'Huge shortage': Anthropic CEO Dario Amodei turns to Elon Musk amid AI chip crunch

Silicon Valley’s artificial intelligence ambitions are running into a practical obstacle. There is not enough computing capacity to satisfy the demands of the companies building increasingly powerful models. Even well-funded technology giants are struggling to secure the processing resources their researchers need, prompting negotiations across corporate rivalries and raising the cost of developing AI systems.

A report by The Wall Street Journal (WSJ) describes a market in which access to advanced chips has become a source of internal conflict, commercial leverage and unexpected cooperation. Start-ups face particular difficulties as suppliers gain bargaining power, while established laboratories must weigh competing demands from research teams, consumer products and cloud customers.

The strain is also changing relationships between companies that have publicly criticised one another. Anthropic, whose Claude models compete in a crowded market, has reportedly approached several rivals for additional capacity as demand for its services has grown.

Scarce computing power reshapes corporate relationships

One of the most notable developments involves Anthropic and Musk’s artificial intelligence venture, xAI. According to the WSJ, Anthropic co-founder Tom Brown travelled to xAI’s offices in Hawthorne, California, in late March to discuss access to computing infrastructure.

The discussions reportedly led to an arrangement under which Anthropic would lease more than 300 megawatts of computing capacity through a deal linked to xAI, potentially worth as much as $45 billion. The reported terms illustrate how urgent infrastructure requirements can override public disagreements between competing executives.

The timing was notable. Days before the reported partnership, Musk had criticised Anthropic on X, alleging that its AI was biased against several demographic groups. He described the alleged behaviour as “misanthropic and evil” and demanded that the company address it. The reported infrastructure arrangement nevertheless points to a willingness to do business where additional computing resources are valuable.

Anthropic CEO Dario Amodei also contacted Meta’s chief AI officer, Alexandr Wang, earlier this year to explore whether the social media group could supply spare capacity, the WSJ reported. Meta considered the possibility but ultimately declined to provide chips to Anthropic for the time being, according to people familiar with the discussions.

The approaches underline a broader problem: companies investing heavily in their own AI programmes must decide whether scarce resources should support internal development, external customers or competitors seeking to expand their operations.

Google’s internal disputes reflect a wider industry bottleneck

The pressure is not confined to commercial negotiations. Google has also experienced disagreements over how computing resources should be distributed among its research teams.

The WSJ reported that Google spent $2.7 billion in 2024 to bring back Noam Shazeer, a prominent AI researcher, to help lead work on projects including its Gemini models. Despite the substantial outlay, the company did not guarantee the computing capacity he wanted. After resources were redirected from his team to another initiative, Shazeer left for OpenAI, according to the report.

Google co-founder Sergey Brin has reportedly intervened in allocation decisions to direct computing power towards projects he favours. DeepMind chief Demis Hassabis, meanwhile, has expressed dissatisfaction with the resources available for fundamental research, which may take longer to produce commercial returns.

CNBC has separately identified computing capacity as a major source of friction within Google. Although the company is spending heavily on data centres, chips and related infrastructure, demand continues to exceed what is readily available. Commitments to supply infrastructure to outside AI developers, including Anthropic, can add another layer of complexity when those customers compete with Google’s own models.

Outside the largest technology companies, intermediaries are searching for computing resources wherever they can find them. The WSJ described efforts by San Francisco Compute to locate usable server clusters in unconventional settings, including a converted chicken coop in rural America, as well as sourcing hardware in Taipei amid flooding in Manila.

The scramble is also affecting prices. The hourly cost of renting Nvidia H100 processors under one-year contracts has risen by 60 per cent over the past year, according to the report. Cloud providers are increasingly seeking long-term commitments and substantial upfront payments.

For AI developers, access to computing power is therefore becoming as important as research talent and investment. The companies able to secure sufficient infrastructure may be better placed to train and operate their models, while smaller competitors risk falling behind before their products can reach scale. Yet expanding supply takes time, leaving the industry to negotiate over a resource that money alone cannot immediately provide.