JALAPEÑO — KEY FACTS
● What it is: OpenAI's first custom AI chip — an ASIC inference accelerator, not a training chip
● Built with: Broadcom (silicon implementation) + Celestica (rack integration)
● Development time: Concept to tape-out in 9 months — claimed fastest ASIC cycle in high-performance semiconductors
● Cost savings: ~50% cheaper than current AI GPUs (Broadcom CEO Hock Tan, Bloomberg interview)
● Already running: GPT-5.3-Codex-Spark at production target frequency and power
● Deployment: Initial by end of 2026 — gigawatt-scale data centers with Microsoft and partners
● Design assist: OpenAI's own models accelerated parts of the chip design and optimization process
● Purpose: LLM inference only — training still runs on Nvidia hardware
What Jalapeño Actually Is
Jalapeño is an ASIC — an application-specific integrated circuit — designed exclusively for LLM inference. It is not a general-purpose AI accelerator like Nvidia's H100 or GB200, which can handle both training and inference. It is not a repurposed training chip. It is built from scratch around one specific task: running pre-trained AI models in response to user requests at the lowest possible cost and latency at massive scale.
The architecture is designed around what OpenAI knows about how its models actually behave at inference. Richard Ho, who leads OpenAI's hardware program: "We optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models." The practical result: reduced data movement between compute and memory (the primary bottleneck in LLM inference), better balance between compute, memory, and networking resources, and utilization rates closer to theoretical maximum — which is why Broadcom CEO Hock Tan told Bloomberg the chip shows approximately 50% cost savings versus current AI GPUs.
Physically: the chip features eight HBM (high-bandwidth memory) sites surrounding a large compute die. Tom's Hardware describes the die floorplan as highly regular and columnar — consistent with a tiled AI accelerator architecture — though no exact die size, memory configuration, or clock speed has been disclosed. The chip is a reticle-limited design, meaning the compute chiplet is as large as the manufacturing process allows per exposure field.
Nine Months — Why the Development Speed Matters
Typical custom ASIC development cycles for high-performance semiconductors run 18-36 months. OpenAI and Broadcom went from initial design to manufacturing tape-out in nine months — which they describe as potentially the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. OpenAI President Greg Brockman told CNBC the speed was enabled by three things: deep software-hardware co-development with OpenAI's engineering teams, Broadcom's silicon implementation expertise, and critically — OpenAI's own AI models accelerating parts of the design and optimization process.
That last point is the one with the biggest implications. The models OpenAI serves to users in ChatGPT and Codex helped design the chip those same models will run on. Brockman: "The degree to which our models have been able to accelerate it was very surprising to us." If AI can consistently compress chip design cycles from 24 months to 9 months, the implications extend far beyond OpenAI — every semiconductor company with access to frontier AI capability gains a structural advantage in development speed. This is the first documented case of AI accelerating the production of the chips it runs on at production quality.
The Strategic Context — Why OpenAI Is Building Its Own Chips
Before Jalapeño, OpenAI had a structural disadvantage shared by no major AI lab of its scale: it did not own any of its inference infrastructure. Google runs inference on its own TPUs. Amazon runs inference on its own Trainium and Inferentia chips. Microsoft has Azure Maia. OpenAI ran everything on Nvidia GPUs purchased from a supplier whose pricing and availability it could not control.
The financial stakes are significant. At OpenAI's scale — 1 billion monthly ChatGPT users, millions of API calls per second, Codex's 5 million weekly developers — inference cost is the single largest line item in the unit economics. A 50% reduction in inference cost per token does not just improve margins. It determines whether OpenAI can price its products competitively at scale while moving toward profitability ahead of its IPO. VentureBeat notes: "Still, as OpenAI lays the groundwork for a heavily anticipated public offering in 2026, the Jalapeño inference chip may offer some reassurance to private investors and public markets that OpenAI has a plan for digging itself out of the financial hole and moving toward profitability."
The strategic framing from Brockman is explicit: "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 full-stack play — the same move Google made with TPUs, Amazon made with Trainium, and Microsoft made with Azure Maia. OpenAI is the last major AI company to do it. Jalapeño is the first step.
What It Means for Nvidia
Jalapeño is inference-only — it does not replace Nvidia for training, and Nvidia remains OpenAI's primary training hardware partner. OpenAI has also signed agreements with AMD (Instinct MI450 Series GPUs) and Cerebras for additional compute diversity. None of these deals end Nvidia dependence at the training layer. But at the inference layer — where the majority of ongoing compute spend is concentrated at OpenAI's scale — Jalapeño represents a genuine shift away from Nvidia GPUs for a significant and growing workload.
The pattern mirrors what happened at Google and Amazon: custom inference ASICs handle the high-volume predictable workload (serving users), while Nvidia GPUs handle the unpredictable research-intensive workload (training new models). Nvidia retains the high-value frontier training revenue. OpenAI captures the inference efficiency gains. This is not the Nvidia killer story some headlines suggest — it is the natural maturation of an AI company large enough to justify custom silicon for its most predictable workloads.
Deployment Timeline and Gigawatt-Scale Ambition
| Timeline |
Milestone |
| Oct 2025 |
OpenAI-Broadcom partnership announced |
| June 24, 2026 |
Jalapeño unveiled. Engineering samples delivered to Altman and Brockman by Hock Tan |
| Now (testing) |
GPT-5.3-Codex-Spark running on Jalapeño at production target frequency and power |
| Late 2026 |
Small prototype deployment — initial production at limited scale |
| 2027+ |
Gigawatt-scale data centers with Microsoft and partners — compute requiring energy on the order of entire cities |
| Multi-generation |
Jalapeño is generation 1 of a multi-chip roadmap with Broadcom. Technical performance report coming in the next few months |
The Honest Caveats — What We Don't Know Yet
No hard performance numbers published. OpenAI and Broadcom claim "substantially better performance per watt than current state-of-the-art" but have not released benchmark numbers, clock speeds, memory bandwidth figures, or throughput comparisons. A detailed technical performance report is promised "in the coming months." The 50% cost savings figure comes from a CEO interview — not a published technical document. Take it directionally, not precisely.
Late 2026 deployment is "small prototype." Broadcom CEO Hock Tan described the end-of-2026 deployment as "small prototype development." The gigawatt-scale ambition is real but it is a 2027 and beyond story — not 2026. The timeline from "engineering sample" to "production at gigawatt scale" is typically 18-24 months for complex custom silicon.
Training still needs Nvidia. Jalapeño is inference-only. OpenAI's training runs — the most computationally intensive work — still depend on Nvidia GPUs. Frontier model training at GPT-5.x scale requires H100/GB200 clusters that Jalapeño cannot replace. The "reducing Nvidia dependence" story is accurate for inference but not for training.
Frequently Asked Questions
What is the OpenAI Jalapeño chip?
Jalapeño is OpenAI's first custom AI inference chip, co-developed with Broadcom and unveiled June 24, 2026. It is an ASIC — an application-specific integrated circuit — designed exclusively for running large language models at inference. It is not a training chip and does not replace Nvidia GPUs for model training. Early testing shows approximately 50% cost savings versus current AI GPUs. Initial deployment is planned by end of 2026.
Does Jalapeño threaten Nvidia?
For inference workloads at OpenAI scale: yes, meaningfully. For training: no — Jalapeño cannot replace Nvidia H100/GB200 clusters for frontier model training. The same pattern exists at Google (TPUs for inference, Nvidia for some training), Amazon (Trainium/Inferentia for inference), and Microsoft (Azure Maia for inference). Nvidia retains the high-value frontier training business. OpenAI captures inference efficiency gains with Jalapeño.
How is Jalapeño different from Google's TPU or Amazon's Trainium?
All three are custom ASICs designed to reduce inference costs versus Nvidia GPUs. The key difference: Jalapeño is designed specifically around OpenAI's own model architectures and inference patterns — the kernels, memory movement patterns, and serving workloads OpenAI runs every day across ChatGPT, Codex, and the API. Google's TPUs are more general-purpose across Google's diverse workloads. The 9-month development cycle (vs Google's multi-year TPU cycles) reflects how specifically Jalapeño is tuned to one company's exact inference needs.
Will Jalapeño make ChatGPT and Codex faster or cheaper for users?
Cheaper to serve — almost certainly yes, once at scale. Faster for users — possibly, though latency improvements depend on deployment configuration. The 50% cost reduction in inference is primarily an OpenAI margin story, not a "ChatGPT will feel faster tomorrow" story. The user-facing benefit materialises if OpenAI passes some of the savings on through lower API pricing — which becomes more likely as the IPO approaches and competitive pressure from Anthropic and Google intensifies.
Sources: OpenAI official announcement · Broadcom Globe Newswire · Bloomberg — 50% cost savings · CNBC · Related: OpenAI news hub 2026 · June 25 AI news digest · SpaceX acquires Cursor $60B