OpenAI unveiled Jalapeño, its first custom inference ASIC built with Broadcom on TSMC's 3nm process — and it signals a structural shift in who controls the cost of AI.
On June 24, 2026, OpenAI unveiled Jalapeño, its first custom-built AI inference processor. Developed in partnership with Broadcom and manufactured by TSMC on its 3nm process node, the chip moved from initial design to manufacturing tape-out in nine months — a timeline that OpenAI achieved in part by using its own AI models to accelerate chip design and verification, according to reporting by Futurum Group and Tom's Hardware.
The chip is inference-only. It will not train models. That distinction matters enormously for understanding both the technical architecture and the strategic logic behind it.
According to Tom's Hardware's technical coverage, Jalapeño is a reticle-sized ASIC incorporating HBM3E memory. Its architecture prioritizes data movement optimization — balancing compute, memory bandwidth, and networking resources specifically for the inference workload profile rather than the training workload profile, which has different computational characteristics.
Engineering samples are already running workloads in OpenAI's labs, including GPT-5.3-Codex-Spark, according to the OpenAI announcement and corroborated by coverage from AI Weekly. Full deployment at gigawatt scale, in partnership with Microsoft and other data center partners, is targeted for late 2026.
The performance claim is specific: Broadcom and OpenAI state that Jalapeño delivers approximately 50% lower inference cost per token compared to current-generation AI GPUs, according to Forbes and CRN Asia. That figure applies to the workloads Jalapeño is designed for. It says nothing about training — because the chip was never designed to train anything.
Training gets the press coverage. Inference is where the money goes.
Training a frontier model is a one-time or infrequent cost. Running that model to serve millions of requests per day is continuous. For a company at OpenAI's scale — with ChatGPT serving hundreds of millions of users — inference compute costs are the dominant operational expense, not training.
The dependency on NVIDIA GPUs for inference is expensive not because NVIDIA's chips are poorly designed for inference, but because they are designed for both training and inference. They carry compute capabilities and memory architectures that an inference-only workload does not use. An ASIC that strips away everything irrelevant to inference and optimizes the remaining components for that specific workload can achieve significantly better cost efficiency — which is precisely what Jalapeño claims to deliver.
This is not a new insight. Google deployed its first TPUs (Tensor Processing Units) in 2015, specifically for inference. Amazon developed its Inferentia chip for AWS inference workloads, with the first generation shipping in 2019. Microsoft has developed its Maia AI accelerator. Meta has its MTIA chip. The pattern is clear: any company running AI workloads at sufficient scale eventually finds that general-purpose GPUs are the wrong tool for inference, and builds something more specific.
OpenAI is arriving at this inflection point later than its hyperscaler peers — but its nine-month tape-out timeline, if accurate, suggests the chip design process itself is being compressed by AI-assisted design tooling.
OpenAI's choice of Broadcom as its design partner, and Celestica for board and rack system integration, reflects the reality of custom silicon development: building a chip requires not just chip architects but a full supply chain of expertise OpenAI does not have internally.
Broadcom has become the dominant partner for hyperscalers pursuing custom ASICs. Google's TPUs, Meta's MTIA, and now OpenAI's Jalapeño all route through Broadcom's custom ASIC business. This concentration creates its own dependency — but Broadcom's experience with AI ASIC design at scale is currently unmatched outside of TSMC itself.
TSMC's 3nm process is the relevant manufacturing node. It is the same node used for Apple's M3 and A17 Pro chips, and represents the current leading edge of volume production. The fact that OpenAI secured TSMC 3nm capacity for Jalapeño — in an environment where leading-edge capacity is heavily contested — suggests either significant pre-committed purchase volumes or priority access negotiated as part of the broader Microsoft-OpenAI infrastructure relationship.
[UNCERTAIN] The specific nature of the capacity agreement between OpenAI, Microsoft, and TSMC has not been disclosed publicly.
NVIDIA's revenue dependency on hyperscaler AI customers is not a secret. Its data center segment generated $47.5 billion in revenue in fiscal Q1 2026, according to NVIDIA's own earnings report. The bulk of that is driven by demand from exactly the companies — Microsoft, Google, Meta, Amazon, and now OpenAI — that are simultaneously building custom silicon to reduce their NVIDIA spend.
The Jalapeño announcement does not threaten NVIDIA's near-term position. The chip is not yet in production, targets inference only, and will initially cover only a portion of OpenAI's inference workload. NVIDIA remains the dominant supplier for AI training, where no major competitor has emerged.
The structural threat is in the medium term. As each major AI consumer brings inference workloads onto custom silicon, the addressable market for NVIDIA's GPU-based inference products shrinks. The training market will sustain NVIDIA's revenue for years. But the inference market — which scales proportionally with AI usage — is being systematically captured by custom ASICs.
[OPINION] The pace at which this happens depends on how quickly inference ASICs can be designed, validated, and deployed at scale. Nine months from design to tape-out is fast. The validation and integration phases that follow are slower and harder. OpenAI's late 2026 deployment target for gigawatt-scale production is aggressive, and it is more likely to slip than to accelerate.
The following represents the author's analysis and should not be taken as financial or investment advice.
The Jalapeño announcement matters less as a product launch and more as a confirmation that the inference silicon market is now fully contested. Every major AI lab and hyperscaler either has or is building a custom inference ASIC. The era of general-purpose GPU dominance in inference is ending — not because GPUs are bad, but because the economics of purpose-built silicon are impossible to ignore at scale.
[OPINION] For OpenAI specifically, this is a necessary move toward operational independence. Its current inference cost structure is entirely at NVIDIA's pricing discretion. Every dollar of inference cost reduction through Jalapeño is a dollar that can be reinvested in capacity, model improvement, or margin. At the usage scale OpenAI operates, a 50% cost reduction in inference is not a feature — it is a structural shift in the company's financial model.
The nine-month design cycle is the most technically interesting claim in the announcement. If AI-assisted chip design genuinely compresses development cycles from years to months, the barriers to custom silicon entry fall dramatically. Companies that could not previously justify the capital and engineering investment in custom chips may find themselves reconsidering that calculation within the next 18 months.
That is the longer-term implication: Jalapeño is not the end of NVIDIA's inference business. It is the beginning of a market structure where inference silicon is no longer dominated by a single general-purpose architecture.
