Docs: OpenAI and Anthropic have projected profitability to investors with and without training costs, and report inference costs exceeding half of their revenue
Silicon Valley's hottest startups have the same challenge: funding giant computing costs — OpenAI and Anthropic …
Context & Ripple Effects
Earlier investor projections already put OpenAI’s path to profit later than Anthropic’s, with their projected break-even timelines diverging even before this disclosure clarifies how much the answer depends on treating training costs separately.
The new documents shift attention from headline revenue growth to serving economics: inference alone consumes more than half of revenue at both companies, making recurring usage—not just frontier-model development—the central financial constraint.
First-order effects
- OpenAI and Anthropic must show investors two distinct profitability views: one that includes the cost of training new models and one that does not, exposing how sensitive reported economics are to that accounting boundary.
- With inference taking more than half of revenue, each company faces immediate pressure to improve serving efficiency, raise monetization per use, or constrain costly usage before gross margins can expand.
Second-order effects
- Enterprise pricing, product limits, and model-routing choices become more consequential, since heavier adoption can deepen revenue while also adding substantial variable compute expense.
- The cost burden strengthens the appeal of cheaper alternatives and routing tools; related coverage shows customers turning to lower-cost models as AI bills rise, adding pressure on OpenAI and Anthropic to defend pricing and performance.
Third-order effects
- If inference remains this expensive at scale, the competitive advantage in generative AI will increasingly rest on durable serving infrastructure and unit economics, rather than model capability alone.
- Separating profitability from training costs may become a key investor test for AI companies: whether recurring model usage can support itself before continual frontier-model investment is counted.
The trend: Generative-AI competition is moving from a race to build frontier models toward a race to make high-volume inference economically sustainable.