The math behind the AI boom just got a much bigger number attached to it. Sequoia partner David Cahn, one of the first investors to quantify the payback problem back in 2023, now pegs AI infrastructure spending for 2026 at $1.5 trillion, according to TechCrunch AI. Add up three years of hyperscaling and Cahn figures the industry needs to earn roughly $3 trillion to justify all those chips and data centers. What stands out here is that he calls that an underestimate.
Back in 2023, Cahn was reacting to Nvidia’s $50 billion in annual GPU revenue. He ran the numbers on operating costs and operator margins and landed on $200 billion in revenue needed to break even. Then he threw down a challenge: go build AI products worth paying for. The spending kept climbing. The revenue target climbed with it, and rising memory prices plus inference-specific chips are pushing it higher still.
💰 Where the revenue actually stands
The income side of the ledger is growing fast, but the gap is real. Per TechCrunch AI:
- Anthropic is thought to have hit $60 billion in ARR.
- OpenAI reportedly earned $13 billion in 2025, and said in November it had reached $20 billion ARR.
- Both are presumably earning more this year.
Stack that against a $3 trillion payback target and you see the problem. Impressive revenue, enormous distance left to cover.
📉 The 2028 bet
Here’s the timeline that matters. Torsten Slok, chief economist at asset manager Apollo, points out that Google, Meta, Microsoft, and Amazon are all forecasting big jumps in free cash flow by 2028. That’s the year the hyperscalers expect their chip spending to pay off.
Slok’s worry is what happens if it doesn’t. Two trends are already working against the frontier labs:
- Companies are shifting to cheaper open-weight models, often Chinese ones, instead of frontier-lab systems.
- Token prices keep falling. Sam Altman says OpenAI’s latest model is 54% more token efficient on coding tasks.
Efficiency is great news if you’re paying to run AI agents. It’s trickier if you’re the one who built a token factory and needs usage to explode to cover the buildout. Cheaper tokens only pay back the data center if customers burn far more of them.
⚠️ Why this matters beyond tech
Slok frames the downside in blunt terms. “With so much riding on so few names,” he writes, “a slower payoff wouldn’t just be a sector problem, it would risk tipping the economy into recession and the S&P 500 into a correction.” This is significant because the AI trade is now concentrated in a handful of megacap stocks. When that few names carry that much index weight, a missed forecast stops being an industry story and becomes a market one.
Slok and Cahn are looking at the same picture from opposite desks: one an equity-market economist, one a venture investor who’s been tracking this gap since 2023. Neither is calling a crash. Both are flagging that the payoff assumption baked into 2028 is doing a lot of heavy lifting.
🧭 What to do about it now
You can’t move the hyperscalers’ capex, but you can position around the uncertainty:
- If you build products, focus on revenue per token, not raw usage. Falling prices reward apps that turn cheap inference into something customers actually pay a premium for.
- If you buy AI, the cost curve is bending your way. Test open-weight and efficient models before locking into frontier pricing.
- If you invest or plan budgets, treat 2028 as the checkpoint everyone is watching. Watch hyperscaler free-cash-flow guidance the way you’d watch an earnings warning.
The next two years turn Cahn’s challenge from a thought experiment into a scoreboard. Either AI usage scales fast enough to fill a $3 trillion hole, or the market starts asking harder questions about who pays for all this silicon. You can find the full breakdown at the original TechCrunch AI report.