OpenAI Partners with Broadcom to Launch Jalapeño Chip Promising 50% Lower Cost Per Watt than Current GPUs

OpenAI and Broadcom unveiled the Jalapeño on Thursday, marking the first inference chip designed by Sam Altman's startup. The accelerator is described as a purpose-built ASIC, not a generic GPU or an adapted training chip, and was co-developed with Broadcom and electronics manufacturer Celestica in a nine-month cycle from initial design to tape-out, which OpenAI characterizes as the fastest documented advanced silicon program.

Engineering samples are already running production workloads in the lab, including GPT-5.3-Codex-Spark, at the same frequency and power envelope projected for mass production. Early internal numbers indicate cost savings of approximately 50% per watt compared to commercially available state-of-the-art GPUs. Initial deployment is planned for late 2026, with expansion in subsequent years across data centers operated by OpenAI and its partners.

Why Move Away from Nvidia Now

The context revolves around the size of the bill. Estimates from the four largest hyperscalers suggest $700 billion in AI capex by 2026, with Microsoft, Google, Amazon, and Meta each accounting for between $125 billion and $200 billion. OpenAI, which is not a hyperscaler on the balance sheet, currently divides its spending between Microsoft, Oracle, and CoreWeave, with each generated token running on Nvidia's H100 or B200 GPUs. Migrating even 20% of the inference traffic from ChatGPT and Codex to proprietary silicon at 50% lower cost per watt could significantly shift the company's breakeven point.

The contrast with the strategies of Anthropic or xAI is stark. Anthropic tends to prefer purchasing capacity in Google's TPU and AWS's Trainium without its own chip. xAI operates exclusively on Nvidia GPUs. Besides OpenAI, the only other players betting on dedicated ASICs are the hyperscaler companies themselves: AWS's Trainium, Google's TPU, and Microsoft's Maia. For Broadcom, this contract confirms that joint design with customers has become a recurring business after the success of Google's TPU. The company's stock has risen nearly 60% this year, and the order book for custom silicon has surpassed $10 billion.

Who Loses, Who Wins in Each Region

The most obvious assessment comes from the United States, where Nvidia remains by far the preferred supplier for training but is now sharing the inference market with ASICs produced by hyperscalers and their clients. For American CIOs evaluating multi-year contracts with OpenAI, the arrival of the Jalapeño introduces a future price reduction vector, as part of the cost savings is expected to flow into per million tokens fees.

In Taiwan, home to TSMC, market expectations indicate another customer for the advanced ASIC line that the foundry already produces for Broadcom, although OpenAI has not confirmed the manufacturing partner. In Singapore and India, where OpenAI has been expanding inference regions alongside Microsoft Azure, the projected drop in cost per query facilitates local tariff launches in strong currencies. In delivery centers in India, where TCS, Infosys, and Wipro scale development squads around Codex, any price cut per million tokens is quickly perceived in margins within weeks.

In Europe, where sovereignty is the theme, the Jalapeño presents an ambiguous outlook. It reduces dependency on Nvidia but further ties OpenAI's inference to a supply chain that combines design in San Francisco, silicon implemented by Broadcom, and assembly by a Canadian supplier, with the foundry step in a jurisdiction that OpenAI did not detail in the announcement. For a European bank operating Anthropic's Claude via Bedrock and OpenAI's GPT via Azure, the question arises about where the physical data center processing the prompt is located and who owns the silicon.

What We Still Don’t Know

OpenAI has not disclosed the process node from TSMC being used, has not confirmed the initial volume of wafers, nor indicated who else will have access to the chip beyond its own operations and Microsoft. There is also no public comparison against Nvidia's B200 or the upcoming B300 on standardized metrics, only the internal metric of cost per watt. Nvidia had not commented on the announcement by the time this article was published.

The next milestone to watch will be Broadcom's fiscal report in September. If custom silicon revenue doubles again on a quarterly basis, the market will begin to price Nvidia as less dominant in the inference segment of the AI stack. If initial Jalapeño shipments slip from Q4 into 2027, the narrative changes, and Nvidia regains momentum.