OpenAI and Broadcom Unveil Jalapeño: OpenAI's First Custom AI Chip
OpenAI and Broadcom have unveiled Jalapeño, a purpose-built LLM inference ASIC designed in nine months, targeting production deployment by end of 2026.
OpenAI has spent the last few years as one of Nvidia’s biggest GPU customers. That relationship has served them well, but it comes at a significant cost, both financial and strategic. On 24 June 2026, OpenAI and Broadcom made it official: they are building their own silicon, starting with a chip called Jalapeño.
Jalapeño is OpenAI’s first custom AI accelerator, designed specifically for LLM inference. It is not a repurposed training chip or an adapted general-purpose processor. It was built from scratch around how large language models actually behave at inference time, addressing the practical bottlenecks that matter most: data movement, memory bandwidth, compute balance, and networking efficiency.
From Idea to Silicon in Nine Months
The development timeline alone is worth pausing on. A typical ASIC design cycle runs 18 months to two years. OpenAI and Broadcom completed Jalapeño from initial design to manufacturing tape-out in nine months, which they describe as the fastest development cycle ever achieved for a high-performance advanced semiconductor.
That speed came from a tight software-hardware co-design process, Broadcom’s silicon implementation expertise, and something genuinely novel: OpenAI used its own models to accelerate parts of the chip design process. The models that run on OpenAI’s infrastructure today helped shape the infrastructure that will run future models. It is a feedback loop that is hard to replicate if you are working with an off-the-shelf chip from a third-party vendor.
Engineering samples are already running ML workloads in the lab, including GPT-5.3-Codex-Spark, at production target frequency and power. Early results show performance per watt substantially better than current state-of-the-art accelerators.
What Is Actually Inside Jalapeño
On the hardware side, Jalapeño is a large compute die, notably bigger than compute dies found in most inference accelerators and more comparable in scale to training processors. It is surrounded by eight HBM stacks to minimise memory latency. A single 300mm wafer yields roughly 50 to 60 chips.
Broadcom’s Tomahawk 6 networking silicon (1.6 Tbps throughput) is integrated into the inference stack alongside the compute die, and Celestica is handling board, rack, and system-level integration. The goal is to combine the throughput of leading training accelerators with the low latency you want for interactive inference workloads.
A detailed technical performance report is expected in the coming months.
What This Means for You
If you use ChatGPT, the Codex API, or anything else OpenAI powers, Jalapeño is directly relevant to your experience, even if you never think about chip architecture.
Inference is where AI meets real users. Every answer ChatGPT gives you, every API call your application makes, every Codex task that completes is an inference operation. Bloomberg reports that Jalapeño cuts inference costs by around 50%. If that holds at scale, the downstream effects are real: faster responses, lower API prices, more reliable access during peak demand, and the ability to run more capable models without a proportionate cost increase.
For developers building on OpenAI’s APIs, cheaper and faster inference matters directly to unit economics. Products that are currently marginally viable at today’s API pricing become considerably more interesting if costs halve.
OpenAI’s Bigger Play
OpenAI is following a path that Apple, AWS, and Google have each walked before it. Controlling your own silicon means you can optimise every layer of the stack around a single goal, rather than accepting the trade-offs baked into hardware designed for a broader market.
Google has had Tensor Processing Units since 2016. Amazon has Trainium. Apple’s M-series chips are the clearest example of what full-stack control can do for performance, efficiency, and margin. OpenAI is now building toward the same position.
Jalapeño is described as the first chip in a multi-generation compute platform. Broadcom CEO Hock Tan told CNBC that deployments will begin at gigawatt-scale data centres with Microsoft and other partners, with meaningful ramp in 2027 and full-scale operation in the first half of 2028. The end of 2026 marks initial production deployment, not the full rollout.
For OpenAI specifically, this matters ahead of an anticipated public offering in 2026. Custom silicon is a credible path toward improving margins in a business that currently spends enormously on compute. Jalapeño tells investors there is a plan beyond being a permanent customer of Nvidia.
Where Nvidia and AMD Fit
Jalapeño does not make Nvidia irrelevant to OpenAI overnight. The chip is designed for inference workloads, and training large frontier models still requires enormous GPU clusters. OpenAI will continue to be a significant Nvidia customer for the foreseeable future.
What Jalapeño does is reduce dependence at the inference layer, which is where the volume and the cost accumulate as user demand grows. It is also worth noting that the competitive picture will keep moving. Early testing shows Jalapeño ahead of current AMD Instinct MI350 and Nvidia Blackwell-based accelerators on performance per watt, but AMD’s MI400 series and Nvidia’s Rubin-based chips are coming. How Jalapeño stacks up against those will be the more meaningful comparison.
The Short Version
OpenAI built a custom inference chip in nine months with Broadcom, using its own AI models to help design it. Early results look strong. Production deployment starts late 2026, with real scale in 2027 and 2028. The expected payoff is faster, cheaper AI inference, which flows through to every product OpenAI ships and every developer who builds on its APIs.