Jul 16, 2026 · 44m · mad
OpenAI’s Compute Chief: We Can’t Build Fast Enough | Sachin Katti
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, OpenAI's Head of Compute Infrastructure Sachin Katti discusses the massive physical, financial, and technological engineering required to build and power gigawatt-scale AI data centers. He details OpenAI's multi-cloud strategy, custom silicon co-design, energy grid integration, and the physical constraints facing global AI deployment.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 29.9% of the talking time here. How this is scored →
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sachin directly rejects the host's framing of potential overbuilding risk, clarifying that demand far outstrips supply and that physical supply chain delays are his actual operational paranoia.
Hardest push from Matt ▶ 14:59 Challenging overbuilding and demand lagMatt directly challenges Sachin on the risk of overbuilding compute infrastructure, pressing on the inherent delay between constructing physical datacenters and actual demand realization.
Biggest teaching moment ▶ 14:16 Educating on inference versus trainingSachin reframes the host's clear-cut separation between training and inference workloads, explaining that synthetic data generation, post-training, and test-time compute mean much of modern training is fundamentally inference.
Matt holds his own ▶ 28:05 Detailing multi-billion dollar compute ecosystemMatt demonstrates notable expertise by citing specific deal figures, including a $20 billion arrangement with Cerebras, and pressing Sachin on how Stargate fits into OpenAI's broader partner landscape.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Historic Scale and Economics of AI Infrastructure | 4 | 3 | 1 | 1 | Matt cites specific industry spend metrics like $50 billion for OpenAI and $700 billion across AI compute. Sachin confirms the scale of investments and notes the rapid pace of decision-making required inside OpenAI. | |
| OpenAI's Evolution Into Direct Infrastructure Construction | 3 | 5 | 2 | 2 | Matt asks if OpenAI building datacenters represents a corporate pivot or new line of business. Sachin reframes this as developing a new organizational muscle, explaining the physical reality of liquid-cooled datacenters converting electrons to tokens. | |
| Power Grid Integration, On-Site Energy, and Nuclear Energy | 4 | 5 | 1 | 1 | Matt connects their Paris location to France's nuclear infrastructure and asks about datacenter power sources. Sachin explains behind-the-meter gas generation and OpenAI's commitments to fund local grid generation and transmission. | |
| OpenAI's Custom Silicon and the Jalapeño Inference Chip | 4 | 6 | 2 | 2 | Matt prompts Sachin on the strategy behind custom silicon like Jalapeno and whether inference compute now dominates training. Sachin educates Matt on tokens-per-watt efficiency and reframes how post-training synthetic generation blurs the line between training and inference. | |
| Overbuilding Risks vs. Physical Supply Chain Constraints | 4 | 6 | 4 | 3 | Matt pushes on the potential risk of datacenter overbuilding due to long construction lead times. Sachin firmly counters that demand far outstrips supply and clarifies that his main paranoia is physical supply chain bottlenecks hindering expansion. | |
| Local Economic Impact and Closed-Loop Water Recycling | 4 | 5 | 2 | 2 | Matt brings up public relations challenges regarding local community impact and water consumption myths. Sachin explains that closed-loop cooling systems consume minimal net water and highlights tax base and infrastructure benefits for rural areas. | |
| Site Selection Criteria for Next-Generation Data Centers | 3 | 5 | 1 | 1 | Matt asks about datacenter site selection criteria and Sachin's role as Head of Industrial Compute. Sachin outlines key location criteria and describes the internal challenge of allocating scarce compute capacity among competing teams. | |
| Sachin Katti's Career Path Across Academia, Startups, and Industry | 2 | 3 | 1 | 1 | Matt asks Sachin about his unique career path across academia, startups, and major tech firms. Sachin highlights his background as a Stanford professor and Intel CTO, noting how OpenAI uniquely merges research lab, startup, and scale. | |
| OpenAI's Infrastructure Partnerships and Project Stargate | 5 | 5 | 3 | 3 | Matt demonstrates deep industry awareness by naming various compute partnerships and questioning if Project Stargate evolved. Sachin clarifies that Stargate serves as an overall umbrella strategy rather than a single joint venture. | |
| The Abilene Texas Supercluster and Expansion Sites | 4 | 5 | 2 | 2 | Matt inquires about the Abilene supercluster and how datacenter buildouts are financed. Sachin explains that OpenAI operates as an off-take tenant while cloud and infrastructure partners handle the capital and construction. | |
| Accelerated Chip Development and AI-Assisted Hardware Design | 4 | 5 | 1 | 1 | Matt notes the remarkably fast nine-month tapout timeline for the Jalapeno chip. Sachin attributes this speed to experienced talent, Broadcom partnership, knowing workload specifications, and using AI tools to accelerate hardware design. | |
| Interconnect Reliability and the Open-Source MRC Protocol | 3 | 6 | 1 | 1 | Matt asks about the newly open-sourced MRC networking protocol. Sachin uses a city traffic routing analogy to explain how multi-path packet spraying ensures fault tolerance across massive 100,000 GPU clusters. | |
| Industrial Bottlenecks, Hardware Lead Times, and Labor Demands | 4 | 6 | 2 | 2 | Matt brings up hardware bottlenecks like memory and electrician shortages, along with OpenAI's new guaranteed capacity offering. Sachin details long lead times for physical equipment and recontextualizes guaranteed capacity as securing enterprise AI supply. |