If you are active within this market, it’s more than likely that your portfolio has decreased from YTD highs. That holds even more true if you are heavily invested in AI stocks. Most of the damage this month is inside the AI complex. CoreWeave, Nebius, and IREN each fell sharply this week, taking their October losses past 30%. Meanwhile, Meta, Oracle, and Nvidia CDS spreads reached the widest levels on record.

The AI bubble has not burst. The cost of capital has risen. This forces the market to reprice every business that funds its buildout with borrowed money. Multiples carry that repricing. Earnings carried the index through 2026 while the multiple fell. We expect earnings to carry the index through 2027 as well.

Read more

The Federal Reserve Bank of Philadelphia prepared this edition from information collected on or before October 5, 2026. The survey window covers the first three weeks after the FOMC raised the federal funds rate by 25 bps to a target range of 3.75 to 4 percent on September 16.

Read more

Note: The article below was written prior to Venture Global’s recent deal with GE Vernova and Baker Hughes for up to 16 gas-fired power islands to supply data centers. Any price targets or assessments do not include this information. Overall, the deal provides VG with long-term optionality. However, we will wait for more news on customer agreements before offering any further thoughts.

Overview

Venture Global reported Q2 2026 results on August 11. Revenue rose 48% to $4.58 billion, falling short of the $4.68 billion consensus cited by Investing.com. Diluted EPS reached $0.51, slightly beating the $0.50 estimate. Consolidated Adjusted EBITDA rose 79% to a record $2.5 billion. Management raised full-year EBITDA guidance to a range of $8.7 billion to $9.1 billion, up from $8.2 billion to $8.5 billion.

Our Q1 note in July maintained a Buy rating with VG trading between $12 and $13. We argued the stock would transition to trading on operational execution rather than headline noise. Q2 delivered that execution. EBITDA increased 82% sequentially from Q1, and VG has contracted 91% of its available 2026 cargoes. Nevertheless, the share price remains largely unmoved.

Q1 article below.

Read more

Overview

The FOMC raised the target range by 25 bps to 3.75%-4.00% on September 16 in a 12-0 vote. It was the committee’s first rate increase since 2023. Hammack, Kashkari, and Logan dissented for a hike in July and joined the unanimous vote. The minutes show support beyond the voters, with every participant backing the increase.

Participants grounded the decision in elevated inflation, a labor market near full employment, and solid growth. Almost all judged that labor market risks had diminished to broadly balanced while inflation risks stayed tilted to the upside. Many described the hike as risk management against persistent inflation from stronger demand or new supply shocks. Several viewed the policy rate as not restrictive or only mildly restrictive. A couple of participants had raised their estimates of the neutral rate. Most participants assessed that another increase would likely be appropriate by year end.

Inflation

Staff estimated total PCE inflation edged up to 3.8% in August due to higher energy prices. Core PCE remained at 3.4%. Following the revised BEA methodology, August total inflation measured 3.4% and core measured 3.0%.

​Above-target inflation stems from tariffs, geopolitical energy costs, and AI technology prices. Core services inflation excluding housing remains elevated, driven by software and portfolio fees. Recent models show underlying inflation moving sideways above 2%. While 3-month core PCE dropped, this metric often understates second-half inflation.

​Participants view inflation risks as tilted to the upside. Sustained energy costs increase the risk of broader price pressures. Extended exposure to inflation above 2% risks unanchoring wage and price expectations. Businesses are increasingly passing cost increases through to consumers rather than absorbing them via compressed margins.

​AI-related technology costs represent a growing driver of broader inflation. AI investment increases input costs across transportation and raw materials while lifting aggregate demand. Several participants noted the scale of the buildout continues to exceed forecasts.

​Productivity gains from AI present medium-term upside potential for total output. However, timing and magnitude remain uncertain. Input cost pressures from the AI buildout are impacting current core goods prices, whereas productivity offsets remain theoretical.

Labor Market

Unemployment fell to 4.1% in July and August, 0.3 percentage points below its average in the second half of last year. Participation stepped back up in August. Payroll gains picked up notably across a range of industries. The employment cost index rose 3.3% over the year through June. Average hourly earnings rose 3.1% over the 12 months to August, down from the 3.5% pace cited in the July minutes.

A majority of participants judged that the labor market had strengthened slightly. Job gains modestly outpaced labor force growth. Several participants noted low overall dynamism. They cited subdued hiring and layoff rates alongside a low job-finding rate and elevated long-term unemployment. Some participants again highlighted strong wage gains for skilled workers tied to the AI buildout. Conversely, others noted that aggregate wage growth remained moderate and was not driving inflation pressures. Overall, participants viewed labor market risks as broadly balanced.

Market Stability

The Desk reported that 2- to 10-year Treasury yields rose about 35 bps over the intermeeting period. Part of the move reflected a higher expected policy path and strong data. Market commentary also cited geopolitics, uncertainty around the Treasury buyback program, and competition for capital from heavy private AI infrastructure debt. Staff attributed most of the short-end increase to inflation compensation and most of the long-end increase to real rates.

The manager noted that spreads on hyperscaler debt stayed wide given heavy issuance and long duration. Staff reported those spreads narrowed slightly over the period after widening notably from low levels at the start of the year. According to the manager, equity gains this year came entirely from actual and expected earnings, with price-to-earnings multiples declining. Companies benefiting directly from AI infrastructure spending outperformed the rest of the market.

Many participants judged financial conditions supportive of growth despite higher long yields. They cited equity gains and narrow corporate spreads. Staff described credit as somewhat tight for private credit loans and small businesses. Direct lending issuance slowed in July amid subdued retail interest. Housing finance remained restrictive due to elevated mortgage rates.

Stock gains supported spending by higher-income households. Meanwhile, energy costs strained low- and moderate-income households.

Balance Sheet and Process

The Desk paused reserve management purchases, judging reserves to remain ample. The manager stated the Desk sets the monthly purchase pace with no preset path. Money markets absorbed a substantial increase in bill issuance since July with little rate impact. A few participants called for stronger tools, communications, and strategy to address Treasury market dysfunction. They emphasized limiting the Federal Reserve’s overall market footprint.

Our View

The minutes reinforce our stance that policy remains accommodative relative to inflation. Several participants described the prior rate level as mildly restrictive at most. Staff raised its inflation forecast through 2028, and participants widened the upside skew on inflation risk.

As we have stated for the past few months, the Fed is expected to cut rates twice this year. The first cut was a risk-management move signaling the Federal Reserve is willing to curb price growth. A second cut before year-end will solidify that argument.

The majority of inflation is driven by supply-side factors, specifically tariffs and energy, although the AI buildout continues to contributes to demand.

The September payrolls report missed estimates significantly, pointing to potential cooling in the labor market. However, the committee judged labor market risks as balanced prior to that release.

We maintain our call for a second rate hike before year-end, contingent on the Middle East conflict remaining unresolved.

ZCR

Read our disclaimer here.

Amazon reported Q2 2026 results on July 30. Net sales rose 20% to $200.6 billion, ahead of the $196.6 billion consensus. AWS revenue grew 36.7% to $42.2 billion. Jassy said that was the segment’s fastest growth in 18 quarters. Operating income rose 43% to $27.5 billion. Management also raised 2026 capex guidance to about $220 billion, from about $200 billion.

In March, our long case at $199 was built on three main drivers: AWS re-acceleration, robotics cutting retail costs, and AI infrastructure lock-in. Q2 pushed the first and third drivers well past what we assumed in June. AWS growth rose about nine points in a single quarter. The backlog disclosed in the 10-Q reached $496 billion. North America segment margin expanded to 7.9% from 7.5% a year earlier but fell from 9.0% in Q1.

Our Q1 update, linked below, laid out the path to a $260 base-case target over 12 to 18 months. Amazon closed above $260 for most of August and now sits near that level. As the case plays out, let's deep dive into Q2 and see where it gets us.

The Numbers

Revenue: Net sales grew 20% to $200.6 billion from $167.7 billion, with no FX effect. North America sales grew 16% to $116.2 billion. International sales grew 15% to $42.2 billion. Prime Day moved into June this year, from July in 2025. Management said on the July 30 call that the shift will cut Q3 growth by about 400 bps relative to Q2.

AWS: Revenue surged 36.7% YoY to $42.2B ($169B annualized run rate, up from $150B in Q1), marking four consecutive quarters of accelerating growth (17.5% to ~37%). Operating income expanded 64% to $16.6B (39.4% margin; +650 bps YoY or +520 bps ex-energy derivative gains). AWS generated 60.5% of total operating income on 21% of net sales.

AI and Chips: Jassy said the AWS AI business and the chips business have each passed a $25 billion annual run rate. Our Q1 note put those figures at $15 billion and $20 billion. He said both are growing at triple-digit rates. He also said Anthropic and OpenAI have each made multi-year, multi-gigawatt commitments to Trainium.

Backlog: Unrecognized long-term contract commitments reached ~$496B as of June 30 (6.4-year weighted-average remaining life), up from ~$195B YoY. Major 10-Q expansions include OpenAI adding $100B over 8 years (to an existing $38B commitment) and Anthropic adding >$100B over 10 years. Management noted on the call that capacity will remain constrained relative to total demand through 2026 and 2027.

Retail and Advertising: North America operating income expanded to $9.1B (from $7.5B YoY), while International rose to $1.7B (4.1% margin, up from $1.5B YoY). Advertising revenue grew 26% to $19.8B, Third-Party Seller Services rose 16% to $46.8B, and Subscriptions grew 12% to $13.7B. Q2 operating income included $1.2B in one-time benefits ($600M tariff refunds, $600M energy derivative gains); excluding these items, normalized operating income was ~$26.3B (13.1% margin).

The Anthropic Gain: Net income reached $62.6B ($5.75/share vs. $1.68 YoY), driven by a $53.4B pre-tax non-operating gain on Anthropic investments. Per the 10-Q, the Anthropic stake reached a $190.4B carrying value at June 30 ($92.5B nonvoting preferred, $97.9B convertible notes; up from $60.6B total at Dec 31) against $18.0B in cumulative cash invested ($8.0B in notes through 2025, plus $5.0B each in Q2 Series G/H rounds). Amazon also held $28.7B in OpenAI Series C preferred stock at June 30, funding the remaining $21.3B commitment post-quarter end.

Capex and Free Cash Flow: Amazon spent $53.1 billion on cash capex in Q2. Management raised the full-year guide to about $220 billion, citing "the higher cost of memory." That implies roughly $62 billion a quarter in the second half. Trailing twelve-month operating cash flow rose 33% to $161.4 billion. Trailing free cash flow still swung to an outflow of $7.6 billion, from an inflow of $18.2 billion a year earlier.

Balance Sheet: Long-term debt rose to $128.9 billion at June 30, from $65.6 billion at December 31. First-half debt proceeds totaled $67.0 billion. Cash and equivalents stood at $78.2 billion. Marketable securities added another $44.8 billion.

Q3 Guidance

Amazon guided Q3 net sales to $197 billion to $202 billion, or 9% to 12% growth. That range includes an FX headwind of about 80 bps. Operating income guidance is $22.5 billion to $26.5 billion, against $17.4 billion in Q3 2025. The midpoint implies 41% growth in operating income on about 10.5% revenue growth. In June, Amazon guided Q2 operating income to $20 billion to $24 billion. It delivered $27.5 billion, or $26.3 billion excluding one-time items.

What Changed?

In June, we cited a $15 billion AI run rate and a $20 billion chips run rate. Both now exceed $25 billion. The Anthropic stake has moved from an estimate in our Q1 piece to a $190.4 billion line on the balance sheet. Amazon has also invested $50 billion in OpenAI, which has committed $138 billion to AWS. Jassy now sees AWS as “very possibly” a trillion-dollar annual revenue business “in time.” On September 30, Amazon signed a licensing deal with Synopsys worth more than $1 billion to support Trainium and Graviton.[Benzinga].

On August 31, the FTC and 22 states sued Amazon, alleging hidden surcharges in its ad auctions that cost advertisers tens of billions of dollars. The stock fell 2.5% that day.

Risks

• Free Cash Flow Headwinds: Planned 2H capex (about $62B/quarter) significantly exceeds trailing quarterly operating cash flow (about $40B/quarter). Management noted infrastructure spend creates near-term cash drag as data centers are built prior to generating revenue.

• AI Lab Concentration: Anthropic and OpenAI represent the largest disclosed backlog expansions and a combined carrying value of about $240B. A funding squeeze at either entity threatens both backlog commitments and balance sheet valuation.

• Regulatory Exposure in Advertising: An FTC complaint targets the core auction mechanics powering Amazon’s $19.8B quarterly ad segment (+26% YoY).

• Input Price Inflation: Memory cost increases forced a $20B upward revision in full-year capex guidance, posing ongoing unit cost risks for new capacity.

Catalysts

• Q3 Earnings: Operating income beating guidance top end ($26.5B) would repeat Q2’s trend, while AWS growth accelerating past 36.7% would mark a fifth straight quarter of acceleration.

• Anthropic Public Listing: An IPO would establish a liquid market price for Amazon’s $190.4B carrying stake.

• Trainium Adoption: Securing named external, non-equity customers for Trainium would validate Amazon’s merchant silicon strategy.

• Leo Network Commercialization: Deployment reaches 400 in-orbit satellites, targeting initial commercial service launch in late 2026.

Updated View

We set the $260 target in March and reaffirmed it in our Q1 review. That target predates 37% AWS growth, the $496 billion backlog, and a Q3 operating income guide 41% above last year. We are raising our 12 to 18 month base-case target to $280.

The math starts with Q3 operating income guidance of $24.5 billion at the midpoint. For the following three quarters, we assume 25% year-over-year growth. That sits well below the 43% Amazon reported in Q2. It also leaves room for depreciation from the $220 billion build. The result is next-twelve-month operating income of about $118 billion. At 24x, the operating business is worth about $2.84 trillion. We add $240 billion for the Anthropic and OpenAI stakes at carrying value and subtract about $27 billion of net debt. Across there state diluted shares we land at about $280.

The consensus street target is $330 across 59 analysts. We remain more conservative in our valuation due to negative free cash flow drag and potential upward pressure on Treasury yields.

Conclusion

In March, the central question was whether enterprise AI demand would translate into AWS revenue. A backlog approaching $496 billion confirms strong demand. However, Amazon is now funding a $220 billion capital expenditure program through $67 billion in new debt amidst negative free cash flow. Furthermore, its two largest backlog expansions stem from AI labs in which Amazon holds $240 billion in equity and convertible stakes. Given regulatory exposure from the FTC lawsuit and yield-driven debt refinancing pressure, we maintain a more conservative stance than Wall Street.

We move Amazon from Strong Buy to Buy, with an updated $280 price target.

Rating: BUY

ZCR

Read our disclaimer here.

Rubrik reported Q2 FY2027 results on August 27 for the quarter ended July 31. Revenue of $427.3 million cleared the top of management’s June guidance by $30.3 million. Non-GAAP EPS of $0.20 came in at four times the $0.05 guidance ceiling. Management then raised every full-year guided metric for the second consecutive quarter.

Our Q1 note argued that execution was running ahead of our March thesis. Q2 extended that lead across ARR, operating leverage and guidance. Gross margin and free cash flow margin both fell year over year. Neither change breaks the thesis. Low-margin hardware sales in other revenue pulled down gross margin. A smaller deferred revenue build pulled down the FCF margin.

RBRK closed at $114.47 on September 30. That price sits about 35% above the $85 bull case our EV/ARR framework produces on the raised guidance.

The Numbers

Revenue: Total revenue reached $427.3 million, up 38% YOY. Subscription revenue grew 37% to $407.2 million. Material rights contributed $4.7 million, down from $14.2 million a year earlier. Excluding material rights, revenue grew 43%. Regionally, Americas revenue grew 33% to $299 million, while international revenue grew 52% to $129 million.

ARR: Subscription ARR reached $1.66 billion at quarter end, up 33% YOY. Net new subscription ARR grew 35% YOY to approximately $96 million. Cloud ARR grew 39% to $1.48 billion, with adjusted net new Cloud ARR (excluding customer migrations) growing 20%. Subscription net retention remains above 119%.

Customers: Customers contributing over $100,000 now account for 88% of subscription ARR. That cohort grew 23% to 3,084 customers. The $1 million+ customer cohort grew by more than 57%.

Gross Margin: GAAP gross margin compressed to 78.4% from 79.5% a year earlier, and non-GAAP gross margin slipped to 81.0% from 81.6%.

Profitability: Non-GAAP operating income swung to $33.5 million (7.8% margin), against a $4.4 million loss in Q2 FY2026. Non-GAAP net income was $44.7 million, or $0.20 per diluted share, heavily beating Street consensus of $0.04. Subscription ARR contribution margin expanded 460 bps to 14.0%. GAAP operating loss narrowed to $71.9 million YOY, though it widened sequentially due to a $27.6 million jump in stock-based compensation.

FCF: Operating cash flow was $76.8 million (18% margin). Free cash flow rose to $65.7 million (15% margin, compared to 19% in both Q2 FY2026 and Q1 FY2027). First-half FCF reached $139.3 million. Hitting the raised full-year guidance requires $184 million to $194 million in second-half FCF.

Balance Sheet: Rubrik ended the quarter with $1.75 billion in cash, equivalents, and short-term investments. Accounting for $1.13 billion in convertible senior notes, net cash sits at roughly $615 million. Deferred revenue accounts for $1.93 billion of the company's $3.35 billion in total liabilities.

FY2027 Guidance

Management raised its full-year subscription ARR target to $1.880 billion to $1.885 billion, from $1.854 billion to $1.862 billion in June. Choudary put the implied growth rate at about 29%. Revenue guidance rose to $1.685 billion to $1.693 billion, from $1.638 billion to $1.648 billion. Non-GAAP EPS guidance rose to $0.47 to $0.53, from $0.25 to $0.35. FCF guidance rose to $323 million to $333 million, from $293 million to $303 million. The subscription ARR contribution margin target moved to about 15.5%, from about 14%.

For Q3, Rubrik guided revenue to $429 million to $431 million and non-GAAP EPS to $0.07 to $0.09. Choudary said the revenue range implies 23% growth, or 31% excluding material rights.

What Changed?

Net New ARR Outlook: Choudary said the new guidance lifts full-year net new ARR by $24 million to $25 million. He put implied net new ARR growth for the year at 14%, up from 7%. CEO Bipul Sinha added that the full-year net new ARR outlook has risen by $45 million since the fiscal year began. Both executives said the Strata acquisition contributes zero ARR to the quarter or to guidance. The entire raise is organic.

Product Expansion: Agent Cloud has moved from shipping code in Q1 to more than 15 paying customers, according to Sinha. Choudary said guidance assumes minimal Agent Cloud ARR this year. Rubrik also added Agent Identity as a fourth Agent Cloud pillar. It launched Agent Cloud for Anthropic's Claude Code.

AI Tailwinds: Sinha called Mythos and other frontier AI models a major inflection point for the industry. He said customers concerned about Mythos-class attacks now come to Rubrik for machine-speed recovery.

Migration Runoff: Choudary said cloud migrations are largely complete this year. Management said non-cloud demand is now growing in regulated industries, government and sovereign-sensitive markets. Choudary added that non-cloud business carries a higher margin because customers self-host. Separately, Rubrik committed £375 million ($500 million) to the UK and named London its EMEA headquarters.

Risks

• Back-Half Net New ARR: Citi analyst Fatima Boolani said on the call that guidance implies net new ARR growth falling from the 20s in the first half to mid-single digits in the second. Choudary said Rubrik plans and pays its sales force on annual net new ARR. He also cited pipeline confidence.

• Cloud Migration Runoff: Adjusted net new Cloud ARR grew 20%, against 35% for total net new subscription ARR. With migrations largely finished, Cloud ARR growth now depends on new logos and expansion within existing customers.

• Cash Conversion: FCF margin fell 400 bps year over year to 15%. Deferred revenue added about half as much first-half operating cash as it did a year earlier. Full-year guidance requires $184 million to $194 million of second-half FCF.

Catalysts

• Q3 FY2027 Results: Quarter ends October 31. Net new ARR above Q2 levels (~$96M) would counter the guidance-implied 2H deceleration.

• Agent Cloud ARR Disclosure: First explicit Agent Cloud ARR metrics would shift product valuation from pure optionality to fundamental revenue multiples.

• Strata Contribution: Current guidance models zero Strata ARR. Any FY2028 contribution provides pure upside to the organic base.

• Rubrik Flex Momentum: Introduced at June Forward as a unified platform license, Flex already anchored a top Q2 expansion deal with a large US healthcare system.

Updated View

Our March framework applied 7x and 10x EV/ARR to the $1.88 billion ARR we expected Rubrik to reach. The new guidance midpoint of $1.8825 billion sits slightly above that target. At 7x, EV comes to $13.2 billion. Adding about $615 million of net cash and dividing by the 230 million shares in Rubrik’s Q3 guidance gives a base case near $60. The same math at 10x yields about $85.

Sell-side targets moved higher after the print. According to TipRanks, Citi raised its target to $130 from $96 on August 28, BTIG to $125 from $109, and Scotiabank to $120 from $114. Wedbush also raised its target to $120 from $90 on September 11.

At the September 30 close, EV equals about 14x FY2027 ARR guidance. We set our 7x and 10x multiples in March, during broad SaaS multiple compression. The market has since rerated Rubrik past our bull multiple. Two consecutive quarters of organic guidance raises justify a higher multiple. Supporting 14x forward ARR requires second-half net new ARR that current guidance does not assume.

Conclusion

Rubrik has met every operating condition in our March thesis for two consecutive quarters. The open question is the entry price. At about 14x forward ARR, the stock already discounts a second half well above guidance, which implies mid-single-digit net new ARR growth. It also discounts Agent Cloud revenue that management has kept out of guidance. A Q3 beat on the scale of Q2's would validate the rerating. A Q3 print inside the $429 million to $431 million range would leave the stock 35% above our bull case.

We downgrade Rubrik to Hold from Strong Buy. The business has earned a higher multiple than the one we set in March, but not the 13x the market pays today. Potential re-entry target would be $85.

Rating: HOLD

ZCR

Read our disclaimer here.

Part 1 set the AI buildout against Britain’s railway mania of the 1840s and the American fiber boom of the late 1990s. In both cycles, capital deployed fast while funding shifted from operating cash toward debt. Late-stage lenders absorbed the losses once capacity outran demand. This piece answers the question Part 1 left open: is the current buildout a bubble, and how should the ZCR portfolio sit against it?

Measured against GDP, the AI surge is smaller than railway mania, though far more compressed. Spending across the largest data center operators approaches $750 billion in 2026, roughly four years after the cycle began. Morgan Stanley projects $570 billion of global AI-related debt issuance in 2026 as the funding mix moves toward leverage. Nvidia has assembled a structured-financing platform to bring more capital in.

That transition introduces three risks: a hyperscaler funding gap, a correlation breakdown between chip suppliers and big tech, and redundant capital in parallel compute clusters. The sections below test the buildout against the historical record and reach a final assessment on whether we are in a bubble.

What is a Bubble?

Because the word gets used loosely, it needs a definition before the analysis starts. Charles Kindleberger defined a bubble as an asset price deviation that fundamentals cannot explain.

Kindleberger and Hyman Minsky built a five-stage model of how bubbles form and unwind. Minsky laid its foundation in his 1986 book Stabilizing an Unstable Economy. The model runs through displacement, boom, euphoria, profit-taking, and panic.

Displacement is the shock that creates a new profit opportunity, whether an invention, a policy change, or a rate cut. In the boom, expanding credit lets capital chase the new market. Euphoria arrives when price extrapolation becomes self-reinforcing and participants assume the trend runs indefinitely. Profit-taking begins when insiders reduce exposure while the wider market still believes the thesis. Panic is the reversal, usually triggered by one specific failure that forces a rush into cash.

The sections below apply that framework to the AI build-out.

Previous Cycles

In Britain’s railway mania, displacement combined steam railway technology with the low Bank of England rates of the early 1840s. During the boom, 272 separate Acts of Parliament authorized 9,500 miles of track in 1846 alone. Euphoria drew middle-class savers, clergy, and literary figures into railway scrip bought on partial payment terms. After the Bank of England tightened in 1845, the railway share index topped out in 1846. Railway shares fell by roughly two-thirds between 1846 and 1850 while companies called the remaining 90% due on partly paid shares. Those calls wiped out households that had committed their savings. Average dividends on ordinary shares reached 1.83% in 1850, against the 10% promoters had advertised.

Commercialization of the internet was the displacement for the telecom and dot-com cycle, which ran the same model on a shorter clock. Telecom equipment spending grew 18% a year from 1996 to 2000, rising from $62 billion to $135 billion annually. Euphoria carried the Nasdaq to its March 2000 peak. Cisco traded above 100 times forward earnings and briefly ranked as the world’s most valuable company at roughly $500 billion. The Nasdaq then fell 78%. WorldCom and Global Crossing both filed for bankruptcy in 2002, with an estimated 85% to 97% of the fiber laid during the boom left dark for years.

Japan ran the same sequence through real estate and equities rather than physical infrastructure. Its displacement was the September 1985 Plaza Accord and the Bank of Japan rate cuts that followed. The Nikkei 225 rose from roughly 13,000 in late 1985 toward 26,000 by 1987, while commercial land prices rose more than 300% across Japan’s six largest cities. Euphoria peaked on December 29, 1989, when the Nikkei reached 38,915. Tokyo Stock Exchange market capitalization briefly passed $4 trillion, roughly 1.5 times Japanese GDP. Stocks traded near 60 times earnings, against about 15 for the S&P 500. Corporate Japan leaned further into the run through zaitech, the practice of booking speculative land and stock gains as operating income. After the Bank of Japan tightened in 1989, the Nikkei lost nearly half its value within a year. Land prices then fell for more than a decade, in an aftermath that became Japan’s Lost Decade.

The AI Build-out

Capital expenditure among the largest data center operators approaches $750 billion in 2026, up from under $450 billion in 2025. S&P Global Ratings estimates that the 2026 spend by Alphabet, Amazon, Meta, Microsoft and Oracle will equal 38% of their combined revenue.

Through 2025 the build-out ran mostly on operating cash flow. In 2026 it runs increasingly on debt. Hyperscaler bond issuance rose from an annual average of $40 billion in 2020 to $121 billion in 2025. Morgan Stanley projects roughly $570 billion in global AI-related debt issuance for 2026. Goldman Sachs projects $400 billion in hyperscaler investment-grade issuance for 2027 alone.

Alphabet, Amazon, Meta, Microsoft and Oracle had issued about $220 billion of bonds in 2026 by August 10, according to BNP Paribas data. Median spreads on the two- to four-year paper of Amazon, Alphabet, Meta and Oracle widened to 40 bps over Treasuries. Apollo reported that order cover on hyperscaler bond sales fell from nearly five times in February to below two times in July. Amazon’s first sterling sale on September 9 raised £4.25 billion at roughly 2.5 times cover. Analysts read that as weaker demand than Alphabet drew for its own sterling debut in February.

The Buffett Indicator stood at 229.9% on August 6, 2026. That compares with a 62% baseline in 1989 and a dot-com peak near 150% in 2000. The reading also exceeds Japan’s 1989 peak of roughly 151%, a figure corroborated between 150% and 154% across three independent sources.

On August 10, 2026, Nvidia signed memorandums of understanding with six financial institutions. The agreements brought insurance, pension, and structured credit capital into the build-out for the first time. None of that capital held direct exposure to the sector before.

J.P. Morgan Private Bank tracked the 20-day rolling correlation between semiconductor and hyperscaler stock returns. The figure fell from above 0.5 in early 2025 to close to negative 1 by mid-2026. Over the same period, the four largest hyperscalers kept building parallel compute clusters for overlapping customers, echoing the duplicate lines of the railway era.

AI-related investment drove roughly 74% of US GDP growth in the first quarter of 2026, per the Bureau of Economic Analysis. Consumer spending contributed almost nothing. The BEA’s second estimate, released August 26, left second-quarter real GDP growth unrevised at a 1.5% annualized rate, down from 2.1% in the first quarter. Equipment and intellectual property products led the quarter’s investment gains, with the software and research spending tied to the AI buildout.

ChatGPT’s November 2022 launch was the displacement in Minsky and Kindleberger’s terms. Capital deployment now outpaces prior infrastructure cycles, while the financing structure shifts from internal cash toward external credit, as it did in 1846 and again in 2000. On both measures, the build-out sits in the boom stage of their model.

Valuation Comparison

The S&P 500 Shiller CAPE ratio stood at roughly 41 to 42 in August 2026. That is the second-highest reading in the index’s history and the highest since 1999, when the ratio touched a record 44.2. Only eighteen months in a series running back to 1881 have registered higher, all of them in 1999 and 2000.

Forward price-to-earnings divides a company’s current price by its expected earnings per share over the next twelve months. It measures what investors pay today for a dollar of expected profit. The Nasdaq-100 traded at roughly 25 times forward earnings in July 2026, against approximately 60 times at the March 2000 peak. The S&P 500 sits at 20.0 times, against 24.4 times in March 2000.

Nvidia trades at roughly 17 times fiscal 2028 earnings estimates, well under a fifth of the multiple Cisco carried at its 2000 peak. Micron trades below 6 times forward fiscal 2027 estimates. The Nikkei’s 60 times earnings in 1989, set against an S&P 500 multiple near 15 at the same moment, ran more than double where the Nasdaq-100 trades today.

Hyperscaler post-issuance leverage runs between 0.4 and 0.7 times even after the 2026 debt surge, against an investment-grade corporate average near 3 times. That comparison comes from Investing.com’s May 26, 2026 analysis of the AI bond wave. Global Crossing entered bankruptcy in January 2002 carrying $12.4 billion of debt against $22.4 billion of assets. WorldCom’s $30 billion of bonds went largely to zero. Hyperscalers today sit nowhere near that profile, though their move from equity funding toward debt funding matches both earlier cycles.

The Case For

Valuation: The CAPE ratio sits at its second-highest level in 150 years.

Financing: The financing structure has moved from cash toward debt at a pace matching or exceeding the railway and telecom cycles. S&P cut Oracle to BBB- on July 9, one notch above speculative grade, as its data center plan burned cash. Moody’s holds a negative outlook on the company. The Financial Times reported on September 18 that banks had failed to syndicate $18 billion of Oracle-linked data center debt to a wider investor base. The banks now hold more of it than planned.

Earnings quality: Depreciation assumptions at several large operators may flatter reported earnings by a material margin. If so, today’s multiples apply to overstated profits.

Circularity: Nvidia’s proposed $100B investment in OpenAI failed to materialize, with the company ultimately contributing $30B to OpenAI’s March round. In August, Nvidia agreed to guarantee up to $105B for an OpenAI-leased SB Energy data center campus in Ohio, where it will supply compute starting in 2028. Initial backstop negotiations reached $250B before scaling down below $120B.

To support its compute demands, OpenAI executed a $300B cloud agreement with Oracle, which routes tens of billions back to Nvidia for chips. Separately, OpenAI’s 6GW agreement with AMD included options for up to 160M AMD shares.

Nvidia maintains direct equity in core reseller partners CoreWeave and Nebius. Its exposure includes a guarantee to purchase up to $6.3B in unused CoreWeave GPU capacity through 2032. Nscale recently disclosed $44.6B in GPU service agreements with Anthropic and a $1.0B equity issuance to Nvidia.

Across major hyperscalers and AI infrastructure firms, direct cross-equity stakes approximate $46B, underpinned by $879B in multi-year purchase commitments. This vendor-financed circular capital structure echoes the framework that preceded the collapse of Lucent and Nortel in 2001 and 2002.

The Case Against

Multiples: Forward multiples on the companies building this cycle sit well below dot-com and Japan-era peaks. Nvidia’s multiple runs at less than a fifth of Cisco’s 2000 level.

Current profits: The businesses funding the build-out earn real money today, which the pre-revenue dot-com companies of 1999 did not.

Leverage: Hyperscaler leverage remains a fraction of investment-grade norms even after the 2026 debt surge.

Contracted demand: KKR’s November 2025 research found data center construction proceeding against signed offtake contracts, unlike the speculative, uncontracted building that defined the fiber boom. The same research found no vacant speculatively built sites.

The Profitability Question

The claim that AI is unprofitable is only half true: infrastructure providers are highly profitable, while model developers operate at a loss. Evaluating AI's real financial viability requires looking at both net income and capital expenditures together.

For the quarter ended June 30, 2026, Amazon reported AWS segment operating income of $16.6 billion on $42.2 billion of segment revenue, against $10.2 billion a year earlier. That is an operating margin near 39%. AWS grew 37%, its fastest rate in eighteen quarters, and reached a $169 billion annualized run rate. Andy Jassy said in the July 30 release that Amazon’s AI and chips businesses had each passed $25 billion run rates.

Google Cloud more than tripled operating income in the same quarter, to $8.81 billion from $2.83 billion a year earlier, on $24.77 billion of revenue. Operating margin rose to 35.6% from 20.7% as cloud revenue grew 82%. Backlog reached $514 billion, up more than $50 billion sequentially. Management expects to recognize just over half of it as revenue within 24 months. Two years ago Google Cloud barely cleared breakeven.

At Microsoft, fiscal 2026 ended on June 30 with $331.8 billion of revenue and $155.2 billion of operating income, up 21%. Azure passed $100 billion of annual revenue for the first time, growing 41% for the year and 43% in the fourth quarter. Commercial remaining performance obligations grew 84% to $678 billion. Microsoft said the entire $51 billion sequential increase in commercial bookings came from customers other than large AI model companies.

$96.2 billion of revenue for the quarter ended July 26, 2026 put Nvidia up 106% year over year, with data center revenue at $89.0 billion. GAAP gross margin held at 75.0%. Operating income reached $63.7 billion and net income $59.7 billion. The company guided the October quarter to $108.0 billion. Jensen Huang framed the quarter in the August 26 release with a single line: “compute is revenue.” On these audited segment results, selling AI compute is a profitable business.

OpenAI lost $20.9 billion on $13.07 billion in revenue in 2025, while Anthropic continues to operate at a loss with projected profitability delayed until 2028 or 2029. Because both developers purchase vast amounts of compute from cloud providers and chip-makers, a material share of infrastructure-layer profits is directly funded by venture capital rather than sustainable operating cash flow.

Hyperscalers are spending more on AI infrastructure than their operations generate. Alphabet, Amazon, and Oracle recently posted negative free cash flow as capital expenditures reached record levels. Microsoft remains the main exception, keeping cash generation positive.

At the same time, reported profits are misleading. A large share of net income at major tech firms now comes from paper gains on private AI startups. These accounting marks inflate reported earnings, masking the fact that physical infrastructure is consuming cash faster than the business brings it in.

The Listing Risk

Anthropic and OpenAI are preparing massive initial public offerings that will test model-layer valuations. Anthropic filed a confidential S-1 following a $65 billion round at a $965 billion valuation, with revenue exceeding $65 billion and a potential public debut in late 2026. OpenAI targets a 2027 listing while generating over $40 billion in annualized revenue, operating at a $14 billion projected annual loss, and holding more than $600 billion in cloud commitments.

These filings present three major market implications:

• Price discovery: A live quote for model builders will force every hyperscaler to revalue its private equity stakes. Under GAAP accounting, a weak debut directly impairs reported tech earnings, while a successful listing inflates them further.

• Capital supply: A $100 billion offering will drain liquidity from equity markets. As seen in recent large-scale tech debuts, massive supply can depress broader secondary market valuations during periods of peak sentiment.

• Regulatory disclosure: Public S-1 filings will reveal audited compute commitments, gross margins, and hardware depreciation schedules for the first time. These metrics will establish the benchmark for evaluating all underlying AI infrastructure spend.

Conclusion

The current AI infrastructure expansion does not fit the classical definition of a financial bubble. Based on Kindleberger’s test, an asset bubble requires price appreciation that underlying fundamentals cannot justify. Today, chip-makers and cloud providers report audited profits and expanding segment margins. Valuation multiples also sit far below comparable historical peaks.

The real structural risk lies in capital flow dynamics across the supply chain. Nvidia generates the highest gross margins in the sector by selling hardware to hyperscalers. Hyperscalers capture strong segment margins by reselling compute capacity to enterprises and AI labs. Meanwhile, AI model labs operate at severe losses. They rely on private venture capital to subsidize end-user pricing. Each layer remains profitable only while the layer below secures fresh funding. As private capital transitions into public equity markets, financial strain moves upstream. Unlike the late-1990s fiber optic boom where builders absorbed direct losses, current downside risk rests primarily with the customer base.

Five structural conditions determine whether this expansion escalates into a bubble. Circular financing has already arrived. Chip suppliers now guarantee customer leases, moving a step beyond 1990s vendor financing. Debt is displacing operating cash faster than in prior cycles. Credit markets are responding with wider spreads and thinner order books. Late institutional capital, including pension, insurance, and structured credit funds, is entering the stack after primary equity commitments have peaked. Redundant capacity is also being built as hyperscalers construct parallel clusters for overlapping clients. High utilization rates keep overcapacity an emerging risk rather than a settled fact. Finally, AI capital investment is now large enough to influence monetary policy and federal funds rate decisions.

No single factor confirms a systemic bubble today. Together, these dynamics show an infrastructure boom advancing toward late-cycle conditions faster than underlying cash flow generation can support.

That assessment does not argue for avoiding the sector. Railway mania, the telecom boom, and Japan’s asset bubble each produced permanent losses alongside durable winners. In every case, the infrastructure outlived the capital structures that financed it.

The ZCR portfolio during this time, as always, seeks quality over chasing hype. We favor contracted revenue and clean balance sheets over leverage. AMZN and NVDA sit on one side of the trade, while Oracle and Allbirds are on the other. When assessing companies, it is important that you look not only at current valuations, but at future DCF models and overall TAM. For the Mag 7, data center stocks, and software layer plays, ensure you conduct thorough research on leadership and the income statement and balance sheet. We understand that in this cycle there will be winners and losers. We just have to ensure the sizing of our winners offsets the losers we happen to pick. We are also starting to build a cash position moving into Q4, as we will be doing a year-end rebalance and will further dive into in a subsequent post.

We assume roughly three to four years of runway before there is a major reckoning within this space. Once you understand your runway, you can size appropriately. The three to four year time-frame is just a working thesis as to how this cycle plays out. It is mainly driven by depreciation assumptions, refinancing, and compute commitments.

Remember that Railway Mania bankrupted the speculators and enriched the freight operators. The fiber boom bankrupted WorldCom and enriched Google. Invest prudently and understand what you own.

The AI future will be built. The question is who goes broke building it.

ZCR

Read our disclaimer here.

For some time we have called for one to two hikes prior to year end. However, we favored a hold at the latest meeting. We disagree with the Fed decision, but that is why they are governors and we are not. While we disagree on the immediate action, long-term views align. Market participants and the SEP now share our baseline of two total hikes by…

Read more

Amkor Technology packages and tests semiconductors that other firms design and fabricate. The company ranked second globally among outsourced assembly and test providers in 2024, at $6.32 billion of revenue and 15.2% share, behind ASE Technology. FY2025 revenue reached $6.7 billion. The stock closed at $46.38 on September 15, 2026, against a mar…

Read more

The Beige Book is the Federal Reserve’s summary of current economic conditions across its twelve Districts, published eight times per year. The Federal Reserve Bank of Minneapolis prepared this edition from information collected on or before August 24, 2026. The survey window therefore closed before the Jackson Hole symposium a…

Read more

Kevin Warsh delivered his first Jackson Hole keynote as Fed Chairman on Friday. The date marked his hundredth day in office. Warsh opened with a joke about hiking trails and former Vice Chairman Don Kohn. He then moved into the substance he had spent weeks declining to preview. The speech confirmed what his July 29 press conference had signaled, Warsh will not fill the vacuum left by dismantled forward guidance.

Warsh devoted a meaningful share of the address to artificial intelligence as a factor of production. He did not frame it as an inflation threat. He cited annualized token sales exceeding $100 billion at the two largest AI labs. That figure has grown more than 500 percent in a year, he said. Warsh left the productivity question open. He does not yet know whether AI lifts productivity broadly across the economy. He also does not know whether AI complements labor or replaces it. A new Fed task force on productivity and jobs will study these questions.

On forward guidance, Warsh made his clearest case yet for retiring the practice. He wants it reserved for genuine crises, not normal policy conduct. He argued that oversharing policy deliberations creates a hall-of-mirrors problem. The Fed reads market prices for signals in return. Both sides grow more likely to miss a genuine turn in the economy. He cited 2021 directly as an example. Guidance issued that year, he said, slowed the Fed’s response to rising inflation. His preference is a Fed that offers clear signals and principles.

Warsh’s economic assessment matched the hawkish undertone of his principles. Business capital expenditure has grown near 9 percent over four quarters. That is the fastest pace since 2021, he said. AI-related buildout accounts for more than half of that increase. S&P 500 profits grew over 20 percent year over year. Credit spreads sit near historic lows. The July Senior Loan Officer survey showed easier lending standards for commercial and industrial loans. Warsh called the 4.1 percent jobless rate consistent with full employment. That leaves inflation as the Fed’s dominant concern.

Warsh then delivered his thoughts on inflation. Warsh cited PCE inflation at 3.7 percent over twelve months. The six-month reading runs hotter, at 4.1 percent. CPI measures are running comparably high, he said. Warsh disaggregated the PCE basket’s 199 components. Over the past year, 54 percent showed price increases above 3 percent. That is down from post-pandemic peaks near 77 percent. It remains well above the pre-pandemic norm of 32 percent. Warsh closed with a line likely to anchor coverage of the speech: he is committed to “a discipline, not to a decision.”

The Market Reaction

Treasury markets read the speech as hawkish within minutes. The 2-year yield rose more than 6 basis points to 4.298 percent. The 30-year yield eased 2 basis points to 5.168 percent. The 10-year held near 4.676 percent. BMO rates strategist Vail Hartman called the speech deliberately hawkish and said it should settle doubts about the Fed’s resolve on inflation. Equities showed a far more muted response. The S&P 500 slipped roughly 0.1 percent shortly after the speech. The Nasdaq Composite eased 0.3 percent over the same stretch.

Friday's move suggests bond investors viewed the speech as reassurance of the Fed’s resolves, however they also understand that the case for a hike is approaching.

CME FedWatch pricing moved with the curve.

Sep 26

The probability of a hike to 375-400 basis points at the September 16 meeting rose to 59.7 percent. That leaves just 40.3 percent odds the Fed holds at 350-375.

Dec 26

December pricing shifted further out along the curve. Market participants now assign a combined 89.1 percent probability to at least one hike by the December 9 meeting. It puts 40.1 percent odds on two hikes, to a 400-425 range. Another 10.7 percent prices three hikes, to 425-450.

Conclusion

This was a much needed speech from Chairman Warsh. In our August 2 piece, we were direct in our criticism of how he handled the prior FOMC press conference. He was ambiguous then and failed to offer any reaction function. Friday gave market participants the clarity they had been waiting for. Warsh confirmed that forward guidance is being retired. He backed that decision with the resolve his July remarks lacked. He then went into great detail on his thoughts of the current economy and policy stance.

Warsh gave no explicit September commitment. He spent more time building the inflation case than the AI-growth case. Market participants understood the signal here. FedWatch pricing now reflects it directly, with odds favoring a near-term hike.

Our standing call from August 2 still holds. Policy remains accommodative relative to realized inflation. We maintain that the Fed hikes once or twice before year end. A hold in September could be defensible if upcoming PCE and CPI data soften. But the Fed risks its credibility if it fails to deliver a hike before year end.

ZCR

Revenue reached $96.2 billion, up 18% sequentially and 106% YOY. Non-GAAP earnings per diluted share came in at $2.22 versus a $2.09 estimate. GAAP diluted EPS reached $2.46, up 128% YOY.

The Quarter

Data Center revenue hit $89.0 billion, up 117% YO…

Read more

​Every major infrastructure cycle follows the same playbook. A new technology unlocks an unpriced profit opportunity, sparking a capital rush that far outpaces overall real demand. The build out then explodes ahead at a fast pace. Then the inevitable shakeout clears the field, leaving behind the backbone for the next era of growth—built on the funds of initial equity holders who footed the bill.

Britain’s Railway Mania of the 1840s is the classic example. In 1846 alone, Parliament authorized 272 separate railway acts covering 9,500 miles of track. By 1847, railway capex reached roughly 7% of British GDP which soaked up nearly half of the nation's total domestic investment. As Andrew Odlyzko noted in his analysis of the period, railways became the largest capital-intensive industry of the 19th century. The tracks got laid and the infrastructure transformed the economy, but the original equity holders were decimated before ever seeing a return.

The late 1990s telecom boom followed the same pattern but on a faster timeline. Communications equipment spending rose from $62 billion in 1996 to $135 billion in 2000. That was an 18% annual growth rate, based on a 2025 KKR analysis. Capex grew unchecked, misaligned with demand. The Nasdaq fell 78% from its peak, and telecom capital spending collapsed. The fiber laid during that boom became the foundation of the modern internet, but capacity sat idle for years before demand caught up.

The AI Buildout

The AI buildout started in November 2022 when OpenAI launched ChatGPT. Hyperscaler capital spending turned sharply upward six months later in mid-2023. Combined capex among Alphabet, Amazon, Meta, Microsoft, and Oracle grew 72% per year from that point. By the fourth quarter of 2025, those five companies spent $140.6 billion in a single quarter. The shift from experimentation to full infrastructure deployment took about eighteen months.

2026 Infrastructure

Capital spending at the fourteen largest public data center operators is on track to approach $750 billion in 2026, up from under $450 billion in 2025. Data center capacity under construction globally passed 23 gigawatts by late September 2025, with three quarters of it in the United States. Analysts raised 2027 spending estimates for the fourteen largest developers by 56% between August 2025 and February 2026. The buildout is not slowing down and forecasts are still rising.

Prior Infrastructure Cycles

Scaled against the economy, the AI buildout remains smaller than the 1840s railway mania. Railway investment consumed 6% to 7% of British GDP at its peak. During speculative runs, some estimates put peak spending as high as 15% to 20%. Michael Roberts and Andrew Coyle published these figures in December 2025 research. Telecom capex during the late 1990s fiber boom topped out near 1.5% of U.S. GDP before collapsing. AI data center investment currently sits near 1.2% of U.S. GDP, while total information-processing investment runs near 4%.

However, the AI cycle is far more compressed than both. Railway mania unfolded over two decades, and the fiber boom built out over six years. The current buildout reached its intensity in under four years.

Capex

Amazon, Microsoft, Alphabet, and Meta will spend roughly $730 billion on AI infrastructure in 2026, based on Wall Street estimates. KKR places total U.S. AI capital outlays at roughly 5% of GDP. KKR also cites McKinsey research predicting nearly $7 trillion in global data center capex by 2030. That total approaches the combined annual GDP of Japan and Germany. Citigroup raised its forecast for cumulative Big Tech AI infrastructure spending through 2029 to $2.8 trillion, up from $2.3 trillion.

Goldman Sachs projects total AI capital spending will reach $765 billion in 2026. Comparing this figure to projected U.S. GDP of $32.4 trillion puts AI capital spending at roughly 2.4% of total GDP.

FCF to Financing

Through 2025, this buildout ran mostly on corporate cash. Capex as a share of operating cash flow at the largest hyperscalers rose from roughly 40% in 2023 to nearly 70% by mid-2026. Stripping out dividends and buybacks, AI infrastructure now consumes more than 90% of Big Tech’s aggregate FCF. Morgan Stanley estimates a $1.5 trillion financing gap between hyperscaler free cash flow and planned AI capital spending through 2028. Hyperscalers carry an estimated $1.65 trillion in off-balance-sheet AI obligations, according to a recent 24/7 Wall St report.

That gap is already visible on hyperscaler balance sheets. Alphabet issued $20 billion in bonds in February 2026. This sale included a rare 100-year sterling tranche. Alphabet then raised its 2026 capex guidance to between $175 billion and $185 billion. Meta completed a $25 billion investment-grade bond sale on May 1, 2026. On the same day, Meta raised its full-year capex guidance to between $125 billion and $145 billion. Meta followed with a separate $30 billion financing package for a Louisiana data center. Debt made up $27 billion of that package. Across the group, Alphabet, Amazon, Meta, Microsoft, and Oracle issued roughly $121 billion in bonds in 2025. By comparison, their annual average was $40 billion in 2020. Morgan Stanley projects global AI-related debt issuance will reach approximately $570 billion in 2026. Debt markets priced $236 billion of that total by late May. That represents four times the prior year’s pace.

In early August, Nvidia announced memorandums of understanding with six financial institutions: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The agreements aim to mobilize more than $500 billion in third-party capital for AI infrastructure. CNBC reported the deal treats Nvidia compute as a fungible entity, similar to commercial real estate or toll roads. It creates an asset class that investors can borrow against for yield. Insurance, pension, and sovereign capital can now access the AI buildout through structured credit and equity vehicles built around GPU fleets. Investors no longer rely solely on hyperscaler equity. Direct balance-sheet borrowing by hyperscalers and third-party capital from Nvidia's platform close the financing gap from opposite directions.

Future of Capex

The 1996 to 2000 fiber buildout grew capital spending at 18% a year. The current AI capex cycle grew at roughly 72% a year since mid-2023. Railway mania at its most speculative absorbed close to half of Britain's domestic investment. AI infrastructure now absorbs more than 90% of Big Tech's free cash flow. Both historical comparisons show a single sector consuming most of the available investable capital in its economy. However the funding mechanisms do differ here. Public share subscriptions and bank credit financed the railways. Corporate and telecom debt powered the fiber buildout. The AI buildout ran on operating cash flow until 2026, when it turned to structured third-party financing. Open debt markets have now loosened capital constraints on the buildout.

Risks and Parallels

Three risks stand out here when we look at historical data.

The Financing Gap

First is the financing gap, and the current numbers show it widening in real time. As covered, major hyperscalers are turning to debt markets to fund their buildout. Alphabet, Amazon, Meta, Microsoft, and Oracle issued $121 billion in U.S. corporate bonds in 2025. Estimates for 2026 global AI debt issuance reach $570 billion. For 2027, Goldman Sachs forecasts $400 billion in investment-grade bond issuance from hyperscalers against $1.14 trillion in capex. That puts roughly 35% of 2027 spending on debt.

Total lease and debt commitments across these five companies reached $969 billion. Off-balance-sheet data center leases account for $662 billion of that total, which equals 113% of their on-balance-sheet debt. Bank of America estimates AI capex will consume 94% of net operating cash flow in 2025 and 2026, up from 76% in 2024. Morgan Stanley projects a $1.5 trillion shortfall between hyperscaler free cash flow and planned AI capex through 2028. The buildout now depends on external capital markets, repeating a pattern seen in railway and fiber cycles.

History shows who absorbs that dependency when it breaks. Global Crossing filed for bankruptcy in January 2002 with $12.4 billion in debt and $600 million in annual interest. WorldCom’s filing five months later erased $30 billion in bonds, mostly held by pension funds, while its equity value collapsed from $150 billion to under $150 million. During the 1840s railway mania, ordinary share dividends averaged just 1.83% in 1850, far below the 10% promoters promised. In every cycle, late capital absorbs losses that early capital priced out. The institutional credit entering Nvidia’s August 2026 platform occupies that same position. It arrives after hyperscaler cash flow has been fully committed.

Inverse Correlation

The second risk is the correlation breakdown. J.P. Morgan Private Bank tracked the 20-day rolling correlation between semiconductor returns and hyperscaler returns from March 2025 through August 2026. The metric fell from above 0.5 to nearly -1 by mid-2026, showing the two groups moved in opposite directions. Goldman Sachs Research documented a similar shift as average stock correlation across major hyperscalers fell from 80% to 20%. Investors rotated away from debt-funded infrastructure names toward companies linking capex directly to revenue.

The telecom cycle saw similar correlation events. Equipment makers Lucent and Nortel traded in step with carriers before diverging sharply when demand lagged. Lucent fell roughly 97% from its peak by October 2002, while Nortel dropped 95%. Vendor financing accelerated that divergence. Lucent committed $8.1 billion in customer loans, Nortel extended $3.1 billion, and Cisco promised $2.4 billion. These loans inflated reported equipment revenue until customers ran out of cash. Nvidia’s vendor financing commitments, including its $100 billion OpenAI investment in September 2025, reflect a similar pattern of a chip supplier financing its own customer demand.

The correlation breakdown between suppliers and customers is where markets start pricing who absorbs a demand shortfall.

Redundancy

The third risk is structural, echoing the 1840s railway mania. Parliament authorized separate, overlapping rail lines between the same industrial towns during the 1840s boom. Redundant capacity helped society in the railway case, but it permanently compressed returns for the equity holders who paid for it.

The fiber crash demonstrates another example of the long-term cost of overbuilding. Between 1996 and 2000, carriers buried about $2 trillion of fiber, which was around 80 to 90 million miles. Sadly, only about 5% was lit by 2001. Traffic did not catch up until Verizon’s FiOS rollout around 2005 and the iPhone in 2007, leaving a five- to seven-year gap between peak build and actual use. With so much dark fiber already in the ground, wholesale long-haul prices fell roughly 55% in 2004 alone as survivors lit existing strands instead of building new ones.

This structural risk of redundancy is repeated today. Currently, Microsoft, Google, Amazon, and Meta build parallel, redundant compute clusters. They buy identical Nvidia chips and compete for overlapping enterprise workloads.

However, KKR disagrees here, arguing data centers differ fundamentally from dark fiber. Extra fiber strands carried negligible marginal costs and sat dark without maintenance. Data centers incur immediate power and cooling costs upon completion. KKR notes zero vacant sites built on pure speculation, highlighting that idle GPU capacity imposes carrying costs dark fiber never did.

Conclusion

The AI buildout is executing the classic infrastructure playbook at unprecedented speed. Hyperscalers have exhausted their internal cash flows and turned to debt markets, while third-party capital enters through structured GPU financing. As seen in the 1840s railway mania and the late-1990s fiber crash, this shift from internal equity to external leverage marks the point where financial risk spreads to late-stage investors.

Three structural flags have emerged: hyperscalers are funding a widening gap with debt and off-balance-sheet leases, chip suppliers and tech giants are decoupling, and companies are still building redundant, overlapping compute clusters.

Unlike railways or fiber, AI is a general-purpose technology capable of driving broad productivity gains across every sector. That economic potential may cushion a demand shortfall, but it does not eliminate the carrying costs of overbuilt capacity. In Part 2, we examine whether these widening credit gaps and parallel buildouts mean the AI infrastructure cycle has officially entered bubble territory.

ZCR

Read our disclaimer here.

Overview

We previously covered the July 29 FOMC decision, which ended in a 9-3 hold with three hawkish dissents. The Chairman left markets with ambiguity and a hesitant press conference. Three weeks later, the minutes provide a clearer look into what the committee was thinking.

Nine members held the target range at 3.50%-3.75%. Hammack, Kashkari, and Logan dissented in favor of a 25 basis point hike. The Committee kept the June commitment to “deliver price stability” in the statement, and almost all members agreed the language should stay.

Inflation

Total PCE inflation reached 4.1% in May, with core PCE at 3.4%. Fed staff estimated June total PCE eased to 3.7% and core to 3.3%, driven largely by falling energy prices. Staff noted this trajectory aligned with their June projections.

The board remained divided on the path ahead. Several noted tariff pass-through had largely completed. A few reported that business contacts were absorbing input costs through compressed margins, but warned that Middle East conflicts or new supply shocks could push prices higher. Many flagged re-escalating regional conflict as the primary upside risk to inflation.

Staff language was explicit: risks to the inflation forecast were “skewed to the upside,” with a distinct possibility that elevated inflation proves persistent.

AI Inflation

The minutes attribute part of inflation directly to the AI buildout, naming it alongside tariffs and Middle East energy costs. Staff noted that core goods inflation rose partly due to AI price pressures. AI-related capex is now feeding directly into broader price pressures. Participants pointed to sharp price increases in data center inputs like chips and steel. They also extended the list to consumer items, including smartphones, hardware, software, and electricity.

The committee remains split on the overall impact of the trend here. Several participants noted that AI price effects remained confined to select categories. Others argued that AI investment is already having broader effects on prices, or would likely do so soon, by pushing up aggregate demand. A few said it was still too early to know whether AI mainly shifts relative prices or drives inflation more broadly and persistently. Some noted that eventual productivity gains could lower costs and expand supply, though no consensus exists on when that offset occurs.

Labor Market

Unemployment held at 4.2% in June, remaining largely unchanged for two years. June payroll growth slowed, though first-half monthly gains averaged above the 2025 pace. Hourly earnings grew 3.5% year-over-year, down 0.4 percentage points from a year earlier. Several participants viewed this moderation as consistent with returning inflation to 2%.

Participants described a two-track labor market. Strong demand and sharp wage gains for skilled trades tied directly to the AI buildout, including electricians, machinists, and engineers. This specialized surge stood out against an otherwise flat national picture.

Broader AI labor impacts remain limited. Displaced workers have been offset by buildout-related hiring, and fears of mass layoffs have not materialized. However, the committee continues to cite the long-term employment effects of AI as a major source of forecast uncertainty.

Market Stability

Staff characterized system vulnerabilities as notable and asset valuation pressures as elevated. The reasoning centers on a thin equity risk premium, which leaves investors with little compensation for taking on equity risk. This leaves virtually no cushion if growth disappoints or rates stay elevated. Staff explicitly noted that the premium sat at a level "that has only been lower in recent history during the dot-com bubble."

Staff flagged elevated vulnerabilities across nonbank financial institutions. Hedge fund leverage reached near all-time highs across major strategies, concentrated in the largest funds. Life insurers also expanded exposure to riskier, less liquid assets.

Participants connected these valuation risks directly to AI infrastructure financing. They highlighted the growing share of AI capital spending funded by debt, including credit from nonbank lenders and regional banks. This commentary marks the Fed's clearest acknowledgment yet of the credit-quality and circular-financing risks detailed in our recent bear case coverage.

Process Update

Warsh proposed reducing the number of annual FOMC meetings from eight to six. He argued the shift would allow more data to accumulate between rate decisions and give staff more time to focus on strategic issues. The committee reached no decision on the proposal. The Chairman confirmed that meeting schedules will remain unchanged for the rest of 2026.

Our View

The minutes reinforce our post-press conference stance that policy remains accommodative relative to inflation. Holding rates becomes harder to justify as the Middle East conflict continues. Staff warnings on upside inflation risks, three sitting dissents favoring a hike, and market pricing for a September move all point in the same direction.

We maintain our call for one to two rate hikes before year-end, contingent on the regional conflict remaining unresolved.

ZCR

Read our disclaimer here.

On June 22, 2026, we opened a controlled test. We funded an automated trading bot with $1,000 in risk capital and asked one question: Can automation find an edge we cannot match by hand?

We built an automated trading system pairing AI with a Robinhood execution connection. The idea was that Python would detect technical setups and AI would validate risk and place trades for us within the RH brokerage.

We closed the test this month. Here is why it failed.

API Costs

Building the bot required initial capital. Operating it created continuous costs. Scanning live data feeds across multiple tickers generated high API fees. Sending order instructions to the broker also added costs. Overall these operational expenses reduced net profits.

Bad Fills

When we trade manually, we watch the charts, wait for indicators to confirm a setup, and act. We set limit prices or take market orders using live data. Execution happens right away. Our bot failed to match our manual speed. It scanned the market, ran rule sets, pinged the broker, and waited for a response. This process forced the bot to act on stale price data. The latency between our decision and actual execution reduced our profits on every trade.

Lag Time and Freezes

Our bot relies on API calls to the broker to execute trades. Countless times, it identified a preferred entry setup, only for lag or freezes to ruin the entire setup. On average, calling the broker, verifying the entry, and filling the order took 45 to 60 seconds. In active trading, this is a century. Active trading platforms execute trades near instantaneously usually in under five seconds.

The lag and freezes introduced by having the bot call the broker made the bot mediocre.

Decision Making

We tested two hard-coded rule sets to detect market signals. The first used a combination of ATR, EMA, and RVOL. We later shifted to an ORB strategy.

Next, we moved to a self-learning model. The bot ran during market hours, scanning and executing trades autonomously. At the end of each day, it reviewed its activity and daily P/L to learn from its errors.

This approach failed over time. The bot repeated bad trades and never learned from its mistakes. It got caught in loss loops during sudden news events and VIX spikes. It over-indexed on past data and seemed to have zero awareness of geopolitical events or shifting sentiment. It lacked something essential: human intuition.

The Verdict

We invested $1,000. We made six. Overall, the experiment was a disappointment.

Yet the test answered our core question. Automation did not find an edge we lacked. It created new costs and new delays that manual trading avoids.

This result does not mean the approach is impossible. Someone smarter will come along and eventually build a self-learning bot or hardcode rules that trade at high-frequency speeds. If API costs and token efficiency improve, a high-frequency trading bot could complement an active trader.

That is what the entire AI build-out aims to create: artificial intelligence that makes human-level decisions instinctively at high speed without supervision.

However, if that actually occurs, wouldn’t the market immediately solve for it?

ZCR

NBIS closed at $187.97 today, down 1.01%. That is more than 70% abo…

Read more

The reaction function

A 9-3 decision with three hawkish dissents marks a sharp break for a Committee that reached unanimity just months ago. The vote highlights a clear split over how to weigh persistent, above-target inflation against a largely unchanged labor market.

The statement explicitly cites Middle East conflicts and energy supply shocks as key drivers of inflation. This begins to incorporate geopolitical risk into the Fed’s reaction function. Because monetary policy targets demand by adjusting borrowing costs, the Fed effectively absolves itself from attempting to resolve supply-side disruptions directly. A rate hike does nothing to reopen pipelines or clear energy shipping lanes.

By officially acknowledging supply-side drivers, the Committee implicitly concedes that a critical component of current inflation sits outside its toolkit. This inclusion serves as both a realistic admission of policy limits and a preemptive explanation for prolonged disinflation, signaling that monetary levers alone cannot solve this phase of the inflation problem.

Transparency at the Fed

The market forgets that Fed policy was not always parsed in real time. For most of its history, the central bank did not even announce rate decisions on the day they were made. Meeting minutes and policy rationale were routinely withheld for years. Congress forced the Fed’s hand. Through FOIA and Humphrey-Hawkins reporting, lawmakers systematically unwound decades of institutional opacity. The modern regime of press conferences, dot plots, and forward guidance is a recent development. It has trained investors to treat Fed commentary as an oracle rather than what it actually is: an institutional baseline that is revised constantly and frequently wrong.

Warsh clearly understands this dynamic. He has repeatedly warned markets to “play the ball, not the referee” while actively dismantling the Fed’s guidance apparatus. He has questioned whether post-meeting press conferences should continue past this year. At the same time, he has rejected forward guidance and stripped legacy market-management language from the FOMC statement.

His underlying thesis is accurate: markets have overindexed on Fed guidance that was never designed to bear such weight. The dot plot is merely a snapshot of current conditions, not a policy commitment.

During the press conference, Warsh acknowledged that five years of persistent inflation created market skepticism that a few weeks of policy cannot undo. His critique here is valid. The Fed’s projections have missed the mark repeatedly over a multi-year stretch. Yet market participants consistently treat those forecasts as authoritative, reallocating portfolios around projections that are routinely overturned by revisions.

However, dismissing over-reliance on Fed guidance does not shield Warsh from critique. Modernizing the Fed’s outdated analytical framework with outside perspectives is a sound initiative. Bringing in outside experts to study AI’s effect on productivity or inflation is a reasonable idea on its own. The issue lies in execution.

Rather than offering interim clarity, Warsh repeatedly deflects tough questions by citing these task forces, leaving markets stranded in ambiguity. Every time he faces a hard question about policy today, he points to a task force for tomorrow. Without transparent selection criteria for members or a clear mandate on how the Fed will use their findings, investors cannot evaluate the outcome. As these reviews drag on, every pressing policy question risks being brushed aside with a generic promise that a task force is studying it.

Elevating five task forces over the standing Committee shifts institutional authority. It may ultimately prove to be a smart modernization of Fed thinking. Or it may simply be a chair hiding behind process while policy drifts.

The market’s initial verdict was not clarity, but confusion. AllianceBernstein called the press conference internally contradictory. Principal Asset Management labeled it one of the most confusing in recent memory. Some analysts noted that while the market heard a dovish tone, the prepared remarks laid the groundwork for a rate hike. That inconsistent messaging is what’s driving near-term uncertainty. It is one of the reasons the long end of the yield curve is spiking. Credit investors and vigilantes are telling the market they want rate hikes and clarity.

The 30 YR

The 30-year Treasury yield surged to 5.267%, touching its highest level since 2007. The move was instantaneous, jumping roughly 10 basis points in real time while Warsh was still speaking.

The short end moved in the opposite direction. Two-year yields fell in real time as Warsh’s comments signaled a dovish lean, prompting traders to price in a near-term rate pause. The long end refused to follow, as 30-year yields spiked that same day. Pricing in persistent inflation, energy risks, and real yield pressures, the long end drifted higher, acting as a direct rebuke to the message being delivered. This dual-direction move illustrates the uncertainty that the market faces. The short end is trading on Warsh's stated preferences, while the long end is pricing what the macro data demands.

The 30-year leaves no room for ambiguity. It forces investors to simultaneously price three decades of growth, inflation, and institutional trust. A sharp yield spike during a press conference signals that institutional money did not gain clarity. Instead, investors demanded a higher term premium to hold duration under an unproven, ambiguous policy framework.

Conclusion

Reduced communication does not automatically erode central bank credibility. Warsh makes a valid point regarding the limits of forward guidance. Identifying market over-reliance on central bank guidance, however, is distinct from managing it effectively. Warsh concluded the press conference without resolving policy ambiguity. Rather than clarifying the immediate decision, he relied heavily on five future task forces, driving the 30-year Treasury yield to a year-to-date high. The bond vigilantes are making it clear they remain unconvinced by his execution.

We maintain that current policy is accommodative rather than neutral, making a continued hold increasingly difficult to justify. Escalating conflict in the Middle East continues to support elevated energy prices, while labor market data stays resilient. Combined, these factors erode the case for holding rates.

As long as the conflict resumes, we project the Fed will hike once or twice before year-end.

ZCR

First-quarter earnings and subsequent operational milestones warrant a…

Read more

The drone and autonomous defense sector has become one of the most crowded, fastest-moving corners of the market this year. It has also become one of the weakest recently. Four smaller names dominate the conversation among investors looking for exposure: Ondas (ONDS), Red Cat Holdings (RCAT), AEVEX Corp (AVEX), and Kratos Defense (KTOS). We revi…

Read more

The Q1 Numbers

VG reported Q1 2026 revenue of $4.599 billion, up 59% year over year, and an EPS of $0…

Read more