Alphabet raised its 2026 capital expenditure forecast to $205 billion last week, then its stock fell 7%. Google Cloud grew 82% in Q2 — the fastest growth among the major cloud providers since 2020 — and revenue beat every estimate on the table. The earnings report was strong on every operational metric.
The market didn't care about cloud growth. It cared that Alphabet turned free cash flow negative for the first time in the company's history, and that 2027 capex "will increase significantly," per CFO Anat Ashkenazi.
The hyperscaler AI story hasn't changed. Cloud demand is accelerating, not slowing. But the financing story has, and that changes the risk calculus for every company in the supply chain. We've moved from Phase 1 — where hyperscalers spent from internal cash with essentially unlimited runway — to Phase 2, where they're borrowing to build infrastructure and the market is demanding proof that each dollar spent converts into durable returns.
Quick Backtesting Tool
The numbers tell a Phase 2 story
Here's where the four major hyperscalers stand after Q2 2026 earnings, all reported within a single week in late July:
| Company | 2026 Capex Guidance | Cloud Growth (Q2) | Free Cash Flow Status |
|---|---|---|---|
| Alphabet | $195–205B (raised) | +82% (Google Cloud) | Negative for first time |
| Amazon | ~$220B (raised from $200B) | +37% (AWS) | Negative, flipped in Q1 |
| Microsoft | ~$190B (held) | +43% (Azure) | Positive, but via accounting shifts |
| Meta | $130–145B (narrowed higher) | N/A (no external cloud) | Remains positive |
Combined, the four companies are projected to spend roughly $760 billion on infrastructure in 2026 — up about 77% from 2025. That's the largest capital expenditure cycle in technology history. But the structure of that spending has shifted in ways the market is only beginning to price in.
Alphabet and AmazonAMZN-- have both turned free cash flow negative. Amazon's trailing 12-month cash from operations was $161.4 billion, but net spending on property and equipment hit $169 billion, creating a $7.6 billion shortfall. MicrosoftMSFT-- is staying positive through accounting maneuvers — shifting data center leases from finance to operating leases and extending data center useful lives from 15 to 25 years. CFO Amy Hood told investors the company will remain free cash flow positive in fiscal 2027, but that guidance now requires more trust than it did last year.
Meta is the outlier. It remains cash-flow positive despite spending $31.1 billion on capex in Q2 alone, nearly double the year-ago period. And unlike its peers, MetaMETA-- has no established external cloud business. Instead, it's monetizing excess compute through what Zuckerberg described as an auction system similar to its ad platform, and is in talks for what could be a $10 billion compute lease to Anthropic over two years.
What these numbers mean is simple: the hyperscalers are no longer self-funding their AI buildout from internal cash generation. They're running on debt. Alphabet's long-term debt jumped 111% to $98 billion in the first half of 2026. Amazon's rose 81% to $119 billion in just one quarter.
Demand hasn't broken. It's accelerating.
The operational side of the story is the opposite of what "AI narrative risk" headlines imply. Cloud growth is accelerating across every major provider.
Google Cloud grew 82% in Q2, accelerating from 63% in Q1, and beat analyst estimates of roughly 64%. Alphabet's contracted future revenue backlog swelled to $514 billion — the strongest forward-demand signal in the company's history.
AWS grew 37%, its fastest pace since the end of 2021, accelerating for the fifth consecutive quarter. The segment added $4.6 billion in revenue quarter-over-quarter — approximately 80% more than its previous largest sequential increase. AWS operating margins expanded to 39.4% from 32.9% a year ago, despite heavy AI investment.
Azure grew 43% in the fiscal fourth quarter, surpassing consensus of roughly 40%. Microsoft Cloud revenue surpassed $214 billion for the full fiscal year, up 27%, and nearly 90% of cloud revenue now comes from customers outside frontier model companies — meaning the growth is enterprise-driven, not concentrated in a handful of AI labs.
Put plainly: the hyperscalers are building capacity because they can't keep up with demand, not because they're trying to front-run a market that's already saturated.
The growth rate cliff that actually matters
This is where the narrative shift is real. UBS sees hyperscalers' capex growth slowing to 25% in 2027 and 6% in 2028. These are growth rate projections, not absolute spending collapsing to zero. A 6% growth rate on a $760 billion base is still $45 billion in additional spending.
But here's the thing that changes the investment calculus for chip suppliers: the market has priced in the 76% growth trajectory. Semiconductor stocks have more than doubled over the past year, and Bank of America's July survey found 82% viewed semiconductors as the market's most crowded trade — with none holding short positions.
When capex growth decelerates from 76% to 25%, chip revenue growth decelerates too. The question isn't whether hyperscalers keep spending. It's whether the growth rate they're priced for survives into 2027 and 2028.
Nvidia, sitting at a $5.4 trillion market cap with 70.7% revenue growth, 64% operating margins, and 89.4% ROIC, is still running an extraordinary business. But at 21.4 times trailing sales, the stock is implicitly assuming that hyperscaler capex continues accelerating at roughly current rates for the foreseeable future. If the 25% capex growth rate in 2027 translates to materially lower chip revenue growth for suppliers, that multiple has work to do.
What management is actually saying
The most useful signal isn't in the headlines. It's in how each hyperscaler is structuring its spending to manage the risk.
Andy Jassy at Amazon broke capex into two buckets: long-lived assets (data centers and land, committed two years before servers arrive, monetizing for 30+ years) and short-lived assets (servers and networking equipment, purchased months before service, breaking even in under three years with a useful life of five to six years). "If demand is insufficient, spending is halted," on the short-lived side. This decoupling is a deliberate risk management move. It means Amazon can commit to the infrastructure that takes years to build while retaining the flexibility to pull the plug on chip purchases if demand slows.
Microsoft is doing something similar. CFO Amy Hood characterized capex as heavily pivoted toward short-lived assets — CPUs and GPUs — which have shorter lead times. The company is also shifting lease classifications and extending data center useful lives to manage the free cash flow narrative.
Alphabet is the most transparent about the pressure. Ashkenazi cited demand for Google Cloud AI infrastructure exceeding available compute as the driver for the capex raise, and warned that supply constraints will persist into 2027. The 2027 capex "will increase significantly" was the line that mattered — not the Q2 beat, not the cloud growth, the forward trajectory.
Zuckerberg at Meta framed it differently. He described Meta as a "full-stack technology company" owning its data centers, infrastructure, chips, and low-level software, and warned that relying on another lab's frontier model is a "policy and business risk" Meta will not take. Meta is building four new revenue streams beyond advertising — transactional business services, compute leasing, enterprise APIs, and consumer subscriptions — and Zuckerberg said the company is receiving offers for compute at a "significant premium" over cost.
The common thread across all four companies is conviction in continued AI demand. The difference is how each is positioning for the inevitable slowdown in capex growth rates.
The rotation is already happening
Active managers are moving capital in response. Edmond de Rothschild's Alexis Boss significantly reduced semiconductor exposure and increased positions in Amazon, liquid cooling, cybersecurity, and software. Alberto Conca at LFG+ZEST reduced memory chip and semiconductor equipment holdings and bought put options on selected semiconductor names.
The Philadelphia Semiconductor Index has retreated nearly 18% from its June high, while the equal-weighted S&P 500 gained 11% over the same period. The rotation isn't complete — chip-focused funds drew record $10 billion in net inflows through May — but the direction of institutional capital is shifting from the suppliers to the buyers.
This makes structural sense. The hyperscalers have the pricing power to demand ROIC discipline from their chip suppliers. They're the ones signing five-year contracts with AWS, Azure, and Google Cloud. When capex growth slows, the buyers have more leverage than the sellers.
Where this leaves capital allocation
The debate is not whether hyperscaler AI spending continues. It's whether the growth trajectory that chip suppliers are priced for survives the transition from 76% capex growth to 25% to 6%. And whether the hyperscalers themselves can generate returns on $760 billion of incremental spending that justify the debt they're taking on.
I believe the demand story remains intact. Cloud growth is accelerating across every major provider. AWS margins are expanding despite heavy investment, not compressing. The backlogs are the strongest in these companies' histories. Enterprise adoption of AI inference is still in early stages, as Jassy noted.
But the financing story has changed, and that matters for how I think about allocation. The hyperscalers are moving from self-funding infrastructure builders to debt-financed construction projects. Corporate debt cover ratios have fallen below 2 times, down from nearly 5 times in February, according to Reuters. The Bank for International Settlements has warned that disappointment in returns could trigger a sudden pullback in financing.
For Nvidia specifically, I still believe the company will extend its dominance in AI accelerators for years. The CUDA ecosystem remains the training infrastructure standard. But the inference market — where efficiency and cost matter more than the CUDA moat — is where margin pressure will first appear. And the 2027–2028 capex deceleration is a real headwind for revenue growth assumptions baked into the current multiple.
What would change my thesis? If hyperscaler capex guidance starts getting cut rather than raised, that would signal demand softening and I'd reassess positions across the supply chain. If AWS, Azure, and Google Cloud growth decelerates sequentially for two or more quarters, that would be the operational confirmation that the spending surge is losing momentum. If chip supplier revenue growth materially underperforms consensus in the face of 25% capex growth, that would prove the multiple compression risk is real.
For now, the story hasn't broken. It's just entered a phase where the hyperscalers need to show that $760 billion of spending converts into $760 billion worth of revenue — and the market is finally asking for the receipts.
That's not a reason to sell. But it is a reason to think about opportunity cost. When the semiconductor trade is the most crowded position in global markets, and the growth rate cliff is two quarters away, the capital that's rotating toward hyperscalers and software beneficiaries is probably right to do so.
