OpenAI and Broadcom Unveil "Jalapeño" AI Inference Chip to Diversify Silicon
Hyperscalers like OpenAI are diversifying hardware with custom chips, while major power deals and skilled labor shortages shape the ongoing AI infrastructure build-out.
The story
OpenAI has officially unveiled its first custom AI chip, named "Jalapeño," developed in collaboration with Broadcom. The application-specific integrated circuit (ASIC) is designed specifically for large language model (LLM) inference workloads, a strategic move to optimize performance and reduce OpenAI's reliance on general-purpose GPUs from companies like Nvidia.
Deployment of the Jalapeño chips is targeted for the end of 2026, with an anticipated power requirement of 10 gigawatts for future operations. This venture into custom silicon allows OpenAI to gain greater control over its infrastructure stack, aiming for improved efficiency and cost-effectiveness in running its AI models, such as ChatGPT.
Broadcom's CEO, Hock Tan, stated that Jalapeño performs on par with Nvidia's Blackwell GPUs and Google's TPUs for relevant workloads, though independent benchmarks are still pending. This development signals a growing trend among major AI players to invest in purpose-built hardware to meet the escalating demands of AI inference, potentially reshaping the competitive landscape of the AI chip market.
Silicon
Jalapeño
Maker: OpenAI (with Broadcom)
What: Custom ASIC designed for LLM inference workloads, aiming to match performance of Nvidia Blackwell and Google TPUs.
For Whom: OpenAI for its AI models like ChatGPT, to enhance speed, reliability, and accessibility of advanced AI.
Dragonfly C1000 CPU
Maker: Qualcomm
What: Chiplet design with over 250 cores, optimized for power efficiency and AI agent workloads, offering more than 2x better performance per watt compared to existing server CPUs.
For Whom: Data centers, with Meta signing a multi-generational agreement to deploy these technologies.
The build-out
Project Kilby
Who: Microsoft and Chevron (via Energy Forge One LLC and Joulent)
Scale: Approximately 2.67 GW of natural gas generating capacity for a co-located power facility and data center complex.
Where: Reeves County, West Texas, Permian Basin.
Supply & policy signals
TSMC Chairman C.C. Wei warned that shortages of skilled workers and water resources in Taiwan are becoming significant constraints for semiconductor capacity.
Implication: These resource pressures threaten future wafer capacity growth at the world's largest contract chipmaker, impacting manufacturing schedules and long-term supply planning.
A coalition of telecom, automotive, medical device, and retail groups warned the Trump administration that surging AI data center demand is driving sharp increases in memory prices.
Implication: Expanding AI data centers are consuming a growing share of memory production capacity, leading to reduced supply and higher costs for non-AI industries.
Nvidia's share of the data center Ethernet switching market climbed to 21.5% in Q1 2026, up from less than 4% two years prior.
Implication: This growth highlights a shift where networking is increasingly sold as part of a tightly integrated GPU-plus-networking package optimized for AI factory workloads, reshaping market competition.
What we'll be watching
- Qualcomm's commercial sampling of Dragonfly C1000 CPU racks in FY26 and High Bandwidth Compute (HBC) Gen 1 with AI250 in mid-2027.
- Chevron's Final Investment Decision for Project Kilby, expected by the end of 2026.
- OpenAI's initial deployment of its Jalapeño chips by the end of 2026.
- NVIDIA GTC Berlin conference, scheduled for October.
- Intel's next earnings update, anticipated on July 23.
Reporting + analyst voices: grounded via Google Search at publish time.