AI's profitability crisis

The loud story is capability: another model, another benchmark, another campus. The quiet story is a spreadsheet. Amazon, Alphabet, Microsoft, and Meta have guided combined 2026 CapEx toward roughly seven hundred billion dollars and up — mostly AI infrastructure — while the pure AI software businesses those dollars feed still report revenue in the tens of billions, not hundreds. That gap is what people mean by a profitability crisis. It is not that nobody uses AI. It is that usage and cash do not yet line up the way the buildout assumes.

CapEx ran ahead of the P&L

Cloud and ads giants can fund GPUs from balance sheets that consumer startups never had. They are doing it all at once. Mid-2026 earnings season pushed Big Four CapEx guidance into the mid–high seven hundreds of billions for the year, after earlier talk nearer six hundred. Training and inference both chew silicon; transformers and interconnects stretch years; power and permitting now set the pace as much as Nvidia allocation. Our reports on megawatts and interconnection and the data-center freeze cover the physical and local walls on the ground. The financial wall is simpler: depreciation and cash CapEx hit this decade; the product revenue that is supposed to pay for it is still ramping.

Circular money and double-counted demand

A lot of "AI revenue" is the same dollar counted twice. A lab sells API access; it then buys cloud and GPUs from the same hyperscalers that invested in it. Investor capital — not operating profit — still funds a large share of frontier training. When OpenAI or Anthropic posts a hot run rate, that is real customer spend mixed with a financing loop: equity and cloud credits in, compute and marketing out. Strip the loop and organic demand from households and cash-generating enterprises is smaller than the infrastructure story implies. That does not make the products fake. It means backlog and valuation stories can outrun cash collection.

Who prints money — and who waits

Unit economics are uneven by layer. Chip vendors and many cloud AI SKUs convert scarcity into margin today. Frontier labs grow revenue fast and still spend heavily on inference, talent, and the next training run; consumer chat is cheap to offer and expensive to serve at scale, while enterprise contracts pay more per token but take longer to land. Wrapper apps on top of someone else's API often look like growth businesses with rented gross margins. Open-weight models (see our open-weight report) pressure API pricing from below. The crisis headline is blunt; the accurate map is a stack: silicon and cloud often profitable, model labs racing toward breakeven on different timelines, thin wrappers exposed first when CFOs audit spend.

Enterprise ROI is the slow fuse

Pilots are easy; production savings are not. Surveys and consulting decks keep finding the same pattern: wide GenAI experimentation, fewer measured P&L wins, lots of seats that never become workflow. That "pilot purgatory" matters more than any single lab valuation. If enterprises treat copilots as optional chrome, the revenue ramp that is supposed to amortize campuses slips. If agents start owning closed-loop tasks — support tickets, code merges, claims — the math can close. Right now the market is priced for the second path while many buyers still live on the first.

What a crisis is not

This is not a brief that AI stops working. Demand for translation, coding assistance, search, and customer support automation is real. Nvidia's data-center run rate and hyperscaler AI SKUs prove someone is paying. A profitability crisis means the timing and capture of profits are contested: CapEx now, uncertain returns later; circular financing that can unwind; open weights that compress API rents; regulation and power that raise the cost of every watt. Bubbles and buildouts can both be true — overbuild relative to near-term cash, underbuild relative to a decade of demand.

The point

Don't treat "AI" as one business. For any product, ask which layer of the stack it sits on, who owns the GPUs, and whether its customers are consumers, enterprises, or other AI labs. The Language Models, Optimization, and Cloud GPU categories map that stack. AI is making money, but unevenly, and the current spending only pays off if revenue catches up before the hardware wears out.