Thesis
Open-source models are unlikely to kill AI compute demand. The bigger risk is that they squeeze the revenue earned on each unit of compute. The chips still get used, but they earn less, while the debt and leases that paid for them stay fixed.
How the squeeze works
Open-weight models have repeatedly closed much of the gap to frontier models within months, at a fraction of the price. Once “good enough” is cheap, closed labs can charge a premium only where their edge clearly matters.
That pushes inference toward a low-margin commodity. Usage may even rise as cheaper AI makes more tasks worthwhile, but revenue per GPU-hour falls. Returns shift to AI users, distribution owners and scarce inputs like power, not to model builders or compute landlords.
Analogy: open models turn intelligence into tap water. The water utility does not get rich; the farms and factories using cheap water do.
Why it matters for Oracle
Oracle is exposed on two fronts at once. Its costs are fixed and its revenue depends on premium AI pricing holding up.
- Fixed costs: free cash flow was negative $23.7 billion in fiscal 2026, and Oracle raised $43 billion of debt that year (Motley Fool). Michael Burry counts $288 billion of signed data-center leases not yet on its balance sheet.
- Pressured revenue: cheaper inference lowers what each chip earns, making debt service harder.
- Customer risk: commoditization also weakens labs like OpenAI, Oracle’s key customer, whose ability to pay depends on premium economics too.
The result is a squeeze from both sides: the chips are busy, but the cash they generate may not cover what was borrowed to build them.
Counterarguments
The frontier keeps moving, so open models may always trail on the hardest, most valuable work, such as complex coding agents. Enterprises also pay for reliability, security and support, not just raw capability. And if cheap models drive a large enough jump in usage, volume could offset lower prices.
Rebuttal: the premium lane may be profitable but too small
These counterarguments assume businesses will keep paying for frontier intelligence. Usage data suggests most economically useful work is moving to cheaper models that are good enough.
- Volume is shifting fast. US models’ token share on OpenRouter fell from about 70% in June 2025 to about 30% by June 2026 (OfficeChai).
- Demand is highly price-sensitive. When OpenAI discounted two new models on OpenRouter, their token usage rose 13.8x (OpenRouter). That is how buyers treat a commodity.
- Tokens and dollars have split. Chinese open-weight models carry over 45% of OpenRouter tokens, while Anthropic holds about 12.3% of tokens but a larger dollar share through premium pricing (Digital Applied).
The labs have lost volume, not yet the money. But a premium lane can be profitable and still far too small: an airline cannot fill a thousand jumbo jets with first-class passengers alone. Capacity was committed for the whole market at premium economics.
The foundational constraint: frontier labs escape commoditization only if their capability lead is worth more to customers than the price gap to open models. Open models trail by months, so that lead must keep regenerating at rising training cost. Otherwise inference competes on economic output per dollar of tokens.
Caveat: OpenRouter skews toward price-sensitive developers and misses much direct enterprise spending through cloud contracts.
Note: this memo was drafted by Claude, made by Anthropic, a closed lab with a direct stake in the premium holding. Weigh the counterarguments accordingly.
Supporting charts
Volume is moving to cheaper models
Chinese-origin models went from about 1% of OpenRouter tokens to over 45% in 18 months, while US labs’ share fell by more than half (OfficeChai, Digital Applied).
Capacity is being built for premium economics
The four largest hyperscalers guide to $720–745 billion of 2026 capex, about five times 2023 levels for the three calendar-year reporters shown (MLQ.ai). This spending assumes compute keeps earning today’s prices.
The fixed costs are growing faster than the companies
Signed but not-yet-started data-center leases now total nearly $1.2 trillion across the five hyperscalers, per Michael Burry’s reading of their filings. Oracle’s $288 billion is almost as large as Microsoft’s despite Oracle being far smaller. Moody’s, using a narrower method, counted $662 billion in February 2026 (MLQ.ai).
These payments are owed whether or not compute prices hold up, which is the core of the squeeze.
What to watch
- GPU rental prices per hour: falling prices signal commoditization.
- Lab revenue growth versus token volume growth: a widening gap means each token earns less.
- Share of enterprise workloads moving to open-weight models.
- Oracle’s quarterly free cash flow and new debt or equity raises.
- OpenAI’s funding rounds and any renegotiation of compute contracts.