OpenAI Unveiled Its First Custom Inference Chip in Nine Months

OpenAI on June 24 unveiled Jalapeño — its first custom inference chip, co-developed with Broadcom in nine months from design to tape-out — and declared early tests show performance per watt "substantially better than current state-of-the-art." The announcement lands the same week Wall Street sold off AI stocks over capex doubts and Ramp data showed Anthropic overtaking OpenAI in paid enterprise subscriptions. OpenAI is no longer just a Nvidia customer. It is a chip designer with a gigawatt-scale roadmap — and an IPO story that now depends on owning the stack, not renting it.

What dropped:

• OpenAI and Broadcom named the chip Jalapeño, calling it the first "Intelligence Processor" in a multi-generation compute platform.
• Engineering samples are already running ML workloads in OpenAI's lab at production target frequency and power — including GPT-5.3-Codex-Spark.
• Broadcom handles silicon implementation and Tomahawk networking; Celestica builds the rack systems.
• Initial deployment targets the end of 2026, scaling toward 10 gigawatts of custom accelerators by 2029 under the October 2025 Broadcom partnership.

Why inference, why now:

Training gets the headlines. Inference pays the bills.

Every ChatGPT answer, every Codex task, every API call burns rented GPU hours. OpenAI designed Jalapeño specifically for LLM inference — memory movement, kernel patterns, serving architecture — not as a general-purpose GPU adapted from older workloads.

Richard Ho, who leads OpenAI's hardware program, said the team optimized "around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models."

OpenAI President Greg Brockman framed it as economic infrastructure: "The world is moving to a compute-powered economy."

The nine-month sprint:

OpenAI claims the Jalapeño program is the fastest ASIC development cycle ever in high-performance semiconductors — nine months from initial design to manufacturing tape-out.

Part of the speed came from the company's own models accelerating chip design and optimization. The same AI serving users is helping build the hardware that will serve future models.

Broadcom CEO Hock Tan delivered the first engineering samples personally to Altman and Brockman. On the earnings call three weeks earlier, Tan said OpenAI silicon was delivered and "on track for production late 26" as part of a 1.3-gigawatt 2027 commitment inside the broader 10-gigawatt deal.

The market context:

The Jalapeño reveal arrives into a brutal week for AI equities.

On June 23, the Nasdaq fell 2.2% and South Korea's Kospi plunged 10% as investors questioned whether hyperscaler AI spending can generate returns before rising rates make borrowing unbearable.

Broadcom, by contrast, rallied on the news — its custom-silicon business already generated $10.8 billion in AI semiconductor revenue last quarter, up 143% year-over-year.

The irony is sharp: the company helping OpenAI escape Nvidia dependence is the same one fabricating custom chips for Google, Anthropic, and Meta.

What OpenAI said — and what it didn't:

OpenAI says Jalapeño will deliver better performance per watt than current state-of-the-art. A detailed technical report is promised in coming months.

Those are lab claims, not independent benchmarks. Investors should treat them as directional until third parties verify.

OpenAI is still a multi-vendor shop — Nvidia for training, AMD for capacity, Cerebras for inference overflow. Jalapeño does not end those relationships. It gives OpenAI leverage in every renegotiation.

Axios reported Hock Tan's blunt assessment: "At the end of the day, you cannot, should not rely on some other third-party GPU to do it for you, because it's such a key part."

OpenAI is also considering whether the architecture eventually handles training — not just inference.

The IPO stakes:

OpenAI's listing narrative was already under pressure.

Ramp's June AI Index shows Anthropic at 41% of paid US business AI subscriptions versus OpenAI's 39.5%. Applications chief Fidji Simo reportedly told staff the company is on "code red."

Jalapeño is the counterpunch — proof OpenAI can verticalize the cost structure Wall Street is punishing. If custom silicon cuts inference costs materially, the $770 billion hyperscaler capex bill becomes survivable instead of suicidal.

If it doesn't, Jalapeño is just another line item in the largest infrastructure bet in corporate history.

What to watch:

• OpenAI's promised technical report on Jalapeño performance versus Nvidia and AMD inference hardware
• Whether Microsoft and partner data centers deploy Jalapeño racks at commercial scale by year-end
• Broadcom's 1.3-gigawatt 2027 OpenAI shipment ramp — and whether financing structures hold
• Nvidia's response as inference — not training — becomes the margin battleground

Convina's view: The model was never the moat. The stack is. Google, Amazon, and Meta learned this years ago; OpenAI is late but not too late. Jalapeño does not dethrone Nvidia — it weaponizes the oldest rule in semiconductors: if your product depends on someone else's silicon, you are always one price hike from crisis. Nine months to tape-out is a statement of intent, not a victory lap. The chip has to ship, scale, and cut costs at gigawatt volume before the IPO underwriters can call it proof. But the direction is unmistakable: frontier AI is becoming a full-stack industry, and renters are becoming competitors.