Core theses · Updated 2026-11-02
The bubble sits in the middle layer: 7-to-10-year debt on 2-to-3-year GPUs
01The claim
The AI bubble sits in the middle layer that must keep refinancing: debt-funded compute lessors buy GPUs whose economic life may be 2 to 3 years with 7-to-10-year debt, and the losses land on creditors.
02Mechanism: why it happens
The seven-dimension scorecard can name a layer because two of its dimensions measure refinancing capacity, and refinancing capacity differs by layer. Valuation, funding, supply, narrative and exits measure market temperature, which every layer shares. Fundamentals read the scissors gap between capex and revenue; liquidity reads the 30-year Treasury yield and credit spreads. These two carry the highest weights. Temperature says the market is hot. Refinancing capacity says which layer cannot survive the heat, so the answer lands on a layer.
The liquidity dimension triggered on August 17, 2026: the 30-year US Treasury yield touched 5.31%, crossing FutureX Capital's preset threshold; the next day it reached 5.336%, the highest since 2007 (Public market data).
Heat is shared across layers; where the loss lands is not. The August reading lit six of seven dimensions: valuation, funding and narrative overheated, supply warm, fundamentals diverging, liquidity triggered; exits read as divergence and do not count. The application layer funds itself from ARR. In the model layer, equity absorbs failure: a funding attrition rate above 90% means shareholders pay. In the middle layer, AI clouds and compute lessors buy compute with 7-to-10-year debt, so losses fall on creditors. We define the bubble as the layer where creditors pay.
China's model layer drew only 22 investments in 2025, RMB 9.4 billion in total, down 52.9% year on year; fewer than ten of a hundred companies were funded, an attrition rate above 90% (36Kr/Zero2IPO, full-year 2025, Reported).
The middle-layer bubble comes from a maturity mismatch. Bonds that finance compute run 7 to 10 years; GPUs carry an accounting life of 5 to 6 years and an economic life that may be only 2 to 3. The debt outlives the collateral's place at the frontier. The last cycle had no such mismatch: fiber in the ground keeps its performance for twenty-five years. Credit is also sinking one rung at a time: hyperscaler operating cash flow in 2023-24, customer contracts as collateral in 2025, GPU-backed loans after that, single-project SPVs in 2026. Each rung down, the money rests on a weaker credit.
Amazon shortened the depreciation life of part of its servers from 6 to 5 years, cutting net income by $677 million within nine months; Meta extended its servers to 5.5 years over the same period, adding $1.96 billion to net income. Same asset, two answers (company filings, Verified).
Fundamentals and liquidity carry the highest weights because together they decide whether the middle layer can refinance. Fundamentals use only the verifiable numerator: capex is guidance the companies give themselves, while model-layer revenue is a market estimate, and an unreliable denominator should not carry an argument. We watch the second derivative of capex; the turn comes when growth itself slows. Liquidity reads the 30-year Treasury yield, which moves earlier than any valuation multiple: when the long end rises, 7-to-10-year compute bonds reprice first and the refinancing window tightens first.
Even the strongest credit rung has started refinancing: on July 22, 2026 Alphabet reported Q2 capex of $44.9 billion, about twice the prior year; free cash flow of minus $5.9 billion, its first negative quarter since the 2004 listing; and $20.3 billion of new debt issued in the quarter (company filings, Verified).
Value sits outside the middle layer because the moats on either side need no refinancing. A closed model's moat, like a compute lessor's, is capital expenditure, which shrinks first when the refinancing window tightens. An open model's moat is a living community. An application's moat is a customer paying for outcomes: retention follows results and gross margin climbs as token prices fall. Each step down the cost curve moves value from the model layer to the application layer. The middle layer pays the capex; the revenue lands one layer up. That is the other face of the scissors gap.
Per company disclosure, Dify went from 30,000 GitHub stars and zero revenue to $10 million ARR in 18 months, turned profitable in early 2026 and earns 57% of revenue overseas; Genspark reached $100 million ARR nine months after launch and $200 million at 11 months (as of August 2026, unaudited).
03Evidence
- 01Liquidity trigger: on August 17, 2026 the 30-year US Treasury yield touched 5.31%, crossing FutureX's threshold; the next day it reached 5.336%, the highest since 2007. The long end has not eased since: on October 5 the 10-year yield closed at 5.31%, its highest close since May 2002. (2026-08-17 / 2026-10-05 · Public market data · US Treasury market data; Federal Reserve H.15 (via FRED); FutureX report agi-capital-cycle-2026)
- 02The numerator of the scissors gap: the four largest hyperscalers spent $410 billion on capex in 2025, guide $725 billion for 2026 and carry a $920 billion consensus for 2027, a 77% rise from 2025 to 2026. In 1996-2001 telecom investment grew 18% a year while OECD telecom revenue grew 7.2%, falling to 1.6% in 2001, after which telecom investment shrank for seven straight quarters. (2026-09-03 · Reported · Company guidance and consensus estimates; OECD; Open Source Breaks the Deadlock slides, p. 15; FutureX report big-tech-ai-capex-2026q2)
- 03Where the second rung of credit sinking lands: Oracle's remaining performance obligations rose from $138 billion (FY25Q4) to $638 billion (FY26Q4), then to $664 billion in the FY27 first quarter ended August 31 and reported on September 10, 2026, up $209 billion year on year, with more than $30 billion of new AI cloud contracts booked in the quarter. Free cash flow was minus $23.7 billion for FY26 and minus $5 billion for the FY27 first quarter; Moody's warned debt could reach four times earnings. (2026-09-10 · Verified · Oracle 8-K Exhibit 99.1, FY2027 Q1 https://www.sec.gov/Archives/edgar/data/0001341439/000119312526387905/orcl-ex99_1.htm; Open Source Breaks the Deadlock slides, p. 14)
- 04Collateral and securitization: packaged data-center securitizations grew from $1.3 billion in 2022 to about $26.5 billion in 2025, with JPMorgan projecting $30-40 billion for 2026; CoreWeave carried $35.6 billion of interest-bearing debt at June 30, 2026. The BIS 2026 annual report warns that AI circular-financing terms are poorly disclosed, that the same asset may be pledged more than once, and that the repricing shock to credit markets could compare with 2008. (2026-06-30 · Reported / Verified (BIS annual report) · JPMorgan; CoreWeave earnings; BIS 2026 annual report; Open Source Breaks the Deadlock slides, pp. 12, 14)
- 05Third-party capital enters: on August 10, 2026 Nvidia signed a memorandum with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to build an independent financing platform mobilizing more than $500 billion of third-party capital for data-center construction and hardware purchases; Jensen Huang called compute an investable asset class. Nvidia closed down 2.86% that day. (2026-08-10 · Verified · NVIDIA company statement; FutureX report agi-capital-cycle-2026, ai-compute-infrastructure-2026)
- 06The demand-side numerator and the cash-flow denominator: on August 26 Nvidia reported data-center revenue of $89 billion for the quarter to July 26, up 117% year on year. CFO Kress put top-five hyperscaler capex at about $800 billion for 2026 and $1.3 trillion for 2027, with Nvidia's own supply and capacity commitments to its suppliers, mostly for memory, up from $119 billion to $279 billion in a single quarter. In the same quarter Alphabet's free cash flow was minus $5.9 billion, its first negative quarter since the 2004 listing. (2026-08-26 · Verified · NVIDIA earnings and call; Alphabet earnings; FutureX report big-tech-ai-capex-2026q2)
- 07Three faces of consensus: the ten largest US stocks make up about 37% of S&P 500 market value, above the 27% at the 1999-2000 peak (Goldman Sachs); global AI funding in Q1 2026 reached $255.5 billion, with three companies taking 67.3%; China's model layer drew only 22 investments worth RMB 9.4 billion in 2025, down 52.9%, with fewer than ten of a hundred companies funded (36Kr/Zero2IPO). (2026-Q1 · Reported · Goldman Sachs; 36Kr/Zero2IPO; Open Source Breaks the Deadlock slides, p. 12; FutureX report agi-capital-cycle-2026)
- 08One instance of an unreliable denominator: OpenAI reportedly told investors its annualized revenue was about $50 billion at the end of September, below the $68 billion widely circulated the month before, a figure that included partners' gross revenue. On October 8, the day the news broke, Oracle closed down 5.48% and CoreWeave fell 7.77%. When the model layer's revenue basis shifts, the middle layer that supplies its compute reprices first. (2026-10-08 · Reported (first reported by the FT, confirmed by CNBC) / Public market data · Financial Times, CNBC; FutureX report big-tech-ai-capex-2026q2)
04The strongest counterargument
Gavin Baker, Managing Partner and CIO, Atreides Management (August 23, 2026, X post, Verified)
His call is shortage, not bubble: a partly self-inflicted compute shortage running through 2028. His challenge to bears: name one quantitative metric in your business that is getting worse. His numbers: his firm's internal AI spend is about 100 times higher in August than in March and still roughly doubling monthly; rents on multi-thousand-GPU B200 clusters rose 50 to 60% in seven months. Demand is real, older cards keep repricing upward, and the bubble case has no deteriorating numerator to point at.
Our reply: We cannot name a deteriorating demand metric, and he is right to ask; that is why our fundamentals dimension reads diverging, not deteriorating. We also side with shortage: chips, power, memory, data, the constraint keeps migrating. The difference is the dial. Shortage speaks to demand; the bubble speaks to whose credit the money rests on and at what tenor. Fiber was short before it was glutted in 1996-2001, when telecom investment grew 18% a year against revenue growth of 7.2%. A shortage priced with 7-to-10-year debt against assets with a 2-to-3-year economic life is the mismatch. His data on older cards repricing upward is the strongest rebuttal we face: if older GPUs prove an economic life beyond three years, the mismatch shrinks. That is now one of our falsification conditions.
Colette Kress and Jensen Huang, CFO and CEO of Nvidia (August 26, 2026 earnings call; August 10, 2026 company statement; Verified)
Capex has real demand behind it, and the numbers can be checked: top-five hyperscaler capex of about $800 billion in 2026 and $1.3 trillion in 2027; Nvidia's own supply commitments to its suppliers, mostly memory, up from $119 billion to $279 billion in a single quarter; Nvidia's data-center revenue of $89 billion in the quarter to July 26, up 117% year on year. Huang's answer to the circular-financing charge: AI-factory compute is becoming an investable asset class, demand comes from real business use, and each project is diligenced independently by institutions.
Our reply: Purchase commitments are the numerator. They prove the capex line is still accelerating, and we have never disputed demand. Our question is what price the middle layer pays for that demand and at what tenor it borrows. The hyperscalers placing the orders are turning cash-negative: Alphabet's Q2 2026 free cash flow was minus $5.9 billion, the first negative quarter since its 2004 listing; Oracle's fiscal 2026 free cash flow was minus $23.7 billion, and Moody's warned debt could reach four times earnings. The August 10 platform answered 'who pays' with third-party capital. That moves risk from equity to credit, along a path closer to the 2000 telecom fiber cycle. Nvidia closed down 2.86% that day; the market has yet to choose between 'new asset class' and 'money passed hand to hand'. Real demand and a middle-layer bubble can hold at once; they did in 1999.
Brad Gerstner, founder of Altimeter Capital and co-host of BG2 (August 8, 2026, Reported)
Field data runs opposite to the story that open source hollows out the frontier: revenue is concentrating at the leading labs, and the intelligence gap may widen over the next two to three years. Open models grow token volume, while the economics keep flowing to frontier labs, which are paying 3 to 5 times market rates to lock in compute. On this view the middle layer sells a scarce resource to buyers who can pay; value sits at the frontier, and the bubble case has the direction wrong.
Our reply: Gerstner counts profits; we count loads. Both are right in different units. His evidence lands on the second rung of our credit-sinking chain: the compute contracts frontier labs lock at 3 to 5 times market are what the middle layer pledges as collateral. Oracle's remaining performance obligations rose from $138 billion to $638 billion in a year and to $664 billion in the quarter reported on September 10, 2026, up $209 billion year on year, while free cash flow for that quarter was minus $5 billion. The credit behind those contracts depends on the labs' ability to keep refinancing: Anthropic is reported to be finalizing a $15 billion revolving credit line, and its listing timetable keeps slipping; Bloomberg reported on October 1 that it could list in mid-November at the earliest. When lab refinancing slows, the middle layer's contract collateral reprices first. Profit concentrating at the frontier and the middle layer living on the frontier's refinancing are two faces of the same fact.
05What would change our mind
- ▸If by June 30, 2027 the combined 2027 capex guidance of Amazon, Microsoft, Alphabet and Meta still sits above $920 billion while each of the four posts positive free cash flow for two consecutive quarters, the fundamentals dimension goes back from diverging to in step, and our middle-layer bubble call drops one grade.
- ▸If the 30-year US Treasury yield falls back below 5.0% before March 31, 2027 and holds there for two consecutive quarters while AI-related credit spreads narrow in step, the August 17 liquidity trigger is logged as a false positive and the threshold is recalibrated.
- ▸Watch through end-2027: if the GPU-backed loans and data-center securitizations issued in 2025 pass through the 2-to-3-year economic-life window with no downgrade and no covenant trigger, and legacy compute contracts renew at or above their original rates, the maturity-mismatch mechanism loses weight and the location of the bubble has to be rewritten.
- ▸If Oracle posts positive free cash flow for two consecutive quarters within fiscal 2027 (ending May 31, 2027) and raises no new equity or debt in the meantime, the reading on the second rung of credit sinking is withdrawn.
06What we do, and what we do not do
- ▸For founders: do not sign long compute contracts at consensus prices. Split by workload: move batch, offline and cost-sensitive loads to open models, keep the hardest steps on a frontier model, and build the stack so either can be swapped.
- ▸For founders: in an application-layer raise, investors read three numbers first: distribution, retention, and whether gross margin climbs as token prices fall; then the four moats. A company funding itself from ARR is not tied to the middle layer's refinancing window.
- ▸For investors: the bubble is the layer where creditors pay. Any structure that buys 2-to-3-year assets with 7-to-10-year debt and re-lets the compute is one FutureX Capital stays out of, however steep the rent curve looks today. Pre-IPO rounds at 50 to 100 times sales have become the norm; we stay cautious there and do not chase.
- ▸For investors: three sets of signals to watch: the 30-year Treasury yield and AI credit spreads; the second derivative of hyperscaler capex growth and their free cash flow; Oracle's remaining performance obligations and cash flow. We do not forecast model-layer revenue and do not issue an exit call on the whole industry. This is not investment advice.
07Open questions
- ?How far does the $500 billion-plus third-party financing platform Nvidia announced on August 10 with six asset managers push back the top signal? Loan-to-value, tenor, residual-value assumptions and a rating method for GPUs are still unpublished for the first projects, so we cannot yet score this rung.
- ?Exits read in opposite directions in the two markets: zero AI IPOs overseas in Q1, 155 Chinese listings in the first half. Should one threshold become two? Bloomberg reported on October 1 that Anthropic could list in mid-November at the earliest; whether a single debut flips the US exit reading at once is something we have no precedent for.
- ?Nvidia announced on September 3 that it will acquire Hugging Face for about $12.9 billion (Per company disclosure, expected to close in the first half of 2027). Once the largest open-source distribution point belongs to the largest compute supplier, does 'an open moat needs no refinancing' still hold in full? Whether the neutrality pledge is kept is something we are tracking without an answer yet.
- ?Insurers and pension funds account for roughly 70% of global private-credit commitments (BIS 2026 annual report). If compute assets reprice, whose balance sheet the shock reaches after the SPV, and how fast, is something we have no model for.
08Changelog
- 2026-11-02Editorial revision: the title now states the claim; orchestration and data were taken out of the middle-layer definition, in line with the other theses; three facts already verified in our reports were added (the 10-year Treasury yield on October 5, OpenAI's end-September run rate, Anthropic's listing timing); the holdings list was removed; evidence grades standardized.
- 2026-09-25First published.
Related reports and answers
- Report · AGI Is Here, Capital Is Losing Its Bearings: Valuations Doubled Before Any Public Test →
- Report · Big Tech's AI Capex Heads to $730 Billion, but Markets Reward Only Contracted Spending →
- Report · Compute's Bottleneck Moves to Memory: Micron Earns $37.7 Billion in One Quarter →
- Report · AI Exits Now Stand on Three Legs, but Mega-IPO Pricing Is Tightening →
- What is FutureX Capital's seven-dimension bubble scorecard? (6 of 7 lit as of August 2026) →
- Which layer of the AI stack is the bubble in? →
- Why does FutureX Capital call open source the chokepoint of AI value? →
- What is outcomes-as-a-service pricing? (results-as-a-service, outcome-based pricing) →
Other core theses
- Applications and open-source chokepoints hold AI's value; the model layer cannot →
- The cost curve has a new author: Chinese open models at 1/40th the price →
- Verifiability times feedback cycle sets the order; each window lasts 12 to 18 months →
- Models are replaceable parts; app profits rest on four moats and outcome pricing →
- 14x apart in capital, 2.7% in capability: the US and China in one cycle →
- Embodied AI: 2026 is the hardware year, and 80 points equals zero →
This essay is FutureX Capital's judgment and argument. Facts and examples come from published research, public talks, and public reporting, graded per the site-wide standard (verified / reported / per company disclosure / public market data). Judgments change with evidence; changes are logged above. Nothing here is investment advice or an offer.