According to multiple people familiar with the matter, three executives who participated in the launch of OpenAI's first "Stargate" data center project have resigned and may leave the ChatGPT development company in the near future. One of the people familiar with the matter said that Peter Hoeschele, an OpenAI executive who played a key role in promoting the implementation of "Stargate", has resigned.
Two other sources said that Shamez Hemani, who is responsible for computing strategy and business development, and Anuj Saharan, another head of OpenAI’s computing department, have also announced their resignations to colleagues.
A person familiar with the matter said that the three departing executives will join the same new company, but the name of the company has not yet been determined.
The infrastructure team has always been at the core of the biggest challenge faced by OpenAI and its competitors: how to obtain sufficient computing power to meet the skyrocketing demand for AI products at a cost that startups can afford. OpenAI has stated that it plans to invest more than US$600 billion in computing power construction in the next five years; this spending plan has prompted it to reach cooperation agreements with a number of chip and cloud service giants such as Oracle and Amazon Cloud Technology.
Before the departure of this executive, OpenAI had completely restructured its infrastructure business. At the end of last year, the company hired former Intel executive Sachin Katti as head of computing power and infrastructure.
An OpenAI spokesperson said in a statement: "We thank Peter, Shameez and Anuj for their contributions to OpenAI and wish them all the best in the future. Sachin Karty has recently joined and is responsible for leading the company's industrial computing department, which is steadily advancing its work in the process of building the required infrastructure for the next generation of AI systems at scale."
Last year, the team finalized a cooperation agreement for 8 gigawatts of computing power capacity in the next few years, falling short of the 10 gigawatt target set when the $500 billion "Stargate" was announced in January 2025. But even so, the scale of this computing power is still very large. As of the end of last year, OpenAI's available computing power was approximately 2 gigawatts, equivalent to the power generation of two nuclear power plants.
Saharan announced his resignation on OpenAI's Slack work group on Thursday, saying, "It is a rare experience in this life to be able to work with everyone to repeatedly build the world's largest computer."
Founding core member of "Stargate"
Heschel, Hermani and Saharan are all members of the founding team of Stargate. This project is a very ambitious plan of OpenAI, aiming to help companies build and operate large-scale data centers. The plan, announced at a public live broadcast at the White House shortly after Trump took office, is a core initiative in OpenAI’s early efforts to gain long-term control over the infrastructure required for the training and operation of advanced AI models.
However, last year, OpenAI finally gave up on this self-built model due to financing difficulties and difficulty in finalizing the proposed joint venture structure with SoftBank and Oracle. Instead, companies are increasingly relying on a partnership model, choosing to lease large amounts of computing power from cloud service and infrastructure providers rather than directly holding related assets.
After Katy joined, many executives, including Herschel, who had previously reported directly to OpenAI president and co-founder Greg Brockman, were placed under Katy's management, and Katy reported to Brockman. Heschel has since been responsible for an internal team called the "Ecology and Cooperation Department of Industrial Computing Power" (EPIC).
According to people familiar with the matter, OpenAI has no plans to find a successor for Heschel's position.
Insinuations against rival Anthropic
Although OpenAI's "Stargate" strategy has been adjusted in the past year, the company believes that its aggressive investment strategy has given it an advantage over the competition. In a recent internal memo disclosed to investors, OpenAI mocked rival Anthropic, saying it was "obviously lagging behind in the development curve" in terms of computing capacity.
OpenAI said that as of the end of last year, Anthropic had a computing capacity of 1.4 gigawatts, while its own computing power reached 1.9 gigawatts. At the same time, OpenAI stated that it plans to significantly increase the scale of computing power. The total capacity will reach the mid-single-digit gigawatt level by the end of this year and exceed 10 gigawatts in 2027.
In contrast, OpenAI expects Anthropic's computing power capacity to be 3 to 4 GW in 2026 and 7 to 8 GW at the end of 2027. (Anthropic has not disclosed its computing power plans, but the company has reportedly discussed seeking at least 10 gigawatts of computing power capacity in the next few years.