On June 24, OpenAI and chip giant Broadcom jointly released its first self-developed AI inference chip——Jalapeño. The debut of this chip marks OpenAI's official entry into the "self-developed silicon" track, taking a key step to reduce dependence on NVIDIA GPUs.
What is Jalapeño?
Jalapeño is aASIC chip designed for large language model inference(Application Specific Integrated Circuit). Unlike general-purpose GPUs, ASICs are optimized for specific workloads and can achieve higher throughput and lower latency with the same power consumption. OpenAI positions it as an "Intelligence Processor" and emphasizes that this is not a general-purpose chip, but a dedicated hardware designed from scratch for AI inference scenarios.
Greg Brockman, president of OpenAI, revealed that the chip hasIt only took 9 months from architectural design to tape-out.——This speed is extremely rare in the semiconductor industry. The secret is: OpenAI uses its own AI model to assist the chip design process, forming a positive feedback loop of "using AI to design AI chips".
Why do we need to develop self-developed chips?
The answer is simple:cost and controllability. OpenAI disclosed in its 2025 audited financial report that its annual operating costs are as high as US$34 billion, of which inference computing power costs account for the bulk. Processed onceChatGPTRequests require considerable GPU computing resources. Self-developed inference chips can:
- Dramatically reduce reasoning costs:Specialized chips far outperform general-purpose GPUs in performance per watt. OpenAI officials stated that Jalapeño’s “performance per watt was significantly better than the existing state-of-the-art solutions” in early testing.
- Get rid of supply chain dependence: NVIDIA GPU supply has been tight for a long time and prices are high. Self-developed chips give OpenAI a "spare tire."
- Full stack optimization: Collaborative design can be achieved from model architecture to underlying hardware. Greg Brockman emphasized that OpenAI "operates the entire technology stack" - models, products, data centers, chips - and each layer can be optimized for the same goal.
Division of labor and cooperation
The roles of each party in this collaboration are clear:OpenAIResponsible for underlying architecture design and AI-assisted optimization;Broadcom(Broadcom) Responsible for silicon implementation and network hardware; Canadian electronics manufacturing services providerCelesticaResponsible for circuit board and rack system integration.
The cooperation between OpenAI and Broadcom will be officially announced in October 2025, but the previous internal research and development had lasted for 18 months. As one of the world's largest ASIC design service providers (it once designed TPU for Google), Broadcom provides OpenAI with mature engineering capability support.
Impact on industry structure
The emergence of Jalapeño has added a heavyweight member to the "self-developed AI chip club". Currently this club includes:
- Google: TPU series, has been iterated to the sixth generation
- Amazon: Trainium and Inferentia series
- Microsoft：Maia series
- OpenAI(New): Jalapeño, focused on reasoning
For NVIDIA, although its GPU is still "irreplaceable" in the field of AI training, in the inference market - the largest computing power consumption scenario for future AI applications - self-developed chips are eating away at share from all directions. The addition of Jalapeño makes this trend even more irreversible.
Summarize
Although Jalapeño is still in the engineering sample testing stage (running inference workloads of models such as GPT-5.3, Codex and Spark), its significance goes far beyond the chip itself. It is OpenAI's declaration to build full-stack control "from sand to application" - no longer satisfied with just training models, but to control every layer of AI infrastructure. If Jalapeño performs as expected, OpenAI will have a compelling cost-optimization story for its upcoming IPO.
📊 Follow AI Dash, learn about the latest developments in AI tools and infrastructure.