OpenAI's Jalapeño: First Custom Inference Chip with Broadcom

OpenAI Builds Its First Chip

On June 24, 2026, OpenAI unveiled Jalapeño — its first custom-built AI inference processor, co-developed with Broadcom. The chip represents a major strategic shift: OpenAI is no longer just a software and model company. It is now building hardware.

Jalapeño is an Application-Specific Integrated Circuit (ASIC) designed from scratch for large language model inference. Unlike general-purpose GPUs from Nvidia or AMD, which handle many types of workloads, an ASIC is tuned for a narrow set of tasks. That narrower focus makes it cheaper and more efficient for specific AI workloads, though less adaptable than a GPU.

Early results, according to OpenAI president Greg Brockman, show "significantly better performance-per-watt than current state-of-the-art alternatives." Bloomberg reports the chip cuts inference costs by approximately 50%. For a company that spent $34 billion in operational expenses in 2025 — with $19.18 billion going to research and development infrastructure — even small reductions in inference costs could dramatically improve the bottom line.

Why an ASIC Instead of GPUs?

GPUs are general-purpose processors. They can handle training, inference, graphics, scientific computing, and more. That versatility comes at a cost: GPUs include circuitry for workloads they may never run, and they move data between memory and compute units more than necessary for LLM inference specifically.

An ASIC eliminates that overhead. Jalapeño starts from a blank slate, designed exclusively for modern LLM serving. OpenAI says the architecture is shaped by its experience running large-scale AI products and is meant to reduce unnecessary data movement while better matching compute, memory, and networking resources.

The tradeoff is clear: an ASIC is less flexible. If the AI workload changes dramatically, the chip may not adapt. But for a company like OpenAI that runs the same class of workloads — serving large language models to millions of users — the efficiency gains are worth the reduced flexibility.

The Partnership: OpenAI, Broadcom, Celestica

Jalapeño is the product of a three-party collaboration:

The partnership was officially announced in October 2025, after 18 months of prior collaboration. OpenAI and Broadcom had been working together since at least early 2025, with the public announcement coming after significant progress had already been made.

Nine Months From Schematic to Tape-Out

The development timeline is extraordinary. Jalapeño moved from early schematics to fabrication readiness in just nine months. In the semiconductor industry, new processor development cycles are typically measured in years. Nvidia's Blackwell generation took over two years from announcement to production. Google's TPU development cycles span multiple years.

OpenAI attributed this speed to a deep software-hardware co-development process that actively used its own AI models to accelerate parts of the chip design. Sources close to the firms told VentureBeat that prior-generation OpenAI models were used, though the company declined to specify which ones.

OpenAI president Greg Brockman told CNBC's David Faber that the chips were "designed from end to end in nine months with help from the company's AI models." He added: "The degree to which our models have been able to accelerate it was very surprising to us."

This is a meta-use of AI that deserves attention: OpenAI used its own language models to help design the silicon that will run future language models. It's a recursive acceleration loop — AI designing the hardware that runs AI, which then designs better hardware.

Performance and Cost

OpenAI has not released detailed benchmark numbers, but the available information points to significant improvements:

• ~50% inference cost reduction — Bloomberg reports the chip cuts inference costs by approximately half compared to current state-of-the-art alternatives
• Significantly better performance-per-watt — Brockman's description on CNBC
• Significantly better performance-per-dollar — Brockman's description on CNBC
• "Outstanding" early performance — sources close to the company, per VentureBeat

OpenAI has already begun testing at least one prior-generation model, GPT-5.3-Codex-Spark, on the chips at production workload levels, though in a test environment. The company received an early physical sample on the day of the announcement.

Brockman posted on X: "Performance per watt looking incredible." This suggests the efficiency gains are real and substantial, not just marketing claims.

Deployment Timeline

OpenAI's rollout plan is aggressive:

• End of 2026 — Initial deployment of Jalapeño chips across active data centers
• 2027+ — Expanding to gigawatt-scale data centers with Microsoft and other partners
• Multi-generation platform — Jalapeño is described as the first product in a multi-generation compute platform

The gigawatt-scale target is significant. A gigawatt-scale data center requires energy on the order of a small city. OpenAI and Microsoft have been building massive AI infrastructure, and purpose-built chips are a critical component of making that infrastructure economically viable.

Broadcom CEO Hock Tan told CNBC that compute demand from the company's six customers is "simply insatiable." He added: "It's just much more than we can address, and this is not just '26, not '27, we're seeing that same and even elevated demand in '28 as well."

Building the Full Stack

Jalapeño is not an isolated product. It's part of OpenAI's strategy to build the entire computational stack behind its models and products. From the company's announcement:

"OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience. Because OpenAI operates across the stack, each layer can be optimized around the same goal: making its models faster, more reliable, and more affordable for users."

This is a fundamental shift in OpenAI's identity. The company started as a research lab focused on artificial general intelligence. It evolved into a product company with ChatGPT. Now it is becoming an infrastructure company — building the hardware, software, and data centers that power its models.

OpenAI joins Google (TPUs), Amazon (Trainium/Inferentia), and Microsoft (Maia) as a full-stack AI company. But there's a key difference: those companies built chips primarily for their cloud businesses. OpenAI is building chips for its own inference workloads, with the potential to make them available to external AI firms as well.

The Economics: Why This Matters

OpenAI's financial situation makes this move urgent. Audited financial documents revealed that while OpenAI generated $13.07 billion in revenue in 2025, its total operational expenses ballooned to $34 billion, resulting in an operating loss of nearly $20.92 billion.

The primary culprit: compute infrastructure. Research and development costs — driven largely by the infrastructure required to train and serve massive language models — accounted for $19.18 billion, or approximately 56% of the company's entire spending footprint. OpenAI reportedly paid Microsoft over $10.59 billion just for R&D and compute infrastructure last year.

Custom inference chips address a critical part of this equation. While training costs are likely the larger expense, inference is where the margins are won or lost at scale. ChatGPT serves billions of requests per month. Even a small per-request cost reduction compounds into massive savings.

Brockman told CNBC that OpenAI "cannot get compute fast enough." Custom chips give the company more control over its compute supply chain, reducing dependence on Nvidia's GPUs and the associated supply constraints and pricing power.

Industry Impact

Jalapeño's announcement has implications across the AI industry:

For Nvidia: OpenAI is one of Nvidia's largest customers. A move to custom inference chips reduces that demand. However, Nvidia's dominance in training means OpenAI will likely continue buying GPUs for model development. The inference chip is a complement, not a replacement.

For Broadcom: The company is positioning itself as the go-to partner for custom AI silicon. With Google, Amazon, Microsoft, and now OpenAI all using Broadcom for custom chip development, the company is becoming the foundry of choice for AI ASICs. Its stock climbed following the announcement.

For the AI industry: If Jalapeño is made available to external AI firms — as both companies' press releases suggest — it could democratize access to efficient inference hardware. Smaller AI companies that can't afford their own chip development could benefit from OpenAI's infrastructure investments.

For Microsoft: As OpenAI's primary infrastructure partner, Microsoft stands to benefit from gigawatt-scale data center deployments. The partnership extends beyond chip design to data center construction and operations.

Open Questions

Several important questions remain unanswered:

• Exact performance numbers: OpenAI has not released detailed benchmarks comparing Jalapeño to Nvidia GPUs or other ASICs
• Manufacturing scale: How many chips can Broadcom produce per quarter? Can supply meet OpenAI's demand?
• External availability: Will Jalapeño be sold to other AI companies, or is it exclusively for OpenAI?
• Software stack: What tools and frameworks will developers use to run models on Jalapeño? Is it compatible with existing AI software?
• Training vs. inference: Will OpenAI develop custom training chips, or continue relying on Nvidia GPUs for that workload?

Conclusion

Jalapeño marks OpenAI's transition from a pure software company to a full-stack AI infrastructure company. The chip is purpose-built for LLM inference, designed to cut costs by approximately 50%, and developed in a blistering nine-month cycle accelerated by OpenAI's own AI models.

The strategic implications are significant. Custom inference chips reduce OpenAI's dependence on Nvidia, improve its unit economics, and position the company to scale its infrastructure more efficiently. The partnership with Broadcom and Celestica gives OpenAI access to world-class semiconductor design and manufacturing expertise.

But the most interesting aspect may be the meta-use of AI: OpenAI used its own language models to help design the silicon that will run future language models. It's a recursive acceleration loop that could become a template for how AI companies approach hardware development going forward.

As Brockman put it: "By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access." Whether that vision becomes reality will depend on execution — but the first chip is a strong start.