OpenAI's Jalapeño Chip Hits Silicon 9 Months After Design Kickoff
OpenAI's first custom silicon, Jalapeño, goes from blank slate to running GPT-5.3 workloads in nine months — and it's just the start of a multi-generation compute platform targeting gigawatt-scale deployment.
Takeaways
- OpenAI ships Jalapeño: its first custom AI chip, purpose-built for LLM inference, co-developed with Broadcom and already running GPT-5.3 workloads in the lab.
- Record-breaking speed: the chip went from blank slate to tape-out in nine months — OpenAI claims the fastest ASIC development cycle ever for high-performance semiconductors.
- Performance claim: early testing shows substantially better performance per watt than current state-of-the-art; full technical report coming in months.
- Gigawatt ambitions: part of a multi-generation platform targeting 10 GW of deployed compute by 2030, with first deployments planned by end of 2026.
- Cost survival strategy: OpenAI projects over $200B in operating expenses through 2029 — owning inference silicon is its primary lever to avoid burning cash on Nvidia GPUs forever.
- Microsoft is the wildcard: the $18B first-phase financing requires Microsoft to commit to ~40% chip purchases; that agreement has not been publicly confirmed.
OpenAI just crossed a threshold that most AI labs only dream about: it has designed and taped out its own chip. Jalapeño, OpenAI's first Intelligence Processor, is a purpose-built LLM inference accelerator co-developed with Broadcom and manufacturing partner Celestica. Engineering samples are already running ML workloads in the lab , including GPT-5.3-Codex-Spark , at production target frequency and power. This is no longer a roadmap slide. It is silicon.
What Jalapeño actually is
Inference is the phase where a trained model generates responses to real user queries , every ChatGPT answer, every Codex completion. It is fundamentally different from training, and general-purpose GPUs were not designed around it. Jalapeño is a blank-slate design for modern LLM inference, not a general-purpose accelerator adapted from earlier AI workloads.
The architecture reflects that philosophy at every level:
- The goal is to combine the power and throughput of today's leading AI accelerators with latency closer to the fastest specialized inference systems.
- The architecture reduces data movement and balances compute, memory, and networking resources to achieve realized utilization much closer to theoretical peak performance. In plain terms: most chips waste a significant fraction of their theoretical compute because data can't move fast enough. Jalapeño was designed to close that gap.
- Broadcom's silicon implementation and networking technologies, including Tomahawk networking silicon, help bring the platform to large-scale production.
- The chips are expected to feature a systolic array architecture optimized for AI inference, high-bandwidth memory (HBM, possibly HBM3E or HBM4), and manufacturing on TSMC's 3nm process node.
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. A full technical report is expected in the coming months.
The team behind it
Jalapeño was delivered to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom President and CEO Hock Tan and President Charlie Kawwas. The hardware program is led by Richard Ho, who previously led Google's TPU initiative. OpenAI hired former Lightmatter chip engineering lead and Google TPU head Richard Ho as the head of hardware.
Ho laid out the company's full-stack hardware strategy for the first time during a closed-door session at Stanford University, arguing that GPUs were never designed for the coming era of AI agents, forcing OpenAI to reclaim low-level control of AI compute through custom silicon and systems. His framing: "We're not building a chip. We're building a system."
Nine months from nothing to silicon
The speed of this tape-out deserves its own headline. Jalapeño was co-developed from initial design to manufacturing tape-out in just nine months, and the custom AI accelerator program represents what OpenAI believes to be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors.
For context: traditional chip companies typically require five to seven years from architecture to tape-out, and Google's first-generation TPU took roughly three years. AI research moves extremely quickly, and the chip design process is typically 18-24 months from inception to tape-out. OpenAI compressed that by an order of magnitude.
How? That speed reflects deep software-hardware co-development with OpenAI's engineering teams, Broadcom's silicon implementation expertise, and the use of OpenAI models to accelerate parts of the design and optimization process. The same models being served to users helped design the hardware that will run future models , a feedback loop OpenAI is leaning into hard.
Ho noted internally that "we can see how the model will change over the next six to nine months," meaning the hardware team is not designing chips for today's models, but doing preparatory design for models that don't yet exist.
The backstory: from financing snag to shipping silicon
The path here was not smooth. The collaboration, internally codenamed Project Nexus, was formally announced in October 2025 with an ambitious target: deploying racks of AI accelerator and network systems targeted to start in the second half of 2026, to complete by end of 2029.
But just weeks before today's announcement, the project appeared to be in trouble. According to a report from The Information citing an internal memo, the project was stalled because Broadcom told OpenAI it would only finance the first phase , a massive buildout requiring 1.3 gigawatts of data-center capacity , if Microsoft agreed to purchase roughly 40% of the custom chips. The financing snag stemmed from a fundamental disagreement over data-center infrastructure strategy: OpenAI envisioned specialized data centers optimized for its custom silicon, while Microsoft preferred to stick with its standard, versatile data-center designs, making Microsoft hesitant to commit.
This condition exposed OpenAI's deep reliance on Microsoft and cast uncertainty on the entire chip plan, which could ultimately cost up to $180 billion. The project's first phase alone carried an $18 billion financing requirement. Today's announcement suggests those negotiations have been resolved , at least enough to get engineering samples running.
The strategic math
Why does OpenAI need its own chip so badly? The numbers are stark. The company is projected to burn through over $200 billion in operating expenses through 2029, and owning its own chips , rather than relying on costly Nvidia hardware , is a central pillar of its strategy to manage those ballooning expenditures.
Owning its own chips rather than being dependent on Nvidia is a primary strategy to manage its bloating expenditure plans. The project represents OpenAI's efforts to break free from its costly reliance on Nvidia hardware to run its AI models. The inference cost problem is particularly acute for products like Codex. The performance challenges have been particularly visible in Codex, OpenAI's AI-powered coding product, where speed is a key requirement for professional users. Altman has previously said that customers place a premium on fast response times for coding tasks.
The full-stack flywheel OpenAI is building looks like this:
- Custom silicon optimized for OpenAI's specific model architectures reduces inference cost per token
- Lower cost enables more aggressive pricing and broader access
- More usage generates more revenue, which funds the next chip generation
- Each chip generation is co-designed with models that don't yet exist, compressing the feedback loop
Who wins, who loses
Gartner predicts that by 2026, compute spending on inference , the running of AI , will overtake spending on training as the technology matures. That shift is exactly the terrain Jalapeño is designed to capture. OpenAI is positioning itself to own the most cost-sensitive part of the AI stack at precisely the moment it becomes the dominant cost center.
The clearest loser in this narrative is Nvidia , at least at the margin. For Nvidia, the deal adds to a growing list of partial defections among major AI customers exploring in-house silicon. For Broadcom's custom AI chip customers, which include Meta, Google, and TikTok parent ByteDance, the partnership might offer more efficiency and lower costs in the long run, as well as a decreased reliance on Nvidia. That said, these XPUs can't completely replace Nvidia GPUs for all things. "There's definitely a trend, although I don't think any of them are going to completely abandon GPUs," noted Gareth Owen of Counterpoint Research.
Broadcom is a clear winner. Broadcom CEO Hock Tan said the company has six core custom chip customers, including Anthropic, Google, Meta, and OpenAI, which are driving the company's growth in AI revenue. AI revenue more than doubled in the second quarter on an annual basis to $10.8 billion, and the company expected AI revenue to triple in the current quarter to $16 billion.
What this unlocks
The most underappreciated part of this announcement is what it makes possible architecturally. Because OpenAI controls both the model and the chip, it can co-optimize in ways that are simply not possible when you're renting someone else's hardware. Kernel implementations, memory access patterns, networking topology, serving schedules , all of it can be tuned to the same objective function.
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. This is the same playbook Google ran with TPUs , and it took Google roughly a decade to get there. OpenAI is attempting to compress that timeline dramatically.
The second-order effect is leverage. The OpenAI-Broadcom chip could give the ChatGPT developer more leverage in pricing talks with Nvidia, whose H100 and forthcoming Blackwell GPUs still dominate the AI training market. Even if Jalapeño never fully displaces Nvidia hardware, its existence changes the negotiating dynamic.
What's 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; and Celestica's board, rack, and system expertise.
The full plan aims to deploy a total of 10 gigawatts of AI computing power by 2030, reducing reliance on Nvidia's expensive GPUs. For scale: 10 gigawatts of capacity is comparable to the electricity needs of over 8 million U.S. households or five times the output of the Hoover Dam. A detailed technical performance report is expected in the coming months. The key open question is whether Microsoft will formally commit to purchasing chips at the scale the financing structure requires , that decision will determine how fast the gigawatt-scale ambitions actually materialize.
For developers and API users, the near-term implication is straightforward: if Jalapeño delivers on its performance-per-watt claims, it feeds directly into faster responses and lower API costs. The chip is not a research project. It is already running production workloads. The race to own the inference layer of AI just got a new contender.