AI Capex Hits Cash-Flow Negative in 2026 as Wall Street Eyes 2027 Peak
Aggregate hyperscaler capex is set to exceed $800B in 2026, crossing above combined operating cash flow for the first time. Bank of America projects free cash flow swings to -$144B by 2027. Wall Street consensus is coalescing around a 2027 capex peak, and the second-order effects for energy markets

Wall Street consensus has shifted: hyperscaler spending is outrunning cash generation, and analysts are already pricing a deceleration.
Key takeaways
- Aggregate capex across five major hyperscalers is projected to exceed $800B in 2026, surpassing their combined operating cash flow for the first time, per NBER Working Paper 35865 by Columbia's Stijn Van Nieuwerburgh.
- Bank of America projects aggregate hyperscaler free cash flow swings from roughly +$180B in 2025 to roughly -$64B in 2026, deepening to -$144B in 2027 and -$186B in 2028, with FCF margins dropping to -5.4% and -5.8% in those years.
- Wells Fargo's Ohsung Kwon told clients this week that the market will likely start pricing a 2027 capex peak "after 3Q EPS and midterms," a signal that Wall Street is already positioning for the deceleration.
The AI infrastructure buildout has crossed a threshold that changes the investment calculus. For the first time in 2026, the capex spent by Oracle, Microsoft, Amazon, Meta, and Alphabet will exceed what those five companies generate in combined operating cash flow, according to Van Nieuwerburgh's NBER Working Paper 35865. The gap doesn't close soon. Bank of America projects the group runs collectively cash-flow negative through at least 2028, with free cash flow dropping to roughly -$64B in 2026, -$144B in 2027, and -$186B in 2028.
That is a structural shift in how the most profitable technology companies on earth are financing their growth.
The Numbers Behind the Inflection
The pace of spending is stark. Aggregate hyperscaler capex rose from roughly $97B in 2020 to more than $400B in 2025. Bank of America now puts the 2026 figure near $860B, with a projected path toward $1.2T in 2027.
A Reuters analysis of LSEG consensus data captures the imbalance precisely: by 2027, hyperscalers will generate roughly $340B more in operating cash flow than in 2025, but capex rises by roughly $534B. That is approximately $1.57 of additional capex for every $1 of additional operating cash flow.
Alphabet's free cash flow turned negative in Q2 2026 for the first time since its 2004 IPO. The company raised its 2026 capex guidance to between $195B and $205B. Gil Luria, head of technology research at D.A. Davidson, put it plainly to Fortune: "The fact that we crossed over to negative cash flow was kind of a negative milestone, and people are reacting to that. Maybe there was a sense that we'd never get to that point, and we just did."
Van Nieuwerburgh's paper frames the structural problem directly: "Internal cash generation remains substantial, but it is no longer sufficient to finance the projected pace of investment without greater reliance on external capital or financing structures that shift assets and obligations away from the operating companies' balance sheets."
That last clause matters. SPVs, leases, private credit facilities, and sale-leaseback structures are already absorbing liabilities that don't appear cleanly on hyperscaler balance sheets. The spending is real; the obligation is increasingly distributed into opacity.
The full U.S. AI buildout tab, per the NBER working paper, runs to nearly $9 trillion through 2032, averaging 3.2 to 3.6% of GDP per year.
What a Distribution Phase Looks Like
The falsifiable thesis here is this: a spending cycle that requires continuous external capital infusions to sustain itself, while liabilities migrate off balance sheets into structured vehicles, has entered its distribution phase rather than its growth phase.
The trigger that disproves it is straightforward. If aggregate hyperscaler free cash flow recovers to positive territory by end of 2027, rather than the projected -$144B, the cycle is self-funding and the thesis is wrong. That would require AI-generated enterprise revenue to compound fast enough to close what Sequoia's David Cahn has estimated as a roughly $600B annual gap between hyperscaler AI infrastructure spending and what the AI ecosystem currently generates in actual revenue.
That gap may close eventually. It has not closed yet.
The mechanism driving this cycle is not new. Spending justified by projected future returns, funded with debt and off-balance-sheet structures, with the liability progressively pushed into opacity. When the return doesn't materialize on the projected timeline, whoever holds the structured AI-infrastructure debt at the bottom of the capital stack absorbs the loss.
Hyperscalers have already partially offloaded that exposure. Institutional capital below them has not.
Energy Markets and What Miners Should Watch
For Bitcoin miners and anyone exposed to energy markets, the 2027 capex deceleration Wall Street is now pricing carries concrete second-order effects.
Hyperscalers have been the most aggressive competitors Bitcoin miners have faced for grid access, power contracts, and cheap electrons. An $800B-plus capex wave in 2026 means long-term PPAs, colocation agreements, and nuclear off-takes are being locked up ahead of anyone else.
If that capex wave peaks and decelerates sharply in 2027 to 2028 as analysts now expect, stranded power capacity comes back to market. Power availability and pricing could ease. GPU and server supply chain pressure likely eases with it. Data center construction slows, reducing competition for electrical labor and transformer supply.
Those are structural tailwinds for miner economics, not guaranteed windfalls, but meaningful inputs into site development and power contract negotiations over the next 18 to 24 months.
Wells Fargo's Kwon expects market focus to sharpen on the 2027 capex peak "after 3Q EPS and midterms." That is the next catalyst window. Watch Q3 earnings guidance from Alphabet, Microsoft, Amazon, and Meta for any signals of capex moderation or delay, and watch whether the private credit and SPV structures absorbing off-balance-sheet AI infrastructure debt start showing stress.
Update, October 7, 2026
Arthur Hayes made his sharpest public statement yet on the AI capex cycle at the Gamma Prime Investing Conference in Singapore, telling CNBC that humanity is "wasting multi-trillion dollars" on AI data center build-up and that the entire infrastructure boom is heading toward an overcapacity collapse. Hayes sees the boom ending in overcapacity, a crash, and a bailout, with late 2027 or 2028 as the pressure point, when data centers now under construction come online and providers seek payment from AI customers he says largely remain unprofitable.
The second-order trade in Hayes' framing is compute abundance, not scarcity. He called the current AI mania a multitrillion-dollar case of capital misallocation, arguing that what it ultimately produces is cheap and accessible massive compute power. That framing has a direct read-through to Bitcoin mining: if AI datacenter overbuilding saturates grid capacity and compresses wholesale power prices, miners who locked in long-term power contracts at current rates would see margin expansion as spot energy costs fall beneath their fixed costs. Fundamental data from the IEA suggests the massive expansion in data center capacity has created a floor for energy prices, making sub-$0.06/kWh power harder to secure. A capex bust that unwinds that floor is the mirror image of the current miner cost squeeze.
Hayes is not stopping at the thesis. Rather than shorting AI stocks, he said he is positioning for cheaper, abundant compute through Flop, an AI-agent payments and compute marketplace he expects to launch in the first quarter of 2027.
He also argues that if the AI infrastructure bubble bursts and triggers a credit crisis, government and central bank liquidity injections follow, and that influx of cash could benefit Bitcoin and other cryptocurrencies, especially around 2027-2028. The thesis layers the energy deflation trade directly on top of the macro liquidity trade, making it one of the more complete AI-to-Bitcoin causal chains Hayes has laid out publicly.
Update, October 8, 2026
Firmus, an AI data-center operator backed by Nvidia and Blackstone, slashed its IPO valuation by 25%, trimming its equity value from A$44 billion to A$33 billion.
The company had been trying to raise as much as $5.5 billion at A$11 a share, which would have made it one of the largest listings in Australian history. By the time the books closed, the A$11 was gone: per the Australian Financial Review via Reuters Breakingviews, Firmus cut the price to as low as A$8, a 25% haircut, and was "scrambling" to hold it at A$8.25.
The execution gap is the core problem. Firmus cut its IPO price from A$11 to A$8.25 a share after bookbuilding stalled, with only 5% of its sold data-center capacity actually built and running, far below rival NextDC's roughly 25%.
In a note circulated Wednesday, Ten Cap portfolio manager Jun Bei Liu said her firm was steering clear of the raise, calling it an "enormous risk" with 97% of the promised data-center capacity not even built yet and an "unprecedented" lack of detail disclosed to investors.
Meanwhile, Plato, an Australian fund manager running about $6 billion, called Firmus a "screaming short," citing 30 "red flags" ranging from valuation to senior management.
Reuters Breakingviews framed the $4 billion neocloud IPO as a pinprick in the AI bubble. This is not an isolated data point. Two weeks prior, SoftBank's SB Energy pulled the marketing of its own giant data-center IPO.
Firmus is one of several neoclouds heading to public markets globally, with UK-based Nscale and US firm Lambda also planning to float in the coming months -- each of them now reading this pricing signal very carefully.
Update, October 9, 2026
A $20 billion accounting gap in OpenAI's revenue figures sent shockwaves through chip stocks on October 8, wiping out billions in market value across Nvidia, CoreWeave, Oracle, and others.
The Financial Times reported that the company told investors its annualized revenue is "approaching $50 billion," well below the $70 billion figure that had circulated a week earlier, a number devised via "attempts by OpenAI's own investors to produce a direct comparison with Anthropic's annualised revenues."
OpenAI and Anthropic calculate annualized revenue differently, with Anthropic counting sales made by its cloud partners while OpenAI does not.
Nvidia erased $169 billion in market value and CoreWeave fell 8% after OpenAI's revenue came in $20 billion below the figure circulating a week prior.
Broadcom and AMD fell sharply alongside the chip sector, exposing how supplier spending plans rest on OpenAI's unaudited, undisclosed revenue. The episode crystallizes a core vulnerability in the entire AI capex stack: the demand assumptions propping up hundreds of billions in infrastructure commitments are built on figures from a private company that releases no financial statements and whose revenue methodology no investor can audit independently.
The revenue shortfall undercuts the central premise of the AI trade, that demand for generative AI services is growing fast enough to justify hundreds of billions in capital expenditure. OpenAI has committed to spending tens of billions on computing infrastructure, including a $500 billion Stargate data center venture with Oracle and SoftBank. If revenue is tracking $20 billion below projections, the math behind those commitments becomes harder to defend.
The market grew concerned that the commercialization pace of AI model developers might lag behind the expansion rate of investments in data centers, chips, and power infrastructure, prompting investors to proactively reduce risk exposure to the AI supply chain. That is the Sequoia gap made visible in a single trading session.
Update, October 14, 2026
The AI buildout stall is now quantifiable. Between 30% and 50% of all U.S. data centers planned for 2026 have been canceled or delayed, a range that has not moved in the industry's favor through August 2026. Out of 12 GW of AI data center capacity announced for this year, only about 5 GW is under active construction. Communities across the United States blocked or delayed more than $130 billion in AI data centers in just the first three months of 2026, with lawmakers introducing more than 300 data center bills in the first six weeks of the year and 14 states floating outright moratoriums on new construction. The political layer is hardening fast. Texas Governor Greg Abbott called a halt to major data center projects pending an ERCOT audit into energy use, water consumption, and community impact, with a warning that failure to comply would result in a project being banned from the grid. New York State imposed a one-year moratorium on data center development for sites of 50 MW and over in July, following significant pushback from communities statewide. In the UK, a Stargate project for an 8,000-GPU Nscale data center was halted in April 2026 by OpenAI, which cited power costs and regulatory hurdles. The supply chain math compounds the permitting problem: shell-and-core construction costs have climbed to $11.3 million per megawatt in 2026, up from $7.7 million in 2020, after more than 50% of data center projects worldwide logged delays of three months or longer in 2025.
The credit market has been pricing the structural rot for weeks. Equities just caught up today. ZeroHedge flagged on October 13 that CoreWeave's G-spread on its 2032 bonds had breached the distressed threshold of 1,000 basis points for the first time ever, with the yield on the same paper topping 15% and the bonds trading below 80 cents on the dollar, a level known as deep distress.
The 2032 bond yield had climbed from 9.625% at issuance in June to a record close of 14.5% on Tuesday before continuing higher Wednesday. Oracle unsecureds were trading like deep junk, and Meta's Project Beignet was approaching record-wide spreads as the data center bond rout spread across the sector.
CoreWeave stock fell nearly 15% Wednesday, less than a day after Jim Cramer told CNBC viewers it was the only neocloud he trusts, with the slide following the bond selloff.
The selling spread to other AI cloud stocks, with Nebius falling around 10% and Oracle dropping another 7%. The underlying leverage is the reason. Total debt reached about $52 billion at the end of June, up from about $15 billion a year earlier.
The company's interest expense rose to $640 million in the second quarter from $267 million a year earlier, and it lost $626 million in the quarter.
D.A. Davidson analyst Gil Luria said on the Prof G Markets podcast that if the AI spending boom turns, "the equity value is going to be zero," estimating that CoreWeave borrows at around 9% and earns about 1% on its investments. Rather than pulling back, CoreWeave's co-founder Brannin McBee told the Financial Times the company is chasing insurance capital as the next funding pool, describing chip-backed debt as "accelerating debt and great debt."
The inversion that Arthur Hayes flagged is now structural, not theoretical. Miners who held energized, permitted power sites are watching the competitive dynamic shift in their direction. The U.S. interconnection queue holds roughly 2,600 GW, about double the country's installed capacity, with median waits from request to operation exceeding five years for projects completed in 2025. ERCOT's large-load queue data shows data centers at 87% of 410 GW, and in PJM, projects entering service in 2025 averaged more than seven years from request to operation. That queue backlog means the power that miners already hold is increasingly irreplaceable on any timeline that matters to a hyperscaler. The capex mirage is cracking at the physical layer and the credit layer simultaneously, and the bond market was screaming about it for weeks before equity investors noticed.
Update, October 15, 2026
CoreWeave's debt stack is now pricing distress in real time. Shares of the neocloud nosedived 15.5% on October 14, while the yield to maturity on its 9.625% bond due in 2032 hit 14.8% and its 9.75% bond due in 2031 reached 14.9%. By early October 15, CoreWeave's 2032 unsecured bonds were trading around 78 cents on the dollar, yielding just over 16% on Tradeweb per Reuters, with five-year credit-default-swap spreads also exceeding 1,000 basis points according to S&P and LSEG data.
CDS spreads sat at 368 basis points as recently as October 6, hit 643 by end of that week, and nearly doubled again nine days later.
The balance sheet numbers behind the move are stark. CoreWeave reported $35.1 billion in debt and $16.5 billion in lease liabilities as of June 30, 2026.
Interest expense alone hit $640 million last quarter, more than a quarter of revenue. The core structural problem is depreciation timing against fixed debt: CoreWeave uses aggressive GPU useful-life assumptions of six years versus an industry estimate of two to three years, while already carrying $1.3 billion in depreciation against $2.57 billion in revenue.
Gartner has warned that roughly 80% of neoclouds could shut down by 2030 if infrastructure spending outruns AI revenue, a finding that bondholders are clearly weighting into their bids.
CEO Mike Intrator pushed back hard. He dismissed the market moves in a memo obtained by CNBC, writing that "the market misunderstands and acts irrationally" and that the company is "confident in our backlog which more than supports all of our debt obligations."
The memo cited a $104.2 billion revenue backlog.
A backlog is a promise of future revenue, not cash in hand, and its value depends on customers following through and capacity being delivered on schedule -- bondholders pricing the 2032 notes below 80 cents appear less convinced the timing works out. D.A. Davidson's Gil Luria, already quoted in this piece on Alphabet's cash flow turn, was more direct: if the AI spending boom turns, "the equity value is going to be zero," he said on the Prof G Markets podcast. The Sequoia gap is no longer an abstraction -- it is now a bond yield.
Update, October 16, 2026
The AI capex debt pile is now repricing in real time. According to the Financial Times, from specialized financing bonds for large data centers leased to Meta, Microsoft, and Nvidia, to debt issued by emerging cloud companies like CoreWeave, prices have all fallen into a pressured range.
Some bonds funding data center construction for companies like Meta, Nvidia, and Oracle have fallen below par, while CoreWeave's bonds have dropped to below 80 cents on the dollar, reaching levels typically considered distressed, with yields nearing 15%.
The bonds issued by Beignet Investor, a special-purpose vehicle jointly funded by Blue Owl and Meta for the Hyperion project, have fallen from above par in early July to around 92 cents, and bonds for two projects developed by Hut 8 in Louisiana and Texas, the latter leased to Nvidia, have dropped to approximately 94 cents and 92 cents, respectively. The sell-off is broad, not idiosyncratic. The cost of insuring CoreWeave's debt against default has climbed to around 1,000 basis points, a level that implies a nearly 63% chance the company defaults within five years, up from about 30% a year ago.
The loss-making cloud company carries about $50 billion in debt and has become a focal point for concerns that the data-center boom is being fueled by credit, straining newer players and pressuring cash flows even at giants such as Alphabet and Meta.
Allianz CIO Ludovic Subran named the dynamic directly. He stated that data center bonds are entering a "Darwinian moment," with weaker credit projects facing existential challenges, and that widening credit spreads reflect how markets previously underestimated the execution risks of data center debt, exacerbated by power supply constraints and construction delays.
Hyperscaler bond issuance went from roughly $17 billion in 2024 to $109 billion in 2025, and $194 billion in the first half of 2026 alone, with Goldman Sachs tracking nearly $500 billion of AI-related debt issued this year. That supply has now run into a demand wall. The off-balance-sheet structures the base article flagged as absorbing obligations in opacity are not insulated from price discovery. The bond market just started doing it for them.
Update, October 21, 2026
Ramp's October AI Index, titled "Cracks in the AI Thesis Part 3", puts hard numbers on the demand side of the capex bubble. Just 1% of companies account for 80% of all business spending on OpenAI and Anthropic, according to Ramp, which tracks payments across more than 70,000 U.S. businesses. Ramp lead economist Ara Kharazian noted that spending on the two AI labs is more concentrated than in any other software category the firm tracks. The composition of that 1% makes the concentration more alarming: 80% of business spend on OpenAI and Anthropic comes from 1% of firms, and within that top 1%, 80.4% are venture-backed, compared to 13.1% of median spenders.
Kharazian suspects 70 to 80% of revenues at AI companies come from other AI companies that raised venture money. "This is not a circular revenue problem but it is a revenue quality problem." A dollar from a manufacturer that found a way to do more with fewer people is worth more than a dollar from a startup burning through its Series B. Same line on the income statement.
The spending stall is the new signal. Token volumes rose over the measured period, but token spend declined. The slowdown has been unusually stubborn. August and September together produced two straight months of virtually no growth in token spend.
The top 1% of firms spent $9,568 per employee per month in September, down 3.1% from its July high of $9,873.
Total dollar spend did eke out a new high in early October, but cleared its previous August peak by a paltry 1.5%.
Kharazian identifies the culprit as deflation: "the main driver is price declines at the frontier, in addition to efficiency and quality improvements for models across intelligence levels, that are driving the blended price per million tokens down."
That dynamic closes the loop on the infrastructure thesis. The hyperscalers are spending toward a projected $1.2T in capex by 2027 on the premise that AI revenue compounds fast enough to close the gap. Ramp's data shows the cohort actually paying for frontier models is a thin slice of venture-funded tech companies whose own spending is now plateauing, with token price deflation eroding the dollar value of every additional token consumed. Kharazian's warning is direct: if the trade doesn't work out, "we have a cascading effect of large customers simultaneously pulling back their AI spend." That is the demand-side mirror image of the capex bust thesis, and it now has transaction-level data behind it.
Update, October 24, 2026
A new analysis by Patrick Feeley published via Substack and picked up by ZeroHedge puts a financing-structure frame on the AI capex story that has been missing from the balance-sheet discussion. The core argument: private credit is financing the AI buildout against collateral that depreciates on a short clock, leased to a handful of tenants. When the first large tenant walks, lenders own stranded GPUs. Washington has already begun treating compute as strategic infrastructure and is unlikely to let those facilities go dark, meaning data center credit becomes what mortgage credit became after 2008: private gains, public backstop.
The numbers behind the financing gap are concrete. Morgan Stanley forecasts roughly $2.9 trillion of global data center capex through 2028, with hyperscaler cash flow covering about $1.4 trillion and leaving a $1.5 trillion external financing gap, of which private credit is expected to fill about $800 billion. That shift is already visible in public debt markets: the marginal dollar is moving away from public markets, with demand coverage on hyperscaler bond deals falling from nearly five times in February to below two times in July. Meanwhile, the guarantees underpinning the private credit stack are themselves circular. Nvidia carries residual value guarantees capped at $105 billion for its initial commitment behind OpenAI's leases at the Ohio PORTS campus, payable to the lessor SB Energy if OpenAI defaults, with total guarantees including other partners reaching $108.5 billion.
OpenAI has agreed to reimburse and indemnify Nvidia for any payout, a promise that is only worth something in the world where the guarantee is never called.
The systemic read-through is the part that maps directly onto the existing article's off-balance-sheet concern. The allocator who believes he bought senior, diversified, asset-backed credit actually owns correlated exposure to Nvidia's roadmap and four tenants' capex budgets, dressed as infrastructure.
Nothing on the public record shows a federal backstop for data center debt is planned or has been requested, but the precedent is Fannie Mae and Freddie Mac: private backstops including Nvidia's $105 billion and Meta's $41.5 billion of residual value guarantees are already real, and that is how a public one starts. The FSOC already counts the exposure inside regulated balance sheets: approximately $445 billion of bank loan commitments to private credit funds as of Q2 2025, with JPMorgan projecting data center securitization heading toward 7 to 10 percent of combined ABS and CMBS issuance, the profile of a systemic asset class.
Update, October 26, 2026
Anthropic's publicly filed S-1 prospectus with the SEC puts a precise number on the cumulative damage: net losses have added up to roughly $132.5 billion since 2023, broken out as $1.5 billion in 2023, $8.3 billion in 2024, $42 billion in 2025, and $80.7 billion in the first half of 2026 alone. That is now a public, auditable figure, not a leaked draft or an analyst estimate, and it is the single largest disclosed loss crater from any frontier AI lab.
The headline number requires unpacking. The net losses include changes in fair value of convertible notes, derivative liabilities, and SAFEs of $1.0 billion, $5.4 billion, and $34.3 billion for the years ended December 31, 2023, 2024, and 2025, respectively, and $72.4 billion for the six months ended June 30, 2026, driven primarily by increases in the company's enterprise valuation. The accounting mechanics inflate the headline, but the underlying operational burn is real and accelerating: total operating expenses reached $12.65 billion in 2025, with compute and infrastructure accounting for $7.33 billion of that, roughly triple the 2024 level.
The forward commitment stack is the number that should unsettle AI capex bulls most. Anthropic has committed at least $518 billion to six computing partners over about a decade, and roughly 80% of that cannot be cancelled or must be paid regardless of usage.
The company's own filing acknowledges that its ability to remain at the frontier depends "in meaningful part" on its ability to generate and sustain substantial revenue. With the S-1 now public and a roadshow reportedly targeting the week of November 9, public market investors will price that bet in real time -- a stress test the entire AI capex stack has not yet faced.
Update, October 28, 2026
The credit market just drew a hard line on AI infrastructure debt. Bloomberg reported today that banks have shelved the planned bond sale for the Anthropic-tied Nexus Data Centers campus in Hubbard, Texas, a 2,000-acre project that had been backed by a Morgan Stanley-led consortium with $15 billion in construction financing. The bonds were expected to receive a speculative-grade rating despite Google's backing, which only becomes effective after the data center is fully built, and bond buyers apparently decided construction risk plus junk paper plus AI demand uncertainty is too much to absorb right now.
The shelved Nexus bond deal is not an isolated event. A separate Bloomberg report published yesterday revealed that a parallel Anthropic financing, U.S. lenders weighing a bond sale of up to about $21 billion tied to Anthropic's chip leases, using a special-purpose vehicle called Project North Star to borrow money, buy chips developed by Broadcom and Google, and lease them to Anthropic, with Anthropic's lease payments intended to service the SPV's debt, nearly fell apart before it was even announced. Broadcom's CEO spent nearly two months trying to get banks to commit over $40 billion to the largest AI debt deal ever, and given rising interest rates and growing difficulties in markets for AI ventures, some banks felt they weren't being compensated enough and considered pulling out altogether, forcing 11th-hour talks with bank executives over a weekend, with Broadcom agreeing to higher loan rates to secure their participation.
Banks also got more comfort around guarantees from Broadcom that will protect them from losses in case Anthropic defaults.
What the Nexus shelving adds, on top of yesterday's chip-deal drama, is a clear read: AI infrastructure bond issuance is now hitting active resistance across two different deal types, project-finance real estate debt and chip-lease SPV paper, simultaneously. Developers usually repay construction debt by issuing new debt once a project is done, a model that depends on lenders and bond buyers staying willing to fund the next round. That willingness is now in question at scale. The structured vehicles the base article flagged as absorbing off-balance-sheet liabilities are running into the same credit friction as the operating companies themselves, the opacity is not insulating the debt, it is concentrating the risk into instruments the bond market is starting to walk away from.
Update, October 29, 2026
Alphabet's Q3 2026 earnings, filed with the SEC on October 28, put a hard number on the cash-flow deterioration this article has been tracking. Alphabet reported negative free cash flow of $9.07 billion during the third quarter.
Capital expenditure reached $51.36 billion in the quarter, more than double the $23.95 billion of a year earlier, while cash from operations fell to $42.29 billion from $48.41 billion.
The heavy spending has pushed Alphabet's free cash flow further into negative territory, with Q3's negative $9.1 billion worse than the negative $5.9 billion posted in Q2. That is two consecutive quarters of negative free cash flow, and the trajectory is deepening, not stabilizing.
The negative $9.07 billion result fell much further than the negative $2.70 billion analysts had expected.
Alphabet covered the gap with $28.7 billion of new notes and $12.4 billion of new shares, bought back no stock, and ended September with total assets of $1,003.3 billion against $595.3 billion at the end of 2025. The machine is now structurally dependent on capital markets to stay operational. As one analyst put it, "Alphabet is now funding the AI buildout with shareholders' money instead of its own cash."
The company left its full-year 2026 capex guidance range of $195 to $205 billion unchanged and continues to expect capex to increase significantly in 2027, with further details to come on the Q4 call.
Alphabet shares slumped more than 10% after the results dropped.
Meta also reported negative free cash flow last week after raising the bottom of its full-year capex range to $135 billion, and its shares sank on the same dynamic. The negative cash flow inflection is no longer a single-company event -- it is the peer group.
Update, November 2, 2026
Anthropic's amended IPO prospectus filed Monday with the SEC delivered the most complete single-company accounting of the AI capex spiral yet. Q3 revenue reached $16.96 billion, up roughly 1,127% year over year, while the net loss fell to $23.29 billion from $66.3 billion in Q2. The narrowing looks better until you read the footnotes: the Q3 headline loss was largely driven by a $21.9 billion noncash revaluation charge on convertible notes, meaning the financing structure itself is generating the number. On an operating basis, Anthropic posted a $567 million loss in Q3 after a single profitable quarter of $974 million operating income in Q2, one quarter in the black followed immediately by a return to red.
The nine-month totals are the number that reframes everything. For the first nine months of 2026, Anthropic reported $33.24 billion in revenue and a net loss of $104 billion, including $94.3 billion in noncash convertible note fair-value charges. That compares to a $25.1 billion loss over the same nine-month stretch in 2025, a four-fold acceleration in one year. Losses are running at roughly three times revenue on a headline basis. Compute and infrastructure expenses rose $14.4 billion over nine months, a 329% increase, with revenue growing fast and compute costs growing faster.
The commitment disclosure is the number that anchors all of it. Anthropic's amended filing lists about $895.4 billion in purchase commitments and lease agreements, substantially all for computing and data centers, with about $70 billion coming due within the next 12 months, nearly seven times the roughly $132 billion in cash and short-term investments the company held at end of September. About $447.5 billion of that total is expected to be incurred before the end of 2030. The counterparty list is a who's who of the infrastructure stack: Broadcom at $161.2 billion, Google at $121.3 billion, SpaceX's AI unit at up to $120.5 billion, Amazon at $110 billion, Microsoft at $44.9 billion, Nscale at up to $44.6 billion, and Lambda at $41.3 billion. The total is still understated. Anthropic plans another $125 billion to $150 billion of Google chips and more than $20 billion with AMD, neither of which carries a fixed dollar commitment, so neither counts toward the $895.4 billion.
The prospectus also flags a structural mismatch that makes the commitment stack harder to escape: many agreements contain minimum-purchase or take-or-pay obligations regardless of utilization, while Anthropic's revenue is primarily consumption-based and much is not under long-term contract. The same neoclouds flagged in the October 8 update, Nscale and Lambda, appear directly in Anthropic's commitment schedule, tying the IPO-market stress already visible in the Firmus pricing fiasco to the demand side of the ledger. This is the Sequoia gap rendered as a balance sheet: locked-in supply costs on one side, variable consumption revenue on the other, and a forward commitment stack nearly seven times available cash.
Update, November 11, 2026
The broken-promise ledger on OpenAI's datacenter buildout is now explicit. On November 5, 2026, OpenAI published a National Community Compact, a set of public promises for the data centers it plans to build across the United States. The document arrives against a backdrop of serial delivery failures. OpenAI's original commitment to spend $1.4 trillion on AI infrastructure before the end of the decade was recently more than halved to a still-hefty $600 billion.
Bernstein puts known datacenter contracts at roughly $614 billion through 2030, and thinks even that figure is unlikely to be fully realized, with construction delays, cancellations, and flexible hardware purchases likely to shrink payments or push them out. Meanwhile, OpenAI's $500 billion Stargate Project in Abilene, Texas, backed by SoftBank, Oracle, and others, showed no significant physical progress as of April 2026 , and Bloomberg reported that Oracle and OpenAI scrapped plans to expand that flagship AI datacenter in Texas after negotiations dragged over financing and OpenAI's changing needs. On the chip-deployment side, OpenAI and Nvidia said in their September release that the first gigawatt of Nvidia systems would be deployed in the second half of 2026, but experts said that timeline would be tough in the best of circumstances.
The political pressure is now doing what market pressure alone could not. The compact announcement follows voters passing bans and other restrictions on data center construction in nearly 40 communities.
Residents of Effingham County had already sued their local officials in September over a secret deal with OpenAI , and the new community promises are not legally binding. The compact covers only campuses OpenAI directly owns and operates. That exclusion is significant: SB Energy builds, owns, and operates the larger Ohio campus that OpenAI leases, and Oracle supplies Stargate capacity including the Abilene site, and the compact does not say whether leased sites and joint ventures are covered.
The pattern is consistent: OpenAI announces headline numbers, the physical and financial reality comes in below those numbers, and the company responds with a new set of public commitments. The November compact gave no sites, no dates, no megawatts, and no dollar totals, and that silence tells you as much as the pledge does.
The actual capital collected has significantly lagged pledges. For the AI capex trade, the read-through is straightforward: the single largest named demand anchor for the entire infrastructure stack keeps shrinking its commitments while adding political friction to the construction pipeline. Every project that slips or cancels is capex that does not deploy, revenue that does not arrive, and a wider gap between the story sold to investors and the infrastructure actually getting built.
Update, November 16, 2026
SemiAnalysis cut its AI accelerator model on November 13, just 11 days after its prior release, slashing 2028 global AI accelerator capex by 28%, with 2029 and 2030 projections falling 41% and 40% respectively. The firm reduced its 2030 global AI accelerator spend estimate to $923 billion, down from $1.531 trillion, and lowered its 2028 outlook from roughly $1.05 trillion to $758 billion.
The downward revisions are driven by tightening capital markets and a slowdown in expansion among frontier AI labs, with updated forecasts also cutting projections for total power demand and chip shipments starting in 2028.
The cuts are largely driven by price deflation rather than volume collapse, and Nvidia absorbs the vast majority of the reduction, while Broadcom and Amazon's internal ASIC efforts remain largely insulated.
Nvidia accounts for 66% of the decline in silicon wafer shipment area, indicating this round of adjustments to AI hardware demand is not being evenly borne by all suppliers. On the analyst side, Nvidia price targets were cut hard: HSBC slashed its target to $170 from $365, UBS lowered its target to $240 from $300, and Mizuho cut its target to $220 from $250.
SemiAnalysis also slashed its 2028 HBM demand forecast by 30% and cut its 2030 power consumption demand forecast by 34%. That last number is the direct read-through for the energy narrative. If a tier-1 semiconductor research firm is now modeling a one-third reduction in AI data center power draw by 2030, the grid pressure that has compressed Bitcoin miner margins on power contracts begins to look like a ceiling rather than a permanent floor. JPMorgan added its own pressure point, saying Nvidia GPUs will account for only 39% of AI compute purchased by OpenAI and Anthropic going forward, against 83% of overall AI chip revenue from 2023 to 2025. The forecast revision cycle has started, and SemiAnalysis just provided the first hard numbers.
Update, November 19, 2026
The $500 billion financing structure Wall Street assembled to backstop Nvidia chip purchases has collapsed in practice before it ever started. Wall Street's biggest banks and asset managers pledged up to $500 billion to help Nvidia customers finance chip purchases earlier this year, but no capital has been deployed to date, according to the Wall Street Journal.
The entire arrangement, announced in August, rested on a few-page memorandum of understanding, not binding commitments.
Nvidia was willing to cover only up to 25% of project costs, and as lenders grew more cautious on AI financing, they pushed Nvidia for more protection. That is a structurally weaker backstop than rival Broadcom offered: Broadcom provided roughly 85% cover on Anthropic's Big Sky deal, against Nvidia's 25%.
The leadership vacuum at Nvidia is accelerating the freeze. Investors are growing increasingly skeptical the program will reach its goals after Jensen Huang's abrupt departure, and interim CEO Colette Kress has expressed skepticism about Nvidia's financing commitments, saying during a recent earnings call that the company will be more disciplined in financing customers.
Credit markets have grown more cautious toward AI borrowers with low credit ratings since early August. No chip financing deals are expected to close under the partnership before 2027.
The implications for the energy and power narrative tracked in this article are direct. Between the Democratic sweep of Congress and Huang's departure, the AI bubble has rarely looked closer to bursting, and the wellspring of capital from Nvidia, for both hyperscaler construction ambitions and energy startups, could start to ebb. If the financing mechanism designed to push chip demand forward into the neocloud tier has zero capital behind it, the marginal data center customer -- the same customer competing with Bitcoin miners for grid capacity -- loses its primary funding bridge. The overcapacity crash Arthur Hayes outlined in October now has a named, documented mechanism: half a trillion dollars in pledged capital that exists only on paper.
Update, November 20, 2026
Michael Burry amplified a damaging Wall Street Journal investigation this week, sharing its findings on X: Wall Street's biggest banks and asset managers had pledged up to $500 billion to help Nvidia customers finance chip purchases earlier this year, but no capital has been deployed to date. The $500 billion headline figure that defined Jensen Huang's summer financing tour was always a nonbinding aspiration. The firms agreed in August to explore ways to organize the funding, but their commitments were set out in a nonbinding memorandum of understanding, so they did not have to raise a specified amount.
The structural snag is straightforward. Nvidia's proposed credit guarantees covering as much as 25% of project costs became a sticking point, with lenders seeking stronger protections before proceeding.
Broadcom had agreed to a much bigger guarantee, covering about 85% of the Anthropic financing , which sets a benchmark Nvidia has been unwilling to match. Jensen Huang, who pushed for the plan, resigned last week, and his interim replacement Colette Kress had already questioned some of those guarantees as CFO.
People familiar with the talks told the Journal they expect the first deal in 2027 at the earliest.
Burry's read-through here is straightforward and consistent with positions he holds. He maintains a bearish position on Nvidia via $140 strike put options expiring in February 2027. The $500 billion commitment evaporating into a nonbinding memo with zero capital deployed is exactly the kind of paper-promise infrastructure this story has been tracking: aggregate AI capex commitments built on financing structures that exist in press releases before they exist in deployed dollars. In effect, the $500 billion was a fundraising target for outside capital, not money already raised. That is not a footnote. That is the entire story.
Sources
- NBER Working Paper 35865, "Financing the AI Buildout," Van Nieuwerburgh (October 2026)
- Brookings Papers draft, "Financing the AI Buildout"
- Reuters / LSEG analysis, "AI investment boom puts Big Tech's free cash flow under pressure" (July 22, 2026)
- CNBC, "Best stocks for cash flow amid AI spending" (September 30, 2026)
- Fortune, "AI is heading toward a cash-flow negative inflection point, and spending will slow in 2027, Wall Street thinks" (October 7, 2026)
Frequently Asked Questions
Why does AI going cash-flow negative matter if the hyperscalers are still posting large profits?
Profitability and free cash flow are different things. A company can report strong net income while its capex consumes more cash than its operations generate. Free cash flow is what funds buybacks, dividends, debt repayment, and organic R&D.
When capex exceeds operating cash flow, the company must issue debt, sell assets, or use off-balance-sheet structures to keep spending. That is not inherently fatal, but it changes who bears the risk and on what timeline. The BofA projections show hyperscalers running -$186B in aggregate FCF by 2028. That is not a rounding error; it is a financing problem.
How does a 2027 AI capex peak affect Bitcoin miners and energy markets?
Hyperscalers have been the dominant bidders for grid capacity, long-term power contracts, and data center construction resources since 2023. A meaningful deceleration in their capex spend in 2027 would release power capacity back into the market, ease competition for electrical infrastructure, and potentially lower the cost basis for miners looking to expand. The effect won't be immediate, but site developers and miners negotiating multi-year power agreements in late 2027 and 2028 may find meaningfully better terms than those available in 2025 and 2026.
Is this the dot-com bubble repeating?
The structural parallel is real: reflexive capex ahead of proven ROI, financed with external capital, with liabilities increasingly obscured in structured vehicles. But the failure mode is more likely telecom overbuild 2001 than pure dot-com vaporware. The underlying technology generates real usage and real cash flows, just not enough to cover the investment at the current pace. The risk is not that AI produces nothing. The risk is that the return timeline is long enough to cause significant balance-sheet stress in the institutional capital stack below the hyperscalers before the productivity dividend arrives.


