OpenAI has developed its first custom inference processor, Jalapeño, with Broadcom. The ASIC is designed specifically to serve large language models, targeting the cost and data-movement bottlenecks that arise when hundreds of millions of people expect immediate responses.
The chip signals a strategic change. OpenAI is moving beyond model and product development toward control of silicon, networking and serving software—the layers that determine whether enormous usage can become a sustainable business.
Inference cost makes specialised hardware attractive
The source estimates that keeping ChatGPT responsive cost $8.4 billion last year and could reach $14 billion this year. OpenAI has also made very large long-term compute commitments relative to current revenue.
General-purpose accelerators offer flexibility but include supplier margins and features not always used efficiently by interactive language workloads. Jalapeño balances compute, memory and networking around OpenAI’s own model roadmap.
OpenAI designed the architecture, Broadcom handled silicon engineering and integrated Tomahawk networking, TSMC manufactures the chip in Taiwan, and Celestica builds boards and racks. Early samples reportedly run frontier workloads at target frequency and power.
Reducing data movement is central because processors often wait for weights and intermediate results rather than perform arithmetic. Higher realised utilisation can lower cost per answer even without claiming the largest theoretical benchmark.
Vertical integration creates a powerful feedback loop
Owning more of the stack lets OpenAI tune model kernels, memory, network scheduling and application behaviour together. Better efficiency can reduce serving cost, support faster products and fund another infrastructure generation.
The company also says its models helped automate portions of chip design, contributing to a nine-month path from blank slate to tape-out. That creates a literal loop in which current AI helps build hardware for future AI.
Specialisation carries risk. A chip optimised for today’s architecture may become less useful if model design changes dramatically. OpenAI needs enough programmability and software support to absorb new attention, memory and quantisation techniques.
Software migration is another hidden cost. Compilers, kernels, debugging tools and observability must let engineers use the chip without maintaining a separate fragile code path. OpenAI can justify that investment through its own volume, but infrastructure partners need documentation and predictable interfaces before the processor can scale across data centres.
Operational success depends on yield, reliability, deployment tooling and sustained utilisation—not a laboratory sample. The first data-centre installations are planned before the end of 2026.
Jalapeño diversifies rather than eliminates dependency
Google and Amazon have spent years developing custom accelerators, while Meta and Microsoft are scaling their own efforts. OpenAI enters late but has unusually detailed visibility into one of the world’s largest inference workloads.
The chip can reduce reliance on Nvidia, yet the supply chain still includes Broadcom, TSMC, memory vendors, rack manufacturers, Microsoft and power infrastructure. Vertical integration shifts dependencies; it does not make the company self-sufficient.
OpenAI should compare full economics across workloads, including development cost, underutilised capacity and the value of retaining general-purpose hardware for new models.
Gigawatt-scale deployment also makes energy and cooling part of the product calculation. A cheaper token that requires vast new electricity and water capacity carries external costs beyond the chip budget.
Jalapeño matters because inference has become the recurring cost beneath every AI interaction. Model leadership can attract users, but infrastructure efficiency determines how many of those interactions a provider can afford to serve.
If the processor performs reliably at scale, OpenAI gains leverage over both cost and roadmap. The company will have become not only a customer of the AI hardware market, but a full-stack infrastructure operator whose software ambitions increasingly shape the physical data centre.
That leverage will be visible only after sustained production utilisation, not during a launch demonstration.
