The AI infrastructure spending race has structural foundations, and the demand math is catching up to the capital

The headline figure is arresting on its own. Meta, Amazon, Microsoft, Alphabet, and Oracle are on track to spend $720 billion on AI infrastructure in 2026 alone, according to a guest expert on The Diary Of A CEO. Marc Rowan puts a related number at $800 billion in capital expenditure from just four public companies. Public reporting from Marketplace and Goldman Sachs corroborates the general range, with estimates clustering above $700 billion for the year. Whatever the precise figure, the order of magnitude is settled. The question worth asking is whether the demand side can bear the weight.

Martin Casado, the a16z general partner, offers the ratio that makes the spending concrete: Microsoft, Meta, and Google are each on track to spend over 50% of their revenue on capex this year. That is a commitment of a kind that changes balance sheets, not just strategies. Casado also notes that the number of companies building frontier models will likely settle at somewhere between three and six, spending anywhere from $200 billion to $2 trillion per year on model development. The range is wide because no honest observer knows the answer yet, but the floor of that range is already being approached.

At the company level, Anthropic’s CFO Krishna Rao makes the intensity of compute procurement visible in ways that aggregate figures do not. Rao describes committing over $100 billion to compute deals with Google and Amazon, with another $50 billion expected to follow from deals inked recently. He spends 30 to 40% of his working time on compute. At one point, he signed two double-digit million-dollar compute commitments during a single 20-minute car ride. These are not the behaviors of an organization treating infrastructure as a line item. Harry Stebbings adds that Anthropic’s revenue commitment to Google now represents roughly 40% of Google’s total future cloud backlog, a figure that, if accurate, suggests the dependency runs in both directions.

The demand signal underlying all of this spending is not manufactured. Patrick O’Shaughnessy notes that AI compute workloads are growing 10x per year. Ivan Burazin reports the entire AI infrastructure market is expanding at roughly 40% month-over-month, with that growth rate applying across participants. Dylan Patel adds that TSMC alone is expected to spend $100 billion on capital expenditure by 2028. Gavin Baker pushes the thought experiment further: if TSMC expanded capacity to match Nvidia’s ambitions, Nvidia could sell somewhere between $2 trillion and $3 trillion of GPUs in 2026 or 2027. That figure describes a ceiling that supply constraints are currently preventing, not a projection of what will happen.

Microsoft, Meta and Google are all on in line to spend over 50% of revenue on capex this year. Martin Casado

The supply-side pressure is showing up in unexpected places. Yaroslav Azhnyuk reports that optic fiber cable prices rose from roughly $4 per kilometer to $32 per kilometer in the first months of this year, driven by demand from AI data centers. That is an eightfold increase in a commodity input over a matter of months, a signal that the buildout is straining physical supply chains, not just financial ones.

Scaling laws provide the theoretical spine that justifies continued spending at this pace. Mark Chen notes that those laws have held across nearly 10 orders of magnitude, and sees no reason they should stop holding. Zvi Mowshowitz carries that logic to its operational conclusion: by 2027, progress in AI will be primarily proportional to compute, with researcher talent becoming a secondary factor. Ethan He offers a mechanism for why that shift is plausible. As coding models allow researchers to implement ideas in hours rather than weeks, compute becomes the binding constraint on iteration speed. The bottleneck moves from human time to hardware.

There is a serious counterargument, and it deserves direct acknowledgment. Daniel Priestley estimates that the current subscriber and revenue base would need to grow roughly 45 times to justify the infrastructure investment being made, and that 95% of users currently receiving free AI tools have not shown willingness to pay for them. These are not trivial concerns. Andrew Feldman’s response, implicit in his arithmetic, is that the paying market is larger than the consumer market. He calculates that 47 million software engineers spending $50,000 to $100,000 each per year on AI tokens would represent $5 trillion in token demand from that segment alone. Whether that spending materializes is not yet known, but it reframes the justification question: the bet is less on consumer subscriptions and more on professional and enterprise use cases where the productivity math is already closing.

The structural argument for the spending holds, provisionally. Workload growth at 10x per year, scaling laws intact across a vast range, and a shift in the binding constraint from talent to compute together describe a race where slowing down carries its own risk. Whether the financial architecture built around that race proves sound is a different question, and one the evidence here cannot fully answer. What the evidence does answer is that the participants making these commitments do not believe they are speculating.