AI’s Spending Boom Has Outrun Its Proof

The strangest thing about the AI boom is that it looks most convincing where it tells us the least.

Nvidia is selling chips at a staggering pace. Microsoft, Amazon, Google and Meta are pouring money into data centers. OpenAI is preparing for Wall Street. Workers are being cut while executives talk about agents, copilots and superintelligence. If you stand close to the machinery, the boom looks undeniable. If you step back and ask who is earning better profits because of all this spending, the picture gets hazier.

I think the right verdict is uncomfortable: AI is already useful, but the market is pricing it as if the hard part has been solved. It has not. The current surge looks less like a pure fraud than a parallel infrastructure bubble wrapped around a real technology.

Start with the money. Capital expenditure, or capex, is spending on long-lived assets: chips, servers, buildings, power systems and networking gear rather than normal day-to-day expenses. The AI boom is capex-heavy because modern generative AI runs on GPUs, or graphics processing units, chips built for the parallel math needed to train and run large models. Those chips sit in data centers, which are basically industrial-scale computing plants with power, cooling, networking and security. They run foundation models, the general-purpose AI systems that can be adapted for coding, customer service, search, advertising, document work and robotics.

By that definition, this is an infrastructure race. Nvidia reported fiscal 2026 revenue of $215.9 billion, up 65%, and fourth-quarter data-center revenue of $62.3 billion, up 75% from a year earlier, according to its February 2026 earnings release¹. Meta said it expects 2026 capital expenditures, including principal payments on finance leases, of $115 billion to $135 billion, driven by AI and its core business, according to its fourth-quarter 2025 results². Alphabet told investors it expected 2026 capex of $175 billion to $185 billion, according to its Q4 2025 earnings transcript³. Amazon’s Andy Jassy said the company expected about $200 billion of 2026 capital expenditures across Amazon, citing demand and opportunities in AI, chips, robotics and satellites, as reported by GeekWire⁴.

That is not software economics. It is closer to railroads, telecom fiber or power grids: huge fixed costs first, profits later if utilization stays high.

The optimistic case has real evidence. Microsoft said in late April that its AI business had surpassed a $37 billion annual revenue run rate, up 123% year over year, while Azure and other cloud services revenue grew 40%, according to Microsoft’s FY26 Q3 release⁵. Amazon said AWS’s AI revenue run rate was over $15 billion in Q1 2026, according to Amazon’s own earnings commentary⁶. Alphabet said Google Cloud revenue rose 48% to $17.7 billion in Q4 2025, led by Google Cloud Platform growth across enterprise AI infrastructure, enterprise AI solutions and core products, according to Alphabet’s earnings release filed with the SEC⁷.

There are also measurable task gains. A Microsoft Research experiment found developers using GitHub Copilot completed a programming task 55.8% faster than a control group, according to the study⁸. An NBER paper by Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that a generative AI assistant raised customer-support productivity by nearly 14% on average, with larger gains for less experienced workers, according to NBER⁹.

If that were the whole story, I would call the bubble talk lazy. A technology that speeds up coding and helps customer-support agents resolve more issues is not vapor. Cloud customers are paying for something.

But the investor question is harder than the demo question. A faster coding task does not tell us whether a software company earns more operating income after integration costs, security review, quality control and subscription fees. A cloud AI run rate does not tell us whether the customer buying compute has a profitable use case. Nvidia revenue is revenue for Nvidia. It is not proof that the economy downstream is becoming more productive.

The broadest firm-level evidence cuts against the triumphal story. An NBER working paper based on nearly 6,000 executives across the United States, the United Kingdom, Germany and Australia found that about 70% of firms were actively using AI, but executives reported little own-firm impact over the prior three years, with roughly nine in ten reporting no impact on employment or productivity, according to NBER’s paper¹⁰ and NBER’s digest¹¹. The same executives expected AI to boost productivity by an average of 1.4% over the next three years, according to the NBER paper¹⁰.

That number should stop everyone cold. The market is funding data centers as if AI is near a step-change in output per worker. The firms using AI are mostly reporting that they cannot see it yet.

The financing layer makes me more skeptical, not less. Off-balance-sheet financing means a company can use structures such as joint ventures, leases or special-purpose vehicles so that risk and assets do not show up as ordinary corporate debt in the simplest headline way. That does not make the financing bad. It does make the economics harder to read.

Meta’s Hyperion project is the cleanest example. Reuters reported that Meta struck a $27 billion financing deal with Blue Owl Capital for the Louisiana data-center project, with Meta retaining about 20% equity and Blue Owl-managed funds owning the majority, according to Reuters via Investing.com¹². Reuters also reported that Meta moved to raise up to $30 billion in its biggest bond sale as AI infrastructure costs climbed, according to Reuters via Investing.com¹³.

Then there is circularity. OpenAI reportedly signed a roughly five-year, $300 billion computing-power deal with Oracle, according to Reuters via Investing.com¹⁴. Nvidia and OpenAI announced a letter of intent to deploy at least 10 gigawatts of Nvidia systems, with Nvidia intending to invest up to $100 billion in OpenAI as systems are deployed, according to OpenAI’s announcement¹⁵. That may prove rational. It also means supplier money, customer purchases, cloud commitments and market enthusiasm can reinforce one another before independent final demand is proven.

This is how real bubbles work. They do not require fake invoices. They require a plausible future, scarce assets, cheap storytelling and financing structures that push the reckoning forward.

The labor story is also being oversold. Intuit said it would cut 17% of its full-time workforce, nearly 3,000 roles, while streamlining operations and AI efforts, according to Reuters via Investing.com¹⁶. Reuters reported that Mark Zuckerberg attributed Meta’s planned layoffs to increased capital spending for AI, according to Reuters via Investing.com¹⁷. These cuts may include real automation. They may also be ordinary cost discipline dressed in AI language. Without occupation-level evidence showing which tasks were automated, which roles were outsourced and which teams were simply flattened, layoffs prove pressure, not productivity.

AI governance, meaning rules for safety testing, accountability, disclosure, competition and security, should be tightening as the stakes rise. Instead, oversight is politically brittle. On May 21, 2026, President Trump called off an AI executive order that would have created a framework for vetting national-security risks of advanced AI systems before public release, because he worried it could weaken America’s AI edge, according to the Associated Press¹⁸. Axios reported that a draft order had contemplated a voluntary framework for labs to share advanced models with the government before public release, according to Axios¹⁹. Investors are being asked to fund a buildout whose risks span labor markets, power grids, cybersecurity and market concentration, while disclosure remains patchy.

The strongest counterargument is simple: general-purpose technologies take time. Electricity, computers and the internet required years of complementary investment before productivity statistics fully captured them. I buy that. I also think it misses the present danger. The issue is not whether AI will matter. It will. The issue is whether current valuations assume profit curves that corporate users have not yet demonstrated.

My threshold is clear. I will become much more bullish when companies disclose AI-attributable revenue, renewal rates, utilization, depreciation, power costs, lease obligations and customer concentration in a way that lets investors calculate returns after the full bill. I will become much more confident on productivity when broad adopters outperform matched non-adopters in operating income and output per worker, not just in anecdotes about faster documents or cleaner code.

Until then, I would treat the AI boom as a real productivity option financed like an infrastructure mania. The next indicator to watch is not Nvidia’s next chip order. It is whether Microsoft, Amazon, Alphabet and Meta can show rising AI margins after depreciation in 2026 and 2027, while non-tech customers report measurable productivity gains above 2% to 3%. If those numbers do not arrive, the market will learn a brutal lesson: useful tools can still be overbuilt.