OpenAI
2026-06-26
OpenAI and Broadcom Tape Out First Inference ASIC Jalapeño in 9 Months, Targeting NVIDIA Dominance
Summary
Key Takeaways
On June 24, 2026, OpenAI and Broadcom unveiled their first custom ASIC, Jalapeño, designed for LLM inference. The chip, built on a dedicated ASIC architecture, achieved tape-out in just 9 months, an industry record. OpenAI handled the architecture design, Broadcom the silicon implementation and networking hardware, and Celestica the board and rack integration. Engineering samples have successfully run OpenAI's GPT-5.3-Codex-Spark model at production-grade frequency and power. Fabricated on TSMC's 3nm process, Jalapeño is optimized for Transformer inference, targeting a 50% cost reduction. Mass deployment, paired with gigawatt-scale data centers, is slated for late 2026. This move signals OpenAI's strategic transformation from a pure model company to a full-stack AI infrastructure provider, following Google TPU, AWS Trainium, and Anthropic. The chip leverages Broadcom's ASIC expertise and its unique advantages in high-speed Ethernet interconnects for AI clusters, bypassing NVIDIA's GPU route.
Why It Matters
This is a control plane shift from NVIDIA's GPU ecosystem to a Broadcom-centric ASIC+Ethernet stack. OpenAI's Jalapeño aims to defend against NVIDIA's pricing power and CUDA lock-in by moving inference clusters from proprietary NVLink/NVSwitch to Broadcom's Tomahawk/Jericho Ethernet fabric, likely leveraging RoCEv2 or IPU for control. However, the narrative hides two critical engineering traps. First, ASIC inflexibility: Jalapeño is hardwired for Transformer inference. If future model architectures (e.g., Mamba, MoE variants) diverge, the chip risks rapid architectural depreciation, unlike NVIDIA's programmable GPUs. Second, Broadcom's hidden lock-in: Replacing NVIDIA's NVLink with Broadcom's Ethernet stack merely substitutes one vendor dependency for another. Broadcom's Tomahawk 5 switches and PCIe Gen5/CXL interconnect will deeply bind OpenAI's custom NICs, creating a new proprietary silo. This is not an open revolution but a shift of control from one silicon giant to another.
PRO Decision
【Vendors】 Competitors (NVIDIA, AMD, Intel) should aggressively highlight Jalapeño's ASIC inflexibility. Emphasize NVIDIA's CUDA programmability and architecture evolution (Hopper to Blackwell), which can adapt to future non-Transformer models (e.g., Mamba, MoE variants), while Jalapeño is hardwired. NVIDIA should accelerate low-cost inference solutions (e.g., L40S, GH200) and benchmark tail latency advantages of NVLink domains over Broadcom's Ethernet fabric for large-scale distributed inference.
【Enterprises】 CIOs and architects must perform a zero-trust audit. Beware of Broadcom's hidden lock-in: assess if Jalapeño clusters mandate Broadcom's Tomahawk switches and custom NICs, blocking alternatives like Arista 7800R or Cisco Silicon One. Demand a clear cross-architecture portability path from OpenAI, including support for standard frameworks (e.g., PyTorch/TensorRT-LLM) and explicit depreciation risk if model architectures shift.
【Investors】 See through the PR. Jalapeño won't disrupt NVIDIA's revenue short-term, as it serves only OpenAI and carries high ASIC NRE costs and 3nm tape-out expenses. Long-term, the TCO inflection point for ASIC vs. GPU hinges on model architecture stability. If Transformers are replaced, Jalapeño becomes a stranded asset. Focus on Broadcom's Tomahawk 5/Jericho 3 switch penetration in AI clusters, which offers greater investment value than Jalapeño itself.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)