Same Capex, Different Denominators: What Q1 2026 Hyperscaler Earnings Said
Meta, Microsoft, Alphabet, and Amazon all guided to roughly the same scale of 2026 AI capex during Q1 2026 earnings (April 29-30). Three of them got rewarded. Meta lost roughly 9% of its market cap on a single guidance line. The capex numbers were comparable. The denominators underneath the numbers weren't. Microsoft's Azure backlog hit $392 billion. Google Cloud's backlog reached $460 billion. AWS reaccelerated to 28% growth on $37.6 billion of quarterly revenue. Meta has a $145 billion 2026 spend and a thesis that personal superintelligence will eventually monetize through Reels and Instagram. The market priced exactly that asymmetry.
I wrote about the macro version of this three weeks ago in the $650B zero-ROI disconnect. That post argued AI capex was running ahead of measurable productivity, and that builders were downstream of the spending in a useful way. Q1 2026 earnings didn't change the macro picture. They changed the price tag. The 2026 hyperscaler capex number rolled from $650B to roughly $725B in three weeks. The financing structure underneath it migrated further off balance sheet. And the market started naming a single line item, asymmetric risk on superintelligence, the way it once named subprime exposure.
The Number That Moved the Tape
Meta raised its 2026 capex guidance from $115-135B to $125-145B. Stock fell roughly 9% after hours. Revenue beat by 33%. Operating margin was 41%. EPS crushed consensus. None of that mattered. The single guidance line moved a $1.6 trillion company down 9%.
Zuckerberg's framing on the call was explicit: on track to deliver personal superintelligence to billions of people, tied to the first model from Meta Superintelligence Labs. The capex line was framed as the cost of getting there. JPMorgan cut Meta to Neutral, price target $725 from $825, citing limited visibility into AI product monetization against the rising infrastructure bill. The 2026 number was the second consecutive guidance reset higher. The original January 2026 range of $115-135B already represented a near-doubling from 2025.
Compare that with the rest of the field reporting the same week:
| Company | 2026 Capex Guidance | YoY revenue | Stock reaction |
|---|---|---|---|
| Meta | $125-145B (raised) | +33% | -9% |
| Alphabet | $180-190B (raised from $175-185B) | +22% | Up |
| Microsoft | ~$190B for FY26 | +18% (Q1 FY26) | Mixed |
| Amazon | ~$200B (reaffirmed) | +17% | -3% |
Combined hyperscaler 2026 capex now lands at roughly $725 billion, up from the $650B figure that anchored analyst models in mid-April. In three weeks, $75 billion of additional 2026 capex got committed across the four names. Alphabet alone ate $35.7B of capex in the quarter and its CFO told the call that 2027 will significantly increase from 2026.
The market rewarded Alphabet for the same scale of spending it punished Meta for. The difference was receipts.
Build-to-Contract vs Build-to-Thesis
Google Cloud printed 63% revenue growth at $20.0B in the quarter, with a $460B backlog. Microsoft Azure grew 40%, with commercial RPO up 51% to $392 billion. AWS reaccelerated to 28% growth, the fastest in three years, on $37.6B in revenue.
Three of the four hyperscalers can point at a signed contract for every dollar of capex. The Microsoft $392B backlog has roughly a two-year weighted average duration. The capex is being consumed by demand that's already in the books. Amazon CFO Andy Jassy told the call that AWS must invest ahead of demand because we typically lay out cash for land, power, buildings, and hardware 6 to 24 months before we start billing customers. Uncomfortable, but still a build-to-contract argument.
Meta has Reels and Instagram ads. The Andromeda ad-targeting system is real and working. Internal Meta data shows +22% ROAS for Advantage+ adopters and 6% retrieval recall improvement. Ad-tech improvements are not an infrastructure-business revenue line item the way Azure or Google Cloud are. Meta is the only one of the four hyperscalers without a third-party cloud business to underwrite the spend. The market priced exactly that.
The other side of the asymmetry is concrete. Daily users declined for the first time in Meta's history. Reality Labs is still absorbing $20B+ of annual losses. The new $145B capex line is rising faster than ad revenue can grow into it. The downside scenario is specific: if superintelligence monetization arrives in 2028+ (Zuckerberg's stated horizon), the company eats two to three years of capex compression with a deteriorating user base and a Reality Labs hole still open. The upside is contingent on a research bet whose timeline isn't visible yet. Meta's argument is build-to-thesis. Same capex number as the others, completely different denominator.
The Financing Structure Migrated Off Balance Sheet
The 2025 version of this story was hyperscalers funding capex from operating cash flows. The Q1 2026 prints made it explicit that the model has been migrating to debt and off-balance-sheet structures, and the migration is accelerating.
Meta's bond filing the day after earnings: up to $25 billion in investment-grade debt across six tranches, 5 to 40 year maturities. The order book hit $96 billion against the $25B raise. Almost 4x oversubscribed. Demand for AI-infra-adjacent investment-grade debt isn't the question. The question is what's behind it.
A week before earnings, Meta closed a $27 billion private bond transaction for its Louisiana data center, 80% owned by Blue Owl Capital, with much of the debt initially absorbed by PIMCO. Joint-venture structure. The data center sits in a separate vehicle. The debt is private. The asset doesn't sit on Meta's balance sheet the way a wholly-owned facility would.
Zoom out and the pattern is industry-wide. JPMorgan projects $30-40 billion of annual data center asset-backed securities and CMBS issuance in both 2026 and 2027. Morgan Stanley expects $250-300 billion of 2026 issuance from hyperscalers and related joint ventures. Coatue's $70B fund launched a separate venture, Next Frontier, to buy land for AI data centers, potentially tens of billions in scale, with Anthropic as a target customer.
The Bank of England named what this means at the system level. In its April 2026 Financial Policy Committee record, the BoE wrote that approximately half of AI infrastructure capex over the next five years is expected to be financed externally, mostly through debt, and that deeper links between AI firms and credit markets mean that should an asset price correction occur, losses on lending could increase financial stability risks. The BoE flagged equity valuations in technology AI as close to the most stretched they have been since the dot-com bubble. More than 65% of S&P 500 Q1 2026 returns came from seven companies.
Three weeks ago the macro framing was $650B from hyperscaler cash flow. After Superweek, the framing is $725B+ with a meaningful share moving through ABS, CMBS, private credit, and JV vehicles, sitting in pension funds and IG credit allocations. The asset is the same. The risk distribution is different.
The OpenAI Anchor Got Rewritten
Half the demand-side justification for hyperscaler capex traces back to OpenAI's compute contracts. Q1 2026 reset that anchor twice: once via OpenAI's actual numbers, once via the Microsoft contract rewrite.
The numbers first. OpenAI's revised projections show losses of approximately $17 billion in 2026, $35 billion in 2027, and $45 billion in 2028, against a $25 billion ARR run rate at the start of 2026. Cumulative spending through 2029 lands around $115 billion. The company is targeting a 2026 IPO filing with a 2027 listing at a valuation potentially up to $1 trillion. In April it raised $122B at an $852B post-money valuation co-led by SoftBank ahead of that filing, meaning the trillion-dollar mark is the public-market ask.
Then the Microsoft deal. On April 27, Microsoft and OpenAI dismantled the exclusive partnership. OpenAI can now sell across AWS and Google Cloud. Microsoft retains a non-exclusive license through 2032. Revenue share from OpenAI to Microsoft continues through 2030. The AGI clause was removed, the bizarre provision that required Microsoft to determine when OpenAI hit AGI as a contract trigger. OpenAI separately expanded its AWS agreement by $100 billion over eight years on top of an existing $38 billion commitment.
The Microsoft renegotiation matters for the capex story for two reasons. OpenAI's $1.15 trillion in committed infrastructure spending is now distributed across Broadcom ($350B), Oracle ($300B), Microsoft ($250B), Nvidia ($100B), and AMD ($90B). The single largest AI customer is intentionally diversifying suppliers. And this is the structure that produces the circular-financing critique. Nvidia invests in OpenAI. OpenAI commits to Oracle. Oracle buys 400,000 GB200 chips from Nvidia for ~$40B, and capital makes a loop. Over $800B of these arrangements now exist. The risk is that demand assumptions baked into hyperscaler capex models are partially supplied by hyperscaler-funded customers.
There's also the trial. Musk v. OpenAI began jury selection April 28, with $134B in damages and a reversal of the for-profit conversion at stake. A Musk win would require unwinding the corporate structure that Microsoft and SoftBank funded against, and would land before any IPO filing. Probability is low. It is not zero.
Capital Migration as Asset Class
Concurrent with Superweek, the broader fund-raising tape printed its own signal. Anthropic is in talks to raise approximately $50 billion at a $900 billion valuation, which would put it ahead of OpenAI's $852B post-money. The company's annualized revenue ran from ~$9B at end of 2025 to ~$30B by end of March 2026. Google announced up to $40B into Anthropic: $10B at a $350B valuation, $30B more on performance milestones, plus 5 GW of dedicated TPU capacity.
Starcloud, the company sending Nvidia chips into orbit to run inference in space, raised a $170M Series A at a $1.1B valuation in March 2026, 17 months out of YC. Starcloud-1 carries an H100 in low Earth orbit and trained an LLM there. Starcloud-2 will fly a Blackwell B200.
Individually, none of these data points support a thesis. The cluster does. Capital is migrating into AI infrastructure as an asset class: equity, credit, real estate, satellite. The capex story isn't four hyperscalers anymore. It's a financial product category, and the BoE's April note treated it as one.
The Skeptic Case Got Sharper
In February, MIT economist Ricardo Caballero published NBER working paper w34722, Speculative Growth and the AI Bubble. The argument is interesting because it doesn't reach for the obvious bubble call. Caballero's model says elevated AI valuations support rapid capital accumulation through a feedback loop where rising capitalist wealth lowers the required return, and that elevated valuations may be the actual mechanism that transitions the economy to a higher-capital equilibrium. The fragility is in coordination. The equilibrium holds only as long as beliefs hold. A coordinated loss of confidence triggers a self-fulfilling reversal.
Goldman's counter is that the Magnificent 7 trade at a median 24-month forward P/E of 27x, roughly half the equivalent valuation of the late-1990s leaders. Nvidia is at 41x forward. Cisco at the dot-com peak was 200x. PEG ratio is 1.7x today versus 3.7x at the 2000 peak. Tech-sector aggregate forward P/E is roughly 30x against 50x peak. And the companies driving this rally are profitable in a way the dot-com leaders weren't. Nvidia alone reported over $120 billion in net income for fiscal 2026.
Both arguments are correct and they describe different objects. The Goldman P/E case is about the equity in the named hyperscalers. The NBER and BoE concerns are about the capital structure underneath: the debt vehicles, the SPVs, the circular financing, and the share of demand baked into the bull case that's supplied by the same companies doing the supplying. Hyperscaler equity is not in a 2000-style bubble. The financing layer underneath might be in a different class of risk than the equity layer suggests.
Layoffs Are the Funding Mechanism
Meta will begin company-wide layoffs on May 20, cutting approximately 8,000 employees in the first wave with additional cuts planned for the second half of 2026. Total workforce reduction lands in the 16,000-employee neighborhood when the second tranche hits. Zuckerberg told an internal town hall the layoffs are about capex, not AI productivity. The capital pool that pays headcount is the same pool that pays for GPUs, and the GPUs won.
The math is unambiguous. Meta has committed $600B+ to U.S. data centers and AI infrastructure through 2028. Headcount reduction at this scale frees up roughly $3-5B of annual operating expense to feed back into capex absorption. Across the broader sector, tech layoffs are running well above 2025 levels. Amazon cut 16,000 in January. The pattern shows up earnings call after earnings call. Humans out, GPUs in, market generally rewards the trade.
The April 14 piece called this the human cost of GPU shopping. Q1 2026 made the trade explicit. The capex commitments are large enough that headcount becomes the variable cost line item that absorbs the volatility.
What This Tells a Builder
The macro frame from the April 14 piece is unchanged. $725B in hyperscaler capex is being spent. Most of it lands on infrastructure, not applications. Builders are downstream of the spend, not exposed to the underwriting risk. API access at $200-2,000/month is unaffected by whether Meta's superintelligence thesis monetizes in 2027 or 2030.
What Q1 2026 added is the financing wrinkle. The infrastructure being built increasingly sits in vehicles that distribute risk into credit markets, pension funds, and securitized products. The April 14 piece sketched two scenarios. Dot-com replay: correction, infrastructure remains, builders inherit cheap inputs. Or 2008-style systemic event: debt sits in the wrong places, correction propagates. After Superweek the second scenario has a more detailed mechanism. Meta's $27B Louisiana SPV with PIMCO holding the paper. JPMorgan's $30-40B/year ABS/CMBS projection. The BoE's approximately-half-externally-financed estimate. The OpenAI revenue numbers that anchor a meaningful share of the demand model.
For a builder, the practical implications haven't moved. Build for back-office automation, not customer-facing AI transformation programs. Use cheap, fast inference. Treat model costs as falling. Ship small, scoped projects. The infrastructure overinvestment subsidizes your access regardless of which scenario plays out. That point holds.
What changed is the time horizon to watch. The macro frame had a 5-10 year resolution window. Q1 2026 narrowed parts of that. The Microsoft-OpenAI rewrite, the $25B Meta bond, the Anthropic round, the Musk trial, and OpenAI's IPO timeline all land within 12-18 months. If a correction arrives, it likely arrives during that window, not on a multi-year fade.
The capex line item got its own name in the sell-off. When markets start naming a single line item, asymmetric risk on superintelligence, they're signaling that a category of spending has become large enough to be a balance-sheet question, not a tech-sector question. Q1 2026 was the quarter that signal printed. The next four quarters tell the rest.
Research conducted May 1, 2026. Sources from Q1 2026 hyperscaler earnings releases (April 29-30, 2026), Bank of England Financial Policy Committee record (April 2026), NBER working paper w34722 (Caballero, January 2026), and post-earnings analyst commentary through May 1, 2026.