OpenAI ramps global AI infrastructure with new exec

OpenAI hired Sachin Katti to lead industrial compute as it seeks rapid global expansion of data center capacity amid memory chip shortages, power grid limits, and local opposition.

Katti, ex-Intel CTO and Stanford professor, will oversee OpenAI’s industrial compute teams.

OpenAI said its computing capacity more than tripled in 2025 to about 1.9 gigawatts, and Sam Altman, CEO, said he aims to produce a gigawatt of new AI infrastructure every week.

The company is expanding data centers across the US, Europe, Middle East, and South America.

OpenAI has multibillion-dollar deals with Cerebras, AWS, and Broadcom for AI chips.

OpenAI still depends on Nvidia for training and plans to deploy about 1 gigawatt using Nvidia’s forthcoming Vera Rubin AI accelerators later this year.

Katti said OpenAI is cautious about supply chains and geopolitics while expanding.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

OpenAI’s Texas pivot follows future chip roadmaps

OpenAI skipped leasing the planned Abilene, Texas data center expansion to match new buildouts with next-generation chip schedules 1.
Power for the expansion was not expected to be ready for another year, and OpenAI hopes to roll out Nvidia’s Vera Rubin chips by then instead of the current Blackwell generation 1.
Nvidia says the Vera Rubin platform delivers 10 times the performance per watt of Grace Blackwell 2.
OpenAI says it has “more than half a dozen” sites in development across several US states, including a location it is building with Oracle in Wisconsin 1.

OpenAI’s multi-vendor plan tracks a squeeze in advanced packaging

OpenAI’s supplier mix responds to a bottleneck in CoWoS (Chip-on-Wafer-on-Substrate), an advanced chip packaging technology that connects high-performance chips with memory at very high bandwidth 3.
In 2026, TSMC (Taiwan Semiconductor Manufacturing Co.), the world’s largest contract chipmaker, is expected to reach total effective CoWoS output of about 1.15 million wafers, according to one analysis 3.
That analysis estimates Nvidia could get nearly 60% of that output 3.
OpenAI orders custom chips through Broadcom and buys AI-compute capacity from Amazon Web Services (AWS), though it remains unclear how, or whether, those deals secure access to partners’ separate CoWoS allocations 3.