OpenAI x Broadcom — The OpenAI Podcast Ep. 8

OpenAI and Broadcom are teaming up to design our own chips—bringing lessons from building frontier models straight into the hardware. In partnership with Broadcom and alongside our other partners, we’re creating the next generation of AI infrastructure to meet the world’s growing demand. In this episode, OpenAI’s Sam Altman and Greg Brockman sit down with Broadcom’s Hock Tan and Charlie Kawwas to announce a new partnership focused on custom AI chips and systems that could redefine what’s possible in computing. Chapters 00:00 Announcing the partnership 03:06 The scale of AI infrastructure 06:03 Collaboration and innovation in chip design 08:49 Historical context and future vision 12:10 Role of compute in AI development 15:01 Optimizing for specific workloads 18:02 Journey towards AGI 21:00 Future of AI and compute capacity 23:50 Wrap-up and future projects

Andrew MaynehostSam AltmanguestHock TanguestGreg BrockmanguestCharlie Kawwasguest

Oct 12, 202528mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

OpenAI and Broadcom build custom chips for massive AI scaling

OpenAI and Broadcom announce a partnership to co-design a custom chip and an integrated system (chip, racks, networking, and software) optimized specifically for OpenAI’s AI workloads.
They plan to begin deploying roughly 10 gigawatts of additional data-center capacity starting late next year, rolling out rapidly over the following three years—on top of existing infrastructure partnerships.
A central thesis is that end-to-end vertical integration can increase “intelligence per watt,” lowering cost per token and unlocking new products (e.g., always-on personal agents) that would otherwise be compute-prohibitive.
Speakers frame AI infrastructure as a civilization-scale utility requiring global collaboration, open standards, and continuous specialization of hardware for distinct workloads like training vs. inference.

IDEAS WORTH REMEMBERING

5 ideas

They are building a full-stack “transistor-to-token” platform, not just a chip.

Altman emphasizes optimizing across chip design, rack architecture, networking, and algorithms to gain major efficiency improvements that translate into faster, cheaper inference and better product performance.

10 gigawatts is enormous—and still insufficient for the long-term vision.

The group positions 10 GW as a major near-term expansion but “a drop in the bucket” relative to future demand if AI becomes an always-available utility for billions of people and increasingly capable agentic systems.

Inference demand is expected to expand faster than efficiency gains.

Altman notes a repeated pattern: a 10× optimization can trigger 20× demand, implying that cost reductions and latency improvements will be rapidly absorbed by new use cases (code, video, automation, agents).

Specialized silicon will diverge by workload: training vs inference needs differ.

Tan highlights that training favors high compute throughput (TFLOPS) and networking for clustered scaling, while inference often benefits more from memory capacity and bandwidth per unit compute.

AI is already helping design the next generation of AI hardware.

Brockman describes using OpenAI models to propose optimizations and reduce chip area and schedule risk—often surfacing expert-known ideas faster, enabling teams to keep iterating up to deadlines.

WORDS WORTH SAVING

5 quotes

“[The AI infrastructure build-out] is the biggest joint industrial project in human history.”

— Sam Altman

“We’re defining civilization’s next generation operating system.”

— Hock Tan

“Ten gigawatts… is a drop in the bucket compared to where we need to go.”

— Greg Brockman

“[Think] from… etching the transistors all the way up to the token that comes out when you ask ChatGPT a question.”

— Sam Altman

“What we want is the most intelligence we can get out of each unit of energy.”

— Sam Altman

Partnership announcement and timeline10-gigawatt deployment scaleVertical integration: transistor-to-token optimizationWorkload-specific chip design: training vs inferenceNetworking and systems design at cluster scaleAI-assisted chip optimization and design accelerationCompute abundance vs compute scarcity on path to AGI

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.