Key Highlights:
OpenAI has unveiled the Jalapeño AI chip - its first custom inference processor, built with Broadcom.
The chip went from design to tape-out in just nine months - a semiconductor industry record.
OpenAI's own AI models helped design the chip. The same models that serve users built the hardware that will run them.
Deployment begins late 2026, with gigawatt-scale data centres planned with Microsoft and other partners.
- OpenAI lost $20.92 billion in 2025. The Jalapeño AI chip is as much a financial lifeline as a technical milestone.
OpenAI has entered the hardware race. On June 24, 2026, the company unveiled the Jalapeño AI chip its first custom-built inference processor. Broadcom co-developed it. The chip arrived physically at OpenAI's offices, handed to CEO Sam Altman and President Greg Brockman by Broadcom's top leadership. That handover carried a clear message: OpenAI is no longer just a software company.
For years, OpenAI depended entirely on Nvidia's expensive GPUs. The Jalapeño AI chip is OpenAI's first serious move to change that.
Also Read | OpenAI Acquires Ona - A Cloud Startup to Power "Codex"
What Is Jalapeño AI Chip?
The Jalapeño AI chip does one thing. It runs AI models for users fast, efficiently, and at lower cost.
OpenAI is clear: this is not a general-purpose processor. It is not a repurposed training chip either. It is a purpose-built inference ASIC, designed from scratch for one job. Inference happens every time a user types into ChatGPT and gets a response. It runs billions of times a day. It costs OpenAI enormous sums. The Jalapeño AI chip targets exactly that cost.
The architecture tackles the biggest inference bottlenecks data movement, compute-memory balance, networking efficiency. Broadcom contributed its Tomahawk networking silicon. Celestica handled board and system integration. The result: a chip built to run AI models closer to its theoretical hardware limits than anything currently available.
Also Read | OpenAI Acquires Weights.gg: Why This Voice-Cloning Deal Matters for the Future of AI
Nine Months. A Record. And AI Built It.
Most high-performance chips take years to develop. The Jalapeño AI chip took nine months from first schematic to fabrication readiness. OpenAI and Broadcom only announced their partnership in October 2025. A working chip arrived by June 2026. That is a remarkable pace for the semiconductor industry. The secret? OpenAI used its own AI models to accelerate the design process. The models that power ChatGPT helped engineer the chip that will run ChatGPT faster.
OpenAI President Greg Brockman told CNBC the speed caught even the engineers off guard. "The degree to which our models have been able to accelerate it was very surprising to us," he said. This is AI building better AI infrastructure. The loop is now real.
The Financial Case Behind the Silicon
The Jalapeño AI chip is not just a technical achievement. It is a financial necessity.
In 2025, OpenAI earned $13.07 billion in revenue. It spent $34 billion. The operating loss hit $20.92 billion. Of that, $19.18 billion - 56% of all spending went to compute infrastructure. OpenAI also paid Microsoft over $10.59 billion for R&D and compute access alone.
Those numbers are unsustainable. Custom silicon is how companies like Google and Amazon brought their compute costs under control. OpenAI is now following the same path.
With a public offering expected in 2026, the Jalapeño AI chip sends a signal to investors. OpenAI has a plan to cut costs. It controls more of its own stack now. Profitability is no longer just a promise.
Also Read | From Open AI to Biotech: Big Moves by Google, OpenAI & Anthropic
OpenAI Retains Nvidia as Core Infrastructure Partner
The Jalapeño AI chip does not replace Nvidia. Not yet. Possibly not ever, for certain tasks.
In February 2026, Nvidia finalised a $30 billion investment in OpenAI. The deal includes 10 gigawatts of computing using Nvidia's next-generation Vera Rubin platform. Amazon committed $50 billion in the same round. That includes two gigawatts of AWS Trainium capacity. OpenAI also holds agreements with AMD and Cerebras.
The division of labour is straightforward. Nvidia handles training the heavy, intensive work of building AI models. The Jalapeño AI chip handles inference the high-volume, cost-sensitive work of serving those models to users.
Even small savings on inference can shift OpenAI's economics significantly. That is the bet.
The strategy is clear: use Nvidia for training, use Jalapeño to slash the cost of serving.
What Comes Next
Jalapeño is the first step in a multi-generation compute platform designed for initial deployment by the end of 2026 and expanding in the years ahead, combining OpenAI-designed accelerators with Broadcom silicon implementation, networking, and connectivity technologies.
Also Read | Why Is Anthropic Calling Claude Tag a New Way to Work With AI?
Broadcom CEO Hock Tan outlined the scale of ambition without holding back. "We will start seeing it really ramp up in '27 and really going full tilt in first half '28," Tan told CNBC.
The ultimate goal of the OpenAI-Broadcom partnership is to deploy gigawatt-scale data centres facilities requiring as much energy as a small city with Microsoft and other partners starting in 2026.
For Broadcom, the payoff is already visible. Broadcom shares demonstrated an 18% year-over-year increase in the first part of 2026 and a nearly 7X boost since the end of 2022.
The message from San Francisco is unmistakable: OpenAI no longer wants to just build the smartest models. It wants to own the silicon they run on. Jalapeño is the first proof that it can.
