OpenAI’s partnership with Broadcom marks a major step toward custom AI silicon, giving the company greater control over inference costs, computing infrastructure, and its reliance on Nvidia GPUs.
OpenAI and Broadcom unveiled Jalapeño on June 24, 2026, marking OpenAI’s entry into custom silicon. The chip, which the company is calling an “Intelligence Processor,” is an application-specific integrated circuit designed from scratch for large language model inference, the process of serving AI model responses to users in products like ChatGPT and Codex. Engineering samples were physically delivered to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom President and CEO Hock Tan at a ceremony on Wednesday.
The chip was designed in just nine months, from initial architecture to manufacturing tape-out. It is manufactured by TSMC on a 3nm process, the same node used by Apple’s M4 chip and Nvidia’s Blackwell line. Celestica, a Canadian electronics manufacturing firm, is handling board and rack system integration. Engineering samples are already running ML workloads in OpenAI’s labs at production target frequency and power, including a model the company identified as GPT-5.3-Codex-Spark.
Early internal testing shows Jalapeño delivering roughly 50% lower inference cost compared to current GPU alternatives, according to Broadcom CEO Hock Tan. OpenAI has not yet released independent benchmark data; a detailed technical report is expected in the coming months. Initial prototype deployments are planned for late 2026, with a full production ramp in 2027 and 2028 across data centers operated by Microsoft and other infrastructure partners.
Why OpenAI Built Its Own Chip
The financial motivation is not complicated. Nvidia commands an estimated 75% profit margin on its high-end GPUs. OpenAI, by contrast, keeps roughly 33 cents of profit on each dollar of revenue after accounting for its infrastructure costs. Running ChatGPT, the API, and agentic products like Codex at scale means inference compute is one of the largest line items on OpenAI’s cost structure. A chip purpose-built for that specific workload, rather than a general-purpose GPU optimised for a broader range of tasks, is the most direct path to reducing that cost.
Inference workloads have specific characteristics that make them a strong fit for a custom ASIC. They are repetitive and predictable: the system is serving responses to LLM queries at very high volume, and the memory movement patterns, kernel operations, and networking requirements are well understood. Nvidia’s GPUs carry substantial compute capacity that inference workloads never use. Jalapeño, by Broadcom’s description, was designed to minimise unnecessary data movement and balance compute, memory, and networking resources specifically around transformer-based model serving.
The second motivation is supply chain control. Relying entirely on Nvidia means exposure to allocation constraints, long lead times, and pricing power that OpenAI cannot influence. A custom chip manufactured through Broadcom and TSMC creates an alternative supply chain and, over time, reduces negotiating leverage that Nvidia has held over the entire industry.
The Broadcom Partnership and Development Speed
The partnership between OpenAI and Broadcom was announced publicly in October 2025, with an initial commitment to deploy 10 gigawatts of OpenAI-designed AI accelerators between 2026 and 2029. The Jalapeño announcement, eight months later, is the first concrete output of that agreement.
The nine-month design cycle is the number that will get the semiconductor industry’s attention. High-performance ASIC development typically takes years. OpenAI attributes the speed to three factors: tight software-hardware co-development between OpenAI’s engineering teams and Broadcom’s silicon implementation group; Broadcom’s existing expertise and manufacturing relationships; and the use of OpenAI’s own AI models to accelerate parts of the design and optimisation process. Greg Brockman noted on CNBC that the degree to which their own models were able to accelerate the design was “very surprising” even internally.
Broadcom brings specific institutional knowledge to this partnership. The firm helped engineer Google’s Tensor Processing Unit line, one of the most successful custom AI silicon programmes in the industry, and has built similar ASICs for other hyperscalers. Its familiarity with TSMC’s manufacturing process and high-performance networking integration, including Broadcom’s own Tomahawk networking silicon, is a core part of why OpenAI selected it as a partner rather than attempting to go directly to a foundry with a design.
What Jalapeño Is and Is Not
Jalapeño is not a training chip. The compute-intensive process of building frontier models from scratch will still require Nvidia GPUs, and OpenAI has not suggested otherwise. What Jalapeño targets is the inference side, the ongoing cost of serving billions of user queries across ChatGPT, the API, and future agentic products. That is the highest-volume, most predictable portion of OpenAI’s compute spend and the part that benefits most from purpose-built silicon.
The chip will not be sold to external customers. It is being built for OpenAI’s own infrastructure, with Microsoft expected to absorb a significant share of initial production as OpenAI’s primary cloud and data centre partner. Broadcom CEO Hock Tan confirmed prototype-scale deployment in late 2026, scaling through 2027 and 2028.
The architecture uses a systolic array design with eight high-bandwidth memory stacks. Broadcom claims its performance is competitive with Nvidia’s Blackwell GPUs and Google’s TPUs on inference tasks. Those claims are vendor-reported from pre-production samples and should be treated as such until independent benchmarks are available.
What This Means for Nvidia
Jalapeño does not displace Nvidia from OpenAI’s infrastructure. Training runs, research experiments, and workloads that do not fit a fixed inference architecture will continue to require general-purpose GPUs. Nvidia remains the dominant supplier for that category of work across the entire industry, and that position is not under immediate threat from a single inference ASIC.
What changes is the economic picture at the margin. If OpenAI successfully moves its highest-volume, most cost-sensitive workloads onto a chip that costs half as much to run per inference token, it reduces the total volume of Nvidia hardware it needs to buy over time. Multiply that across the rest of the industry, where Meta, Amazon, Google, Microsoft, and now Anthropic are all pursuing similar strategies through Broadcom, Marvell, and other custom silicon partners, and the long-term pressure on Nvidia’s pricing power becomes clearer.
The Broader Industry Pattern
OpenAI is the latest, not the first. Google has run its own TPU programme for over a decade. Amazon has Trainium and Inferentia. Meta has built custom accelerators through Broadcom and Marvell. Microsoft launched its Azure Maia 100 accelerator in late 2023 and upgraded to the Maia 200 in January 2026, which is already powering some of OpenAI’s GPT workloads on Azure infrastructure. Anthropic is reported to be weighing a chip programme of its own.
The common thread across all of these programmes is the same: compute efficiency is becoming a competitive moat. As model architectures converge and research techniques become public knowledge through papers and open-source releases, the labs that can serve intelligence most cheaply at scale gain a structural advantage. Custom silicon is the most direct way to extract that efficiency from the hardware layer.
What OpenAI’s announcement adds to this picture is validation that a frontier AI lab can design a competitive inference chip in under a year, with significant assistance from the AI models it is trying to run. That is the detail worth watching, because it implies the cost and timeline of chip development itself may compress as AI tools become more capable at hardware engineering tasks.
Frequently Asked Questions
What is Jalapeño and what does it do? Jalapeño is OpenAI’s first custom AI chip, unveiled on June 24, 2026, in partnership with Broadcom. It is an application-specific integrated circuit designed specifically for LLM inference, the process of serving AI model responses to user queries. It is not a training chip and will not replace Nvidia GPUs for model development or general-purpose AI workloads.
How much cheaper is Jalapeño compared to existing GPU solutions? Broadcom CEO Hock Tan stated in an interview that early testing shows approximately 50% lower inference cost compared to current GPU alternatives. OpenAI has not yet published independent benchmark data. A full technical report is expected in the coming months, and these figures should be treated as vendor-reported pre-production numbers until verified.
Who manufactures the chip and when will it be deployed? Jalapeño is manufactured by TSMC on a 3nm process node. Celestica handles board and rack integration. OpenAI plans small prototype deployments by the end of 2026, with full production rollout expected in 2027 and 2028 across data centres operated with Microsoft and other infrastructure partners.
How quickly was Jalapeño developed? OpenAI and Broadcom completed the chip from initial design to manufacturing tape-out in approximately nine months, which both companies describe as one of the fastest ASIC development cycles ever achieved in high-performance advanced semiconductors. OpenAI’s own AI models were used to accelerate parts of the design process.
Does this mean OpenAI no longer needs Nvidia? No. Jalapeño targets inference workloads only. Training frontier models and handling diverse GPU-dependent workloads will still require Nvidia hardware. Jalapeño reduces OpenAI’s dependence on Nvidia at the inference layer, which is the highest-volume part of its compute spend, but does not replace Nvidia across the stack.
Is Jalapeño available to other companies? No. Jalapeño is being built exclusively for OpenAI’s own infrastructure. It will not be sold to external customers. Microsoft, as OpenAI’s primary data centre partner, is expected to receive a significant share of initial production for deployment in Azure infrastructure.
Sources
- OpenAI: OpenAI and Broadcom Unveil LLM-Optimized Inference Chip
- Broadcom: OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor
- CNBC: OpenAI unveils first chip as part of Broadcom deal in effort to ‘build the full stack’
- TechCrunch: OpenAI unveils its first custom chip, built by Broadcom
- Bloomberg: OpenAI, Broadcom Unveil Jalapeno AI Chip Promising Faster, Cheaper Model Runs
- VentureBeat: OpenAI unveils first custom AI inference chip, Jalapeño, with Broadcom
- OpenAI: OpenAI and Broadcom announce strategic collaboration
Stay in the Loop
For more stories, breakdowns, and unfiltered takes on what is really happening in Indian and global business and tech, follow TheFounder Nation.
Instagram: thefoundernation
We cover what the mainstream business press won’t.
© TheFounder Nation | All rights reserved
