’ meetings and then repeats back what they hear, so Standard Capital cuts out the middleman and lets founders talk to each other directly.

This translates to quarterly in-person sessions with five or six other portfolio companies at a similar stage. It’s a Series A version of the YC view that the best advice for founders comes from other top founders.

Founders name their own valuation

Arguably the biggest challenge to conventional Series A investing, Standard Capital doesn’t engage in traditional negotiation. The founders propose the round size and price, and the fund is either in or it isn’t, for a minimum of 10% (up to 20% if founders want more capital and support). It removes the most corrosive part of the investor relationship of underpricing the founder, and lets the founder price themselves. They set the “bar” for acceptance, and the measure against which Standard Capital will view the founder and the company.

It’s early, but there are signs that the model is working. Now two quarterly cycles in, Standard Capital’s portfolio has grown to eight companies, including Archil and Bolto, which closed $11M and $12M respectively, and Starcloud, the fastest YC company on record to reach unicorn status. Starcloud went on to raise a $250M extension at a $2.3B valuation.

While it’s only a small sample, the same thesis of fast decisions, built on trust and a founder-first approach that paid off in YC, is already starting to play out with Standard Capital.

Has AI changed Series A?

Standard Capital is pitched as an AI-native firm. We have seen how AI adoption has changed the traditional founder and company journey, with startups now reaching higher ARR in a fraction of the time it used to take. We saw Loveable hit $20M and Bolt hit $20M within 3 months, and of course Cursor achieved $100M in rapid time.

Many of those numbers are on par with Series A-stage valuations, which raised the question for Dalton on where the round should sit in the timeline. For companies hitting bigger numbers more quickly, it could make sense to raise a Series A earlier, especially from a dilution perspective. Banking a large seed round followed by a traditional Series A a year or more later, from a traditional VC with no prior relationship, often means giving up 15-20% in a single round to investors starting from zero. Comparatively, raising a smaller seed round followed by a quicker Series A from people who already know the business keeps more of the company and skips a fundraising cycle that adds little beyond capital. Scale AI and Legora, both in batches Dalton worked with, did exactly this, and it worked well.

This is a direct parallel to how we think about our own funnel. Demo Day could become for Series A something close to what YC Roaster already is for us at seed: the earliest touch point and the groundwork for the relationship that lands the access. Being on the ground early, inside the ecosystem, before a founder is making cap table decisions, is the whole game, at either stage.

Of course, it comes with a caveat because not all companies with the numbers on paper have the company underneath. AI has increased ARR, but also changed what it means, and the strength of the product-market fit and the wider company will determine whether that Series A should come sooner or later.

More broadly, the concept also raises a question about the future of stages at all. If the gap between seed and Series A keeps compressing, at what point does “stage” become redundant? A stage-less model of venture, with capital and relationship as a continuous line rather than rounds, could look similar to what Standard Capital is moving towards. A thought worth returning to…

Whether this travels further

We’ve seen just how effective YC’s model is, and how the founders who go through it look for a similar thing once they leave. It’s a reasonable hypothesis to imagine something similar at Series A. Standard Capital is the strongest attempt of this yet, and if it works there, then the next question will be to see how far the model could travel beyond it. Who’s to say something similar can’t be built for Series B, C, or all the way to IPO?

But replicating it convincingly requires the understanding of what made YC work in the first place. That could well be the constraint that stopped other firms from building something similar, and it’s the advantage Standard Capital has: the knowledge of what worked at the seed stage and what founders are looking for at the next stage, and the strength of already sitting inside the ecosystem with the trust to carry on the legacy.

So YC proved a model. The question is, can it be replicated at the later stage? Standard Capital will be the first to discover the answer.

Visit losbtercap.com for more

Dan Gray’s recent X post reopened a conversation the ecosystem periodically circles back to:

X post/thread:

Is YC’s Startup School event hosted in a football stadium a demonstration of the vital “frontier expansion” work that builds visibility and encourages more founders, thereby reducing the scarcity constraint?

Or is it a contradiction of the selective filter that “made YC special”, as Scott Stevenson put forward:

Both sides of the argument may have weight, but the work required to reduce the constraints on progress is also more multifaceted than just the frontier expansion work. The size of the talent pool is not the only “conversion bottleneck” the ecosystem has.

Even if we continue to open YC’s doors ever wider, the impact on conversion constraint can only go so far if there is equal scarcity in what is waiting downstream, close enough, and with meaningful access to the people and knowledge when the funnel opens. If a founder leaves Demo Day without access to that kind of investor, then they are arguably just as constrained as they were before they entered the batch.

Opening the Doors Wider

The most recent Global Entrepreneurship Monitor report shared some key insights into the state of entrepreneurship in the US today. According to the report, one in five American adults is either starting or running a new “venture”, although formal business ownership has been declining for years. The US remains one of the most “entrepreneurial” countries in the world, and yet businesses are failing, which suggests that the people Gray says are missing from the funnel are out there, but they are not in the place that gives them what they need to succeed.

YC is one of the main mechanisms doing this work, by providing both the capital and the ecosystem that turns ambition into successful companies. YC is also doing the “frontier expansion” work Gray is calling for by actively displacing the credentialing signals that most investors lean on, and giving the market confidence to back people it wouldn’t otherwise trust.

So YC is expanding the talent pool, and there is value in taking YC “on the road” if it encourages more founders to apply to YC. However, simply expanding the founder pool is not the solution to improving the founder conversion rate.

Stevenson X response:

Stevenson’s point is that not everyone who wants to be a founder is necessarily cut out to see that ambition through to the end. This is exactly the role of YC’s filter. It tests the product, the company, the numbers, but it also tests the founder.

72% of founders report that entrepreneurship has affected their mental health, and the intensity of the YC program means that it both expands the frontier while filtering the talent pool to separate those with just ambition and those who have everything else required for success.

However, beyond filtering the founder pool, there is still a degree of work that contributes to the founder conversion rate that sits downstream of the access point. Getting more people into the funnel doesn’t do much good if what they need on the other side is just as scarce, and what those founders need is more than just capital.

The Other Bottleneck to Founder Conversion

This is where the greater constraint lies, and Gray alluded to it in an earlier post:

It doesn’t really matter how many founders funnel through YC if there is only a limited pool of investors on the other side that YC is genuinely enthusiastic to work with, and so converting more people into founders only closes half the gap. The other half is whether those founders land with investors who actually understand the pace, the relationships, and the trust required to be useful to them.

A peer-reviewed paper from the Journal of Finance entitled “Whom You Know Matters: Venture Capital Networks and Investment Performance,” showed that better-networked VC firms achieve significantly better fund performance. More importantly here, the portfolio companies of better-networked VCs are also significantly more likely to survive to subsequent financing and eventual exit.

It seems like an obvious fact: being backed by an investor who sits inside the right network becomes a measurable advantage for whether the company makes it to the next round. But the mechanism that makes this happen goes deeper, because what makes that network effective is the work that the investor or the firm has done when creating it.

An investor can exist inside a network by showing up as a name on a syndicate list or a logo on a cap table, but that isn’t the same as the meaningful relationships built through the unglamorous work of operating inside the ecosystem. It is the hands-on presence, not the listing, that creates value for a founder.

Understanding batch dynamics comes from sitting inside them as they happen, repeatedly. Knowing which partners are backing which themes, or who else in the batch is building something adjacent worth knowing about, is not written down somewhere that a founder can look it up. These are insights gained only through being inside the ecosystem, each cycle, and watching it play out. Further, that advantage doesn’t expire when the check is cashed in, because the relationships that gave an investor a real read on a founder’s batch are the same ones a founder will draw on eighteen months later when they need an introduction to a Series A lead or their next customer.

This is also the frame Lobster Capital has built itself around.

Our own allocation success rate moved from 90% in Fund I to 100% in Fund II, including in rounds led by Tier 1 firms, and the driving motor underneath that growth was because we spent time and effort building the knowledge that is useful, and so the relationships that were already in place before the round opened.

Our YC Roaster exists for the same reason - it puts Lobster Capital inside a founder’s process before they’ve submitted an application, a unique access point that few at Demo Day have.

These elements create an experience that turns our network position from a label into something a founder has experienced by the time they need it, and it’s the difference between an investor a founder has heard of vs one they’ve dealt with.

What this means for LPs

Dan is right to challenge the view that we have solved the challenges of improving founder discovery and access, but it is not the only bottleneck. With the increased size of YC cohorts, and the move to more batches across the year, there is meaningful improvement here. However, without the right investors, and enough of them, at the other end, that work is significantly dampened.

An LP looking to generate strong returns needs to look beyond the company and the founder, and consider how the resource they are backing to help generate that return is positioned. A fund sitting downstream, but on the outskirts, will catch some good founders, and return some good growth. The fund that is sitting downstream, with trusted and deep-rooted access, is the type of fund that attracts the top 2% of the batch, and also the one that has the resource to make potential a reality.

This is where Lobster Capital sits: deeply embedded, with the kind of structural access that turns our top 2% investment thesis from a strategy to a fact.

Visit lobstercap.com for more.

Lobster Capital recently invested in Pocket, a device that clips onto the back of your phone and turns conversations that software cannot reach into searchable notes. In seven months, Pocket grew from $1M to $100M in annualized run rate and put more than 270,000 devices into the world. That trajectory is remarkable, but the more important signal is what it reveals: voice is becoming the next computing interface.

The Software Layer Is Solved

The digital meeting capture market is becoming increasingly crowded. Granola, Fireflies, and Otter have all rolled out various solutions across transcription and summarized notes for recorded meetings. Granola, specifically, has raised $192M at a $1.5B valuation in under two years, with 250% revenue growth in a single quarter. Yet for that valuation, the company has minimal public reviews and a tiny measurable user base.

It doesn’t seem to make sense, but it reveals two things:

1. The market is significant

2. The gap is wide

Every professional conversation generates value, and Granola has proven that there is appetite for capturing it. It also left part of the market untouched. Each of these products has a fundamental limitation. They only work when you are on a screen, in a scheduled meeting, with a bot or app running. Anything that is said outside of these planned conversations is missed.

David Haber at a16z discussed this missing layer earlier in June. The highest-value context in any organization, he argues, lives in conversation, and a lot of those are “offline”, during in-person meetings, a product review, or customer calls. These are the conversations that provide the feedback that changes the roadmap. The gap that the existing software products are missing is the size of the rest of the professional world, and it is where the next hardware layer is building.

Hardware Captures what Software Misses

Pocket isn’t competing with Granola (honestly, why would you try?). Instead, it’s built for what Granola structurally cannot reach by focusing on the conversations that are happening off-screen. Software has already proven the demand, and Pocket, alongside a handful of others, is now building the hardware to capture the rest of it.

Pocket’s device clips to the back of your phone and captures offline conversations, phone calls, and in-person meetings and then produces structured notes. Oasis recently launched a smart ring for private dictation, which shares the same concept as Pocket.

If you needed further proof that this is where spatial computing is headed, look no further than Meta’s 2025 acquisition of Limitless, an AI startup that built always-on audio capture pendants. Meta is now targeting 10 million wearable device sales in H2 2026. When a $1.3T company acquires its way into a category, it is an affirmation that the category is on the move.

Pocket’s growth makes the demand for voice capture hard to ignore. The company has gone from $1M to $100M in annualized run rate in seven months and has put more than 270,000 devices into the world. DoorDash was an enterprise customer ahead of any fundraising, and Forbes also named Pocket one of the 21 most promising startups in YC’s W26 batch. Its recent $11M Series A led by Accel is further market confirmation of the transition away from screen-first interfaces toward ambient capture.

What Happens when Recording Leaves the Meeting Room?

At work, recording has become the default without most of us realizing it. We have become used to the fact that our emails, Zoom calls, and Slack messages are all logged, because the convenience of it vastly outweighs any premonitions or discomfort that we might have first felt. The move to voice capture away from the screen is simply the extension of that default into physical conversation.

Further, as companies increasingly draw AI agents into the core of their company and processes, the ability to capture what makes the company tick is even more important if we want those agents to be truly useful. Most of a company’s real culture and decision-making happens in the hallway conversations nobody is recording, rather than in Monday’s catch-up call.

Yet, recording conversations outside the boardroom still feels a little uncomfortable to most people. That apparent contradiction, and the question of where the boundary line is, is stopping a lot of investors from engaging seriously with the category. It feels like a philosophical problem: who consented? How are those recordings being used? What happens to them afterward?

In reality, these are questions that relate to product design. Founders building these types of products must understand the implications of consent and build contextual integrity into the product from day one, as this will create the trust that keeps people using them.

What This Means for Early-Stage Evaluation

Pocket, Oasis, the Meta Ray-Ban glasses, and the Limitless pendant are all different versions of companies building on the same architectural transition: voice is becoming the input layer for the next computing platform.

Oasis launched its new smart ring built for private dictation at the end of June and sold out the first batch within 24 hours. Meta’s hardware roadmap is the large-company confirmation of the same thesis. While the keyboard and screen are the current interface, they are not the permanent one, and the companies building ambient voice infrastructure now are building the plumbing for whatever comes next.

For early-stage evaluation, this raises new questions for LPs. The standard hardware diligence checklist was built for a different kind of hardware company. Applied to voice-capture AI devices, it produces the wrong answers. The signals that separated Pocket from the failures are specific, visible at entry, and replicable.

1. A clear Market gap

Does the product capture something software structurally cannot? Or is it a hardware wrapper around a software-solvable problem? Pocket captures offline conversations, filling a structural gap that Granola left open.

2. Behavioral attachment

Does the product attach to existing behavior, or demand new behavior? Pocket clips to the back of a phone. Humane asked users to abandon their phone entirely. The model isn’t asking users to change how they move through the world and instead integrates with our existing and deeply ingrained behaviors, enabling a natural adoption curve.

3. Consent architecture

Is there contextual integrity built deep into the product, or was it added as an afterthought? Given the regulatory landscape that we now live in, setting up the guardrails for who records what and with whose knowledge should be a basic due diligence task.

4. Prior shipping experience

Has the founder already survived the supply chain, firmware, and manufacturing gauntlet? In hardware, prior experience is a notable differentiator. Akshay had already shipped Omi, one of the world’s largest open-source AI wearables, before Pocket existed. He turned down the Google VR team to build in this space. Hardware is a very different sector, and prior experience should be a prerequisite.

How We Pocketed Pocket

Our path into the round began with the podcast. Akshay knew Lobster Capital through our content and the way we think about YC companies, early signals, and what separates durable momentum from a launch spike. That connection started the conversation and gave us the chance to get close to the team early.

Once we did, the signals a hardware company should be showing were all there. Pocket is led by two founders with prior experience building and shipping hardware at scale. They had nailed the two most commonly overlooked aspects of building a hardware product: ease of adoption and a specific, proven market need.

Voice capture, just like note-taking, is becoming commonplace, and Accel’s markup of Pocket confirms that this is where the spatial computing sector of hardware is heading. With a careful look at exactly who is building hardware, why they are building it, and their go-to market and adoption strategy, investors can pocket their own Pocket from the growing number of hardware companies we are likely to start seeing in this space.

Visit lobstercap.com for more.

After many years investing exclusively in YC companies, the question of “Why YC?” is nothing new to me. However, I am noticing that it is coming up more frequently right now than it has before.

We’ve made the case for why the “YC is declining” narrative doesn’t hold up already, and more than once, but this is a different point. The LP question isn’t as focused on YC criticism so much as the factors that might change Lobster Capital’s stance on “YC-only”, and not the top 2% of every good accelerator.

In a wide industry with plenty of competitors, it’s a fair question. From my view, YC has set a high standard, and nothing has come close to matching that yet, let alone clearing it.

A YC strategy shift and a new opportunity

The ecosystem has decided this is a conversation worth having, and there are a few likely reasons for this. Slow Ventures, a real institutional firm with partners like Sam Lessin on the bench, was, as of this spring, planning a mini-conference to bring several accelerators together and help founders figure out which one fits them. In the words of SV’s GP, Will Quist, YC “crushed it,” but “it’s not a universal fit for all founders.” An institutional firm building a conference around one question is a clear indication of industry thinking.

On top of that, the firms that were previous YC’s co-investors are now running alternative programs instead. For the last decade, Sequoia and a16z both backed YC, and a large number of their companies came out of Demo Day, co-investing alongside YC. However, two changes in more recent years have altered the co-investor dynamic and opportunity.

YC’s 2022 move to a standard $500K check made it more expensive to co-invest at the pre-seed and seed stage. Then, YC Continuity, the $700M fund that supported alumni companies at later stages, was discontinued under Garry Tan as part of a refocus of YC back to the early stage. Combined, those two decisions pulled YC away from the growth stage while also making early access pricier.

Sequoia and a16z responded by building their own versions: Arc and Speedrun. These two programs are not imitations of YC directly because they are offering something different and on a different scale, but they are both firms with capital to deploy across a company’s whole life, beyond the pre-seed and seed stage, thereby filling the gap YC stepped back from.

It’s worth mentioning Pear (PearX) and Greylock (Greylock Edge), who are doing versions of the same thing, because it is an indication of more firms that may be growing impatient with the system YC offers, regardless of the high success proof.

Beyond those two structural changes, YC has also undergone a shift in focus and has gone hard into AI under Garry Tan. Garry commented that 90% of the W26 batch is using AI as the primary system or engine running their product, and wider batch reports show AI-native services and AI-enhanced software as the biggest categories. It’s a focus that matches the current market, but it also means that the entry decision looks different for a founder building outside that framing, much more so than it did a few years ago. While this doesn’t touch on what makes YC valuable once you’re inside it, it may be why some founders are starting to shop around before they apply. It is also the gap some of the newer programs are pitching into.

It’s worth a brief word on Delve here too, since it’s part of the reason for the conversation. When I’m asked whether a Delve-style scandal proves YC has lost control of its own bar, my honest answer is no, they cannot control 800 startups. No one could. It’s not a defense of what happened, but it is a fact about running a program at that scale. YC’s size is both the source of its edge and a risk and criticism.

The historic field: a decade of outperformance

Before looking at the newest competitor, it is worth a look at YC’s track record against the incumbents. PitchBook’s comparative data across the major programs shows a decade of YC outperformance. YC’s unicorn creation rate runs at least 4.5-5.8% depending on cohort window, against 2.2% at Techstars, 1.8% at MassChallenge, 1.5% at 500 Global, and 0.3% at SOSV. The cohort exit value mirrors this success, with YC’s 2012 cohort alone generating $93.3B in exit value, against Techstars’ best cohort at $9.8B. Even when Techstars converged on YC’s own deal terms in an attempt to close the gap, the outcomes didn’t follow, which goes to show that the program that generates the best results is not just about the terms.

There is more than ten years’ worth of data on YC’s performance against programs of the same age, which is a strong rebuttal to any claim that YC isn’t the strongest player in the field.

The newest cohort: too early to score

HF0, Speedrun, South Park Commons, and Station F all come up in these conversations too, but for most of these it is too early to give a verdict, and they all mostly operate in a different field.

Station F is more of an incubator than an accelerator, with no fixed cohort structure or timeline towards a Demo Day equivalent. It is running a different shape than YC to begin with, which makes it hard to score it against the same bar. Speedrun is bigger and better-capitalized, but “bigger” doesn’t mean “better,” and it’s too early to say which side Speedrun will come out on.

South Park Commons also doesn’t compete on the same axis as YC. It’s built for founders who haven’t settled on an idea yet, a “-1 to 0” program rather than YC’s compressed “0 to 1” execution sprint. There’s no fixed three-month clock forcing a fundraise before a team is ready. As one alumni put it, the program does “a good job of not forcing it,” working with founders until they’re actually convicted on the pitch, which is an entirely different stage of the founder journey.

HF0 is a very different model again. Founded in 2020 by Dave Fontenot and Emily Liu, the framework is built on subtraction, not addition, designed to hone in founder focus without distraction. Additionally, HF0 refused first-time founders entirely from its batches. Every cohort is made up of people who’ve already built and sold something before. Consequently, the talent density is higher than an average YC batch, but at the same time, the talent has, in most cases, already been through YC before anyone at HF0 had seen it. That’s the difference: HF0 is where you admire them, and YC is where you back them.

Neo: real numbers and a strong competitor

Unlike the other challengers, Neo is one with real numbers attached to it, and they are numbers that are beginning to attract attention. The Wall Street Journal confirmed Neo’s Fund I returned 7.5x net, and Fund II is tracking at 6.5x net. Neo’s first institutional check was into Cursor, which SpaceX acquired in June at a $60B valuation. Similarly, Kalshi, another early Neo investment, raised at a $22B valuation.

However, Neo and YC are also different models. Neo selects just 30 Scholars a year following a process of 350 hours of interviews, and with scholars often mentored for years before they’ve even started a company. The level of interpersonal detail is closer to a relationship-first talent identification model than it is the YC approach of a systematic filter across a large applicant pool. Neo’s model is also structurally more flexible than a straight accelerator because Neo doesn’t stop at the first check the way a batch program does and also participate in later-stage, follow-on, and growth rounds too. This makes it a broader-spectrum capital provider, even though its origination model is narrower than YC’s, and also fills the same gap we mentioned earlier.

So yes, Neo is producing some strong numbers and some strong companies that are a real challenge in the playing field, but it is also not a direct competitor to YC, meaning that Neo isn’t taking away from YC’s edge in the way the numbers might have you believe at first glance.

The bar a program would have to pass

The honest answer to “why only YC” is that the standard it has set is as yet unmatched. It’s the same standard I’d apply to YC itself if I were building this filter from scratch today.

A program earns a place in our underwriting when three things are true:

1. Independently verifiable outcome data

Neo’s fund-level return multiples came from Neo itself, relayed by the Wall Street Journal as an independent confirmation, but still fundamentally a number the fund reported about its own performance. However, the Cursor acquisition is a real, closed, SEC-filed transaction, and so a stronger category of evidence than a fund multiple, however credible the source repeating it.

YC’s own evidence base is stronger still because there are two decades of vintages, and multiple batches within each, tracked independently by outside firms across enough cycles to be true distribution. YC clears this bar most completely. Neo clears more of it than it did before Cursor’s sale, but the fund-level numbers alone still don’t, yet.

2. Enough throughput to run a systematic filter

Our entire model depends on comparing enough companies against a consistent bar (the $1.6M ARR average at entry). For that bar to mean something statistically, not anecdotally, there needs to be repeated results, not just the occasional outlier. A ten-to-thirty-company cohort can produce a great company, but it can’t generate the volume our filter is built to run against, and so the size of the selection pool also matters.

3. The pre-decision access

Lobster Capital’s edge isn’t a filter run from the outside, but the filter we have built from a structured approach to visibility and early access. YC Roaster, the Lobster Capital YouTube channel, our Substack - all of it is a compounding system we’ve been building steadily inside YC for years that now works to give Lobster Capital the access that wins the top 2%. When an LP asks why we haven’t yet branched out, I think they are more asking whether expansion into something like Neo or HF0 is possible on the same scale we have with YC.

Branching out to a new program only makes sense somewhere we could plausibly build that same access to gain the same level of access and results. For Neo, that access would be structurally difficult to replicate because Neo’s is a multi-year, hand-built relationship with a tiny number of founders before they’ve built anything at all. That’s a different kind of access than the one we’ve spent years building inside YC’s much larger, faster-moving pipeline. Extending our filter to Neo would be starting a different business with a different skill set.

For HF0, it would be structurally easier because the relationship-dependent model is closer in spirit to how we operate. However, it’s built exclusively around repeat founders, which is a different kind of venture math of to the new founder, and not the one our traction filter is built to test. Plus, at roughly ten companies a batch, there isn’t the volume to run a systematic filter against, even if we wanted to.

So let’s consider the field against these criteria honestly: the historic peers don’t clear the first two bars at YC’s level, and haven’t in fifteen years of trying. Neo clears the first bar outright, but it doesn’t clear the second by design because its model is depth over volume, and the third is yet untested. The newest cohort hasn’t had time to clear any of them yet. YC clears them all, the third one specifically because we chose one program to focus the effort on and build the system, instead of spreading out too thin and without conviction.

I’m not necessarily saying “YC forever”, but right now there is a clear bar, and nothing has come close to beating it yet.

Visit lobstercap.com for more.

We buy traction. Lobster Capital’s entire strategy is built on this fact. We pay $25-35M for a company doing $1.7M in ARR on average because revenue at the early stage is the closest thing to proof that we can pinpoint. This filter is also the reason we’ve won the deals we’ve won. However, there is a new category of company that is threatening the reliability of our traction filter.

In the YC Summer 2026 Request For Startups, Tom Blomfield called for a “Company Brain”, where AI agents can pull knowledge from fragmented sources, structure it, and turn it into something a company can use. It would, in essence, act as “a living map of how a company works”.

Garry Tan had already open-sourced his own version of the idea, “GBrain”, and it’s now running thousands of pages and people autonomously. The current YC batch also has a company building a “brain”, Cerenovus, which aggregates every file a business produces into a single knowledge graph, mapping the connections between people, processes, and tools to surface where operations are breaking down.

This is a new category of startup that is rapidly taking hold, and it has implications for early-stage investment.

Specifically, what does a company like this mean for my traction filter? A thin AI wrapper over GPT can generate revenue fast, which means that the number I’ve always trusted might no longer be the same signal it once was. A company doing the real hard work of earning deep integration into a customer’s operations might grow more slowly on paper while building something far more defensible. Both show the same ARR line, but the number alone won’t tell me which company found a shortcut to the number and which earned it.

I’ve spent years learning to spot real traction versus manufactured traction, but the “company brain” companies are throwing out a different problem, where the traction is real, but the durability can’t yet be seen, and it also looks very different.

Where the Numbers Miss the Full Picture

Certain types of knowledge resist being written down, known as tacit knowledge. Michael Polanyi showed that most companies, at their core, are built from multiple different versions of tacit knowledge in a compilation of tribal memory, accumulated judgment, and the specific expertise of specific people. In short, a part of what makes a company a company is inside the person, not the business documents.

Or, at least, it was before AI. AI has begun to breach that barrier, not by being told tacit knowledge, but because we are now using it to gather up these knowledge sources and make business calls by inferring behavioral traces at scale.

Eve Bouffard runs YC’s design team, and over the past few months she’s rebuilt how she works. Now, every meeting transcription, mood board, and brief is collected into a single file she calls “soul.md”, which is then handed to an AI agent that generates a dozen finished design directions in one go. She called it “her own version of AGI”, a term the industry is now using to describe a theoretical form of AI that can gather, understand, and use the tacit knowledge that was previously offline. This is a prime example of the “company brain” model at work.

The efficiency and power of the AGI model is impressive. For a specific event project, Eve Bouffard fed her soul.md the meeting notes, written manifesto, and the date of the launch party, and the agent turned out a full project outline, plus an animated barcode tying back to the physical object and an interactive map of the city that nobody had asked for or even mentioned. The agent had absorbed the manifesto’s tone and the project’s specifics closely enough to invent something nobody in the room had thought of themselves…

This is the exact point of a file like soul.md and the idea behind a company brain: to absorb the full context of a company and think as if it were another employee. It works precisely because the agent has access to the exact kind of knowledge a company’s real expertise is made of.

However, Eve’s model has an additional layer above it that doesn’t get handed over to the agent. The file, transcripts, and automation are the application layer, and on top of that, Eve still makes the final creative decisions that shape the public brand herself. It’s a distinction I keep coming back to when thinking about Company Brain companies. With today’s technology, some of what a company knows, even parts of the tacit knowledge, can now be captured and consumed by AI agents. However, there will always remain some knowledge within the person who serves as the downstream judgment call of all the information the technology collects, the way Eve does.

This is where durability is found for the company brain category, because that layer cannot be faked or hidden by a thin AI wrapper. A wrapper can generate a knowledge graph, but it will struggle to make a judgment call on things that were never written down anywhere that the agent couldn’t access. If that component of depth of context and the human reasoning layer are lacking, then there’s nothing to stop a customer from switching to whichever other wrapper is cheaper that month.

So, two companies can look identical on paper with the same revenue and growth rate, but only one may have captured the fast part while the other has captured the durable part. My filter was built to price proof of demand, which doesn’t serve to tell those two companies apart.

Do we change the filter?

There are two clear options for how we approach these companies now. One is to hold the line and keep pricing ARR the way we always have and accept that some of these will now be mispriced. The other approach is to adjust the method, but only for this category.

For Company Brain companies specifically, ARR stops being sufficient on its own. What I want to know now when underwriting is whether that deep integration exists: the evidence that someone has spent real time inside a customer’s operation and built the model with the human layer intact.

The more practical test we can apply is whether what the customer has built with this company would survive an attempt to leave. This is not the same as churn, but a harder, more specific test of whether what’s been captured is rebuildable by a competitor’s agent in days, or whether it’s accumulated in a way that can’t be reconstructed on demand. For the company brain category, this is the necessary extra assessment layer.

Traditionally, our traction filter acts as a proxy for whether or not a company is hard to walk away from. For most companies, ARR is a reasonable stand-in, but not for Company Brain companies. A thin wrapper can generate the same revenue number as genuine embedded infrastructure, so the proxy breaks precisely where the filter needs it most.

ARR alone no longer clears the bar, and now it has to be paired with evidence of a human layer that a wrapper can’t fake.

A Soul still Needs a Brain

The brain needs the soul. The soul needs the brain. There’s already a soul.md inside YC, and it works because it belongs to one person and it’s as good as her judgment, not better. A Company Brain company is trying to build the same thing at the scale of an entire organization, based on the judgment of dozens of people instead of one.

Whether that holds, or whether some part of it only ever lives in a person and never makes it into the file, is the question I will now be asking before I write the check.

Visit lobstercap.com for more.

The Vintage diversification question is probably one of the most common conversations I am having with LPs at the moment, and for good reason. Vintage year is the single most important variable in VC fund performance, more so than the manager, brand, strategy, or sector. The difference a few years can make is significant.

However, many LPs are directing their concerns solely at fund managers, which is neither accurate nor fair. Vintage diversification is an LP-layer responsibility, not a fund-layer one.

A single-vintage fund, run with the right deployment discipline, can perform well across any cycle. It would be impossible for a single fund to remove vintage risk completely (this is why we have multi-vintage Fund of Funds).

My job as Lobster Capital’s GP is to run the best possible single-vintage YC vehicle. To do so, Lobster Capital implements precise steps to reduce vintage sensitivity within a single fund and acknowledge where the responsibility sits when it can’t be fully eliminated.

The Vintage Concern

The vintage diversification question has always existed in venture capital, cropping up with a renewed vivacity every few years, normally in line with how the broader market is performing at the time. In a capital-rich environment, the competition intensifies, and entry pricing is elevated. Conversely, a constrained market creates opportunities to deploy at more attractive valuations.

We saw this in the funds deployed in 2006-07 vs 2009-11, where the early vintages underperformed by 400–600 basis points1 of net IRR. Similarly, the 2019 vintage top-quartile funds are tracking 2.5-3.0x TVPI2 and 25%+ net IRR, while the 2021 vintage median sits at 0.95x TVPI and -4% net IRR.

These are from the same asset class, just two years apart. With those numbers, ignoring vintage diversification would be a serious oversight, and so LPs are right to be raising the question. However, the market effect on a vintage is also not fully within a GP’s control. Most of the funds that dropped in 2009-11 were managed by the same GPs with the same teams, and so cannot be pinned on a manager skill problem.

That said, the inability to fully control vintage risk does not mean it should go unaddressed. Lobster Capital’s structure carries three specific features that partially dampen vintage sensitivity, and being inside YC provides a fourth that generalist funds simply don’t have.

Lobster Capital’s Vintage Risk Mitigations

1. Lobster Capital’s traction filter works as a pricing anchor

The 2021 vintage is the clearest proof that entry price is the dominant variable in vintage outcomes. Funds that entered at 20x+ ARR multiples paid 3-5x more for the same companies as those entering at 7-8x in 2019, and the TVPI gap reflects that directly. By contrast, the 2022 vintage, deploying at 40-60% discounts to 2021 peaks, is already showing 20-30% early outperformance at the same stage. With disciplined entry pricing, funds can build a strong vintage even in a bad market.

Lobster Capital enters at demonstrated ARR averaging $1.7M and with entry valuations of $30M on average. These are proven numbers that anchor the entry price in real revenue, rather than a misleading surface narrative. Traction over narrative is more than a quality filter. We use it as a partial pricing hedge.

2. YC batch cadence

YC now runs multiple batches per year (4). A single Lobster Capital fund samples the YC outlier pool across multiple market moments within a single deployment period. While not vintage diversification in the traditional sense, this creates a temporal spread within a focused strategy.

The outlier company that drives the fund’s return could emerge from any of those cohorts. With exposure across all the batches within a single fund, Lobster Capital investors have a greater probability of capturing the outlier regardless of which specific market moment it enters, which means that the fund is not entirely dependent on the conditions of any single demo day.

3. YC acts as a pre-selection layer

YC has backed 20% of all companies valued over $5B since 2012. In 2025, it was the second most active unicorn investor in the world with 36 new unicorn deals. Lobster Capital’s YC focus means that we are operating within a pool that has already been filtered for founder quality at the highest level. Macro cycles still affect outcomes, but the baseline quality of the underlying companies is structurally higher than the broad seed market, which reduces the floor on downside scenarios across any given vintage.

The Limits of GP Control

The above mechanisms help to dampen the vintage effect. All fund managers should be applying a similar strategy at any given moment, because we can never be sure when the market shifts, or how much the shift will be.

2021 was a stress test for all of venture capital: global VC deployment hit $621B with valuations inflated to 20-40x ARR across the board. Even the most disciplined funds entering at the lower end of that range faced markdowns of 30-50% by 2023 as markets corrected. The median 2021 vintage fund has still not returned capital to LPs.

Our lesson from 2021 is that discipline is vital, but that it acts as a partial hedge, not a complete one. The macro cycle affects follow-on availability, exit windows, and public market comparables in ways that no entry filter fully insulates against. Good GPs absorbed that lesson and have since tightened deployment pacing, raised entry bars, and become more conservative on markup timing. These are strategies designed to manage investment around the variable cycle, not control it.

Where the Responsibility Sits

Vintage diversification has a bigger solution, but it sits at the LP layer, not the fund layer. If a fund starts to try to solve vintage risk by hedging its own entry logic and investing earlier, later, or more broadly than the thesis allows, they start drifting from the discipline that generates returns and that the LP likely chose that fund for in the first place.

The more sophisticated LP programs treat venture as a structured investment across cycles, not a point-in-time allocation. For example, commitment to Lobster Capital across Fund I, Fund II, and Fund III has vintage diversification by giving LPs exposure to multiple YC cohorts across multiple market cycles with a consistent entry discipline across each.

Within that framework, Lobster Capital’s role is to run the best possible single-vintage YC vehicle: maximum conviction, maximum discipline, maximum access to the world’s most powerful unicorn-generator. It is the LP’s job to manage macro risk at the portfolio level.

The vintage concern is valid, but often misplaced. A single vintage fund and the GP managing it can implement certain mechanisms to dampen vintage sensitivity, and certainly should. At Lobster Capital, this looks like our traction filter, YC batch cadence exposure, and the pre-selection quality of the underlying pool.

However, the 2021 cycle demonstrated that no entry discipline is a complete hedge against a macro inflection. The cycle also demonstrated that entry price discipline produces materially better outcomes. When funds enter on traction, rather than promise, they consistently show more resilience when the market turns. The rest comes down to portfolio construction, and that part belongs to the LP.

Visit lobstercap.com for more

Lobster Capital exists on a narrower premise than most venture funds. We are focused exclusively on companies that have gone through Y Combinator.

After founding Leonis in 2017, a group designed to get investors into early-stage deals through SPVs, initially focused on French startups, I repeatedly found myself asking where the very best startups were being built. By 2020, the answer was undeniably clear.

Between 2017 and 2020, Y Combinator went from funding 100 companies per batch to over 200 companies per batch, producing some of the accelerator’s most famous companies to date, including Deel in 2019 and Airbnb and DoorDash IPOs in 2020.

For an investor looking to secure the best deals with the best founders, proximity to that ecosystem was the only logical step.

From 2020, Leonis shifted focus and deployed more than $24M into YC companies, including two seed investments that have since become unicorns, and it was enough to pull me from Paris to San Francisco full-time. Lobster Capital launched in 2023 as a structured fund around the same YC-only conviction.

That specificity is also what invites the most common question Lobster Capital hears from LPs, in one form or another, on nearly every call: if the fund’s whole strategy runs through a single, well-known accelerator, what is Lobster Capital actually contributing beyond access to that accelerator?

Put more directly: are you underwriting Lobster Capital, or are you just underwriting YC?

Why the “YC Index” Can’t Be Bought

An index fund works because the underlying asset is purchasable. For example, anyone can buy the S&P 500 tomorrow, at the market price, with no relationship required. The universal access makes index funds a “beta”, based on passive exposure and available to anyone without additional skill beyond simply existing in venture capital.

YC doesn’t operate like this. With so many investors now queuing up at YC’s doors, capital alone is not enough to gain entry into the batch’s best rounds, and in oversubscribed rounds it becomes the founder’s choice for who gets in. As we discussed in our recent post on how to win YC Deals, this access requires being trusted before the round is competitive, which takes relationship-building, founder goodwill, and time, none of which can be bought.

This is where the beta framing breaks down entirely because calling it “beta” implies it is a return an investor could capture passively if they chose to. However, the only way to capture YC’s best returns is to first do the work of earning access to them, which means any return that flows from it is already, by definition, a product of work done, not a market exposure taken.

YC is the Framework for the Decision on Top

Of course, this doesn’t mean that YC isn’t doing its own work. Our analysis of YC-backed investing found that investors who backed three or more YC startups per batch between 2018 and 2020 saw top performers return 16x, with the top quartile at 8x and median investments at 5x. Even the bottom-quartile outcomes still cleared 3.3x. When you compare that to the fact that in the rest of venture capital, only the top 25% of funds achieve returns above 3x, it is clear that YC is a strong filter for the best companies.

By now, most investors are aware of the strength of YC, but this is not just a question of YC access. Lobster Capital is not investing in every company that goes through the program; rather, we are using it as the first filter for traction.

The investment decision comes from a second filter, Lobster Capital’s own traction thesis, which we apply on top of YC’s selection process. The mean entry ARR across the portfolio sits at $1.74M, and Lobster Capital pays a $25M to $35M entry valuation as a premium, rather than chasing cheaper, pre-revenue positions. We screen the batches and speak with the top 15-20%, and then identify the very top of that selection through our traction filter with a written investment memo for each decision.

This is a discipline that stands independent of YC’s own selection logic. So, YC decides who gets into the batch, and the Lobster Capital traction filter decides who, within that already-strong pool, we actually back.

The Proof in the Portfolio

The clearest test of whether that second filter does anything is to look at what happens to the companies it selects, specifically the ones Lobster Capital wanted but didn’t always get.

In the only fully seasoned picking vintage, W23, 31% of the companies on Lobster’s chase list reached breakout outcomes, and 38-46% reached a meaningful outcome. The estimated baseline for YC companies reaching Series A or beyond sits at roughly 8%, which creates roughly a 4-6x gap on the batch base rate. This is measured against Lobster Capital’s full chase list, including the deals it lost, which therefore also isolates selection skill from portfolio luck.

The entry discipline has also stayed consistent, averaging $1.74M entry ARR across Fund I. This is a measurable cost to entering at that level, and the $25M to $35M is a genuine premium over what less disciplined capital pays for a far weaker signal.

Lobster Capital’s early markup data also tells a similar story, showing that 14 of the 41 memo-backed companies in Fund I (34% of the portfolio) have already marked up, at an average of 2.4x. It’s early and the fund is still young, but it’s a live signal that the subset the traction filter identifies is already outperforming the batch it pulls from.

The additional distinction that separates Lobster Capital treats “access” differently from other funds now pitching “YC access” as their entire thesis.

Access on its own is just proximity, and proximity is the beta claim wearing a different name. Anyone reasonably well-networked can eventually get some access to YC, but that won’t secure the top 2% without the underwriting work of a targeted, ARR-anchored selection layer applied consistently on top of that access, batch after batch.

That distinction shows up concretely in how some of these relationships actually start. Lobster Capital’s relationship with ZeroEntropy’s founder predates its batch by months, and the investment in OneText began before the company entered YC at all. The early relationships are arguably one of the main deciding factors that secured the end investment deal, and are clear evidence that the traction filter is often already running before Demo Day, and before a company starts having those formal investor conversations.

This same relationship-first approach now extends to how Lobster Capital builds visibility before founders ever reach Demo Day. A deliberate social media and content strategy (2 million views per month) keeps the fund present in the conversations founders are already having, so that by the time a round comes together, Lobster Capital is a name they already know and trust. The practical effect is simple but material: we no longer lose deals we want. Access that used to depend on a warm introduction now compounds through a steady, public presence that reaches founders earlier and more consistently than a purely network-driven approach ever could.

Underwriting Lobster Capital, or underwriting YC?

Lobster Capital is not underwriting YC as an index, but as a deliberately selected, memo-documented subset of it. On top of that base, we filter a second time by a discipline that is unrelated to whether or not a company happened to get into the batch.

YC’s first selection filter creates a very high-quality opportunity set, but Lobster Capital’s job is to act on that opportunity to convert it into our own concentrated exposure.

The evidence so far shows that it is working. Our selection is outperforming well above the batch baseline, entry discipline that hasn’t moved, and there are early markups in a young fund that already point in the right direction.

YC may be an alpha on some level, but the true “alpha” comes from access, selection, timing, pricing discipline, and pro rata ownership inside a market where passive exposure is not available.

Visit lobstercap.com for more.

There are really only three core fundamentals that underpin venture:

1. See the best opportunities before anyone else.

2. Pick the ones worth backing.

3. Win the right to invest.

We have spent most of venture’s history thinking that in order to do all three well at meaningful volume, you required a large team. The ability to see and assess more companies, analyze more markets, and generally scale venture was a problem of headcount, to be solved with more analysts, associates, or platform staff. AI has blown up that reasoning.

If AI has changed the way founders are building companies by scaling faster, on leaner models, with smaller teams doing more, then we should expect it to change the way the best investors find and back those founders. The volume problem that once required a large organization can now be solved with the right tools, structured and scaled in the right way.

The modern VC firm doesn’t need to be big. It needs to be good at three things, with the right tools underneath them.

The Enduring Problem of Scale in VC

Scale has always been a central tension in venture capital because the work that actually wins deals doesn’t scale well. Building founder relationships, engaging with LP networks, and building a strong thesis are all inherently human tasks that cannot be systematized without losing what makes it valuable. Conversely, the administrative work of a large fund can be scaled quite easily, but it creates little differentiation to what matters.

In not understanding where the differentiation is, firms try to solve the problem by hiring staff, because a higher headcount means a higher output, right? The data suggests that the converse is true.

Carta’s fund performance data consistently shows that in the upper tiers of VC returns, the smallest funds post the best results with the top decile TVPI. In the 2018 vintage, the 90th percentile TVPI for the smallest funds was 4.03x, against 1.67x for funds over $100M. Similarly, Santé Ventures analysis found smaller funds averaging 17.4% IRR against 9.7% for larger ones.

The conclusion is clear: size does not correlate with outperformance at seed.

If anything, it hinders it. Any fund, large or lean, experiences a degree of deployment pressure, but with a larger fund there is significant drag in the stages that lead to investment. A larger funds has multiple partners, all with different reads of the market and portfolio construction targets that shape which deals get prioritized. For every extra partner, there is an unavoidable additional layer of friction to the selection process.

For a lean fund, the internal approval chain has fewer potential barriers to conviction on a deal. Therefore, there is a compressed execution time from signal to decision, which is a competitive advantage in a market where the best YC companies are often fielding term sheets within days.

See: Finding Opportunities Before Anyone Else

The “See” function of VC is identifying the right opportunities at the earliest possible stage. It is possibly the most important stage because it builds the relationships that enable deals on the investor’s terms, but it has also traditionally been constrained by human bandwidth. An investor can only attend so many events and work so many rooms, especially if relying on in-person interactions.

AI eliminates the bandwidth constraint without removing the judgment that makes sourcing valuable. Outbound sourcing, which previously required significant manual effort per contact, can now be automated at scale with the correct tooling across multiple channels:

• Lead identification

• Signal extraction

• Personalized outreach

• Initial qualification

Industry data puts AI adoption for deal sourcing at 82% of VC firms, showing that the industry is beginning to take note. However, most funds are using AI as a faster search tool, when its true structural advantage lies in the ability to create and nurture relationship infrastructure by reaching more founders, earlier, with more relevance per contact.

Lobster operates this way across several layers:

• The content engine: Lobster Capital is active across YouTube, X, and Substack, LinkedIn, Podcast, generating ~2 million monthly views. The engine creates inbound at scale, so that a significant share of any given cohort has already encountered Lobster Capital before the first direct outreach.

• YC Roaster: A pre-batch sourcing funnel that uses AI to analyze YC pitches, identifying and building relationships with applicants before they have even entered the program.

In each of these layers, the AI handles the volume. The relationships that those introductions create stay personal, overseen and nurtured by me. AI is a tool we use to amplify the human, not replace the judgment underneath it.

Pick: Judgment at Scale

Deciding whether a company is worth backing has historically been the function most resistant to systematization. Judgment is not a concept that seems like it could, or should, be compressed. However, AI has changed the conditions under which a judgment call can be made by providing faster access to a larger volume of data. We can now make judgment calls on a new scale and in a fraction of the time.

AI application has reduced initial screening time from several weeks to days, through market mapping, competitive landscape analysis, and traction benchmarking at speed, consistency, and greater analytical depth. The framework is improving the number of deals evaluated without adding headcount.

Lobster’s traction-first thesis of a $1.6M ARR average is a prime example of smaller teams operating at scale. AI-assisted market monitoring and competitive analysis make the thesis executable across a high volume of companies that a small team would not be able to sustain manually. This is a significant shift for lean teams who can now scale credible assessments with confidence.

A YC-specialist fund can also mold the Pick framework to closely follow the early traction signals that predict outlier performance at the YC level. Lobster Capital’s previous piece on AI coding ability as a due diligence signal is a concrete example of a tailored Pick framework in practice: identifying a specific, evolving assessment criterion that is native to the current founder cohort and invisible to a generalist fund running standard workflows.

With the right AI stack, a lean team can now assess more companies against specific metrics with greater rigor and precision than a larger team running manual processes.

Win: The Right to Invest

Winning is a downstream result of See and Pick being executed well, but it also carries its own structural nuance in terms of scalability. In the current YC market, the ability to win the best allocations relies on relational equity, not just financial. The funds coming out on top are already trusted by the founders before the competitive process begins.

In See and Pick, AI provides direct leverage for scale. At Win, AI is an indirect enabler. The time and bandwidth recovered from automating sourcing and screening can be reinvested into the work that requires the human: founder conversations, early-stage support, and being visible consistently across the batch cycle. In other words, build systems that scale so you can keep doing the things that don’t scale.

Winning deals is one of those things. It requires genuine trust, built over time, person by person. AI creates the space for that work to happen, but it cannot do the work itself.

What about the portfolio?

The traditional justification for large VC platform teams is normally portfolio support: a team for building a talent network, marketing resources, and Demo Day prep work. I would argue that the support that truly makes a difference to founders is specific, personal, and time-sensitive.

Founders need a reliable fallback at the right moment. They need the honest view on the term sheet, an alternate read on a hiring decision, or a warm intro to exactly the right person at a difficult moment.

A lean fund that has built authentic founder relationships is always going to be able to deliver that better than a 60-person firm outsourcing every founder interaction support to a separate team.

This is where the brand argument comes full circle. No content engine accelerates a referral if the founder doesn’t feel the genuine value that makes them tell the next founder you are the right investor to have at their table. Everything Lobster Capital publishes is downstream of doing the work first.

Different Structure, Better Results

A lean team with AI infrastructure executing See, Pick, Win at depth is a very different investment vehicle than the traditional VC fund.

The funds that scaled to combat scale with headcount are now carrying cost structures that are forcing their hand to less selective deployment at compressed valuations. Conversely, lean teams are operating with lower fee drag, higher GP alignment, and the flexibility to remain disciplined at seed-stage YC pricing.

Lean is not the absence of capability. It is the presence of the right capabilities, scaled where it makes sense, and concentrated where they matter.

Visit lobstercap.com for more.

In 1976, a panel of French judges blind-tasted a lineup of wines and ranked a California bottle above the best of Bordeaux.

The French judges were furious, the wine wasn’t poor quality, but their centuries long system for pricing quality assumed geography was destiny. A wine’s worth was set by where it came from, what appellation, what stage of a legacy-hierarchy it occupied. The new precedent of blind tasting broke the frame. Once you couldn’t see the label, the old categories stopped predicting anything.

Venture has its own version of the label, and it’s called stage.

Seed, Series A, growth, each one a fixed idea about what a company is worth and what it’s allowed to become. And right now, the same way those judges did, the industry is staring at a result that its categories can’t explain.

Every quarter, the same complaint cycles through venture Twitter. YC valuations are too high. The default round is now $4M on $40M post. Founders are getting overpriced and seed investors are getting priced out.

I understand the instinct. I also think the complaint is built on a way of thinking about venture that stopped being true a few years ago.

The world that stage-thinking came from

For most of venture’s history, the industry organized itself around stage.

You are a seed firm, or a Series A firm, or a growth firm. Each stage carried its own rules: how much you owned, what valuation you paid, what return you underwrote.

Growth investors wanted a 3x with a liquidation preference protecting the downside. Seed investors wanted 15 to 20% ownership at single-digit-million valuations.

In each instance, the label was synonymous with style, brand and DD style.

That logic has largely held for the best part of two decades while VC has been mostly focused on software. Software itself was expensive to build, the first version of a product cost a fortune in engineering time, and the whole venture model was structured around financing that cost and then selling the result to as many people as possible. Stage was a proxy for how far along a company was in paying down that initial cost and proving the model.

What changed?

But now, as we all know, the cost of building software has collapsed, and with it the base assumption underneath stage-based investing falls apart. You can see it in the data coming out of YC right now.

In the Spring 2026 batch: 196 companies, 395 founders. 95% touch AI and 80% are AI-native, meaning it saying your startup “uses AI” or is “agentic” is a waste of breath.

A year ago an AI agent was a pitch. Now as B2A, MCP and open APIs become essential, have agentic capabilities is a mere expectation.

When building gets this cheap and this fast, companies hit real milestones earlier than any framework expected. Recent batches have companies reaching $1M in ARR before Demo Day and growing double digits week over week.

Which means for YC companies “pre-seed” or “seed” labels no longer really describe where a company is.

Magnitude, not stage

Mike Volpi, who spent years at Index before recently starting Hanabi, framed the shift cleanly in a recent Jack Altman interview - the job now is to ignore stage, observe the size of the opportunity, the magnitude of it.

His example was that you could enter Anthropic at a $60B valuation and still make 10x or more. Twenty years ago a $10B company was a growth asset with growth returns. Today a company can go from ideation to hundreds of billions in a few years and produce venture-like returns the whole way up.

When the magnitude of outcomes changes this dramatically, entry price no longer determines whether you make money. The thing that determines whether you make money is being in the right company at all.

This is exactly what common YC valuation complaints miss.

Rounds in the batch that hold a $40M post are compared to what “seed” “should” cost. That comparison only works if you believe outcomes are still the size they used to be. They aren’t. The companies coming out of these batches are growing faster and reaching profitability faster than any cohort in history.

If the upside reference class moves, we can expect entry prices to move with it.

Why a higher price can be the safer bet

In Fund I our average entry was $30M and the average ARR at the time we invested was $1.7M. For a seed fund, those numbers look expensive. For what we are actually buying, companies with demonstrated revenue and retention, they are rational. We over-index on early traction on purpose, and while we know venture is a long game - both I and our LPs are very happy with early results.

The clearest example from our portfolio is a company I can’t name yet. We got into the hottest AI hardware company of its batch at a $40M entry. A tier-1 fund came in on the same company at $60M. We got the better price not because we negotiated harder but because the founder already knew us and trusted us before the round opened. They finished YC at roughly $27M run rate. Two months later they were at $105M. The “high” YC price was a rounding error against the trajectory.

Stop optimizing for ownership

The other sacred cow of stage-based investing is ownership. The classic rule says you must own 15 or 20% at entry or the deal isn’t worth doing.

But when these principles hit the road they are merely that - the reality of venture is you fight to maintain or grow your position in extraordinary companies over time through pro-rata, and you let disproportionate value creation do the work.

Ownership at entry is a vanity metric. Being in the best companies, and being able to buy more as they prove themselves, is the game.

The label doesn’t matter

I almost think of Lobster Capital as an early traction investor rather than a seed fund. We come in at a higher valuation than classic seed, into companies that already have revenue, retention, and a market pulling them forward. The label “seed fund” doesn’t describe what we do, and that’s fine, because the labels stopped describing reality.

Stage is a hangover from an era when VC math was different. The investors still arguing about entry price are optimizing for a frame of reference the market already left behind - look at magnitude, pay up for genuine signal, and stop being bound by labels.

Visit lobstercap.com for more.

The best YC deals don’t get discovered at Demo Day. They are earned long before it.

At a recent Demo Day, a founder introduced herself by explaining that she had been a doctor, found Lobster Capital’s YouTube channel, learned about YC through that content, applied, and there we were meeting again on the other side. She knew me before she knew YC.

That story is becoming more and more common, as batches go on. Now, when I open a random YC founder’s profile, a large share of the time they already follow me, be that on YouTube, X, or our podcast.

Why does this matter?

Because that prior familiarity fundamentally changes the nature of the first conversation. The founder already knew me and understood what Lobster Capital is, does, and would bring to her cap table. The relationship already existed before Demo Day, and before any conversation with other investors.

In an increasingly competitive market, this relational equity is the new differentiator that enables the best funds to secure the strongest allocations in the batch.

Capital is Now a Commodity

10 years ago, capital and network were enough to gain investors the best companies on the best deal terms. But then ten years ago, there were fewer investors staking out YC Demo Day, and batches were small enough for an investor to work the room.

As of 2026, Y Combinator’s portfolio Net Asset Value exceeds $1.3T, with over 5,600 funded startups and more than 120 unicorn companies. With every batch, more investors have flocked to Demo Day to try and catch the 6.5% of the batch that go on to become unicorns. Founders now have more options than at any point in the ecosystem’s history, which means that capital alone is no longer a differentiator.

When every fund is bidding on the same elite companies with the same green dollars, the traditional notion of VC “value-add” has lost its edge. Today’s founders are looking for something more: knowledge, tools, and indication that an investor’s network will accelerate growth in their specific market. An investor’s track record still matters, but now your following, content, and reputation matter just as much, possibly even more.

As Mike Volpi recently noted: brand is now the primary way a VC summarizes its value proposition to the founder. Reputation, not check size, is what moves the best deals.

This is possibly the biggest challenge facing today’s investors. Building a brand that resonates with today’s founders and drives access to the best deals requires a new approach that most funds are yet to adopt.

Classical Brand-building Fails with this Generation

Insider insights, resources, access, and the organic signals that an investor understands their world; these are where founder priorities now sit. We are still mostly in the era of venture firms treating branding like corporate advertising with sponsored conferences and logos on lanyards, which lands flat with a 22-year-old founder whose first point of call is social media.

The conversation between founders and investors is happening on YouTube, X, and podcast feeds. Lobster Capital is active across them all, generating over two million impressions per month. We deliberately publish on all of these platforms because they reach the distinct audiences we care about: founders at the earliest stages of thinking about YC, and LPs evaluating whether this fund understands the market it claims to operate in. The breadth is intentional: different platforms reach different founder archetypes and investor profiles, and showing up consistently across the full spectrum is a core differentiator, and one of the greatest strengths of our fund today.

This means two things,

1. Publish openly and often, to scale reach.

2. Carefully curate reputation, person by person and reference by reference.

I create assets that solve problems for founders:

• The Podcast: I sit down with YC founders every week to share their real stories. We have not built this to maximize public reach, but instead to be referenced very heavily in the intro calls with YC founders.

• The YC Roaster: a free tool that helps founders refine their YC applications, using real alums to review and score the best submissions. This builds our network with both alumni and applicants even prior to getting in.

• Agents: Custom AI agents bring value across the current cohort, opening opportunities with no strings attached.

• Ecosystem: active across all social platforms, we know this is where the founders and YC partners spend their time.

By doing real favors for real founders upfront, I earn trust and build a relationship. It is not a coincidence that alongside growing the brand, our allocation success rate has moved from 80% in Fund 1 to 100% in Fund 2, including rounds led by Tier 1 firms.

Visibility was placed at the center of the deal infrastructure, and it makes the investment strategy executable.

How this Plays Out

The prior familiarity the founder at Demo Day had with me and the Lobster Capital functions as more than a conversation starter: the relationship is already warm. The Lobster Capital brand was working ahead of the YC pipeline to lay the framework for the genuine relationships that facilitate deals.

This is the logic of presence across the entire arc of a founder’s journey.

• Pre-YC: through content that reaches people before they have applied.

• During the batch: through the podcast, the network, and the day-to-day help.

• Post-YC: through the ongoing relationship.

Even the deals that don’t close in one cycle are accessible in the next, because the brand keeps compounding and the referrals keep stacking.

YC is uniquely suited to this kind of compounding because it is a layered infrastructure of founder touchpoints. Bookface, YC’s private network, includes a curated investor database with peer reviews and ratings of investors they’re considering. Founders use it to build fundraising shortlists and conduct due diligence on investors. YC’s Startup School works at the opposite end, drawing in the pre-application cohort and exposing them early to the funds and investors inside the ecosystem.

A reputation that is built inside that loop, with one good Bookface review or one founder referral, travels much faster and stays valued longer than in the open market.

Inverting the Access Thesis

The analytical work of assessing whether a company is good is relatively straightforward, but building meaningful relationships with founders requires more than a successful portfolio. What separates funds today is whether founders trust you, refer you, and choose you when they have every other option available.

Conventional investment thesis treats selectivity as a mechanism for access: build a sharp analytical framework, apply it rigorously, and let a strong track record attract the next generation of outliers. The Lobster thesis inverts this: build the access infrastructure first, and the selection framework becomes executable.

Lobster Capital over-indexes on YC companies with demonstrated early traction, averaging $1.7M ARR at entry. The fund deliberately pays a valuation premium for proven product-market fit over a cheaper entry on pre-revenue promise because at the traction levels Lobster Capital targets, the signal quality justifies it. YC’s top-tier companies typically raise at $20–25M post-money at Demo Day in oversubscribed rounds. The access to those rounds, and on those terms, is only possible because the relationship was built long before a founder goes anywhere near YC’s doors, so that by the time the company is moving through the batch, there is already a framework of meaningful access.

At Lobster Capital, the pre-application funnel, the content infrastructure, and the traction-based entry discipline are the mechanisms by which the right 6.5% are consistently reachable before anyone else in the room. It is a structured and targeted system, and it positions us exactly where we want to be.

This is the real answer to the picking versus selectivity question. If you are arriving at Demo Day cold, you are competing on price, and losing to funds willing to write uncapped notes to force their way into deals that may not be the best ones anyway. The founders who matter have already been researching, forming preferences, and in many cases are already committed before they walk into the room. By that point, the access work either exists or it doesn’t.

Visit lobstercap.com for more.

When an investor tells a founder to skip Y Combinator, they are not offering friendly advice. They are pitching for a seat at the cap table they are increasingly struggling to land.

Paul Graham made this point directly on X this month:

“All through YC’s history, investors (for obvious reasons) have tried to tell founders that YC wasn’t worth it. In 2010 they just said we sucked. Now, since it’s obvious we didn’t, they’ve had to change the claim: now it’s YC used to be great, but has declined from what it was.”

The argument has changed shape, but the incentive behind it is the same. When Airbnb, Stripe, Dropbox, DoorDash, and Coinbase made the original argument untenable, the rhetoric shifted: bigger batches, more cohorts, AI saturation.

Over 5,000 companies funded, more than $1300B in combined market value is not the track record of an institution in decline. The critics are not wrong that YC is different today, but they are wrong about what it means.

Structural change is adaptation, not decline. Investors who once had an easy path to the best cap tables are now struggling to keep up, and instead of updating their own approach, they are questioning the institution.

Era Comparison is Inconsequential

The Paul Graham era is not a fair comparison to the Garry Tan era. These are two very different iterations of the same compounding system, operating in categorically different markets.

When Paul Graham built YC, reaching product-market fit took years and access to capital was limited. In that market, providing two weeks of structured mentorship alongside a modest cheque was the right method, and it generated astonishing results.

Garry Tan inherited that model in a world where AI has compressed startup lifecycles beyond recognition. The most recent YC batches have reported revenue growth of 10-20% per week, compared to the 2-4% that once defined an exceptional cohort. Today’s founders are using tools that didn’t exist when the alumni they are being compared to went through the program. For the system to continue to work, adaptation was necessary.

Judging one era by the standards of the other is a category error, and for LPs, it is particularly dangerous. It is the equivalent of evaluating a fund manager’s 2026 portfolio decisions against the conditions of 2010, which operated in a different rate environment and a different market structure entirely. You wouldn’t accept that comparison as a basis for capital allocation, and the same logic applies here.

The “Brand” Criticism

The criticism of Tan’s public builder persona, the GStack release, and SXSW remarks about AI-native development also appeared in the thread, and are another example of critics confusing quality with other factors.

GStack’s technical novelty is a separate debate from YC cohort quality. Garry is shipping products with AI tools in real time, gaining understanding of what AI-native founding looks and feels like. In a batch that is over 90% AI-core, taking a hands-on approach and gaining insight into the tools founders are using is arguably a meaningful advantage when evaluating founders. In today’s market, we would go further: active immersion in the tools founders are using should be table stakes for any investor serious about evaluating AI-native companies.

As we explored in our Third Era post, Garry’s version of YC is a logical extension to what Paul built, in a new technological and economic context. YC is evolving from accelerator to operating system, building structural advantages that compound with every cohort rather than ending the relationship at Demo Day.

The Incentive Behind the Narrative

Some of the points cited most often do have an element of truth. The batch sizes expanded significantly from 2022, this has now reduced again, but cohorts have increased to four each year.

AI is also now the default, raising questions about differentiation at the portfolio level. These are real structural changes, but they are not the same thing as quality decline. Paul was quick to challenge this rhetoric:

“If an investor (or a “scout” working for one) tries this line on you, ask them: when, exactly, did YC decline? It can’t be when I retired as president, because we’ve funded lots of great companies since then. So exactly what shark was it that YC jumped?”

Most of the commentators couldn’t pinpoint that moment. The “decline” they are describing is neither YC’s, nor really a decline at all. With structural changes, the signal environment has also changed. Everything is bigger and faster: higher valuations, compressed runways, funding reached before Demo Day. These are challenges of navigation, not of quality, and it doesn’t mean that the talent pool has declined. Instead, the environment has weakened the traditional signals, and that has made the old map less reliable. Investors now struggling to navigate the market have found it easier to question the institution rather than update their methodology.

Venture Capital doesn’t Reward Impatience

The data that drives returns for LPs is understanding whether YC continues to concentrate the highest-quality early-stage companies at a rate no alternative can match. When pressed on specific evidence of decline in the thread, the most common response pointed to a two-year window.

A two-year horizon is not a serious timeframe for judging startup outcomes, and most of YC’s most celebrated outcomes (AirBnB, Stripe) were not obvious at that point. Building a lasting company takes time. The critics are applying a standard that YC’s own defining companies would have failed at the equivalent stage. As another voice in the thread observed, YC’s output since 2020 is still being written.

Still, even on that impossibly narrow measure, the argument is already breaking down. Legora, the YC W24 legal AI company, became the fastest Y Combinator company ever to reach unicorn status, hitting a $1.8B valuation just 13 months after its batch, and reaching $100M in ARR within 18 months of founding. It has since raised $816M in total funding at a $5.55 B valuation. A single counterexample does not prove a thesis, but it does collapse a narrative.

Darwinism at Play

YC’s enduring structural value has always been something more durable than monopoly on outcomes. The ecosystem is a highly refined concentration of top-tier founders, seen before the market can price them. There is no current alternative that combines YC’s advantages at that scale.

YC now operates at the equivalent of roughly 40 seed funds’ worth of market coverage, while simultaneously functioning as capital provider, talent network, customer ecosystem, hiring platform, and alumni investment pool, all within a single integrated system. Approximately 45% of YC companies go on to raise a Series A, well above the industry average. YC alumni invest around 20% of the dollars raised by YC companies at Demo Day. The network effects are continuously compounding, and they are able to do so because of the structural changes.

The market changed, the system evolved, and so extracting signal from that concentration now requires a different approach. The 2010 playbook will not cut the noise in a 200-company AI-cycle batch. Reading the signal requires systematic methodology, learned from continuous proximity to the ecosystem, close tracking of founder quality, and early traction indicators across the full cohort.

The debate about whether YC has declined is the wrong question, and tells you something important about the investors asking it. The market got bigger, faster, and more complex, and the signals got harder to read. The methodology had to evolve, and not everyone has done that work.

For LPs, that distinction is where the opportunity lives. The system that produced the defining companies of the last fifteen years has evolved deliberately and structurally, in direct response to the market around it. Investors who want to stay at the top of the venture capital food chain must also adapt accordingly.

Whether YC remains the most reliable concentration point for early-stage quality in the world is undeniable. The task is knowing how to read it.

Visit lobstercap.com for more.