OpenAI's Jalapeno chip takes aim at Nvidia dependency

Key Takeaways

• OpenAI's Jalapeno chip, designed with Broadcom, is its first custom silicon for AI inference workloads
• The chip reportedly matches Nvidia Blackwell and Google TPU performance, with deployment planned by end of 2025
• Engineers completed the design in nine months, using AI to accelerate parts of the chip development process

OpenAI has unveiled Jalapeno, its first custom AI chip, developed in partnership with Broadcom. The chip handles inference workloads, the computationally intensive process of answering user queries in ChatGPT and similar applications. OpenAI plans to deploy it by the end of 2025, marking the company's first serious move to reduce its dependence on Nvidia's GPUs.

The announcement signals a broader industry shift. Google, Amazon, and Meta have all invested heavily in custom silicon for AI workloads. OpenAI, which reportedly spends billions annually on Nvidia hardware, now joins that race.

What does Jalapeno actually do?

Jalapeno is an inference chip, not a training chip. That distinction matters. Training large language models requires massive computational resources over weeks or months. Inference is the ongoing cost: every time someone asks ChatGPT a question, the model runs inference to generate a response.

With over 300 million weekly ChatGPT users, inference costs dwarf training costs at OpenAI's scale. A chip optimized specifically for inference on OpenAI's model architectures could dramatically reduce per-query costs.

"It will be performant on, we think, all kind of future iterations of LLMs," OpenAI hardware chief Richard Ho told Reuters. The company already has samples running in its labs on the GPT-5.3-Codex-Spark AI model, hitting target power and performance specifications.

How does it compare to Nvidia and Google silicon?

Broadcom CEO Hock Tan made a bold claim: Jalapeno matches the performance of Nvidia's Blackwell chips and Google's tensor processing units. That's a high bar. Nvidia's B200 Blackwell chips represent the current state of the art for AI inference, and Google's TPUs have years of refinement behind them.

The comparison comes with caveats. OpenAI hasn't published benchmarks. "As good as" can mean many things, performance varies across workloads, and Tan has an obvious interest in promoting Broadcom's design capabilities to potential customers.

Still, even parity would be significant. Nvidia commands 80% of the AI chip market and charges accordingly. Any credible alternative gives OpenAI negotiating leverage and supply chain insurance.

Nine months from design to silicon

OpenAI's engineers completed the chip design in roughly nine months before sending it to TSMC for manufacturing. That's fast. Traditional chip design cycles run 18 to 24 months or longer.

The company credits AI tools for accelerating specific aspects of the process. Chip design increasingly uses machine learning for tasks like placement and routing, where algorithms can explore more design variations than human engineers. The irony is obvious: OpenAI used AI to build hardware that will run more AI.

Canadian electronics manufacturer Celestica will build the server systems that house Jalapeno. Both the chips and servers will be exclusive to OpenAI, not sold externally.

Why Broadcom, not in-house?

Building a chip design team from scratch takes years. Broadcom provides design services and intellectual property that would be difficult to replicate internally. Meta, Amazon, and Google have all used Broadcom or Marvell for similar partnerships.

The tradeoff: Broadcom's profit margins on custom AI chips are lower than on its other products, like networking switches. AI chips require large amounts of high-bandwidth memory from suppliers like SK Hynix and Samsung, which squeezes Broadcom's margins. That cost pressure eventually flows to customers like OpenAI.

Reuters first reported OpenAI was exploring custom chips back in 2023. The company's rival Anthropic is now weighing a similar move, according to Reuters sources from April 2025.

The compute shortage driving this decision

AI labs face a severe compute shortage. Nvidia's latest GPUs ship with waitlists extending months. OpenAI, Anthropic, and others compete for limited production capacity while demand for AI services keeps growing.

OpenAI's $500 billion Stargate infrastructure project, announced earlier this year, reflects the scale of compute investment required. Custom chips are one piece of that strategy. Owning the silicon design means OpenAI can optimize for its specific model architectures rather than using general-purpose GPUs.

Jalapeno is explicitly the "first step in a multi-generation chip development plan." Expect future chips to expand into training workloads if inference proves successful.

Frequently Asked Questions

When will OpenAI's Jalapeno chip be available?

OpenAI plans to deploy Jalapeno by the end of 2025. The chip will be used exclusively by OpenAI and won't be sold to other companies.

What is the difference between AI training and inference chips?

Training chips process massive datasets over weeks to create AI models. Inference chips run those trained models to answer user queries in real time. Inference happens constantly and represents the larger ongoing cost at scale.

Will OpenAI stop using Nvidia GPUs?

Not immediately. Jalapeno handles inference only. OpenAI still needs Nvidia GPUs for training and likely will for years. Custom chips reduce dependence but don't eliminate it.

Who manufactures the Jalapeno chip?

TSMC in Taiwan manufactures the silicon. Celestica builds the server systems. Broadcom provided chip design services and intellectual property.

Is Anthropic also building custom AI chips?

Reuters reported in April 2025 that Anthropic is weighing building its own AI chip. No public announcement has been made.

Source: Tech-Economic Times / ET