Big Tech AI Earnings and Capex

April 29 produced one of the most explicit market verdicts of the AI cycle so far. All four mega-cap hyperscalers reported the same evening, all four beat estimates, all four raised capex — and yet the tape sorted them into clear winners and losers. Alphabet rose roughly 7-10% as Google Cloud printed 63% YoY growth to $20.03B with backlog ballooning to over $460B. Amazon rallied on AWS's +28% growth, the fastest pace in 15 quarters, with management citing triple-digit AI revenue growth on Bedrock. Meta fell about 6% as Mark Zuckerberg deflected analyst ROI questions, and Microsoft sat flat despite Azure growing 40% and an AI run-rate of $37B (+123%).

The distinguishing variable was not how much each company is spending — it was whether investors could draw a clean line from capex to revenue. As S&P Global's Melissa Otto put it, Google demonstrated 'an emerging business line that is beating expectations in a pretty competitive environment.' Meta, by contrast, kept asking the market to extrapolate from internal lab signals. The implication for the rest of 2026 is that the market has stopped grading the AI trade as a single basket. Each quarter is now a referendum on whether a specific hyperscaler can convert another tranche of GPU spend into another tranche of cloud or ad revenue, and earnings-day commentary explicitly framed the prints as an 'AI payoff' test rather than an AI vision test.

Microsoft's headline 2026 capex of $190B came in 24% above the $152B analyst consensus and triggered immediate concern about discipline. But CFO Amy Hood's commentary reframed the number in a way that complicates the bear narrative: roughly $25B of the increase is not GPUs at all — it is memory and chip cost inflation passing through the budget. DRAM contract prices rose approximately 95% quarter-over-quarter in Q1 2026, and memory is now projected to consume about 30% of hyperscaler data center spending. In other words, a meaningful slice of 'AI capex' across the industry is really memory-supplier pricing power showing up in the buyer's P&L.

The other half of Hood's message is just as consequential. She warned that even with the additional investment, Microsoft expects to remain capacity-constrained on GPUs, CPUs, and storage through at least 2026. That is a striking admission for a company spending $190B: the constraint is not demand discovery, it is supply. Google's own commentary about being 'compute-constrained' near term and Microsoft's $392B commercial backlog (+51%) tell the same story. The implication is that the capex curve is not being set by hyperscaler ambition alone — it is being set by what Nvidia, the memory oligopoly, and grid operators can physically deliver, which means the $800-900B estimates for 2026 may understate true demand rather than overstate it.

Strip away the slide decks and the question that dominated investor and developer forums is the one Wall Street is starting to whisper: at what revenue level does $400-700B in annual capex actually pencil? The community math, widely circulated, is that justifying $400B of capex at a 25% gross margin and 10% depreciation requires roughly $160B in incremental annual AI revenue. Reported AI revenue in 2025 was on the order of $20B. That is an order-of-magnitude gap, and it is showing up in cash flow today: Amazon's trailing free cash flow collapsed 95% YoY to $1.2B in Q1, and analysts cited in the research warn Big Tech FCF could decline up to 90% in 2026 if the spending pace holds.

The bull rebuttal is real and not trivial. Anthropic's annualized revenue reportedly jumped from $30B to $40B in a single month, and enterprise practitioners on community forums insist demand is supply-constrained, not demand-constrained. Jensen Huang argues '$700 billion' is just a starting point because the world needs vastly more token-generation capacity than today's installed base. But the financing structure is becoming load-bearing: hyperscalers are tapping debt markets to bridge the gap between capex and cash generation, and 57% of economists in a Deutsche Bank survey now flag the AI bubble as the #1 market risk. The bifurcation in the April 29 prints suggests investors are no longer willing to underwrite the gap on faith — they want the revenue line to converge toward the spending line, on quarterly cadence.

The capex story is easy to read as software companies behaving like utilities. The truth is more literal: they are increasingly competing with utilities for the same atoms. Deloitte projects U.S. AI data-center power demand will grow more than 30x to 123 GW by 2035, up from just 4 GW in 2024. That trajectory has put U.S. utilities into what the research describes as a 'massive CapEx cycle' of their own simply to serve hyperscaler demand. Meta's stated ambition to build 'Meta Compute' targeting tens of gigawatts is, in grid terms, the equivalent of standing up multiple nuclear reactors' worth of dedicated load — a build that has to be permitted, transmitted, and energized on timelines measured in years, not quarters.

The second-order effect is concentration. More than 50% of Nvidia's revenue now comes from just five hyperscalers, which means any flinch in AI demand would ripple through semiconductors, utilities, and the broader equity market simultaneously. The 'AI trade' increasingly is the market. And the financing posture compounds the risk: with cash flow collapsing under the weight of capex, hyperscalers are leaning on debt markets to bridge the gap, a dynamic MUFG has explicitly framed as 'financing the AI supercycle.' That makes the build-out sensitive not only to demand but also to credit conditions and grid throughput — variables hyperscalers do not control. The April 29 bifurcation matters because investors are starting to price exactly this fragility: hyperscalers who can pay for the atoms with revenue are being treated very differently from those paying for them with borrowed cash and a promise.