Every AI agent a company deploys arrives with a login, a set of permissions and a new way in. Every model that writes code also writes bugs, and the same class of model finds them faster than any human team. If AI spreads through every company, rising security spending is about as safe a call as you can make.

The market made that call early and loudly. As of the November 2 close, Okta and CrowdStrike have more than doubled this year and Palo Alto has nearly doubled, while the CIBR cybersecurity ETF is up around 35%, roughly three times the Nasdaq-100’s 11%.

What those returns don’t show is how they were made, and CrowdStrike, the sector’s bellwether, is the clearest case. It grew revenue 24% over the twelve months to July while its trailing price-to-sales multiple went from 28x to 50x, so most of its return came from investors paying more for each dollar of sales. That 50x is a trailing measure; on enterprise value over next-twelve-month sales, which I use from here on, CrowdStrike trades at about 37x today. A year ago this was a laggard trade: CIBR trailed the Nasdaq-100 by about 7 points in 2025, and in February the sector sat at its lowest levels since late 2023.

Nobody serious argues that security doesn’t matter. After a full re-rating, the real question is whether these stocks still pay you for the risk.

_{THE CYBERSECURITY PARADOX}

Demand comes first, because it’s real. Gartner puts global security spending at $247.5B in 2026, up 12.7% in constant currency, and at $372.8B by 2030, with securing AI as the fastest-growing slice. That path works out to about 11% a year. The leaders grow at roughly twice that pace and are priced as if the gap lasts for years, a bet that rests on one belief: that AI concentrates the spending in a handful of winners. Three things push against it.

AI is pressuring the detection layer. On February 20 Anthropic released Claude Code Security, a scanner that reasons through a codebase the way a human researcher would, as a limited preview for enterprise customers with free access for open-source maintainers. The Global X cyber ETF fell 4.9% that day, and SailPoint and Okta each dropped about 9%. OpenAI has shipped four cyber-focused models this year, and a mid-October report that it was in talks to buy the security firm Trail of Bits knocked Palo Alto 4% and CrowdStrike 3%. Finding vulnerabilities is turning into a feature of frontier models. That doesn’t make detection worthless, but it puts a ceiling on what a standalone “we find the problem” product can charge.

Platforms are absorbing budget. Google closed its $32B Wiz deal in March, Palo Alto paid roughly $25B for CyberArk, and Microsoft bundles security into its licenses without disclosing what it earns from it (the last figure, from January 2023, was already above $20B). When a buyer can get something good enough with its cloud or Office contract, a standalone vendor has to be clearly better or it loses on price.

The market stopped paying for beats. F5 beat and fell 20% on October 28. Fortinet beat and raised that evening, then lost 20% the next session after service revenue grew only 13% and its service guide was cut. Palo Alto, CrowdStrike and Zscaler fell 2 to 4% in sympathy without reporting anything, which is what a crowded trade looks like.

The strangest part is that the same AI headline cuts both ways. Leaked details of Anthropic’s Mythos model knocked Palo Alto 6% in March, yet by August CrowdStrike’s CEO was crediting it for demand: “The Mythos moment translated into mass-market acceptance that AI adoption needs security.” On October 21, governments took websites offline to patch PHP flaws found with frontier AI, and HACK posted its best week against the S&P 500 since 2014.

The data shows rising spend and expanding multiples. My inference is that detection-only vendors will lose pricing power before it shows up in their reported numbers. Headlines can’t tell you which way AI cuts for a given company, but the business model can.

_{WHO ACTUALLY MAKES MONEY?}

Read more

I spent the whole day digging through the Marvell-Lumentum filings, the deck and the call, and this is the strangest deal I’ve owned both sides of.

Lumentum holders spent months waiting for a buyer with a fat premium for the scarcest lasers in AI. On Monday the buyer showed up, and the stock fell 18.6%.

On October 28, the last close before Bloomberg reported the talks, 4.5 Marvell shares were worth $1,132. That’s a 23.9% premium, and it’s the number in the press release.

Then Marvell fell three sessions in a row, and in an all-stock deal the offer falls with it. By Monday’s close those same 4.5 shares were worth $771, and Lumentum closed at $684.40, at a discount even to that.

Management called it a once-in-a-generation combination, and both stocks got sold anyway. I own both, so I wanted to answer three questions in order: why Marvell wants this, who got the better end, and what I do with each stock now.

First, what Lumentum holders are actually getting.

WHAT LUMENTUM HOLDERS ACTUALLY GET

Each Lumentum share becomes 4.5 Marvell shares, with no collar and no cash floor. Every 10% Marvell drops takes 10% off what Lumentum holders get.

At October 28 prices that valued Lumentum at ~$115B including debt, or ~$102.5B for the equity. Marvell holders end up with ~67% of the combined company and Lumentum holders with ~33%. Matt Murphy stays chair and CEO, three Lumentum directors join the board, and the close is expected in the first half of 2028.

The break fees tell you who’s worried about what. Marvell owes $5.8B if its board backs out or takes a better offer, Lumentum owes $2.9B in the same situation, and if regulators kill the deal, Marvell pays Lumentum $3.2B, about $34 a share.

The premium depends entirely on which price you pick: 23.9% on the headline, 5.1% at Friday’s close, and minus 15.6% if you compare Monday’s package to Lumentum’s pre-leak price.

Then there’s the 12.7% deal spread, what Lumentum holders collect if the deal closes and Marvell doesn’t move. Hold on to that number, because the whole call comes back to it.

So the price looks bad for Lumentum. Before judging it, it’s worth understanding why Marvell wanted this badly enough to give up a third of the company.

WHY MARVELL NEEDS THE LIGHT

The easiest way to see it is to follow a signal out of an AI chip.

Marvell owns the electronics. Its DSPs clean up the signal inside a large share of the world’s high-speed pluggable transceivers, and it designs custom AI chips for the hyperscalers.

Lumentum owns the light: indium phosphide lasers, transceivers, and optical circuit switches, a business growing fast with Google.

Co-packaged optics (CPO) changes the balance between the two. It puts the optical engine right next to the chip, so the signal travels millimeters instead of meters, which saves a lot of power and needs fewer DSPs. The industry’s own forecasters expect linear-drive optics and CPO to eat into DSP sales over the next five years.

The laser is the one part CPO can’t put on the chip. It sits outside the package, it has to be brighter, and only a handful of companies can make it at scale.

So as optics move into the package, Marvell’s piece of the link shrinks and Lumentum’s matters more. Murphy said on the call that tight supply is not the reason for the deal, and I believe him. The real reason is that CPO would slowly shrink his core business without it.

Why not just partner? Because with lasers sold out for years, the company that owns the fab decides who ships first, and no supply agreement gives you that. Marvell is buying the part of the link that survives the transition, before the transition hits its own franchise.

The question is why it had to happen now, and the answer runs through Nvidia.

WHY NOW: NVIDIA’S OPTICS RAMP

Read more

If your grandma owns staples, utilities and dividend payers, she had a perfectly good month and has no idea what you’re complaining about.

Since October 7, the Nasdaq-100 is down about 9%, while the S&P 500 without its tech and communication services stocks is up almost 5%.

Same market, same economy, same Fed, and a 14-point gap in under four weeks. In a recession, grandma’s stocks would be falling too.

Good chart, so I’m using it, but I took the label off. The Kobeissi Letter called this the largest monthly drop on record for AI stocks, and it isn’t close.

The Nasdaq Composite fell 5.3% in October, which wasn’t even its worst month since March 2025, when it lost about 8%. The record still belongs to October 1987 at about -27%, followed by November 2000 at about -23% as the dot-com bubble came apart. Even April 2022 was worse than this, at -13%.

Headlines like that are how people end up selling at the worst possible time.

The October scorecard says the same thing, name by name.

The bottom reads like an AI portfolio, mine included: Intel -46%, CoreWeave -37%, Nebius -34%, Sandisk -33%, Marvell -25%, AMD and Micron -17% each, and Lumentum -13%.

The top reads like grandma’s shopping list: Walmart +19%, Keurig Dr Pepper +18%, Mondelez +13% and Costco +11%, with utilities like Exelon and Xcel right alongside. Even inside chips the split held, as Texas Instruments, NXP and Analog Devices, which sell into cars and factories rather than data centers, all finished higher.

Three things split them, and they run through everything below. Borrowing got more expensive for the companies building AI, oil got more expensive, and the funds that owned the most AI had to sell it.

Long-duration growth stocks feel all three the most. A company earning most of its profit years from now gets hit hardest when the 10-year sits near 5% and its own credit spreads widen, while a soap maker paying a dividend today barely notices, and hedge funds were never crowded into it anyway.

Then Friday made it worse. The Philadelphia Semiconductor Index (SOX) fell 7.2%, its worst day since June, and the iShares SOXX ETF lost about 10%, partly because Intel, down 28.6% on its guide, is one of its biggest holdings. Chips finished October down about 15%.

Last week I asked what would bring the hedge funds back. This week answered half of it: the fundamentals got stronger while prices got worse, and that gap is the whole setup for the months ahead.

But first, the macro.

MACRO OVERVIEW

Hikes. The Fed held at 3.75% to 4.00% on Wednesday in a 10-2 vote, with Lisa Cook and Beth Hammack dissenting in favor of a hike.

Kevin Warsh said the “monetary medicine” was “starting to work some,” but that there are “no firm conclusions,” and he wouldn’t call financial conditions tight.

The data gave him no reason to soften. Third-quarter GDP grew 3.4% annualized against roughly 3.1% expected, and September PCE inflation rose to 3.6%, with core at 3.1%.

December hike odds peaked near 80% on Thursday and ended the week around 72%. Futures price the policy rate near 4.7% by September 2027, about three more hikes.

Yields. The 10-year briefly broke back above 5% around the Fed decision, then closed Friday at 4.94%, down 35 basis points over October. The 2-year closed at 4.56% and the 30-year held at 5.31%.

Same pattern as the last two weeks: when AI sells off, money hides in Treasuries.

But look at what that means. Yields fell 35 basis points in October and AI still had a terrible month, so a lower 10-year on its own wasn’t enough. The level is what matters, and 4.94% is still higher than at any point from 2008 to 2022.

Oil and diesel. Brent’s December contract expired Friday at $117.64, its highest close since April 29, while January, now the front month, settled at $113.43. WTI closed at $104.12, its highest since mid-September and a fourth straight weekly gain. Both rose 14% to 15% in October, and diesel averaged $6.35 a gallon.

What drove it: the first successful limpet-mine attack on a tanker near Hormuz, Houthi claims of strikes on Saudi oil facilities, and Gulf shipments falling to about 15.7 million barrels a day, from more than 24 million earlier in the month.

Iran. Qatar delivered a draft deal on October 25, Tehran called its ideas “unacceptable,” and the mediators went back on Friday with edits. Iran’s reply is expected around November 1.

The sticking point is order. Washington wants the nuclear issue settled first, and Tehran wants the blockade and sanctions lifted first.

Meanwhile, the Wall Street Journal reported that Trump was briefed on strike options on Thursday and that Arab allies were told to prepare for a “short, intense campaign” after the midterms. Trump on Friday: “one way or the other, it’s going to end very soon.”

October had 43 shipping incidents in the region, 30 of them direct attacks on tankers, the most of any month of the war.

So no hike this week, one likely in December, oil at a six-month high and the war’s next phase pencilled in for after Tuesday’s midterms. That’s a lot of weight on risk appetite, and on Friday it gave way.

THE FRIDAY FLUSH

Intel’s guide lit the match: fourth-quarter revenue of $15 billion to $16 billion against $17.1 billion expected, and a PC slump into the first half of 2027.

It took the whole group with it. Qualcomm -11%, AMD -11%, Marvell -10%, Sandisk -13.5%, Micron -5% and Nvidia -4%.

The Nasdaq lost 2.9%, its worst day since June. The SOX closed below its 200-day moving average at its lowest since late July, and the VIX touched its highest since June before closing at 19.4.

Most of these companies didn’t report anything on Friday. Marvell barely touches PCs, Nvidia’s customers just told us they still can’t get enough capacity, and Micron and Sandisk are selling every chip they can make.

When good and bad stocks fall together like this, someone is selling the whole basket, and the positioning data says exactly who.

WHO CONTROLS THE TAPE

Over a few weeks, prices are set by whoever has to trade. Right now that’s the hedge funds, and their books explain most of Friday.

By Morgan Stanley’s prime brokerage numbers, long/short funds have cut 7.8% of their gross exposure from the peak. That’s the biggest de-grossing since the data starts in 2016, bigger than March 2020, 2021 and 2016, and it happened before Friday.

Momentum is where it shows most. Jefferies’ tech momentum basket is down 57% over 60 days, deeper than anything on its chart since 2000. The dot-com unwind and the 2021-22 growth crash both bottomed near 40%.

These are the stocks hedge funds owned the most, and they’re getting sold for being crowded, with nothing in their earnings to justify it.

This is the part that matters most to me. Funds are holding on to their core AI longs and cutting net exposure by shorting the index instead.

QQQ short interest jumped to roughly 82 million shares, the highest since 2022, while SPY short interest fell. They didn’t give up on the stocks. They hedged them.

Index shorts push every tech stock down together, which is why correlation is rising again and why good and bad reports got the same treatment on Friday.

And a hedge is the easiest thing in the world to take off. Funds don’t need new conviction or new ideas to come back. They just buy back the index shorts, and the same mechanics that crushed everything at once lift everything at once.

THREE HYPERSCALERS, ONE VERDICT: SHOW ME THE CASH

Strip the noise out of earnings week and one picture is left: the ad businesses slowed and the clouds sped up. That’s what money moving from the present into the future looks like.

A year ago Google Cloud and Search were growing at roughly the same pace. Now Cloud grows ~102% and Search ~12%.

AWS sped up from 20% to 39% and Azure from 40% to 48%, while Amazon’s ads fell from 24% to 11%.

Meta is the control group, with ad growth steady at 26%. Yet it still fell as much as 7.6% on its own report last week, when its free cash flow turned negative. Strong ads didn’t save it.

Microsoft did the cleanest double beat. Revenue of $92.3 billion and EPS of $4.89 both came in ahead, Azure grew 48% and is guided to 49% to 50%, and it still produced $18.6 billion of free cash flow after $45.2 billion of capex.

Amazon landed in the middle. AWS beat and grew at its fastest pace in nearly five years, while ads and the fourth-quarter guide missed. Trailing free cash flow is now minus $42 billion, and debt is filling the gap.

Google had the worst reaction by far. Cloud doubled, but Search missed, free cash flow went to minus $9.1 billion, and Alphabet sold $12.4 billion of stock while buying back nothing. It lost about $500 billion in a day, the biggest one-day loss in its history.

The market paid for free cash flow and punished the lack of it, no matter how fast the cloud grew.

It isn’t perfectly clean, since Google also missed on Search and Amazon recovered most of its after-hours drop by Friday’s close. But Meta last week and Microsoft this week point the same way.

These companies are crossing from a cash-printing present to a capex-heavy future, and right now the market only gives credit to whoever can cross without borrowing.

The credit market is drawing the same line.

Since last October, the cost of insuring Amazon, Google, Oracle, Meta, Intel and Tesla debt against default went from about 36 basis points to about 120. For Apple, Nvidia, Microsoft, Walmart, Costco, Lilly and Caterpillar it went from about 29 to about 48.

Both groups got more expensive, but the gap between them went from about 7 basis points to more than 70, and most of it opened after June, right as the capex numbers exploded.

Lenders are charging more to the companies that borrow to build, and the stock market is handing out the same treatment.

That’s a filter I’m adding for any new name: can it pay for its buildout from its own cash? If it can’t, I want a lower price to make up for it.

And none of this is slowing the spending down. Street estimates have capex at the biggest AI spenders going from $482 billion in 2025 to $951 billion this year and $1.37 trillion in 2027.

Here’s why that matters for my book. The market is punishing the hypers for spending it, but a big share of every dollar ends up as revenue for the kind of companies I own: memory, optics, chips and power.

One caution. The growth in that spend slows from almost doubling this year to about 44% in 2027 and 10% in 2028. For suppliers, the slowdown matters as much as the total, and it’s a big part of why memory trades like it’s at the peak.

CHEAPER THAN IT HAS BEEN IN YEARS

The selloff took valuations with it, and tech is now cheap by any yardstick I trust.

Tech trades at 18.4 times forward earnings, down from about 31 a year ago and well below its 5-year average of 25.3 and 10-year average of 23.0. That’s cheaper than the April 2025 tariff low, when it briefly touched about 21.

Then there’s Nvidia, the company at the center of all this, at about 29 times trailing earnings. That’s the lowest multiple on this chart going back to early 2023, when it touched more than 150.

On next year’s expected earnings it’s around 18 times, about the same as the S&P 500, for the company selling the most important product in the buildout. And the stock is sitting right on its 50-week average, a line it has bounced off several times this year.

Adjust for how profitable companies are today and the broader market tells the same story. Bloomberg’s ratio of the S&P 500’s forward P/E to its operating margin is down to about 0.8, from about 1.25 at the end of 2025, the lowest in more than a decade.

You’re paying less for each point of margin than at almost any time since 2013.

And credit isn’t breaking. High-yield spreads widened from 267 basis points to 337, a real move, but nowhere near the 578 of 2022, the 820 of 2016 or the 1,013 of March 2020.

Last week I showed the stress at the riskiest AI borrowers. The broad credit market isn’t pricing a crisis.

THE DEMAND SIGNALS THE SELLOFF IGNORED

The tape says AI is in trouble. The usage data says the opposite.

Since July 12, B300 rental rates are up 106%, B200 up 66% and even the 2022-era H100 up 17%. Over the same stretch, the blended price of a token fell 58%.

The chips are getting more expensive to rent while what they produce gets cheaper. That only works if usage is exploding, and it is.

Token spend is up 122% to 160% since May depending on whose data you use, while the price per token fell 67%. Put those together and the tokens running through OpenRouter grew roughly eightfold in six months.

OpenRouter is a small slice of the market, so I treat it as a signal rather than a total, but it matches what Microsoft said on its call: demand still exceeds supply.

And the productivity is starting to show up in the national data. Output per worker in the US sectors that adopted AI first has pulled away from its pre-pandemic trend, and the gap has widened sharply since ChatGPT launched, while the UK, Canada and most of Europe are still tracking theirs.

It’s early and it’s concentrated, but it’s measurable now, and it’s the payoff the whole buildout is betting on.

And adoption has barely started. AI sits at about 20%, roughly where the internet was in the mid-1990s.

The internet passed 50% within a few years of that point, and the stocks that rode it had brutal drawdowns along the way.

We’re at the bottom left of that curve.

THE CATALYST: THE NUMBER THAT TURNS THE AI TRADE AROUND

Read more

Intel reported on Thursday night. By Friday’s close the stock was at $65.76, down 28.6% in a single session, its worst day since July 1974. Roughly $140 billion of market value disappeared in a day. The stock is still up about 78% this year, but it now sits more than 50% below the record it set in June.

What makes it sting is that the quarter itself was good. Data center revenue grew 78%, gross margin came in two points above guidance and adjusted earnings beat by about 30%. Then management guided the fourth quarter well below what anyone expected, said the PC market will shrink again next year and told investors not to count on positive free cash flow in 2027.

I own this stock. I added to it this week, before the report, and I’m down about 25% on the position. So I’m going to be harsh about it, because the version of this story I’d like to believe is the one I trust least right now.

THE QUARTER

Revenue was $16.5 billion, up 21% and right in line with expectations. Adjusted gross margin was 44.1% against Intel’s own 42% guide, and adjusted EPS was $0.52 against roughly $0.40 expected.

Data Center and AI did $7.3 billion, up 78%, Intel’s fastest growth in that business in 15 years, at a 41.5% operating margin. Revenue from custom chips, which Intel reports inside that segment, more than doubled from a year ago.

The PC business went the other way. Client revenue fell 5% from a year ago and 9% from the second quarter to $8.1 billion, as soaring memory prices push PC makers to build fewer machines.

The foundry booked $6.7 billion of revenue, but almost all of it is Intel selling wafers to itself. It lost $1.85 billion. Revenue from outside customers was $411 million, and the CFO said most of that came from a program that won’t repeat.

One accounting note. GAAP EPS of $0.60 came in above adjusted EPS because Intel booked a $1 billion gain on government-linked shares it holds in escrow, which it marks to market every quarter. The same line produced a $12.5 billion loss in the second quarter. Ignore it in both directions.

Then came the guide. Intel expects fourth-quarter revenue of $15 billion to $16 billion against about $17.1 billion expected, adjusted EPS of $0.25 against $0.43, and gross margin of 38%. Client revenue is expected to fall more than 20% from the third quarter. Management now sees the PC market shrinking about 12% this year and another 10% in 2027.

Part of that is an accident. An explosion at Intel’s Fab 28 in Israel on October 14, which killed two workers, shut down a plant that makes mostly server chips. Intel puts the cost at about $300 million of revenue and two points of gross margin in the fourth quarter, with a similar hit in the first quarter of 2027.

Worth noting the order in which Tan listed what held Intel back. First came Intel 3 yield and cycle time, where he said he is “demanding significant execution from the team.” Then the industry-wide substrate shortage. Then the accident. The first item on that list is Intel’s own factory.

THE BEAR CASE: MANAGEMENT KEEPS LOSING THE ROOM

Here is the case against Intel, as harshly as I can make it.

Beating the quarter doesn’t matter anymore. Intel has beaten earnings expectations in each of its last five reports. The stock fell 17% after the January report and 29% after this one, both times because the outlook disappointed. In July it guided above expectations and still dropped 8%. Investors are trading the outlook, and twice this year the outlook came in worse than the quarter that came before it.

The spending plan already slipped. In January, Intel told investors 2026 capex would be flat to slightly down from 2025’s $17.7 billion. It’s now about $20 billion, and 2027 will be “up strongly”, with no number attached. Equipment orders rose about 40% from the second quarter. The CFO said positive free cash flow next year is “probably not” going to happen.

Shareholders are paying for the turnaround. Since August 2025, Intel has sold about $39 billion of stock: $8.9 billion to the US government (paid for with grants Intel had already been awarded), $2 billion to SoftBank, $5 billion to Nvidia and $23 billion to the public in August. The diluted share count is up 17% in a year. Anyone who bought that August offering paid $95 a share. It closed Friday at $65.76. Intel also agreed in April to pay $14.2 billion to buy back the stake it had sold in its Ireland fab, partly with new debt.

The foundry is still a hole. Intel Foundry lost $13.3 billion in 2024, $10.3 billion in 2025 and another $6.35 billion in the first nine months of this year. That’s about $30 billion in under three years. In August, the CFO targeted breakeven by the end of 2027, then said he could see it slipping into 2028.

14A is running out of road. Last year Intel warned it might not build its next process, 14A, without a major outside customer. In January it said customers would decide between the second half of 2026 and the first half of 2027. We’re about a third of the way through that window. On Thursday, Lip-Bu Tan said he was encouraged by progress with “multiple potential customers” on 14A, and didn’t name any. Elon Musk has said Tesla and SpaceX plan to use 14A for their Terafab project, and Apple is reportedly evaluating it, but Intel hasn’t announced a signed 14A customer.

It can’t make enough of what’s working, and part of that is self-inflicted. Tan said he was “disappointed” Intel couldn’t meet demand this quarter. Most of the headlines blamed substrates, but the CFO also said Intel is “still constrained on Intel 3”, and that the first half of 2027 will be tight on both wafers and substrates. A leading-edge wafer takes months to move through a fab, so wafers that weren’t started for servers early enough can’t be rushed out now. Intel is only now taking 18A capacity in Oregon that was meant for Panther Lake laptops and using it for Intel 3 server chips. That’s the right call. In my view it’s also a few quarters late, in the middle of the strongest demand for server CPUs Intel has seen in 15 years.

The capital priorities are hard to follow. Ten weeks after raising $23 billion from shareholders, and days before this report, Intel agreed to buy Classiq, an Israeli quantum software company, for a reported $2.3 billion. Quantum may matter in ten years. Substrate supply matters this quarter, and a slice of that money could have gone into prepaying suppliers for it.

Put together: management beats the quarter and misses the guide, supply planning lagged the biggest server upcycle in years, the capex plan only goes up, shareholders keep getting diluted, the foundry still loses close to $2 billion a quarter, and the customer that justifies 14A hasn’t signed.

STRESS-TESTING THE BEAR CASE

Some of that holds. Some of it is exaggerated by a very bad Friday.

The margin collapse is only partly a one-off. The Israel shutdown explains about two of the six points of gross margin Intel is giving up in the fourth quarter. Without it, the guide would be around 40%. That’s still four points below the third quarter, and management blamed the rest on weaker PCs, underused 18A capacity and lower-margin custom chips. So the accident softens the miss but doesn’t explain it.

Running out of supply is the better problem. Data center demand “continues to meaningfully exceed available supply”, in the CFO’s words. Intel is adding about 40% more substrate capacity, though most of it doesn’t arrive until the second half of 2027. A company that can’t build enough is in a very different spot from one that can’t sell what it builds.

The PC slump is partly a hangover. Windows 10 support ended in October 2025, which pushed a lot of companies and households to replace machines early. My read is that some buyers also ordered ahead once memory prices started climbing, so part of 2026 and 2027 demand was pulled into 2025. The CFO said customers are now cutting the weeks of inventory they hold. That’s a cycle working itself out, and it hits AMD’s PC business too. It doesn’t excuse Intel’s own share loss in PCs, from about 76% of x86 units in early 2025 to about 70% this year.

18A is real now. Yields still aren’t at industry-standard levels, but Tan said they are improving at a double-digit rate. The CFO said 18A costs are down about 40% since the first quarter and that 18A went from under 5% of client CPU revenue in the first quarter to nearly a third. Panther Lake has ramped, and Clearwater Forest, the first 18A server chip, is shipping for revenue. A year ago this was the biggest open question about Intel.

The balance sheet can take it. Intel ended the quarter with $54 billion of cash and short-term investments against $50 billion of debt, generated $1.6 billion of adjusted free cash flow in the quarter even though the full year will be negative, and is still rated investment grade by all three agencies. The US government owns about 8% and Nvidia about 4%. Companies with that profile don’t go under.

There’s a dated promise on the table. Asked about outside foundry customers, Tan said Intel will have “a couple” of customers ready to announce within the next two quarters. He didn’t say whether that means 14A, 18A-P or packaging, so I wouldn’t read it as a 14A win yet. But it’s specific, it has a deadline, and it can be checked by mid-2027.

No customer, no 14A is a discipline. Tan and the CFO have repeated on several calls that Intel won’t build out 14A capacity without customer commitments. That’s frustrating for anyone waiting on a name, but it caps how much Intel can lose on the node. And Intel shipped the 0.9 design kit in October, the version customers use to start serious designs, with the 1.0 kit due by the first quarter.

The foundry is losing less every quarter. Foundry losses went from about $2.4 billion in the first quarter to $2.1 billion in the second to $1.85 billion in the third. That’s still a lot of money, but the direction is right.

So where do I land? The bear case is wrong about survival and right about returns. Intel has the cash, the backers and a working process. What it doesn’t have yet is proof that all this spending earns anything for shareholders.

THE BULL CASE: AGENTS RUN ON CPUS

Read more

Amazon reported on Thursday night, and the stock dropped more than 7% after hours to around $226. By Friday’s close it had clawed most of that back, ending at $240.35, down about 1.4% on the day, after a 4% slide on Thursday before the report.

AWS grew 39%, its sixth straight quarter of acceleration and its fastest pace in nearly five years. Its backlog hit $578 billion. Andy Jassy told analysts he expects AWS to grow faster than 40% in both 2027 and 2028, and said AWS could become a trillion-dollar revenue business within five to ten years.

And at the same time, Amazon’s trailing free cash flow went to minus $42 billion. Its long-term debt has gone from about $66 billion to $173 billion this year, and Jassy said 2027 capex will be “meaningfully more” than this year’s roughly $220 billion.

That’s the whole debate in two paragraphs. Here’s how I worked through it.

THE QUARTER

Revenue was $201.9 billion, up 12% and a hair below the $202.2 billion expected. Amazon says moving Prime Day into the second quarter cut about 4 points from third-quarter growth, so the underlying pace was closer to 16%.

AWS did $46.0 billion, up 39% and slightly ahead of the $45.7 billion consensus. AWS made $17.8 billion of operating profit, a 38.7% margin. AWS is less than a quarter of Amazon’s sales, but it supplied about 60% of the growth in both revenue and operating profit this quarter.

Operating income was a record $28.1 billion, up 61%, and the overall operating margin hit a record 13.9%. The growth rate is flattered: a year ago Amazon took a $2.5 billion FTC settlement charge plus severance costs, so underlying growth was closer to 30%. That came in well above the top of Amazon’s own guidance range of $22.5 billion to $26.5 billion.

EPS was $1.89 against about $1.97 expected, even though operating income beat, because of a $1.2 billion non-operating loss. The decline from $1.95 a year ago is mostly the Anthropic gain lapping: a year ago Amazon booked $9.5 billion of pre-tax gains on that stake.

The misses were in ads and in the outlook. Advertising came in at $19.6 billion against about $20.8 billion expected. For the fourth quarter, Amazon guided revenue of $240 billion to $247 billion and operating income of $28.5 billion to $33.5 billion, both a little below consensus at the midpoint. Amazon has a habit of guiding conservatively and beating its own range, as it did by a wide margin this quarter.

THE BEAR CASE: THE CASH IS GOING THE WRONG WAY

Here is the bear case at its strongest.

The beats are narrow and the misses are in the growth engines. AWS beat by less than 1%. Ads missed by almost 6%. Both parts of the fourth-quarter guide came in short. For a stock priced on AI growth, a 0.7% AWS beat isn’t much to hold onto.

The cash has flipped. Over the last twelve months Amazon generated $168 billion of operating cash and spent $211 billion on capex, net of sales. A year ago it was producing $15 billion of free cash flow. Now it’s burning $42 billion.

Debt is filling the gap. Long-term debt went from $65.6 billion at the end of last year to $173 billion, while cash and short-term investments slipped from about $123 billion to about $114 billion. In nine months Amazon went from roughly $57 billion of net cash to roughly $59 billion of net debt, a swing of more than $115 billion. On top of that it set up a secured facility of up to $80 billion, backed by the hardware it buys and available from the start of 2027 through the third quarter of 2028. Amazon isn’t alone, since Microsoft sold $25 billion of bonds this week, its first US bond sale since 2017, but borrowing against servers is new for Amazon.

And it gets bigger. Jassy said 2027 capex will be “meaningfully more” than 2026. If AWS growth stalls while that bill keeps rising, the depreciation lands on margins for years.

A lot of this spending has a short shelf life. A data center shell can stand for decades. The chips inside it are outdated in a few years. So a growing share of this capex is closer to a recurring cost than a one-time build: once the racks are full, some of the budget goes to replacing them just to stand still. Nobody outside Amazon knows how much of the $220 billion is growth and how much is upkeep, and that split decides what AWS really earns on its capital.

Ads are slowing. Ad growth fell from 26% in the second quarter to 11% in the third, more than halving. Ads are widely seen as one of Amazon’s highest-margin businesses, and they help pay for the buildout.

Put together: Amazon is spending more than it earns, borrowing the difference, and promising to spend even more next year, while the business that funds it is slowing.

STRESS-TESTING THE BEAR CASE

Some of that holds up. Some of it doesn’t.

The ad slowdown is mostly timing. Prime Day, Amazon’s biggest ad event, moved from July into June this year. That flattered the second quarter and hurt the third, and management put the hit to third-quarter ad growth at about 8.5 points. Two other things muddied the comparison. A year ago Amazon started sending its paid search traffic to its own search results pages, where it sells more ads, and the third quarter was the first to lap that boost. And the second quarter had the first season of NBA games on Prime Video, which the third quarter didn’t. Combine the second and third quarters and ads grew about 18%, in line with the high-teens growth Amazon guided for the fourth quarter, and still well ahead of Google’s roughly 12% over the same six months.

Two things keep me from fully buying the excuse. Management only put numbers on the timing effects in the quarter that missed, and Amazon doesn’t disclose how much profit ads make, so we can’t see what the slowdown costs. If ads come in near 10% again in Q4, the bears are right. Until then, I read the 11% as mostly a calendar problem.

The capex number is lumpier than it looks. Amazon spent $77.3 billion on capex in the third quarter before asset sales, about $19 billion, or roughly a third, more than analysts expected, but kept its full-year number at about $220 billion. Subtract the first three quarters and that implies only about $48 billion in the fourth. Some of the third quarter was likely spending pulled forward, which makes the trailing cash burn look worse than the run rate.

The profit side is healthier than ever. Operating income grew 61% to a record, North America’s margin rose to 7.7% from 4.5% (about 6.8% without last year’s FTC charge), and AWS earns 38.7 cents on the dollar. Retail has its own automation story too: Jassy said Amazon now runs more than a million robots in its fulfillment network and sees a chance for them to help with the last 100 yards of delivery. Cash burn caused by spending is a very different problem from cash burn caused by a weakening business.

The debt has assets behind it. Not all of the borrowing went into data centers. Amazon finished funding a $50 billion investment in OpenAI this year, about $21 billion of it after June. And its stake in Anthropic was valued at about $190 billion at the end of June, with Anthropic preparing to list in the coming weeks. That stake alone exceeds all of Amazon’s long-term debt, though it’s illiquid and partly Amazon’s own estimate.

So where do I land? The bear case is right about the cash. It’s real, it’s large, and 2027 makes it bigger. It’s mostly wrong about the business. The ad miss is largely timing, profits are at records, and the capex spike looks partly pulled forward. The question is whether what the cash is buying is worth it.

THE BULL CASE: AWS IS ACCELERATING INTO A BACKLOG

Read more

SanDisk did $11.0 billion of revenue last quarter. A year ago it did $2.3 billion. It now keeps 85 cents of gross profit on every dollar it sells, a margin most software companies would envy, and it bought back $5 billion of its own stock in a single quarter.

The stock closed Thursday at $1,330, about 43% below its June 25 record close of $2,335. Micron tells the same story: a record quarter reported at the end of September, and a stock that was about 28% below its June 25 record close as of last week.

So the question I’ve been getting all week is simple: is it time to sell memory?

Both sides have a real case. The bulls say memory has gone from a commodity that rode the PC and phone cycle to a core input of the AI buildout. The bears say growth is slowing fast, margins can’t go any higher, and this is exactly what the top of every memory cycle has looked like. I went through both, ran the numbers the bears would want to see, and landed on an answer at the end.

THE QUARTER

Revenue came in at $11.0 billion against about $10.7 billion expected, up 23% from the prior quarter and 378% from a year ago. Non-GAAP EPS was $49.51 against $47.15. Gross margin was 85.1% and operating margin 80.4%, with operating expenses at under 5% of revenue.

Data center did $4.8 billion, up 61% in one quarter. Edge, which covers PCs, phones, cars and industrial devices, did $5.8 billion, up 7%. Consumer fell 29% to under $400 million, which management put down to the market adjusting to higher prices.

The cash is the part people underrate. Operating cash flow was $6.4 billion, adjusted free cash flow $5.7 billion, and SanDisk returned $5 billion through buybacks, retiring about 4% of its shares over two quarters, although the diluted share count is still about 3% above a year ago. It has $5.1 billion of cash, no debt, and says it will return all excess cash to shareholders. An 8-for-1 split takes effect on November 20, and all per-share numbers in this piece are pre-split.

For the December quarter, SanDisk guided to $12.2 billion to $12.6 billion of revenue, another record, with gross margin of 84.5% to 85.5% and “modest” price increases.

FROM 12% TO 44%

This is the chart behind the word “structural.” A year ago, data center was $269 million, about 12% of SanDisk’s revenue. Today it’s $4.8 billion and 44%.

That matters because the buyers are different. Consumer and PC demand is price-sensitive and swings with the economy. Data center buyers are cloud companies and AI labs building inference capacity, and they increasingly sign multi-year contracts. The more of SanDisk’s revenue comes from that second group, the less this business behaves like the old memory cycle.

WHY THE AI BUILDOUT NEEDS MEMORY

It’s worth being precise about why the bulls call this structural, because the whole debate turns on it. Every AI system runs on three layers of memory. HBM sits on the accelerator and feeds the model during training and inference. DRAM sits in the server. Flash storage, which is what SanDisk sells, holds the data, the model checkpoints and, increasingly, the working memory of the model itself. You can’t add compute without adding all three.

Training made HBM the bottleneck. Inference and agents are now doing the same to flash. Management said some inference and agent workloads need 60 to 90 gigabytes of flash per CPU, and it expects the memory that stores an AI model’s working context, the KV cache, to make up about a third of the AI data center flash market by the end of the decade. Data center now takes roughly half of all NAND bits, up from about 20% early in the decade, and SanDisk expects it to pass 55% of the market in 2027.

That’s the difference from past cycles. Inference demand grows with usage, every day, rather than ending when a training run finishes. As long as AI usage keeps growing, the need for memory grows with it. That makes the demand structural, even if the price of memory will still move in cycles.

THE SLOWDOWN IN ONE CHART

Read more

And this was the quarter Google Cloud grew 103%.

There’s a second milestone buried in the numbers. For the first time since Alphabet began breaking it out, Search made up less than half of Alphabet’s revenue.

That’s why the debate this week is so loud. One camp says AI is finally eating Google’s core business and the company is squeezing advertisers to hide it. The other says the market just panicked out of the best cloud business in the group. I’ve been bullish on Google all year, so I made myself argue the bear side first.

THE QUARTER

Revenue was $127.6 billion, up 25% and just shy of the $128.2 billion expected. Cloud beat by $1.1 billion to $1.6 billion, depending on whose consensus you use. Almost everything else missed.

Search and other came in at $63.4 billion against about $65 billion, growing 12% when analysts had about 15% in their models. YouTube ads grew 9% to $11.2 billion. Google Services as a whole missed by about $2 billion.

EPS was $2.12, down 26%. A big piece of that is non-operating: an $8.7 billion loss on equity securities, mainly the SpaceX stake, cost about $0.55 a share. But don’t let anyone tell you the whole miss was a paper loss. Strip out the markdown and EPS is still about $2.67 against $3.02 expected, and operating income came in at $40.6 billion against roughly $43.2 billion, about 6% short. That part is the business.

Then there’s the number that really moved the stock: free cash flow of minus $9.1 billion, when LSEG’s consensus had it near break-even. It’s the second negative quarter in a row for a company that used to generate more than $70 billion a year.

THE BEAR CASE: SEARCH IS BEING HOLLOWED OUT

Here is the bear case at its strongest, because it deserves to be taken seriously.

Start with the strangest line on the call. Sundar Pichai said query growth hit a five-year high. Yet Search revenue grew only 12%, down from 19% two quarters ago. If searches are accelerating while revenue slows, each search is earning less than before, or at best its revenue is growing much more slowly.

Google gave four reasons. Currency was the single largest. Retail in Asia Pacific took a hit when the EU scrapped duty-free treatment of low-value parcels, so those sellers pulled back on ads aimed at European shoppers. Media and entertainment normalized after a World Cup-boosted Q2. And AI Mode, the conversational version of Search, is growing fast while monetizing at a lower rate than the classic results page.

The first three should lap within a year, although Ashkenazi expects currency to hurt slightly more in Q4. The fourth is structural. AI Mode queries run about three times longer than classic searches, they earn less per query, and Google hasn’t said how big that gap is. If that gap doesn’t close, more usage means slower growth.

Then come the advertisers. At least one big advertiser went public this week with a sharper accusation: Google is manufacturing growth by charging more for worse clicks. The complaint is that exact-match keywords now trigger ads on thousands of loosely related searches, budgets get overspent, and prices go up regardless of competition. People who have worked in media compare it to legacy TV, which raised the price of each slot for years to hide falling audiences, until buyers decided the return wasn’t there and left.

Add that Apple’s Eddy Cue testified in May 2025 that Safari searches had fallen for the first time, and Q4 Search has tougher comparisons built in, and you get the bear conclusion: Google’s best business is turning into a slower, lower-multiple business, and pricing tricks are buying time.

STRESS-TESTING THE BEAR CASE

I went through the advertiser claims one by one, because a lot of this week’s analysis rests on them.

The pricing levers are real. In the 2023 antitrust trial, a Google ads executive testified that the company adjusted auction pricing to hit revenue targets, raising prices by up to 5%, and as much as 10% on some queries, usually without telling advertisers. So the idea that Google leans on price is on the court record, and it isn’t new.

The “full bid” claim isn’t documented. Google’s help pages still say you pay only what’s needed to beat the ad ranked below you. The catch is automated bidding, which can push the price above a manual max bid, and Google has been moving advertisers onto it for years. That probably explains a lot of what advertisers are feeling.

Close variants are a real, long-running loosening. Exact match has included close variants since 2014, expanded to same-meaning searches in 2018, and there’s no opt-out. It does exactly what the advertisers say: it widens the pool of searches Google can put ads on.

The 2x budget overspend is an old policy. Since 2017, a campaign can spend up to twice its daily budget on a given day, while monthly billing is capped at 30.4 days of budget. It feels like overspending on busy days, though the policy is nine years old.

Volume is the open question. Google says query growth is at a five-year high, Apple said Safari searches dipped in 2025, and Google is the only one with the full data.

Two more facts cut against the “dying media” story. First, raising prices isn’t by itself a sign of decline. Meta grew ad revenue 26% in Q3 with price per ad up 14% and impressions up 10%. The difference is that Meta reports both price and volume every quarter, while Google hasn’t given a clean number on search volume or revenue per search. Second, ad budgets aren’t shrinking. Meta’s numbers say the money is there.

And the “below 50%” milestone is mostly about Cloud. Search dollars still grew 12%, to $63.4 billion. Search’s share fell from 55% to under 50% in two quarters because Cloud’s share jumped from 18% to 24%.

Amazon’s report on Thursday night sharpens the picture. Its ad business also grew 11% and missed by more than Google’s did in percentage terms, about 6%, which Amazon blamed on Prime Day timing. AWS beat. The two companies printed the same pattern: ads light, cloud ahead, and both stocks down.

My read of that chart is twofold. The ad money is still out there, and this quarter more of the new ad spending went to Meta than to Google or Amazon, so Search’s problem looks more like lost share than a shrinking market. And the market isn’t paying for cloud beats while capex keeps climbing, which is why the cash section below matters more than the Search headline.

So where do I land? The bear case is right that Google is leaning harder on price and that AI Mode earns less per search, and it’s right that this matters for the next two years. It’s wrong that Search is shrinking today. The real risk is that pricing headroom runs out before AI Mode ads catch up, and for now that threatens the multiple more than the revenue.

THE BULL CASE: EVERYTHING ELSE IS GROWING

Read more

I sat down with Microsoft’s numbers last night, and this is the report I’ve been waiting for all year.

For months the fear around AI has been the same: the biggest companies on earth are spending faster than they can earn it back. Credit spreads blew out on that fear, neoclouds sold off hard, and Google fell about 10% after hours on an earnings miss while spending around $200 billion on capex.

Then Microsoft reported the same evening and broke the pattern.

It put $45.2 billion into infrastructure in the September quarter, counting finance leases. That’s about half of everything it brought in. And it still grew revenue 19%, grew operating income 21% and kept an operating margin just under 50%.

Azure, the business all that spending feeds, grew 48%. Next quarter Microsoft is guiding to 49% to 50%.

At the end of 2024, Microsoft spent 23 cents of every revenue dollar on infrastructure in cash. Now it spends 43 cents, and revenue growth has sped up from 12% to 19% over the same stretch. If the spending were running ahead of demand, you’d see the opposite.

So the question I wanted to answer is simple: is Microsoft spending too much on AI, or is the market still underestimating how much compute this needs? It matters well beyond Microsoft, because every one of those dollars lands with someone in the supply chain.

THE QUARTER

Revenue came in at $92.3 billion against $90.6 billion expected, up 19%. Adjusted EPS was $4.89 against $4.73.

Operating income hit $45.9 billion against about $44 billion expected, a 49.8% margin. Operating expenses grew only about 6%, so almost every extra dollar of revenue fell straight to profit.

Azure was $33.2 billion for the quarter, up 48%. It’s the first time Microsoft has given a quarterly Azure figure in dollars. Microsoft Cloud reached $63.9 billion, and its growth accelerated from 27% to 30%.

Operating cash flow was $58 billion, up 29%. Free cash flow fell 27% to $18.6 billion because cash capex doubled, from $19.4 billion to $39.4 billion. Microsoft still sent $10.2 billion back to shareholders.

A quick note on the new reporting: almost everything now sits in Agents and Infra, $76.8 billion and up 25%, which holds Azure, Microsoft 365 and the AI products. Devices and Consumer, $15.5 billion and down 3%, holds Windows, Xbox and search. Prior periods were recast, so don’t line these up against old segment tables.

There were two misses. The backlog came in at $677 billion against about $706 billion expected, which I’ll come back to. And capex was $45.2 billion including leases, against Microsoft’s own guide of more than $50 billion.

The capex miss is the interesting one. Amy Hood, Microsoft’s CFO, said some third-party GPU-as-a-service contracts slipped into the second half of the fiscal year, January to June 2027. Microsoft wanted to spend more. Its suppliers couldn’t deliver in time.

The stock closed at $531 and traded up roughly 3% to 6% after hours.

THE GUIDANCE

Results tell you where Microsoft was. Guidance tells you where it thinks demand is going, and every line points up.

Azure at 49% to 50% is the headline. Analysts had about 45% in their models. That’s a fourth straight quarter of acceleration, on a business already doing more than $130 billion a year, while it’s still short of capacity.

Hood pointed to two drivers: engineering work that squeezes more out of the CPUs and GPUs Microsoft already runs, and the third-party capacity that slipped into the second half.

Then there’s capex. Hood guided to about $170 billion for the calendar year, down from about $175 billion in July. Some coverage read that as fiscal 2027, but the math only works for calendar 2026: the four quarters through December add up to about $168 billion if the December quarter lands just above $50 billion.

Don’t read it as a cut in ambition. The July number was lower than April’s only because of an accounting change to building lives, and this one is lower because suppliers delivered late. The December quarter alone is guided above $50 billion, and spending accelerates from there.

The two numbers with the most execution risk are Azure at 49% to 50%, because it assumes capacity shows up on time, and the 65% cloud margin, because every point of Azure growth brings more depreciation. If third-party deliveries slip again, both are at risk at once.

THE $677 BILLION QUESTION

Read more

For two months the market has asked the same two questions about AI: where is the return on all this compute, and who is going to pay for it now that credit got expensive? Neocloud stocks fell, credit spreads blew out, and “bubble” was everywhere.

Then one of the biggest buyers of AI compute on the planet filed its S-1, and one chart answers both. In the first quarter of 2025, every dollar Anthropic spent on compute brought in 40 cents of revenue. By the second quarter of 2026 it brought in $2.08, a 5.2x jump in five quarters. Compute spending went up about six times over that stretch, and revenue went up 29 times.

That 5x is the most important number in this filing for the AI trade, and it’s what this piece is about. The $80.7 billion net loss got the headlines, but it’s almost entirely an accounting mark. and underneath it sits a company with an 85% gross margin, an operating profit in the latest quarter and $131.7 billion in the bank. The real question is how much of that is already in the price at $1.5 to $2 trillion, and what investors are actually buying.

THE NUMBERS BEHIND THE STORY

Revenue. Quarterly revenue went from $0.4 billion in Q1 2025 to $2.0 billion in Q4 2025, $4.7 billion in Q1 2026 and $11.6 billion in Q2 2026. That adds up to $4.6 billion for 2025 and $16.3 billion for the first half of 2026, 14 times the same half a year earlier. The draft leaves the September quarter blank, but it compares Q3 with the last quarter of 2024 and calls it 86 times bigger. and elsewhere it gives that 2024 quarter as $197 million. That implies about $16.9 billion for Q3, close to $68 billion annualized.

Profitability. When the filing regroups its costs by function, the cost of serving customers comes to $2.38 billion for the first half, an 85% gross margin on the company’s own split. For the full half the operating loss was $693 million, a -4% margin against -184% a year earlier, and the second quarter alone made roughly $974 million of operating profit. On the company’s adjusted measure, which mostly adds back stock compensation, first-half operating income was $1.6 billion.

The $80.7 billion loss. Nearly all of it is the price of success. $72.4 billion is the change in fair value of convertible notes, derivatives and SAFEs, and another $8.8 billion is other expense, most of it from agreements to buy notes from partners. As Anthropic’s value went up, the paper it sold to investors and partners before the IPO became worth more, and accounting books that rise as a loss. No cash leaves the company, though public shareholders will feel it as dilution when that paper turns into shares at the listing. In Q2 alone, $1 billion of operating profit sat under a $66 billion paper loss.

Cash. At June 30, Anthropic held $112.4 billion of cash and $19.3 billion of short-term investments, $131.7 billion in total, more than Amazon, Meta, Microsoft or Nvidia. It has raised $141.1 billion over its life and still holds most of it, and the first half generated about $3.9 billion of operating cash. The $127.6 billion stockholders’ deficit on the balance sheet is an artifact of $157.5 billion of convertible preferred stock that converts at the listing.

WHO IS ACTUALLY PAYING?

Read more

AI keeps getting cheaper to use while the chips behind it get more expensive to rent. The average price paid per million tokens has fallen to about $0.98, less than half its late-May high, while H100 rents rose about 3% over the same stretch, B200s about 10% and B300s almost 39%.

Then this month, the cost of insuring CoreWeave’s debt against default spiked. Its five-year credit default swap jumped from 368 to 1,194 basis points in nine days, and Nebius fell 10% to 12% in sympathy, as if the two were the same stock, even though their balance sheets look nothing alike.

Rental prices decide how this ends. Leverage on GPUs works while rents hold, and if they break, we find out what a dark GPU looks like.

_{SAME BOOM, TWO BALANCE SHEETS}

CoreWeave ended June with about $29.8 billion of net debt, up from roughly $6.6 billion at the end of 2024. Nebius sat near $0.5 billion and held net cash for most of 2025. Neither figure includes operating leases, which add $16.3 billion at CoreWeave and $1.5 billion at Nebius.

How the debt is built matters more than its size. At June 30, CoreWeave carried roughly $35 billion of gross debt:

Read more

But first, the macro, which matters even more now. Last week we saw that hedge funds cutting their exposure drove much of this drawdown, so this week the question is what would bring them back and, if possible, when.

MACRO OVERVIEW

Hikes. Next week’s meeting is no longer the risk. Markets give an October 28 hike only about an 8% chance, with roughly 92% on a hold. December is a different story: the odds of a quarter-point hike there climbed to somewhere between 71% and 78% this week as oil pushed inflation fears back up. The data didn’t help either. Jobless claims came in at 194,000, below the 200,000 expected, and the flash factory reading hit 56.0, the highest in 53 months. A strong economy with oil above $100 is exactly what keeps the Fed leaning toward one more hike.

Yields. The 10-year slid most of the week, from about 5.15% on Monday to a close near 4.97% on Thursday, the day chips sold off, before ending Friday right at 5.00%. That’s down from 5.11% the Friday before, and the 30-year eased to about 5.31% from 5.43%. Same pattern as last week: every time AI stocks get hit, money hides in bonds, and for long-duration growth stocks a 10-year back at 5% is one of the few things that went right.

Oil and diesel. Oil went the other way. Brent settled at $116.72 on Friday, its highest close since April 29 and up about 6% on the week, while WTI finished near $103. Most of the move came Thursday and Friday, after Houthi strikes on Saudi Arabia and reports that the US and Israel have drawn up plans for new strikes on Iran after the November 3 midterms. Diesel followed: the EIA national average rose to about $6.28 a gallon, its first weekly increase in four weeks, with California above $8.24.

Iran. Diplomacy and escalation are running side by side. Araghchi said indirect talks continue on a seven-day ceasefire proposal, with Pakistan and Qatar in the middle, while Tehran’s public red lines haven’t moved: sanctions lifted, threats ended and no uranium leaving the country. At the same time, Washington ordered non-essential staff out of Saudi Arabia, and Iranian officials have warned they would target undersea cables, data centers and AI facilities in the region if heavy fighting resumes. Trump has said the US won’t strike Iran before the midterms, the blockade of Iranian ports stays in place, and the reported post-midterm strike plans haven’t been approved.

So oil is back near a six-month high, December hike odds keep climbing and the war has a date circled after the midterms. The S&P 500 only slipped about half a percent on the week, while the Philadelphia Semiconductor Index lost another 3% and was down as much as 6.5% at one point on Thursday after Meta’s results. But here’s what the headlines miss: hedge funds are leaving, while the slower money kept coming in.

Nebius institutional ownership, shares held (thousands). Source: Fintel.io, as of October 21, 2026.

Institutions held about 180 million Nebius shares as of October 21, up roughly 70% since May and the highest level on Fintel’s chart. These numbers come from 13F filings that lag by up to a quarter, so they mostly show the summer and early fall rather than this week, but the direction is clear: the long-only money that sizes positions in years kept adding while the fast money rotated out, and that’s usually who’s still holding when the next leg starts.

THE REAL BOTTLENECK: THE COST OF MONEY

Jensen Huang put it plainly on Thursday: demand for neocloud services is “off the charts.” What’s scarce is the capital to build it, and that’s what the market is pricing right now.

Five-year credit default swaps on hyperscalers and chipmakers, basis points. Source: ICE Data Services, via Bloomberg.

This is what it costs to insure the debt of the companies building AI. Oracle’s five-year CDS went from about 144 basis points in January to roughly 320 now, a record. SpaceX sits near 240, Broadcom near 160, and Meta, Nvidia and AMD are all around 100 to 110, while Microsoft is still near 50. Every line is pointing up at the same time, so the market is charging more to fund the whole buildout, with the most leveraged builders paying the most.

CoreWeave implied yield (5-year plus CDS) vs. the Bloomberg high-yield telecom index during the 1998 to 2002 fiber boom and bust. Source: Fidelity Investments, Bloomberg, Haver.

At the riskiest end, CoreWeave’s five-year CDS finished the week near 1,130 basis points, past the 1,000 line the market treats as distressed. This chart lines up CoreWeave’s implied borrowing cost with high-yield telecom bonds from the late-1990s fiber boom: its implied yield jumped from about 9% to 16%, while telecom junk yields climbed from about 12% to 26.5% before that bust bottomed. The comparison scares people because in 2000 the first warning came from debt, months before the stocks cracked.

Meta gave us the clearest picture of the new math this week. Revenue grew 26% to $64.6B, but capex hit $34.9B in the quarter, free cash flow turned negative, at about -$2.1B, and Meta announced a new class of nonvoting shares, partly to help fund the AI buildout. When one of the strongest balance sheets on the planet starts inventing new ways to raise money, the cost of money becomes the story. Meta fell as much as 7.6% on Thursday and took chips down with it.

Here’s why I don’t think this turns into 2000. CoreWeave is about 4% of projected data center capacity, and its $35.1B of debt is about 1.7% of this year’s global data center investment, per Nomura. Most of the buildout is paid for by companies whose CDS still sits around 100 basis points or lower, and the Treasury Secretary reportedly told a closed-door investor conference this week that he doesn’t see a “systemic problem” with AI. If expensive money slows the weakest builders down, the buildout gets slower and more durable, which is a far better outcome for the suppliers than a 1999-style overbuild.

So the real bottleneck is who can borrow, and at what price, which is exactly where Nebius separates itself.

THE TRADE GOT CHEAPER, AND IT STILL ISN’T A BUBBLE

If financing is getting more expensive, why am I not worried about valuations? Because earnings kept running while prices stopped.

SOX price vs. implied forward earnings, January 2, 2025 = 100. Approximate reconstruction from forward P/E. Chart by Steve (@Future_Ripples), built with Astra from a Bloomberg forward P/E chart. Price data: Nasdaq via FRED.

Since January 2025, the SOX is up about 114%, while its implied forward earnings are up roughly 198%. Earnings nearly tripled while prices only doubled, so chips are cheaper today on forward earnings than when this run started. The chart is an approximate reconstruction, so treat the numbers as directional, but the gap is far too wide to be noise.

S&P 500 price path vs. the 1994 to 2002 analog, and tech forward P/E then vs. now. Source: Fidelity Investments, Bloomberg, FactSet.

The price path looks a lot like the late 1990s, and that’s where the bubble talk comes from. The valuations don’t. At the 2000 peak, tech traded at 58.6 times forward earnings. Today it’s 19.3 times, below even its April 2025 low of 21.6. The whole market tells the same story, with the S&P 500 at 18.5 times forward earnings, below its 5-year and 10-year averages and the cheapest since the April 2025 tariff low. A bubble is prices running ahead of earnings, and right now we have the reverse.

GPU rental rates vs. a six-year straight-line depreciation schedule, first observed rent = 100%. Source: Ornn Data, via a16z, as of October 19, 2026.

The other bubble argument is that GPUs lose their value in two or three years while being depreciated over six. Rents say otherwise: an H100 from 2022 still rents 54% above a six-year schedule, an H200 143% above and a B200 111% above, and even the 2020 A100 is 9% above its book value. If these chips were going obsolete early, rents would be falling below those dotted lines, and they’re going the other way.

The cost of funding the buildout went up, but what the assets earn went up faster, and that’s the opposite of a bubble.

TIMING THE REVERSAL

I can’t time the bottom, and anyone who says they can is guessing. But the clues point to us being a lot closer to a bottom than a top.

Maximum drawdown: Goldman Sachs high-beta momentum basket vs. the S&P 500. Source: Goldman Sachs.

Start with the disconnect: Goldman’s high-beta momentum basket is down about 41% from its peak, while the S&P 500 is barely 2% off its high. Funds hid in the megacaps and sold the fast, leveraged AI names, and a gap that wide is a positioning event far more than an economic one, since in a real recession the index would be falling too.

Share of semiconductors in the MSCI ACWI index by market cap and by risk capital. Source: Bloomberg Finance L.P., J.P. Morgan Flows & Liquidity.

This is why semis fall the hardest, as explained this week. They’re about 13% of the global index by market cap but, because they move so much more than everything else, about 20% of the risk budget. When volatility jumps, risk models force funds to cut and semis go first, whatever the fundamentals say. The same machine runs in reverse once volatility calms down, and the snowball rolls the other way.

Goldman Sachs Prime Book, Global Information Technology: cumulative 2-week net trading flow, standard deviations. Source: Goldman Sachs Prime Services.

And the selling has been historic. From October 5 to 16, hedge funds sold global tech at more than 3 standard deviations below normal, the heaviest two weeks of net selling in this data going back to 2017. Spikes this extreme lined up with the March 2020 and April 2025 lows. They don’t mark the exact day, since this summer’s spike came weeks before September’s final low, but they tell you most of the forced sellers are already out.

Goldman Sachs GIR Risk Appetite Indicator, October 2022 to October 2026. Source: Goldman Sachs Global Investment Research, via Neil Sethi (@neilsethinew).

And the fear has spread well past tech. Goldman’s Risk Appetite Indicator, built from moves across stocks, bonds, credit and other assets, fell to -0.52 after what Goldman called one of its largest drops on record, more than two standard deviations below its average and the lowest since April 2025. In the last four years it has only been this low three times: the SVB scare in March 2023, the yen carry unwind in August 2024 and the tariff crash in April 2025. Each of those landed close to a major low, although in 2024 and 2025 the reading kept falling to around -1.5 before it turned. It’s global too, with Hong Kong tech down 43% from its October 2025 peak. Broad, indiscriminate selling like this is what late-stage capitulation tends to look like.

Markets usually bottom on one last sharp flush, and I wouldn’t be surprised to see one before this turns. What would turn it is on the calendar: Microsoft, Alphabet and the Fed on October 28, Amazon on October 29, the midterms on November 3, and Nebius earnings expected around November 11. Strong capex guides from the hyperscalers, a calmer Fed or any credible step toward a ceasefire could each be the spark. None of that gives me the exact day, but we know what to watch.

WHY NEBIUS IS AN EVEN BETTER BET THAN IT WAS A FEW DAYS AGO

Nebius fell another 4% this week to $158.74, including a drop of as much as 6% on Wednesday, the same day it announced new financing. The market read a new loan as more leverage. I read it as the most important thing a neocloud can show right now: it can borrow, and cheaply.

The deal is a $3.5B senior secured facility with Citi as lead arranger. It runs 364 days and is priced at SOFR plus 2.75%, rising to a maximum of 3.50%. With SOFR around 3.9%, that’s roughly 6.7% today and about 7.4% at the cap. Compare that with the 8.5% to 9.75% coupons on CoreWeave’s unsecured notes, or the roughly 16% the credit market now implies for it. In a week where everyone is asking who can still fund the buildout, a bank just lent Nebius $3.5B at less than half CoreWeave’s market cost.

A one-year facility is a bridge, and management told us where it leads: asset-backed financing on more than $40B of long-term commitments from investment-grade customers. As the COO put it, “Sequencing our financing is how we optimize our long-term cost of capital.” The rest of the stack fits: the August convertibles carry 0.50% and 4.50% coupons, the July GPU facility is priced at SOFR plus 2.50%, the nearest material maturity is a small convertible due in June 2029, and about 70% of new contracts in the first half came with customer prepayments covering 50% to 60% of the capex. That’s a company being careful about how it pays for growth, which is exactly what the market is rewarding now.

Which brings me to the bigger point: not all neoclouds are created equal.

Neocloud financial health: net debt / EBITDA (left) and total debt / equity (right). Latest reported quarters; debt includes lease liabilities. Source: Sergey (@SergeyCYW), Compounding Your Wealth.

Most of the group is still burning cash. On the left, the red bars are companies with negative EBITDA, and only Nebius and CoreWeave earn the operating profit to service their debt. On the right is how much of each balance sheet is borrowed. As of each company’s last reported quarter, Nebius sat at about 1 times debt to equity, the lowest in the group, against 10 times for CoreWeave and 36 times for TeraWulf. That predates the August convertible and the Citi facility, so Nebius is closer to 2 times today by my rough math if the Citi line is fully drawn, still nowhere near CoreWeave. Both earn EBITDA, but one is funded mostly by shareholders and the other mostly by lenders, and that’s the difference the credit market is pricing.

To be honest, I’m usually not a big fan of this business model. Buying expensive hardware with borrowed money and renting it out by the hour can turn into the airline business very quickly once supply catches up. That’s exactly why the balance sheet matters so much, and why Nebius is the one neocloud I own in size. It earns money, it carries the least debt relative to its equity in the group, and when it does borrow, it pays about half of what its closest peer pays. Nebius is just built different. Its chief revenue officer put it simply this week: “We’re not carrying the kind of debt that they are.”

GPU rental prices, trailing 7-day average, January 1, 2026 = 100. Source: Ornn Data, via a16z, as of October 19, 2026.

Meanwhile, the thing Nebius sells keeps getting more expensive. Since January, rents on an H200 are up more than 160%, a B200 about 120% and an RTX PRO 6000 about 65%. Even the H100 from 2022 rents for about 35% more than it did in January, and the six-year-old A100 is still above where it started the year. Rising rents on capacity that’s already built drop almost straight into margin, and they make every one of those customer contracts easier to finance.

And the icing on the cake: the next leg is already lined up. Nebius expects 800 MW to 1 GW of connected power by the end of this year, coming online through the first half of 2027, and it plans to start deploying Vera Rubin late this year or early next, including the Meta deal worth up to $27B that starts delivering in early 2027. Management has said it could sell its entire planned 2027 capacity, and this week its chief revenue officer said that capacity would sell out if it were online today.

The risks haven’t gone away. A 364-day loan has to be refinanced or replaced within a year, the spread steps up if conditions change, and every raise keeps dilution in view. Customer concentration is still an open question, and as long as CoreWeave’s credit keeps making headlines, Nebius will trade with it. I size for those risks rather than pretend they don’t exist.

Put the week together: the share price fell while the company proved it can borrow cheaply, its assets kept getting more valuable and the slow money kept accumulating. That’s why Nebius is an even better bet than it was a few days ago.

MY PORTFOLIO

Read more

Add up the 2026 capex of the big US clouds, the neoclouds and China’s three largest platforms and you get about $1.01 trillion, roughly double 2025’s $510 billion. Next year it climbs to $1.39 trillion, then $1.58 trillion in 2028. (Figures from Bank of America’s capex tracker, a mix of its own estimates and consensus.) That’s total capex for these companies, not a trillion dollars of pure AI hardware, but AI is the reason those budgets doubled.

The revisions are the bigger story. A year ago, the Street had the core cloud group (excluding leases and neoclouds) spending about $496 billion in 2026. Today that estimate is $859 billion, up 73%, and it rose in every update along the way, while the 2027 number more than doubled.

Google, Microsoft, Meta and Amazon report in the coming weeks, so it’s worth seeing who is actually writing the checks.

_{WHO IS WRITING THE CHECKS}

Amazon leads 2026 at $222 billion, followed by Google at $202 billion and Microsoft at $177 billion, while Alibaba, Tencent and Baidu add $54 billion combined. By growth rate, Oracle leads at 136%, then Google at 120%, Meta 98%, Amazon 68% and Microsoft 50%.

At that pace, Google overtakes Amazon in 2027 at 313 billion, more than the whole group spent in 2024, China included (293 billion). In total, the group adds about $500 billion of spending this year and another $377 billion in 2027. Microsoft, Amazon and Meta include finance leases.

_{THE NEOCLOUDS ARE NOW A REAL ITEM}

Neocloud capex jumps from $31 billion in 2025 to $132 billion in 2026, more than 4x in one year, then rises to $174 billion in 2027 and $234 billion in 2028, about what Meta spends that year.

Three names account for about 85% of the 2026 total:

• xAI: $55B in 2026, $72B in 2027, $117B in 2028

• CoreWeave: $33B, $44B, $45B

• Nebius: $24B, $33B, $43B

That gives chip suppliers a second class of buyers big enough to move the order book. The catch is concentration: xAI alone is over 40% of 2026 neocloud spend, and CoreWeave’s budget flattens after 2027 in these estimates.

_{CAN THEY AFFORD IT?}

Start with capex intensity, the share of revenue that goes back into capex. For the big US five it rises from 28% in 2025 to 48% in 2026 and 57% in 2027, then eases to 52% in 2028. At the peak, more than half of every revenue dollar turns into chips, concrete and copper. Meta runs hottest at about 70% in 2027, while the neoclouds spend roughly $2 for every $1 of revenue this year and stay above 100% through 2027.

Cash flow is where it bites. Excluding neoclouds, capex runs at 106% to 109% of operating cash flow through 2028, well past the 65% to 70% telecom carriers reached at the peak of 4G. That pushes the group’s free cash flow negative by about $54 billion this year, $89 billion in 2027 and $113 billion in 2028.

To fund it, the five largest spenders have raised about $370 billion this year, mostly through long-dated bonds. That total includes a $90 billion Alphabet equity offering in June.

My read: these companies are deliberately spending more than their operations bring in, pulling capacity forward with cheap, long-term money. Still, keep the telecom comparison on your desk, because this buildout runs hotter than that one ever did.

_{IS THE DEMAND REAL?}

The first check is backlog. Remaining performance obligations (RPO) are revenue customers have already signed for that hasn’t been booked yet. Set against cumulative 2026 to 2028 capex (each company defines backlog a little differently, so read this as directional):

• Microsoft: $679B of capex against $678B of RPO

• Oracle: $279B against $638B

• Google: $894B against $514B

• Amazon: $834B against $496B

Together those four carry more than $2.3 trillion of RPO, with three caveats. Not all of it is AI, Oracle expects only about 12% of its backlog to convert within a year, and some of it leans on one customer: OpenAI was about 45% of Microsoft’s $625 billion commercial RPO as of December 2025. I left Meta out on purpose, since its $349 billion figure is what Meta owes suppliers, not what customers owe Meta.

The second check is price. If the world had too many GPUs, rental rates for the newest chips would crack first, and they haven’t. A B200 rents for about $5.88 an hour, near its spring peak of $6.10 and up about 28% this year, and H100s have climbed to $2.82. The two-generation-old A100 has slipped to $1.52. That’s normal aging, but it’s the number depreciation bears are right to watch, especially for neoclouds running older fleets.

_{CHEAPER TOKENS, MORE OF THEM}

A token is roughly a word fragment an AI model reads or writes. The estimated market price per million tokens has dropped to about $0.99, down from a June peak of $2.07 and below its $1.49 median. Open-weight models now handle around 60% of US tokens on OpenRouter, at prices about 80% below the proprietary labs.

The bear reading is that cheaper AI needs less infrastructure. I read it the other way: when tokens get cheaper, people use far more of them, and agents are why the jump is so steep. A chatbot answers once, while an agent searches, reasons through many steps, calls tools, runs code, checks its work and retries, burning 10x to 1,000x the tokens of a single chat. Google’s AI infrastructure chief, Amin Vahdat, has said Google needs to roughly double its serving capacity every six months.

The risk sits with the labs. Cheaper tokens squeeze their margins, and they’re among the biggest customers paying for all this capacity.

_{WHERE THE MONEY LANDS}

Read more

Prometheus is Meta’s AI supercluster in New Albany, Ohio, and the plan is to grow it to roughly a million GPUs wired together to run as a single machine. It already spans hundreds of thousands of accelerators across several buildings, and its first phase is training Meta’s newest models today.

The GPU count will grab the headlines, but Meta’s own engineering write-up spends surprisingly little time on the chips. Most of it is about keeping them busy: moving data between them, swapping out the ones that fail, feeding them power without rattling the grid and pulling the heat back out.

At this scale a data center stops behaving like a server room and starts behaving like a factory, with supply lines, spare parts, its own power problem and a full maintenance crew. Each of those needs a supplier, and that is where AI spending spreads once the GPUs are bought.

_{LESS SPACE, MORE POWER}

Before getting to everything around the chips, it helps to know what Meta is installing. Prometheus runs two generations of Nvidia rack-scale systems today, GB200 and GB300, with Vera Rubin arriving next. Each one carries its own Meta codename.

The power row tells the story. The same 72 GPUs went from filling two racks to fitting in one while their power draw kept rising, so each generation puts more watts, more heat and more cabling into every square foot of floor.

On memory, the upgrade is bandwidth rather than capacity. Rubin keeps 288 GB per GPU, the same as GB300, but moves to HBM4 with 19.2 TB/s of peak bandwidth per GPU, and that speed is what keeps a chip from sitting idle while it waits for data.

Meta’s own figure for GB300 over GB200 is a modest 5 to 10%. Real fleets rarely match keynote multiples, and closing that gap is what the rest of Meta’s write-up, and the rest of this piece, is about.

Read more

Longer answer, we need a macro detour first, because almost everything that hit us this week started outside the AI trade.

MACRO OVERVIEW

Hikes. Last week the October hike had moved to December. Wednesday’s CPI tried to drag it back. September inflation came in at 3.7%, above the 3.6% expected and up from 3.4% in August, with core at 2.6% against 2.5% expected. Gasoline jumped 9.1% in the month and did more than half of the work. Odds of a hike at the October 28 meeting doubled to about 42% within hours, then faded to somewhere between 10% and 16% by Friday, while December is now about 89% priced. An October move would be awkward for Warsh: the meeting lands six days before the election, and the Fed has hiked that close to Election Day only once since 1990.

Yields. Here’s the surprise of the week. A hot CPI should have pushed long rates higher, and instead the 10-year slid to about 5.05% midweek and closed Friday near 5.11%, down from 5.26% the Friday before. The 30-year is back around 5.43% after spiking to 5.70% last week. That’s money running into bonds for safety while stocks sold off, which tells you this week was driven more by fear than by the Fed.

Oil and diesel. Brent touched $110 on Friday for the first time since May and settled at $109.86, up 4.9% on the week, while WTI jumped 10.4% to $101.41. Two weeks ago Kpler had Hormuz flows back above pre-war levels, and by October 13 they had dropped to about 68% of normal. On top of that, Ukraine hit another Russian refinery, Ryazan shut down a crude unit, and traders started doubting how much diesel Russia can actually deliver after Washington temporarily lifted sanctions on it. The IEA agreed to release more than 20 million barrels of diesel from emergency reserves, and the US average still ticked up to about $6.33 a gallon, not far from September’s $6.53 record.

Iran. The calm lasted about a week. Iran said it was stepping up strikes in response to the US blockade of its ports and new sanctions, two cargo ships were sunk in the Strait, and a missile aimed at Yanbu, the Saudi port that lets crude bypass Hormuz, was intercepted. The blockade keeps tightening too: as of October 11, US forces had redirected 135 vessels and destroyed 10 IRGC-linked tankers. On the diplomatic side, Araghchi said on Saturday that Iran is open to talks but wants “credible guarantees” against new attacks, pointing back to June’s Islamabad understanding as the template. Trump said last week he won’t strike Iran again before the midterms.

So we got a hot CPI, a hike that flirted with October, oil at $110 and an Iran that’s escalating again. Even so, the S&P 500 only lost 2% on the week, while the Nasdaq 100 had its worst week in about a year and a half and the Philadelphia Semiconductor Index put together its longest losing streak since 2008. Macro set the mood, and the reason AI fell so much harder than the index shows up in who was selling.

Hedge fund flows into global AI stocks, % of gross market value. Source: Goldman Sachs Prime Book, as reported by Seeking Alpha.

This tracks hedge fund flows into global AI stocks on Goldman’s prime brokerage book. Their net buying, the black line, peaked around February and drifted lower all summer. Then, in five straight sessions from October 8 to 14, hedge funds erased every bit of net AI buying since the start of 2025. Most of it was funds dumping longs, about 3.5 times more long selling than new shorting, and the heaviest single day ranked second only to July. Short sellers kept pressing the whole way down.

Goldman’s own read is that this deleveraging could be closer to the start than the end, since gross leverage is still near five-year highs, and that’s the risk I’m watching. The other side of it is that net exposure to AI now sits at the 6th percentile of the past year. Once the fast money has sold, the people left holding are mostly there for the fundamentals, and washouts like this tend to set up the rebound, even if they rarely end on a schedule.

Flows explain the speed of a drop and say very little about whether the reason to own these names has changed. For that, the better question came from CoreWeave, which lost about 25% this week as the market started asking who is paying for all this debt. So let’s zoom out.

THE AI TRADE, ZOOMED OUT

This is UBS’s model for Nvidia’s revenue. Nvidia’s fiscal year ends in January, so FY27 is basically calendar 2026. UBS has revenue nearly doubling to about $420B this fiscal year, growing another 77% to $744B the year after, then topping out near $985B in FY29, which is mostly calendar 2028, before slipping 5% in FY30.

Two things stand out to me. The growth rate already peaked: +95% this year, +77% next year and +32% after that, so the dollars keep climbing long after the percentages start to cool. And the first down year in the model is FY30, which is mostly calendar 2029. If the AI trade were over, the biggest supplier of the buildout wouldn’t be modeled to add another $565B of revenue over the next two years. Whether 2029 turns into a plateau or a cliff is the real debate, and that debate is three years away.

Demand is where bulls have the easiest argument. OpenRouter routed 379 trillion tokens in August, 25 times what it handled a year earlier and 12 times January’s level. OpenRouter is one router that skews toward developers and open models, so treat it as a sample of the market, but the direction matches everything else we see. It’s also why Saturday’s DeepSeek headline didn’t worry me: every time models get cheaper, the token count climbs.

And this is what that demand looks like inside the two biggest labs. Combined compute at OpenAI and Anthropic went from about 0.6 million H100-equivalents at the end of 2024 to about 2.9 million at the end of 2025, almost 5 times in a single year. Inference alone grew roughly 4 times, from about 0.23 million to 0.94 million. The 2026 bars are an illustrative 3x scenario rather than a forecast, so read them as a sense of scale.

The part I care about most is where that compute goes. Pre-training is shrinking as a share, while post-training and inference now take most of the capacity, and both lean heavily on memory bandwidth. More inference means more GPUs rented by the hour and more HBM packed into every one of them, which is good news for both my neocloud and my memory names.

Spending is still going up, just more slowly. Consensus has the five biggest spenders at around $800B this year, and Goldman sees roughly $1.2T in 2027 and close to $1.4T in 2028. Bloomberg has the big four alone on track for about $750B in 2026, more than half of what the entire S&P 500 spent last year. A slower climb still means a bigger check every year, and every dollar of that check lands on someone’s income statement: GPUs, memory, optics, power and the neoclouds renting it all out.

Now the part the bears are right to point at. Hyperscalers used to turn about 100% of their earnings into free cash flow. On BofA’s numbers that’s now below zero for the next twelve months, while the rest of the S&P 500 sits around 85%. They’re spending more cash than they make, and more of the gap is being filled with debt. That’s why the most leveraged names, CoreWeave first, got hit the hardest this week.

My read is that this is a financing question far more than a demand question. Demand has never looked better, and most of the spend comes from companies with some of the strongest balance sheets on the planet. What the market is repricing is who carries the debt and how fast it gets paid back. That’s exactly the lens I use for my biggest position.

WHY NEBIUS IS STILL MY BIGGEST BET

Nebius took one of the hardest hits in the group, down about 27% on the week and closing Friday at $165.47. Part of that is guilt by association: when CoreWeave drops 25% on debt fears, every neocloud gets sold with it. Part of it is fair. Nebius needs roughly $250B to build out 5 GW and has secured about 20% of it, it carries about $2B of net debt after a $5B convertible. If financing dries up, Nebius will feel it.

So why is it still the name I have the most conviction in? Two charts.

B300 GPU rental price index, $ per hour. Source: Bloomberg.

First, pricing power. Renting a B300 hit an all-time high of $7.45 an hour on October 14, up 64% from its $4.55 low last November, after 17 straight days of gains. Nebius raised its own GPU prices by 17% to 21% on October 1. A company that can raise prices in the middle of a selloff isn’t selling a commodity, and every increase drops almost straight into margin on capacity that’s already built.

Second, customers are helping pay for the buildout. On the Q2 deals, prepayments covered 50% to 60% of the capex, Nebius expects those deals to pay back in about 1 year and 10 months, and the core AI cloud already runs at about 50% adjusted EBITDA margins. Borrowing first and hoping tenants show up is a very different business, and it’s why I don’t lump Nebius in with CoreWeave’s debt story.

Then there’s the growth. Nebius exited June at a $3.0B run rate. William Blair has it at about $8B by December, $16.8B in 2027, $30B in 2028 and $72B by 2030, and the slowest year on that chart still grows 41%. Inference adds another leg: Token Factory workloads more than tripled in Q2, the same token wave from the chart above landing on Nebius’s own platform.

Now put that next to the price. At Friday’s close, the whole company was worth about $41B, a little over half of that 2030 run rate.

The risks are real, and I want you to see them too. Hitting the December number means ARR has to grow about 2.7 times in six months, and William Blair’s model is an estimate, not company guidance. The payback math includes capacity that isn’t built yet, and if the market keeps punishing anything that borrows, the next raise will cost more. I’d rather know those risks and size for them than pretend they don’t exist.

That’s why Nebius stays my largest position. A week like this one tests conviction, and the numbers that built mine are still moving the right way: prices are rising, customers are prepaying and the run rate keeps compounding. A lower share price only makes that math more attractive.

MY PORTFOLIO

Read more

Ask most investors where the AI capex boom goes and they will say NVIDIA. But in this illustrative $100 AI infrastructure buildout, accelerators (GPUs and custom AI chips) account for $25. That is only a quarter of the modeled spending, although NVIDIA also participates in other categories, including networking and CPUs.

Think of an AI data center as a factory. The accelerator is the machine on the floor, but it still needs memory to feed it, a network to link it to the other machines, electricity, cooling and a building to sit in. That is where the other 75 goes.

_{THE 100 DOLLAR BUILDOUT}

Half the modeled spending goes to semiconductors and related server hardware, and the other half to infrastructure that makes the systems run. Below, each bucket is opened one level down.

Read more

Cerebras listed on May 14 at $185, opened at $350 and hit about $386 in its first session. Five months later it trades around $146, about 21% below the IPO price and more than 60% below that first-day high.

The reasons are easy to list. Gross margin guidance disappointed after the first public quarter, staggered lockup releases keep putting pre-IPO shares on the market, senior executives have filed notices to sell stock, and the business still leans heavily on one customer.

It’s easy to look at that and call it a broken IPO. Some of those worries, like the unlocks, fade within weeks. Others, like customer concentration, will be with us for years, and I’ll get to them. But the price is reacting to the next few months, while the numbers that matter play out over the next few years, and UBS just put those numbers on paper.

WHAT UBS IS MODELING

One thing to know upfront: UBS was one of the underwriters on the Cerebras IPO and rates the stock Buy, so treat its numbers as the optimistic end of the range.

UBS has Cerebras going from $890M in revenue this year to $3.3B in 2027 and ~$22.6B by 2030. Over the same stretch, operating margins swing from −18% to 46%, and net cash grows from $5.4B at the end of this year to $12.7B by 2030.

That’s roughly 25x revenue growth in four years. It sounds aggressive, but about half of it is already under contract, which I’ll get to in a moment.

THE VALUATION MATH

Using UBS’s diluted EPS, here is what you pay at $146:

· 2027: about 79x earnings of $1.84, its first year of UBS-adjusted profit

· 2028: about 20x earnings of $7.21

· 2029: about 10x earnings of $14.69

· 2030: about 6x earnings of $24.36

UBS built its model on a $196.74 share price on September 28. Two weeks later the stock is about 26% lower, with the same estimates underneath.

WHAT COULD IT BE WORTH?

Here’s my math on UBS’s estimates. I priced each year-end off the following year’s EPS and let the multiple come down over time, since growth slows from EPS doubling in 2029 to +66% in 2030. These are illustrative scenarios, not price targets.

Read more

The memory bear case is decades old: prices rise, manufacturers build, supply floods in, and the cycle ends. It has worked so often that most investors stopped questioning it.

Memory is Layer 6 in my 12-layer map of the AI buildout, and four numbers explain why I keep it heavy. These are J.P. Morgan’s estimates of annual demand growth from 2026 to 2028:

• AI server memory: +92%

• HBM: +63%

• Other server DRAM: +37%

• PC and mobile: -3%

If you are new here: DRAM is the working memory inside every server, phone and PC, HBM is a stacked, faster DRAM that sits next to every AI chip, and NAND is storage.

Phones and PCs drove every past memory cycle, and that demand is now shrinking. The other three lines are growing faster than the industry can add capacity.

_{FOUR REASONS THE CASE GOT STRONGER}

Read more

NVIDIA built the most vertically integrated platform in this industry. GPU, CPU, networking, software, the rack, even the reference design for the building. If you wanted the best AI system of the last three years, you bought it the way NVIDIA designed it.

Today at the Open Compute Project summit in San Jose, NVIDIA handed the industry the rack and compute-tray designs for Vera Rubin, while Meta, Google and Cisco all showed up with announcements built around NVIDIA hardware.

On first read, that looks like a company giving away part of its edge. I read it the other way, and NVIDIA’s spokesperson summed up the logic in one line during the press briefing: the platform is “vertically integrated, but horizontally open.”

Keep that phrase in mind.

_{NO, NVIDIA DID NOT OPEN-SOURCE THE GPU}

First, what NVIDIA didn’t do: it didn’t give away the Rubin chip. What it contributed to OCP is the hardware around it, bundled into a spec called MGX 1.2:

• Compute trays with no fans, cables or hoses. Everything is liquid cooled and modular, which NVIDIA says cuts tray assembly from hours to minutes.

• A liquid-cooled 5,000-amp busbar. The busbar is the metal spine that carries power down the rack, and cooling it lets it carry far more current in the same space.

• Blind-mate liquid connectors. They seal on their own when a tray slides in, which makes swapping a cold plate simpler and lowers the risk of leaks.

• Standard power, cooling and mechanical interfaces for racks of up to 250kW, many times what a typical enterprise rack draws.

Put simply, NVIDIA kept the recipe and published the exact dimensions of the kitchen. Suppliers can now design the ovens, the plumbing and the wiring to fit, as long as they build to NVIDIA’s measurements.

_{WHY WOULD NVIDIA HAND THIS OUT?}

NVIDIA got that exact question on the briefing call. Its answer: the industry’s bottleneck is now how fast AI factories can be manufactured and deployed. With a public spec, equipment makers can start building right away instead of waiting for each customer to finalize its own design, so lead times shrink.

That’s the official reason. My read goes a step further.

Anyone who followed the Blackwell ramp saw the rack, not the chip, become the hard part. Busbars, manifolds, cold plates, power shelves and optics all have to show up on time, and every hyperscaler designing its own version of each one slows the whole chain.

A common spec helps the supply side, since a supplier can engineer one design and sell it to many customers instead of redoing it for each one. It also gives NVIDIA a growing ecosystem already tuned to its systems. That doesn’t lock rivals out, but it makes NVIDIA the easiest platform to deploy at scale.

To be fair, this isn’t new. NVIDIA has contributed designs to OCP for about 13 years. What’s different now is how much of the system it’s handing over and how many suppliers are building to it.

_{AI AGENTS NEED MORE THAN GPUS}

The second announcement is the Vera CPU rack, which NVIDIA says is already shipping.

A chatbot waits for your question, runs the model and answers. An agent keeps working between model calls: it writes and runs code, calls tools, pulls data, spins up sandboxes and checks its own output. The GPU does the thinking, but most of that other work runs on CPUs.

NVIDIA’s answer is a dedicated CPU rack that sits beside the Vera Rubin NVL72, leaving the NVL72 to focus on training and inference. Both are built on the same MGX spec, so they share power, cooling and mechanics.

Take the performance numbers with a grain of salt, since they’re NVIDIA’s own. It claims roughly 2x over Grace on agentic workloads like scripting, compression and code compilation, and 1.8x on Terminal-Bench, which runs real agent tasks in real containers. For context, NVIDIA says CPU core performance has improved about 10% a year over the past seven years. Partners report similar gains: about 2x for Meta on its hyperscale cloud benchmark and for the Texas Advanced Computing Center on its HPC suite.

The bigger point for investors: if agents shift more work back to the CPU, NVIDIA gets to compete for spending beyond the accelerator rack. Winning it is a separate question, since hyperscalers can run that work on x86, their own Arm chips or custom silicon.

_{META PICKS NVIDIA SILICON, KEEPS ITS OWN SOFTWARE}

Meta is adopting Spectrum-6, NVIDIA’s 102.4 Tb/s Ethernet switch chip, to connect its Vera Rubin clusters. That’s 128 ports of 800G, twice the capacity of Spectrum-4.

Meta’s Spectrum-6 switch, N-IceCube, slots into IcePack, its liquid-cooled network rack, and keeps running FBOSS, Meta’s own network operating system. NVIDIA is fine with that. It’s loosening its grip on the software to tighten it on the silicon underneath.

Spectrum-6 also enables multiplane networking, which splits GPU traffic across up to eight separate highways instead of one. When one closes, the network card catches it in hardware and reroutes in under three milliseconds, which NVIDIA says is 30x faster than software-based approaches. With 100,000-plus GPUs something is always failing, so this can decide whether a training run finishes on schedule.

_{CISCO AND NVIDIA SHRINK THE OPTICS PROBLEM}

Read more

MACRO OVERVIEW

Hikes. The October hike is still mostly off the table, but it didn’t go away, it just moved. Odds for the October 27 to 28 meeting sat around 17 to 20% all week, while a December hike is now about 86% priced, on top of September’s move to 3.75 to 4.00%, the Fed’s first hike since 2023. Wednesday’s CPI (October 14) is the last inflation print before the meeting, so that’s the number to watch this week.

Yields. Long rates did the damage this week. The 10-year touched about 5.31% on Wednesday and the 30-year briefly hit 5.70%, its highest level in roughly 24 years, before both eased into Friday with the 10-year closing near 5.25%. Mortgages are back around 7.4% and consumer sentiment dropped to 46.3, its lowest reading since May. Even so, the S&P 500 finished the week up 1.2% and the Nasdaq up 0.6%.

Oil and diesel. Brent finished the week near $104, up about 2%, as tanker attacks around Hormuz kept a risk premium on crude. It slipped toward $103 on Monday after Trump announced an “energy ceasefire” between Russia and Ukraine and said Russia would supply diesel to the US and global markets. Kpler’s provisional data shows flows out of Hormuz back above pre-war levels at the end of September. At the pump, gasoline held around $4.36 and diesel eased to $6.28, down from its $6.53 record on September 22, but fuel is still biting, and Delta cut its guidance this week after its fuel bill rose 62% from a year ago.

Iran. Talking and shooting at the same time, again. President Pezeshkian called negotiations with the US pointless on October 5, while Qatar says mediation continues and Vance says any deal requires Iran to cut its enrichment capacity. At sea it got worse: at least nine tankers were hit in and around Hormuz in the first week of October, and UKMTO raised its risk rating to “severe.” On October 9, Trump said he won’t strike Iran again before the midterms, citing “productive discussions with Tehran,” and by Monday Iran’s foreign minister had reportedly reached out to Riyadh about de-escalation after Houthi attacks on the Saudi capital’s airport.

So in one week the hike moved to December, long yields hit a 24-year high, oil firmed and Iran kept negotiating through a wave of tanker attacks. On paper that’s a rough week for growth stocks, and the AI bucket did take the hit on Thursday, when the FT reported that OpenAI was “only” approaching $50B in annualized revenue and AMD fell 4% while Micron dropped 4.3%. Funny how fast the story changes. A few hours later, Bloomberg reported that OpenAI expects to reach $70B in annualized revenue by the end of the year, so the number people panicked about OpenAI not having is the number it expects to hit in less than three months. The FT story was really about how revenue gets counted, and by Friday the indexes still closed the week green.

Let me show you why I think the market underneath is in much better shape than the headlines.

EARNINGS ARE DOING THE WORK

This is the S&P 500’s forward P/E. It peaked near 23x at the end of last year and it’s back to about 19x today, roughly where it bottomed when the Iran war started, even with the index sitting near its highs around 7,800. Prices didn’t fall. Earnings rose fast enough to drag the multiple down with them.

A bubble works the other way, with people paying more and more for the same dollar of profit. Today the market is paying less for every dollar of earnings than it did a year ago, after one of the strongest rallies we’ve seen.

And earnings season starts this week, with the big banks reporting on October 13. Deutsche Bank points out that the S&P 500 has risen in 21 of the last 23 midterm cycles, measured from one month before the election to two months after. We’re inside that window right now.

AI CLOUDS AND COMPUTE PRICES

The AI bucket sold off this week on a headline about how OpenAI counts its revenue. Meanwhile, the data on actual compute demand keeps saying the same thing it has said all year.

Between June and August, the cost of tokens roughly halved, while the price to rent an H100 went the other way, from about $2.30 to about $2.75 an hour. This is Jevons paradox playing out in real time. When intelligence gets cheaper, people use far more of it, and every one of those extra tokens still needs a GPU to run on.

That demand shows up in what older chips are worth. An H100 rents for about $2.63 an hour today, roughly where it was in April 2025 and about 30% above its low late last year, even though it’s a generation behind Blackwell. The H200 went from about $2.70 in May to $3.29, and the B200 is back at $5.69 after bottoming around $4.40 in January.

The one exception is the A100, whose residual value keeps sliding, which is what a six-year-old chip should do. Everything from Hopper onward is holding its value far better than the “GPUs are worthless in three years” crowd expected, and that matters a lot for the neoclouds’ depreciation math.

And this is where all that demand lands. CoreWeave reached about $2.6B in quarterly revenue in roughly 26 quarters, faster than Azure, AWS or Google Cloud at the same age, and Nebius is already at around half a billion a quarter just seven quarters in. Zoom out to the left chart and both are still tiny next to AWS at roughly $43B a quarter, Azure near $30B and Google Cloud near $25B.

The supply chain is telling the same story. Lumentum’s CEO said on Bloomberg this week that the company is “completely sold out” through early 2029 with “no end in sight,” and that hyperscalers are now sharing the cost of new capacity through long-term commitments.

Cheaper tokens drive more compute, more compute keeps older GPUs valuable, and the companies renting them out are growing faster than the hyperscalers ever did. One headline about revenue accounting doesn’t change any of that.

MY PORTFOLIO, NOW EASIER TO FOLLOW

Read more

AMD lives in Layer 5 of my framework, compute, and now reaches up into Layer 3 with its software and down into Layer 7 with its networking. On the August 4th earnings call, Lisa Su said “agentic AI is creating a new growth vector for server CPUs”.

A chatbot answers one question and stops. An agent takes a goal and works on it for hours: it reads a long brief, plans, calls tools, runs code, checks the result and tries again. The heavy thinking still happens on GPUs, but much of the work around it, such as running tools, coordinating steps, retrieving data and executing code in a sandbox, runs on ordinary CPUs. The context an agent carries needs a lot of memory, and many steps depend on the output of the previous one, so delays compound across the workflow.

That is why comparing one AMD GPU against one rival GPU stopped being useful. The biggest buyers now order a rack of 72 GPUs wired to behave like one computer, with CPUs, networking, cooling and software included, and they commit to it years in advance.

AMD’s version is Helios, and it stacks five things AMD owns or designs:

AMD designs Helios rather than manufacturing and selling the complete rack itself. Its product page calls it a reference design, a blueprint that partners like Sanmina, Celestica, Wiwynn, Dell and HPE build and sell under their own names, while AMD sells the chips inside. That explains the ZT Systems deal: AMD paid $4.4B for ZT in March 2025 and seven months later sold the manufacturing arm to Sanmina for $2.4B in cash plus stock. It kept the architects and sold the factory.

Data Center revenue was $6.7B in Q2, up 107% year over year, and 58% of company revenue, against 42% a year earlier. AMD is now a data center company that also sells PC chips.

_{SECTION 1 COMPANY SNAPSHOT}

_{How AMD went from Intel’s problem to NVIDIA’s}

You only need three eras.

• Survival (2012 to 2016). Su joined in 2012 and became CEO in October 2014. The stock traded under $2 in early 2016. The company bet almost everything on one new CPU core, Zen.

• Share capture (2017 to 2022). Zen worked, and EPYC turned Intel’s server monopoly into a real fight. The Xilinx deal (closed February 2022) added FPGAs and adaptive chips, which became the Embedded segment. Pensando added data center networking.

• Systems (2023 to now). MI300 got AMD into AI accelerators, ROCm matured enough for the frontier labs, ZT’s design team arrived, and Helios pulled it all into one product.

Today AMD reports three segments:

Gaming is shrinking as the console cycle ages, Client is taking share in a softening PC market, and Embedded just posted its best growth in three years. Data Center is where the valuation lives.

_{SECTION 2 FUNDAMENTALS}

_{How AMD makes money}

A single Helios rack has five AMD pieces in it. Three are sold and two exist to sell them.

• The accelerator. Instinct MI455X does the heavy AI math. A rack holds 72, each with 432GB of HBM4 memory.

• The host CPU. EPYC “Venice” (Zen 6, 2nm, up to 256 cores) feeds and coordinates the GPUs. A rack holds 18.

• The network. Pensando “Vulcano” AI NICs move data between racks at 800 Gbps, and Salina DPUs handle storage, security and front-end traffic.

• The software. ROCm is free, because its job is to make the hardware usable.

• The system. The Helios design is a blueprint AMD hands to partners so they build racks around its chips.

The CPU is also a business of its own, and agents are why it re-accelerated this year. Management splits server CPU demand into three jobs: host chips that sit next to GPUs, the general-purpose servers every cloud already runs, and a new one it calls agentic sandboxes. A sandbox is a walled-off server where an agent runs code and calls tools without touching anything else, and production agents increasingly need one. More agents means more sandboxes, and more CPU cores per GPU than chatbots ever needed. Server CPU revenue has set records for five straight quarters, cloud and enterprise each grew more than 70% in Q2, and management expects server revenue to grow more than 80% year over year in the second half of 2026.

AMD is the clearest merchant alternative to NVIDIA, with competitive CPUs, AI accelerators and data center networking under one roof, and Helios is the first product where it designs them as one system.

_{The management team}

Lisa Su has been CEO since October 2014 and Chair since 2022. She holds an MIT PhD in electrical engineering and spent 13 years at IBM’s semiconductor R&D before Freescale. Jean Hu is CFO, Forrest Norrod runs Data Center Solutions, and Mark Papermaster is CTO. When World Labs closes, Fei-Fei Li joins as EVP and chief scientist reporting to Su.

The board has real semiconductor depth: Abhi Talwalkar ran LSI and chairs Lam Research, and John Marren was a top-ranked semiconductor analyst before TPG and Temasek. It also recently added two finance heavyweights, KC McClure (former Accenture CFO) and Tim Ryan (former PwC US chair). That matters for a company now writing warrants, lease guarantees and equity commitments.

For transparency: Su, Hu, Norrod and director Joseph Householder all adopted pre-scheduled 10b5-1 selling plans in Q2. Su’s covers up to 450,000 shares through June 2027. That’s normal after a run like this, but you should know it’s there.

_{Who actually buys from AMD}

Oracle, HUMAIN, Vultr, TensorWave and Cirrascale are also on AMD’s Helios partner page, and EPYC already runs across AWS, Google Cloud, Microsoft and Oracle. Anthropic is using Claude to help optimize ROCm, which matters in Section 4.

A year ago the question was whether hyperscalers would run Instinct at all. Now demand is measured in gigawatts, and management sizes revenue at “double-digit billions” per gigawatt. The last column is the catch: each of the three anchor deals came with AMD giving something back, and Section 7 puts a number on it.

_{SECTION 3 FINANCIAL PROFILE}

_{The AI business pays for itself}

AMD pays for its AI buildout out of a profitable business, with roughly $10B of net cash and no need to raise money. For Q3 it guided about $13B of revenue (±$300M) at a 56% gross margin and $3.65B of operating expenses, which by my math implies non-GAAP EPS of about $1.93.

Three things in the filing sit behind the headline numbers.

GAAP earnings got a one-time lift. GAAP EPS of $1.38 includes $483M of gains on investments in companies that went public during the quarter, about $0.29 a share. Non-GAAP strips that out, and it shouldn’t be annualized.

Cash flow leaned on suppliers. First-half operating cash flow of $5.3B included a $2.2B increase in accounts payable, money AMD owed suppliers and hadn’t yet paid. Receivables and supplier prepayments each rose about $1B. Free cash flow fell from $2.57B in Q1 to $1.56B in Q2 as capex nearly tripled year over year ($808M vs. $282M), and first-half buybacks slowed to $221M from $1.2B a year earlier.

The commitments are growing faster than the balance sheet. AMD has $30.3B of unconditional purchase commitments ($17.4B due in the rest of 2026), $4.5B of data center leases not yet started, $9.5B of new long-term data center leases signed after the quarter, up to $4.1B of guarantees on partners’ data center leases, and the up to $5B Anthropic commitment. That is about $53B against $13.1B of cash. Most of it is spread over years and some of it is contingent, but a chip designer doesn’t usually carry obligations like these.

The number I’d watch is Data Center segment margin. It hit 31% in Q2, before Helios ships in volume, and Jean Hu said data center AI runs slightly below the corporate gross margin average. The test is whether it holds near 30% as Helios ramps from Q4 into Q1.

_{SECTION 4 PRODUCT & TECHNOLOGY}

_{AMD designs the machine and its partners build it}

2025 · MI350 proves AMD can ship. The job was ramping an AI accelerator at scale for the largest labs. Instinct sales more than doubled year over year in Q2.

2026 · Helios proves AMD can sell a system. This is the inflection. The key idea is “scale-up”: how many GPUs can share memory and talk to each other as if they were one chip. NVIDIA’s NVLink has owned this for years. Helios is AMD’s first rack where all 72 GPUs work as a single domain, and it gets there on open standards, including UALink and UALoE, OCP rack specifications and Ultra Ethernet, instead of an NVIDIA-style proprietary fabric. AMD claims up to 15% more throughput at the same rack power and up to 30% more tokens per dollar than the competition, and those are AMD’s own numbers. Initial shipments were due in Q3, with a step-up this quarter and another in Q1.

2027 · Helios 500 has to prove the cadence is real. AMD promises larger scale-up domains over copper and optical links, and calls MI500 the biggest generational leap in Instinct history.

2028 · Helios 600 has to prove it wasn’t one good cycle. AMD has shared few details, and for now the job is to arrive on time.

The yearly rhythm matters more than any single chip. A lab signing a multi-year gigawatt plan needs to know a faster, compatible generation will arrive every year. That predictability has been NVIDIA’s real weapon, and AMD has now laid out a roadmap to match it. The next two years will show whether it can execute.

The memory angle. Memory capacity is AMD’s main selling point, because bigger models and longer context fit in fewer GPUs. In a shortage it’s also one of the costliest parts of the rack and the one AMD controls least. Su said on the call that AMD has good visibility into 2027 HBM supply and can trim memory on mid-sized models where the extra capacity doesn’t pay. That’s good news for the memory makers, and for AMD’s gross margin it’s the variable I’d track.

_{Built for agents: the Cerebras deal}

On July 23rd AMD and Cerebras announced a joint inference system, the clearest look so far at how AMD thinks about agents.

Answering a prompt has two halves. First the model reads the prompt and everything attached to it, which for an agent can be an entire codebase or a week of prior steps. Then it writes the answer one token at a time. Reading rewards memory and throughput, and writing rewards speed, so the two companies split the job: Helios handles the throughput-heavy reading stage, known as prefill, and Cerebras’ wafer-scale chip, a single processor the size of a dinner plate, handles the latency-sensitive writing stage, known as decode.

The claim is up to 5x more tokens per second per watt than a Cerebras-only setup, based on the two companies’ modeling and not yet on a system in production. Cerebras will deploy Helios in its own data centers, and Su said the combined service should reach the Cerebras cloud in Q4 and extend into 2027.

I read it two ways. The first is that AMD is willing to concede the latency-sensitive decode stage to a specialist when that produces a better system, and the open design is what lets that chip slot into Helios. The second reading is about size: no dollar figure, purchase commitment or equity was disclosed, so this is a product story today and not yet a revenue one.

_{The part that decides everything else: ROCm}

A phone with better hardware still loses if your apps don’t run on it. For a decade CUDA was NVIDIA’s app store, and AMD kept launching strong hardware into a store with empty shelves, which is why benchmark wins never turned into share.

What changed this year:

• More than 3 million models now run on AMD out of the box, and leading open models launch with day-zero Instinct support.

• Open-source contributions to ROCm grew more than tenfold in a year.

• ROCm supports PyTorch, JAX, vLLM, SGLang and Triton.

• ROCm.ai lets developers use coding agents (Claude, Codex, Cursor) to port and optimize code for Instinct. AMD says its latest software stack delivers more than 2x the training and 3x the inference performance of ROCm 7.

AMD needs the cost of switching to fall below the price and performance gap, whether or not developers ever come to prefer ROCm. That’s why frontier labs signed first: they write their own kernels and depend least on CUDA. ROCm is now a credible stack at the frontier. What it hasn’t proved is that the long tail of enterprises and startups, who just want things to work, can adopt it with CUDA-like ease. If coding agents can translate CUDA code cheaply, NVIDIA’s software lead shrinks faster than most models assume. That’s my read, and it isn’t proven yet.

_{SECTION 5 GROWTH CATALYSTS}

_{What has to show up between now and 2028}

The near-term ones arrive together. The Q3 report, OpenAI’s first Helios capacity, the Cerebras service and the World Labs close all land between now and year-end, along with the first Venice deployments at cloud providers and Lisa Su’s OCP keynote. By January there will be real evidence on Helios where today there is only guidance.

The 2027 bar is already high. Management’s frame is Data Center revenue more than doubling and server CPUs growing more than 70% off a much larger base. When an analyst floated roughly $30B of 2027 Instinct revenue on the call, Su said that was probably too low. Hitting every date on the timeline still disappoints if the numbers land under that frame.

The last item is my own test for a re-rating. Most of the near-term deployment milestones are tied to demand AMD has already signed, much of it with equity attached. A hyperscale MI500 deal with no warrant, investment or guarantee would show AMD can win gigawatts on product alone, and “second source” would stop being the right label.

_{SECTION 6 MACRO & INDUSTRY BACKDROP}

_{Room for two}

The most common mistake in the AMD debate is assuming it only works if NVIDIA loses.

AMD sizes high-performance and AI compute at close to $2T by 2030, growing about 40% a year. Within that it puts data center accelerators at roughly $1.4T (more than 45% a year) and server CPUs at about $220B (more than 50% a year), and Su said on the call that agentic sandboxes are the smallest piece of the CPU number today and will be the largest by 2030. This is a seller sizing its own market, so discount it. But even taking AMD’s number at face value, 10% of a $1.4T accelerator market is $140B a year from one product line, against the roughly $30B I estimate the whole Data Center segment does this year.

NVIDIA can stay dominant and AMD can still become several times its current size. The thesis depends on the market growing, and NVIDIA keeping its share doesn’t break it.

The real third player is custom silicon: Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia and Meta’s MTIA. Anthropic runs on TPUs and Trainium as well as AMD. Labs spread their orders on purpose, so part of AMD’s gigawatts is the buyer hedging. That gets AMD in the door, and it limits what AMD can charge once inside.

NVIDIA's advantage is that it controls every layer of the machine. AMD's plan is to match it layer by layer, with partners doing the building.

_{SECTION 7 RISKS}

Read more