OpenAI chip design leans on its own models as Luna undercuts open source
OpenAI chip design work now depends heavily on the company's own AI models, and its chief financial officer says proprietary models can undercut open-source systems on price when both are delivered through a cloud provider. Sarah Friar said this week at Goldman Sachs' Communacopia + Technology Conference in San Francisco that the internal Jalapeño chip had reached tape-out nine months after design work began, with OpenAI's own models doing much of the architecture work. Tape-out is the stage at which a chip design is finalized and sent to a fabrication plant.
Friar said chip design, life sciences, and financial services are the areas where enterprise clients most want AI adjusted to their own workflows. She put OpenAI's enterprise revenue growth at 32 percent from June to July, compared with 20 percent annualized growth for the whole business. Revenue is now almost evenly split between enterprise and consumer customers, ahead of the balance the company had targeted for the end of the year.
What a nine-month tape-out gives OpenAI
The nine-month result has implications beyond internal speed. It suggests OpenAI's frontier models can take part in designing chips that will later run those models, pulling model research, chip development, and inference costs into a tighter loop. It also changes OpenAI's position with outside partners: design cycles that often take years can now fit into months, giving the company more leverage with chip vendors and cloud providers. Enterprise buyers get a demonstration that open-weight competitors cannot easily match.
Friar's pricing claim is aimed at the common assumption that open-weight software is the cheap option while models from OpenAI and Anthropic carry a premium. When a model runs through a cloud provider, she said, OpenAI's lower-cost Luna model can run more cheaply than Chinese open-source models used the same way. She cited an 80 percent cut to Luna's price and said usage rose roughly tenfold after the cut.
Friar also said OpenAI is testing outcome-based pricing instead of usage-only billing, tying what customers pay to measurable results rather than raw consumption. Codex, OpenAI's coding assistant, now has 25 million users. The two moves test whether enterprises still need to self-host open-weight systems to control costs, or whether a proprietary model operated by its vendor can be just as cheap while taking on the operational work.
Why this matters
Chinese open-weight developers and Anthropic continue to improve their models, so the open question for buyers is whether Luna's cloud price advantage lasts. OpenAI is repositioning itself from a premium model provider into an integrated, lower-cost supplier, with the nine-month chip tape-out offered as supporting evidence. The revenue and delivery figures OpenAI reports in the coming months will show how durable that positioning is.
Related Articles
- OpenAI's Jalapeño Inference Chip Cuts Costs 50%
- OpenAI Jalapeño chip benchmarks reveal a pricing-power shift in AI inference
- Sarah Friar Unveils New OpenAI Enterprise Strategy
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.