🧔♂️ A friendly human may check it before it goes live. More news here
OpenAI, Broadcom partner on 10 GW AI accelerator systems
OpenAI and Broadcom have announced a partnership to co-develop and deploy custom AI accelerator systems and Ethernet networking solutions, aiming to roll out racks with a total capacity of 10 gigawatts.
OpenAI, based in San Francisco, will design the accelerators and systems, while Broadcom, a semiconductor company headquartered in Palo Alto, will handle development and deployment.
The companies plan to begin deploying these systems in the second half of 2026, with completion targeted by the end of 2029.
The new racks will use Broadcom’s Ethernet and connectivity technology, and will be installed across OpenAI’s facilities and partner data centers.
OpenAI and Broadcom have previously worked together on AI accelerator co-development and supply agreements.
Executives from both companies described the collaboration as important for building large-scale AI infrastructure, but did not disclose financial terms.
🔗 Source: OpenAI
🧠 Food for thought
Implications, context, and why it matters.
OpenAI 10 GW plan faces supply gaps
- Missing pieces cover foundry choice with node and wafer allotment for the 2029 10 GW target 1. It also skips advanced packaging such as Chip-on-Wafer-on-Substrate (CoWoS) and High Bandwidth Memory (HBM) plans 1.
- TSMC (Taiwan Semiconductor Manufacturing Company) holds over 70% of foundry share 2. CoWoS remains tight at TSMC as Nvidia and AMD fill near-term slots, so OpenAI’s 3 nm chip for 2026 must vie for the same lines 34.
- Samsung’s “Stargate collaboration” adds HBM plus DRAM (dynamic random-access memory) supply for OpenAI, yet HBM still runs short as Nvidia and AMD lead allocations 23. OpenAI still buys Nvidia GPUs under a deal near 100 billion 3. Its custom silicon targets inference, while training remains on Nvidia 3.
Recent OpenAI developments
Stay updated on the go with our mobile app.
Get latest insights with smoother, more personalized experience through TIA mobile app.