OpenAI unveiled Jalapeño, its first custom inference chip, co-designed with Broadcom and built to run the company's own models in production. The disclosure puts OpenAI on a list that already includes Google, Apple, and SpaceX — frontier-scale buyers who have decided that one supplier for AI silicon is one too many. Nvidia remains the default training chip for almost everyone in the field, but the inference layer is where the hedge is being built first.
Jalapeño is an inference part, not a training part. That distinction matters: inference is where token costs accumulate at scale, and where a chip tuned to a specific model architecture can extract meaningful performance and power gains over a general-purpose GPU. Broadcom is the silicon partner, the same role it has played for Google on the TPU line for more than a decade.
Google's TPU program is the longest-running case study and has given the company a credible argument that it runs Gemini inference cheaper than rivals running on Nvidia. Apple's transition from Intel to Apple Silicon is the other reference point, and the one cited explicitly in the TechCrunch Equity discussion of OpenAI's move. The Mac transition delivered durable performance-per-watt gains that Intel could not match on a roadmap basis.
Key facts
- 01OpenAI's Jalapeño is a custom inference chip co-developed with Broadcom, the company's first in-house silicon.
- 02OpenAI joins Google, Apple, and SpaceX on a growing list of buyers building chips to hedge Nvidia exposure.
- 03Apple's Intel-to-Apple-Silicon transition is the playbook cited for the performance gains custom silicon can unlock.
- 04The trend is framed as a hedge, not a clean break — Nvidia remains the default training supplier for now.
SpaceX is the more unusual entrant. Custom silicon inside Starlink and inside SpaceX's vehicle stack is less about competing with Nvidia in a data center and more about owning the inference layer in environments where commodity GPUs are not an option. The common thread across all four companies is volume: you build your own chip when you are buying enough of someone else's to justify the engineering bill.
Nvidia's position is not in immediate danger. The company still sells essentially every high-end training GPU that gets shipped, and the CUDA software moat means that even buyers who design their own inference chips keep buying Nvidia for the training side. What is changing is the ceiling. If OpenAI runs a meaningful share of ChatGPT inference on Jalapeño in 2027, that is revenue Nvidia was previously booking.
Broadcom is the quiet winner. The company has spent years positioning itself as the merchant-silicon partner for hyperscalers who want their own chips without building a chip company from scratch. Google was the anchor customer; OpenAI is the highest-profile addition. Broadcom's AI-related revenue has been one of the few credible counterweights to Nvidia's quarterly numbers, and Jalapeño extends that trajectory.
The move also reflects a maturing of OpenAI's infrastructure thinking. Earlier in the company's life, the priority was access to as many GPUs as Microsoft and Oracle could secure. With run-rate revenue now in the tens of billions and a model release cadence that includes the recently shipped GPT-5.6, the calculus shifts toward owning more of the cost structure. We covered OpenAI's GPT-5.6 launch and its aggressive three-tier pricing last week; Jalapeño is the supply-side answer to that demand-side strategy.
Custom chips are not a guaranteed win. Designing silicon is expensive, the first-generation parts almost always underperform the roadmap, and the software work required to make a new accelerator competitive with CUDA is measured in years, not quarters. Google's TPU only became broadly useful outside Google after multiple generations. OpenAI is starting that clock now, not finishing it.
The competitive read is straightforward. Every frontier lab now has a credible path off Nvidia for inference within two to three years: Google has TPUs, Anthropic has access to both TPUs and Amazon's Trainium, Meta is building MTIA, and OpenAI now has Jalapeño. Nvidia's pricing power on inference chips peaks in 2026 and faces real pressure after that. Training is a different story, and the one Jensen Huang will keep pointing to.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
