Compute#AI chips#Model training
OpenAI compute chief Sachin Katti departs November 15, ex-xAI infra head Ruddarraju takes over
OpenAI compute chief Sachin Katti leaves Nov 15 for a new venture, staying as an advisor; ex-xAI infra head Uday Ruddarraju takes over compute.
Sachin Katti, the executive running OpenAI’s infrastructure and compute organization, announced on October 14 that he will step down on November 15 and stay on as an advisor, writing that he has “decided to move on to build something new.” An OpenAI spokesperson confirmed to CNBC that Uday Ruddarraju, the company’s CTO of compute, will take over the entire compute organization. Katti joined from Intel only last November. In his year in the job, OpenAI announced the 3.2 GW Camellia campus in Georgia and a roughly 8 GW-IT project in Ohio, and its first custom inference chip, Jalapeño, posted its first results in August. His exit is the latest in a string of executive departures at OpenAI this year.
Key points
- Katti’s last day is November 15, 2026, after which he remains an OpenAI advisor; his departure post names no next venture.
- Ruddarraju joined OpenAI in July 2025 from xAI, where he led infrastructure engineering and worked on the Grok 3 training run across 100,000 GPUs; he has been CTO of compute since July.
- Under Katti, OpenAI announced the 3.2 GW Camellia project in Georgia and the roughly 8 GW-IT PORTS-Pike campus in Ohio; the GPT-Infra business moves to chip chief Richard Ho.
Background
Newsrooms couldn’t agree on what to call Katti, and the timeline explains why. CNBC described him as “infrastructure and compute lead,” The Information and Unite.AI as VP of compute strategy, the Wall Street Journal as head of computing capacity. The roles shifted twice in three months: he joined OpenAI in November 2025 from Intel, where he was chief technology and AI officer, took over Stargate in March, according to Data Center Dynamics, and moved to the compute strategy vice president role around August, also picking up GPT-Infra.
His departure has company. In July, product and business chief Fidji Simo moved to an advisory role citing a severe exacerbation of a chronic illness; in August, longtime executive Brad Lightcap and revenue chief Denise Dresser left, as CNBC recounts. The compute line had its own drain: data center chief Chris Malone left in August, and Peter Hoeschele, central to Stargate, moved to Meta earlier this year with colleagues Keith Heyde and Shamez Hemani, per Data Center Dynamics. President Greg Brockman told CNBC at the time that because OpenAI is in the spotlight, “every departure gets scrutinized in a way that it doesn’t otherwise” — “I’m a constant, Sam [Altman] is a constant.”
The facts
The handover itself looks orderly. Katti works through November 15 and then advises; Bloomberg reported, via The Next Web, that the business leaders under him keep their jobs. The real change of hands is GPT-Infra, the effort Katti has run since July that applies OpenAI’s models to infrastructure work such as semiconductor design. He wrote that it went from an idea to a business with customers and partners in three months. It now moves to Richard Ho, the company’s chip chief — who said earlier this month that he does not fully trust the models’ design changes.
In his post, Katti described helping orchestrate “one of the largest AI compute infrastructure buildouts in history,” securing double-digit gigawatts of new capacity across projects he listed as Ohio, Portsmouth, Georgia and Trainium, with NVIDIA, Broadcom, AMD, Oracle, Microsoft, Google, AWS, Cerebras and CoreWeave among the partners. Three milestones have OpenAI’s own announcements behind them:
| Project | Announced | Scale | Partners and terms |
|---|---|---|---|
| Camellia, Georgia | 2026-07-22 | 3.2 GW, delivered in phases 2028–2032 | Georgia Power supply; OpenAI pledges no residential rate increases and $80M in community benefits |
| PORTS-Pike, Ohio | 2026-08-17 | ~8 GW-IT, first 800 MW expected 2028 | SB Energy builds and leases under a 20-year deal; NVIDIA invests $1.5B and backs the first 4.25 GW |
| Jalapeño inference chip | 2026-08-25 | 1.5–1.9x more AI work per watt at peak, per OpenAI | Tested on SemiAnalysis’ InferenceX; deployment planned by end of 2026 |
CNBC notes OpenAI is targeting roughly $600 billion in total compute spend by 2030, and the cloud partnerships keep widening — OpenAI already runs hosted agents on AWS. The custom inference chip race isn’t OpenAI’s alone either; Microsoft shipped its own Maia 300 earlier this month.
What Ruddarraju inherits
The compute organization Ruddarraju now runs owns OpenAI’s infrastructure roadmap — the standing guarantee that models can be trained and served. His record is operating training at scale: head of infrastructure engineering at xAI, where he worked on the Grok 3 run across 100,000 GPUs, before Elon Musk’s xAI was acquired by SpaceX earlier this year. He joined OpenAI in July 2025 and became CTO of compute this July; The Next Web points out he was the compute CTO OpenAI had already hired but declined to name during August’s data center shake-up.
He told Bloomberg that Katti has done “an incredible job maintaining the commercial partnerships,” expects a smooth handover, and will now spend more time on the commercial side himself. What he inherits is an almost entirely American buildout: OpenAI paused Stargate UK in April — British electricity prices run more than four times US levels and copyright rules remain unresolved, and the Guardian later reported OpenAI never visited the announced north-east England site. OpenAI’s only announced European site is a 230 MW data center in Norway; the Ohio project alone is about 35 times larger. Jalapeño is due to start deploying within OpenAI’s infrastructure by the end of 2026, with a second generation in development.
What others say
Katti’s departure post, quoted at length by CNBC and Unite.AI, is gracious: “Building AI infrastructure at this scale has given me an incredible view into every dimension of the compute world, from silicon and systems to datacenters, energy, supply chains, financing, and the economics of AI.” He closes: “I’m a builder at heart, and I’m looking forward to building something new.” Ruddarraju’s Bloomberg comments stress continuity, and Brockman has framed the broader departures as growing pains under scrutiny.
Two discrepancies are worth recording. The Information reported the announcement as an X post; CNBC and Unite.AI both quote it as a LinkedIn post, and this entry follows the latter because both outlets quote the text directly. And the sharpest caution comes from inside: Ho’s stated reservations about trusting model-made design changes suggest the GPT-Infra handover is not a formality.
Our take
For developers, nothing about API pricing or release cadence changes this month. Three things are worth watching instead: whether Jalapeño’s first deployments land by year-end as planned; whether GPT-Infra keeps its three-months-to-a-business pace under Ho; and whether a leader whose background is running giant training runs tilts OpenAI’s multi-gigawatt dealmaking toward securing training capacity first.
The honest catch: Katti lasted about a year and his role was reshaped twice in three months, in a department OpenAI expects to spend some $600 billion by 2030, per CNBC. At that scale, continuity on the supply side is itself a risk, and the outside environment is hardening — OpenAI, Blackstone and SoftBank formed a data center lobbying alliance earlier this month as data centers became a flashpoint in the midterm campaigns. What Katti builds next is, for now, his own “something new.”