OpenAI and Broadcom have announced a joint program to create 10 GW of custom AI accelerators and Ethernet‑based networking systems, aiming to power next‑generation AI clusters by 2029.
Custom AI Accelerator Design
The partnership focuses on chips that embed lessons from large language model training directly into hardware. This approach seeks to improve latency and energy use for models such as large language models.
- OpenAI‑designed compute cores optimized for transformer workloads
- Integrated memory hierarchy to reduce data movement overhead
- Programmable interfaces for rapid model iteration
- Security features to guard against malicious AI extensions (internal guide)
- Compatibility with existing GPU and ASIC ecosystems
Ethernet‑Based Scale‑Up & Scale‑Out Networking
Broadcom supplies Ethernet, PCIe, and optical solutions that connect accelerator racks in both scale‑up (single‑rack) and scale‑out (multi‑rack) configurations. The design relies on open standards to keep integration straightforward.
- Multi‑terabit Ethernet fabric supporting low‑latency traffic
- PCIe Gen5 lanes for high‑throughput accelerator attachment
- Optical links for inter‑rack communication beyond 100 m
- Dynamic bandwidth allocation based on workload demand
- Monitoring tools integrated with OpenAI’s management platform
Power and Energy Planning (10 GW Target)
Deploying 10 GW of accelerators requires careful power distribution and cooling strategies. Both companies plan to use energy‑aware scheduling to match compute load with available capacity.
- Power budgeting per rack with 10 kW ceiling
- Liquid cooling loops for high‑density modules
- Renewable energy sourcing options for data‑center sites
- Real‑time energy monitoring dashboards
- Fail‑over mechanisms to maintain operation during power spikes
Deployment Timeline and Ecosystem Impact
The rollout begins in the second half of 2026 and aims for full deployment by the end of 2029. Early installations will appear in OpenAI’s own facilities and select partner data centers.
- 2026 H2: Prototype rack validation and performance testing
- 2027 Q1–Q3: Pilot deployments with partner cloud providers
- 2028 Full‑scale production of accelerator modules
- 2029 Completion of global rollout across major regions
- Implications for AI model development outlined in Choosing the Right AI Model