AI infrastructure spending has been marked up again, but the more important move is in earnings. The market is getting evidence that cloud demand, memory pricing and enterprise AI revenue are rising with the buildout.
The AI capex debate has moved into a harder phase. Six months ago, the bear case was simple enough: Big Tech was about to spend far ahead of revenue, leaving shareholders with expensive data centers and a long wait for returns. Now the spending number has gone up, but so have the earnings attached to it.
That does not mean every dollar will earn its way back. It does mean the clean bubble argument is getting harder to make. The latest round of first-quarter results showed Microsoft, Alphabet, Amazon and Meta lifting or confirming capital plans while cloud revenue and AI run rates also moved higher. This is not what speculative overbuild usually looks like in its early stages. In a classic bubble, estimates rise because the story gets bigger. Here, estimates are rising because customers are still showing up.
According to first-quarter earnings compiled by the Financial Times, Google, Amazon, Microsoft and Meta are now on track to spend about $725 billion in capital expenditure in 2026, up from roughly $410 billion last year. That is an astonishing number, but it is not floating alone. Google Cloud reached $20 billion in quarterly revenue, AWS reported $37.6 billion and Microsoft said its cloud revenue hit $54.5 billion, with AI now large enough to be discussed as a measurable business rather than a product demo.
The first thing to understand is that this is not one kind of spending. The market often talks about AI capex as if every dollar is going into Nvidia GPUs, but the bill is broader than that. It includes data centers, memory, networking gear, power connections, cooling systems, custom silicon and land. Each category has its own shortage, and those shortages are now showing up in corporate guidance.
Microsoft has guided to roughly $190 billion of 2026 capex, well above the $152 billion analyst estimate cited in recent coverage, with CFO Amy Hood attributing about $25 billion of that figure to higher component costs. That is not just a bigger appetite for infrastructure. It is inflation inside the AI supply chain. High-bandwidth memory, advanced servers and electrical systems are becoming strategic inputs in the same way oil once was for industrial expansion.
Meta has also lifted its expected 2026 capital spending range to about $125 billion to $145 billion. That remains the hardest sell to investors because the company is still asking the market to trust a heavy buildout before all of the revenue is plainly visible. But even there, the core advertising business has been strong enough to keep the argument alive. Investors may dislike the scale of spending, yet they cannot say the company is funding it from weakness.
Alphabet is in a different position. Google Cloud is still smaller than AWS, but its growth rate has become one of the clearest pieces of evidence for the AI bull case. When a cloud unit is expanding quickly while management says demand is still constrained by available infrastructure, capex becomes less about faith and more about capacity planning. That is exactly the kind of detail that changes how investors price the cycle.
This changes the startup market
For AI infrastructure startups, the revised capex outlook is not just a Wall Street data point. It sets the tone for fundraising in 2026. If hyperscalers keep spending and earnings keep following, late-stage companies in memory optimization, inference routing, cooling, data center software, observability and power management will have a stronger story to tell.
The old startup pitch was that AI demand would be enormous one day. The better pitch now is that customers already have infrastructure pain, and the largest buyers in the world are paying to remove it. That is a much more practical conversation. A company that can reduce inference cost, improve GPU utilization or shorten data center deployment time is no longer selling efficiency as a nice extra. It is selling relief inside a capital cycle that is already straining budgets.
Nvidia remains the obvious winner, but it is not the only one. AMD, TSMC, HBM suppliers, networking vendors and electrical equipment companies are all tied into the same spending chain. That matters because the earnings story is spreading beyond software platforms. The money is moving into the physical economy of AI, where capacity is built, shipped, powered and cooled.
There is still a real risk here. Depreciation will catch up. Some AI workloads may not support today’s pricing. Enterprise pilots may take longer to become durable software budgets. If demand softens while these assets are still being built, the market will punish companies that treated temporary scarcity as permanent growth.
But the current evidence points to something more complicated than a bubble. Spending has been revised higher, yet revenue and earnings have not broken away from it. They have followed. For investors and founders, that means the next question is not whether AI capex is too large in the abstract. It is whether each layer of the stack can turn scarcity into margin before the next wave of capacity arrives.
Also read: NuExtract3 gives startups a smaller path to document AI • Schneider Electric says India’s AI data center buildout is now real business • AI is turning bug hunting into a security arms race
