OpenAI x Broadcom — The OpenAI Podcast Ep. 8
OpenAI's Sam Altman and Greg Brockman, alongside Broadcom’s Hock Tan and Charlie Kawwas, announce a strategic partnership to co-design custom AI chips and vertically integrated systems. This collaboration aims to build the next generation of AI infrastructure, starting with a 10-gigawatt deployment, to meet the world's soaring demand for intelligence.
The OpenAI-Broadcom Partnership: A New Era in AI Infrastructure
OpenAI and Broadcom have announced a significant partnership, 18 months in the making, to co-develop custom AI chips and, more recently, entire custom systems. This collaboration marks a pivotal step in building the next generation of AI infrastructure. The initial goal is to deploy 10 gigawatts of these new, fully integrated systems starting in late 2025. Sam Altman frames this endeavor as potentially the “biggest joint industrial project in human history,” aiming to create what Charlie Kawwas calls “civilization’s next-generation operating system.”
The Rationale: Full-Stack Optimization for Unprecedented Scale
The core of the partnership is vertical integration, enabling optimization across the entire technology stack. This approach allows the teams to design everything from the silicon level—“from etching the transistors”—all the way up to the final model output—“the token that comes out.”
This holistic design philosophy, encompassing the chip, racks, networking, and algorithms, is expected to yield massive efficiency gains. The primary goal is to maximize the “intelligence per watt” that can be extracted from each unit of energy. By controlling the entire pipeline, OpenAI can deliver:
- Faster Models: Reduced latency for user interactions.
- Cheaper Models: Lowering the cost per token, making advanced AI more accessible.
- Better Performance: Overall improvements in model capability and efficiency.
This increase in performance and reduction in cost is crucial because, as OpenAI has consistently observed, any 10x optimization in efficiency is met with 20x more demand. The 10-gigawatt deployment is an incremental capacity, built on top of existing infrastructure, yet it’s still considered a “drop in the bucket” compared to the future needs for achieving Artificial General Intelligence (AGI).
Custom Silicon for Specific Workloads
A key motivation for designing custom chips is the ability to tailor hardware for specific AI workloads, moving beyond general-purpose hardware. While GPUs are lauded for their flexibility in research, a system highly optimized for a known workload can achieve superior performance and efficiency.
Hock Tan explains the distinction:
- Training Chips: Require immense computing capacity (TFLOPs) and robust networking for large-scale clusters.
- Inference Chips: Benefit more from increased memory and optimized memory access relative to raw compute.
By creating chips specifically for their inference needs, OpenAI can “control its own destiny” and build the most effective end-to-end platform for its models. This was a lesson learned from past frustrations where, as Greg Brockman notes, chip startups often didn’t listen to feedback about the future direction of AI models.
AI-Accelerated Chip Design
A unique aspect of this collaboration is OpenAI’s use of its own AI models to accelerate the chip design process. This meta-level application of AI has already yielded significant results:
- Massive Area Reductions: AI models are poured onto human-optimized components to find novel optimizations.
- Accelerated Timelines: The models can identify design improvements that would have taken human experts months to discover, effectively compressing the development schedule.
Greg Brockman shared an anecdote where the team, facing a deadline, chose to let the AI continue running optimizations without human review, trusting the process to deliver superior results. This demonstrates a future where AI not only powers applications but also becomes an indispensable partner in creating the hardware it runs on.
The Long-Term Vision: Compute as a Global Utility
The speakers draw parallels between this AI infrastructure build-out and historical transformative projects like the railroad and the internet. The ultimate vision is to establish AI compute as a “critical utility” accessible to all 8 billion people on the planet.
This journey is directly tied to OpenAI’s mission. The company’s focus shifted dramatically towards compute after empirically discovering in 2017 that scale was the most effective driver of progress in AI capabilities. The goal is to move from the current state of “compute scarcity,” where even internal teams at OpenAI are constrained, to a future of “compute abundance,” where anyone with an idea has the power to build it.
Technological Roadmap and Timeline
The partnership is moving quickly, with the first silicon expected back soon and deployments beginning in late 2025. Broadcom brings a forward-looking technological roadmap to the table, with innovations that will continue to push the boundaries of compute:
- Advanced Nodes: Moving from 2-nanometer processes to even smaller geometries.
- 3D Stacking: Increasing density and performance by stacking components vertically.
- Integrated Optics: Announcing 100-terabit switching with optics integrated directly into the chip, enabling massive clusters with unprecedented performance and power efficiency.
This collaboration is not just about a single chip but about building a sustained foundation for the future of AI, enabling the development of GPT-5, 6, 7, and beyond on increasingly advanced, purpose-built hardware.