The Capex Conviction Test

Three hundred billion dollars is in motion. The real question is no longer whether AI is real. It is which companies will be proven right — and which will quietly write off their bets.

The numbers

Between mid-2024 and mid-2025, four American companies — Microsoft, Alphabet (Google's parent), Meta (Facebook's parent), and Amazon — said they would spend about $325 billion in 2025. Almost all of it on the chips, buildings, and electricity that AI needs. That is roughly comparable to the GDP of Finland. In January 2025, the U.S. President announced Stargate, a $500 billion AI infrastructure plan with OpenAI, Oracle, and SoftBank.

This essay is about how to think about all this with discipline, not enthusiasm. The argument: there is a short list of questions every CEO should be able to answer about any AI investment. The gap between the CEOs who can answer them and the CEOs running on enthusiasm is now wide enough to show up in stock prices.

Three CEOs, three strategies, three failure modes

Satya Nadella at Microsoft: build everything yourself. Microsoft is building the chips, the data centers, the cloud, the AI models, and the software. It even signed a twenty-year deal to restart a reactor at Three Mile Island, the Pennsylvania site associated with the famous 1979 nuclear accident, because AI data centers need huge amounts of electricity. Failure mode: the servers wear out before the revenue arrives.

Sundar Pichai at Google: improve what you already own. Google has three billion daily users across Search, YouTube, and Android. Gemini does not have to win as a chatbot. It has to quietly make those existing products better. Failure mode: the AI eats its host — if a Gemini summary is so good the user never clicks, Google destroys the ad space it is trying to protect.

Marc Benioff at Salesforce: skip the infrastructure, sell the outcome. Benioff (and ServiceNow's Bill McDermott, Workday's Carl Eschenbach, and others) is not building data centers. He rents AI from the big clouds, wraps it inside workflows customers already use, and charges for results. Failure mode: the AI underneath becomes a commodity faster than the customer relationship becomes a moat.

Three strategies that each make sense. Three different ways to fail. The CEO's job is to know which one your company is actually running.

The three questions every AI investment must answer

1. Are costs falling faster than usage is growing? The price of running a GPT-3.5-quality query fell 280 times over between November 2022 and October 2024 (Stanford AI Index). For tech companies: revenue per query is collapsing. For you: the AI workload you cannot afford today will be cheap in 18 months — so beware long, flat-rate contracts at today's prices.

2. Are you renting compute, or building something you own? In early 2024, the Swedish payments company Klarna said its OpenAI-powered chatbot was doing the work of 700 human customer-service agents. The press called it a breakthrough. But every Klarna competitor could sign the same contract with OpenAI and get the same chatbot. Within a quarter, the capability was table stakes — not an advantage. Klarna's cost savings were real. Klarna's competitive moat was not. Contrast Bloomberg, which built its own AI model on proprietary financial data its rivals don't have. Same technology. Very different defensibility. Ask: could a competitor replicate this by signing the same vendor contract? If yes, you are buying productivity, not a moat.

3. Will the money come back before the assets wear out? AI servers depreciate over five to six years. Corporate AI revenue arrives slowly — gated by procurement and security reviews. A $40 billion training cluster generating its first dollar 18 months after construction has roughly 54 months to earn the money back. That is a sprint, not a stroll.

Every AI investment carries a hidden clock. The CEO who can describe that clock is making a decision. The CEO who cannot is placing a bet.

Three forecasts, with reasoning

1. Big tech AI spending will flatten in the second half of 2026 — unless AI revenue grows faster than 70% year-over-year. Reason: depreciation math. The announced spending cannot be reconciled with the profit margins these companies have promised, unless revenue accelerates.

2. Frontier-AI token prices will fall another 60–80% by end-2026. Reason: three forces at once — more efficient AI designs, new chips from Nvidia and AMD plus custom silicon, and free open models from DeepSeek, Mistral, and Meta's Llama proving frontier capability can be built outside the closed labs.

3. Two or three Fortune 500 companies will take material AI-related write-downs in their FY2026 or FY2027 results. Reason: corporate write-down cycles typically run 24 to 36 months after peak excitement. AI likely passed its first major enthusiasm peak in 2024. And CFO language has shifted from "investing aggressively" to "measuring carefully" — usually the precursor.

For your operating leaders

Ask each business unit head to bring one page on every AI project above a threshold you set ($500K or $5M, your call). The page answers four questions:

• What does this produce that we own? Transcripts and summaries — not much. Proprietary models, embedded workflows — a moat.

• What is our unit cost trend? Plot six months. Flat trend = ask the vendor why.

• What is the hidden clock? When does the value need to arrive? What if it arrives a year late?

• What would make us stop? Every AI bet should have a pre-committed kill switch.

This is not a way of slowing down AI investment. It is a way of speeding up the good projects, by killing the bad ones faster.

The CEOs who will be proven right are not the ones with the boldest forecasts. They are the ones whose forecasts could have been proven wrong — and weren't. It Is What It Is