OpenAI shouts to investors: Computing power has become a “hard constraint” for products; we run faster and harder than Anthropic

Zhitong Finance App learned that this week, OpenAI revealed to its investors that with early aggressive investment in computing resources, the company had a key advantage in the competition with rival Anthropic PBC. Anthropic is currently considering an initial public offering (IPO).

According to a memo sent to some investors, the ChatGPT developer said OpenAI has surpassed Anthropic by “rapidly and continuously” increasing computing capacity to support wider adoption of its software. The memo comes after Anthropic launched a more powerful AI model called Mythos. OpenAI stated in the memo that despite being criticized for being too expensive, its ambitious infrastructure construction allows it to better keep up with the growing demand for AI products.

“This gap is critical because computing power is now a product constraint,” OpenAI wrote in the memo. OpenAI declined to comment.

Founded by a former OpenAI employee, Anthropic's service has sometimes been difficult to maintain online status in recent months due to the launch of new products and a surge in demand driven by public support for its confrontation with the Pentagon. In a memo to investors, OpenAI quoted a report by Strachery blogger Ben Thompson, which stated that computing power limitations may have affected Anthropic's decision to release Mythos only to select partners.

OpenAI said it will have 1.9 gigawatts of usable computing capacity in 2025, three times that of the previous year (1 gigawatt can power approximately 750,000 US households). The company expects this figure to rise to the “low double digit” gigawatt range next year, reaching around 30 gigawatts by 2030. In contrast, OpenAI estimates that Anthropic will have 1.4 gigawatts of capacity by the end of 2025 and will have 7 to 8 gigawatts next year.

OpenAI stated in the memo: “Even at the upper end of the Anthropic forecast range, our growth rate is still significantly ahead and the gap is widening.”

In recent months, Anthropic has increased its investment in physical infrastructure for AI services, including a commitment to invest $50 billion to build data centers in the US. In addition, Anthropic has also expanded strategic cooperation with Broadcom (AVGO.US) and Google (GOOGL.US), a subsidiary of Alphabet Inc., to obtain approximately 3.5 gigawatts of computing power from 2027. Furthermore, Anthropic has a diverse supplier portfolio and cooperates with the three major cloud service providers Google, Microsoft (MSFT.US), and Amazon (AMZN.US).

In response to a request for comment, Anthropic quoted its chief financial officer Krishna Rao's previous statement regarding the deal with Broadcom and Google. “This groundbreaking collaboration with Google and Broadcom continues our prudent approach as we expand our infrastructure,” said Rao. We are making our most important computing power commitment to date to keep up with this unprecedented pace of growth.”

In the past, Anthropic's spending strategy was generally more conservative than OpenAI. OpenAI plans to invest around $600 billion in data centers and chips by 2030, and recently raised $122 billion to support these commitments. Still, considering that OpenAI is not expected to be profitable in the next few years, the pace of its spending is still being questioned. On Thursday, OpenAI said it would suspend an infrastructure project in the UK, citing energy cost issues.

Anthropic CEO Dario Amodei said in an interview at the end of last year about the company's spending plan: “As a company, we will manage as responsibly as possible. I think some players are in 'YOLO'.” In an investor memo, OpenAI described Amodei's views as misjudging the extent to which the market is in demand for more AI products. The company said, “In hindsight, this caution is less about self-discipline than underestimating the speed at which demand is arriving.”