OpenAI Unveils Jalapeño, Its First Custom Chip Built With Broadcom

OpenAI on Wednesday unveiled Jalapeño, its first custom-designed computer chip, co-developed with Broadcom to run ChatGPT and other AI products faster and more cheaply, with details laid out in the full unveiling of the Jalapeño inference processor.

Designed specifically for inference, the work of running an already-trained model to answer a user’s prompt, Jalapeño will be deployed at gigawatt scale inside Microsoft and other data center partners by the end of 2026. OpenAI’s own AI models helped design the chip, and engineering samples are already running OpenAI workloads in the lab. The performance figures OpenAI is putting on the table are anchored to internal early testing, with a detailed technical report still to come. The chip is the first public product of the strategic collaboration OpenAI and Broadcom announced in October 2025.

The Chip Itself

OpenAI calls Jalapeño its first “Intelligence Processor,” an accelerator built from a blank slate for modern large language model inference. The company says the chip is not a general-purpose accelerator retrofitted from earlier AI workloads. The architecture was designed around OpenAI’s roadmap of models, kernels, serving systems and product needs.

Jalapeño is built to run large language models across the industry, including OpenAI’s own. The stated goal is to combine the power and throughput of today’s leading AI accelerators with latency closer to the fastest specialized inference systems. OpenAI says that combination makes the chip a fit for interactive LLM products at scale. Engineering samples are already running machine-learning workloads in OpenAI’s labs at production target frequency and power, including GPT-5.3-Codex-Spark.

The chip was physically delivered to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom CEO Hock Tan and Semiconductor Solutions President Charlie Kawwas. The companies treated the handoff as the public marker of the program.

A Nine-Month Build and a Self-Designed Stack

OpenAI co-developed Jalapeño from initial design to manufacturing tape-out in just nine months, what it called the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. The speed came from deep software-hardware co-development between OpenAI engineers and Broadcom’s silicon team. OpenAI’s own models accelerated parts of the design and optimization process. The same models served to ChatGPT users helped shape the silicon that will run future models, OpenAI said.

OpenAI designed the chip architecture. Broadcom handled silicon implementation and contributed its Tomahawk networking silicon for production, per a joint press release on the LLM-optimized processor. Celestica handles the board, rack and system work that turns bare silicon into data-center hardware. The architecture reduces data movement and balances compute, memory and networking resources to push realized utilization closer to theoretical peak performance. It is the first chip in what OpenAI calls a multi-generation compute platform.

Performance Claims and What’s Still Unproven

The single biggest qualifier on Wednesday’s announcement sits in the way OpenAI framed the chip’s numbers. All of the performance claims rest on internal early testing.

While OpenAI is still measuring final performance, early testing shows that Jalapeño will deliver performance per watt substantially better than current state-of-the-art, the company wrote. A detailed technical report is promised “in the coming months.” OpenAI says the architecture cuts data movement and rebalances compute, memory and networking to push realized utilization closer to theoretical peak performance. The benchmark OpenAI is leaning on is the engineering samples already running the company’s own workloads at production target frequency and power.

Broadcom CEO Hock Tan called the chip “just the beginning of a multi-generation roadmap.” He framed it as the opening move in a long-running collaboration designed to deploy gigawatt-scale data centers with Microsoft and other partners beginning in 2026. The framing matters because the companies have not yet published the independent benchmarks investors and customers will want to see.

Richard Ho, who leads OpenAI’s hardware program, said Jalapeño was designed around the kernels, memory movement, networking and serving patterns that matter most for frontier AI models. His claim is that the chip will execute the company’s most important workloads close to the hardware’s theoretical limits. Neither Ho nor Tan gave a specific figure for performance per watt on Wednesday. OpenAI is promising that figure, with methodology, in the technical report due in the coming months.

• 9 months from initial design to manufacturing tape-out
• 10 gigawatts of custom AI accelerators covered by the OpenAI-Broadcom collaboration
• Engineering samples running GPT-5.3-Codex-Spark in the lab at production target frequency and power
• Performance per watt “substantially better than current state-of-the-art,” based on early testing
• Detailed technical report due “in the coming months”

Where Jalapeño Goes and Who Else Gets a Foot in the Door

Initial deployment is set for the end of 2026 at gigawatt scale, with Microsoft as the first named data center partner and other facilities to follow. The October 2025 strategic collaboration between OpenAI and Broadcom set the volume target at 10 gigawatts of custom AI accelerators, with the rollout running from the second half of 2026 through the end of 2029, per the October 2025 announcement of the 10-gigawatt collaboration. Jalapeño is the first chip in that multi-generation build.

The deployment stack now has three named partners with clearly divided roles. Broadcom contributes silicon implementation and Tomahawk networking. Celestica handles the board, rack and system work that turns bare silicon into data-center hardware. Microsoft and other data center operators provide the floor space, power and cooling.

OpenAI has been explicit about the strategic logic. The company is designing the infrastructure underneath its own models, an arrangement it argues will let every layer, from chip to product, be optimized around the same goal of faster, cheaper, more reliable AI.

• Microsoft: First named data center partner for the initial 2026 deployment.
• Broadcom: Silicon implementation partner and supplier of Tomahawk networking silicon.
• Celestica: Board, rack and system integration partner.

Will It Loosen Nvidia’s Grip?

Nvidia is the elephant in the room. TechCrunch reports that OpenAI’s chip plans have long been seen as a way to reduce dependence on Nvidia’s GPUs, the dominant accelerators behind most of today’s AI training and serving. Jalapeño is the most concrete public step yet on that road.

Other hyperscalers have already walked it. Google and Amazon have built their own AI accelerators to serve similar workloads, per TechCrunch’s reporting. Pre-training of the largest models will most likely continue to run on Nvidia hardware, leaving Jalapeño to capture the inference half of the workload. That is where the most compute hours at OpenAI actually happen, and the half where custom silicon can move the cost line fastest.

Whether the performance claims hold at data-center scale is the next question, and it will land in the technical report OpenAI is promising in the coming months. The packaging step that turns finished silicon into shippable AI accelerators has been running close to capacity elsewhere in the industry this year, AI chip packaging capacity straining to keep up with demand, which will shape how quickly new OpenAI racks can come online. Brockman, in OpenAI’s announcement, tied the chip to a broader bet on AI compute as a general-purpose resource.

The world is moving to a compute-powered economy. Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems.

Greg Brockman, OpenAI’s president and co-founder, said this in the company’s June 24 announcement of the chip.

Frequently Asked Questions

What is the OpenAI Jalapeño chip?

Jalapeño is OpenAI’s first custom AI chip, an inference accelerator designed from scratch for large language model inference. Unveiled in San Francisco and Palo Alto on June 24, 2026, it is the first chip in a multi-generation compute platform OpenAI is building with Broadcom and Celestica.

Who co-developed Jalapeño with OpenAI?

OpenAI designed the chip’s architecture from scratch. Broadcom handled silicon implementation and contributed its Tomahawk networking silicon for the production platform. Celestica is responsible for board, rack and system integration at scale.

When and where will Jalapeño be deployed?

Initial deployment is set for the end of 2026 at Microsoft and other data center partners, with multi-generation expansion through 2029. The two companies first disclosed the broader partnership in October 2025, with deployment then running from the second half of 2026 through end of 2029.

Does Jalapeño replace Nvidia chips?

OpenAI framed Jalapeño as part of a broader strategy to reduce single-supplier dependence. Detailed performance data, drawn from the engineering samples already running in OpenAI’s labs, is still pending in a technical report due in the coming months. The engineering samples are running GPT-5.3-Codex-Spark at production target frequency and power.