Broadcom CEO Hock Tan walked into OpenAI's offices on June 24, 2026, carrying a 300mm silicon wafer. He handed it to Sam Altman and Greg Brockman in person. On the wafer: roughly 50 to 60 ASICs. OpenAI's first custom chip.

The chip is called Jalapeño. It's a reticle-sized ASIC — roughly 840mm², about as large as EUV lithography can print as a single die — built on TSMC's 3nm process with eight HBM stacks around the central die. OpenAI and Broadcom went from initial design to manufacturing tape-out in nine months. For custom silicon, that is fast.

The claim that matters: roughly 50% lower inference cost per token than current NVIDIA GPUs, per Hock Tan. Jalapeño is inference-only — training workloads stay on NVIDIA. The deployment timeline runs late 2026 for prototype testing, production in 2027, scaling through a 10-gigawatt data center commitment through 2029 with Microsoft and other partners.

Greg Brockman said: "By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access."

Strategy statement. Whether it's marketing or real depends on what the chip actually does in 2027.

The Apple analogy, and where it breaks

Multiple outlets called this an "Apple-like move." The comparison is right in one way and wrong in another.

Apple built M1 to own its end-to-end product experience. The advantage isn't just performance-per-watt. It's that the same team that designs the chip also designs the OS, the app stack, and the product shell. Every layer is optimized against the others. The integration is the moat.

OpenAI is trying to own the chip and the model and the product. But the chip is being designed by Broadcom, not by OpenAI. The model training still needs NVIDIA. And the product runs on infrastructure built partly by Microsoft. "Building the full stack" is the aspiration. The actual stack today has a lot of seams.

The other difference: Apple was escaping Intel, which wasn't competing in AI accelerators. OpenAI is working with Broadcom to reduce dependence on NVIDIA, which is actively competing and isn't standing still. H100, H200, Blackwell, Rubin. NVIDIA's roadmap doesn't pause while OpenAI prototypes Jalapeño.

Source spread

• OpenAI — Official Jalapeño announcement — hype. OpenAI's framing: "best inference platform for LLMs," "build the full stack," first in a multi-generation platform. First-party numbers.
• Tom's Hardware — Technical breakdown — builder. Best technical detail: die size, HBM count, node, reticle context. Reliable for the physical spec claims.
• CNBC — "Build the full stack" — builder. Good on Brockman and Tan quotes, lighter on specs.
• TechRadar — Apple-like move framing — skeptic. Raises the expertise gap and the NVIDIA-doesn't-stop problem.

Pros & cons

What's real:

• The physical chip exists. Tan handed Altman a wafer. There is a chip.
• The architecture is coherent. TSMC 3nm, eight HBM stacks, reticle-sized die — this is a reasonable design for the memory-bandwidth-constrained problem of running large language model inference. The specs are internally consistent.
• The 50% cost claim is attributed and specific. Hock Tan didn't say "significantly cheaper." He said 50%. Specific enough to hold someone accountable in 2027.
• OpenAI has the volume to make custom silicon economic. Google's TPUs and Amazon's Trainium work because they run at enormous scale. OpenAI is approaching comparable scale on inference.

What deserves scrutiny:

• Production is 12–18 months away. "Late 2026 prototype" means the real story is a 2027 story. Custom silicon timelines slip.
• The 50% number comes from Broadcom, not an independent benchmark. Lab testing differs from production at scale.
• Training stays on NVIDIA. This is an inference-stack play. For frontier model development, the NVIDIA dependency is unchanged.
• Jalapeño is not for sale. You won't buy one. You'll pay for inference on servers that might use it, eventually, if production works. The relevance to any individual builder is indirect.

Further reading

• OpenAI — Jalapeño Intelligence Processor announcement — primary source
• Tom's Hardware — Jalapeño technical breakdown — best technical detail available
• CNBC — OpenAI first chip, "build the full stack" — executive quotes and strategic framing
• Broadcom investor press release — joint announcement, official figures