In Silicon Valley, the struggle for computing resources has intensified: the heads of companies are personally involved in their distribution, and direct competitors are forced to enter into alliances in order to gain access to additional capacity. The shortage is particularly painful for startups that face multi-year financial commitments.
Image Source: Numan Ali/unsplash.com
Indicative V this plan history II-developer Noah Shazirah (Noam Shazeer), who left Google V 2021 year which retells Wall Street Journal. IN 2024 year company concluded deal by $2,7 billion, providing his return To work over advanced projects, including Gemini. But even This Not guaranteed his teams necessary volume computing resources. Less how through two year Shazir reported one from their teams, What Google reduced dedicated to her power V benefit another project. Soon after this He moved on V OpenAI.
According to WSJ sources, there are heated debates within Google about which projects deserve priority access to computing resources. Company co-founder Sergey Brin sometimes interferes with the formal process of their allocation, redirecting capacity to projects that he considers important.
AI company executives are also personally involved in capacity lease negotiations. At the end of March, Anthropic co-founder Tom Brown visited Elon Musk at the offices of rival xAI to negotiate the lease of computing resources. The unexpectedly high demand for Claude Code led to a lack of capacity and frequent service failures. In May, Anthropic announced agreement with SpaceX, the value of which, as it turned out later, could reach $45 billion within a few years. Musk, who has previously publicly criticized Anthropic, said on social media X that he was impressed with its leadership, adding: “None of them gave me any suspicion of bad intentions.”.
Earlier this year, Anthropic CEO Dario Amodei also had a phone conversation with Alexandr Wang, head of artificial intelligence at Meta✴trying to gain additional computing power. According to the publication’s sources, after internal discussions Meta✴ I decided to refuse for now.
“There’s a real gold rush going on right now.”says Evan Conrad, CEO of San Francisco Compute. His company vets data centers and rents out their computing power to AI developers, from big players like Nvidia to small startups like Standard Intelligence.
Despite announcements of massive computing power on the way, the reality is that far fewer resources are available today, said Nick Frosst, co-founder of Canadian AI startup Cohere. Some projects are cancelled, while others are implemented on a smaller scale than originally planned. The construction of new data centers is also hampered by regulatory restrictions and local resistance.
Due to shortages, prices for computing resources are rising. According to SemiAnalysis, the hourly cost of renting capacity based on Nvidia H100 accelerators under an annual contract has increased by 60% over the past year.
Cloud providers are increasingly demanding multi-year contracts with upfront payments of up to 30% of the transaction amount. Sometimes startups also need to find guarantors who are ready to take over payments if the company itself cannot fulfill its obligations.
Because it’s difficult to forecast demand, some companies are overbooking capacity out of fear of future shortages. Others overestimate the demand for their products and are then forced to resell unused resources. For startups, both options are risky: Excessive lease obligations can lead to bankruptcy, and lack of capacity can limit growth and lead to user dissatisfaction.
Some venture capital firms now reserve computing clusters themselves and lease capacity to their portfolio companies or act as guarantors under contracts with suppliers. Access to computing resources is becoming no less important for development for young AI developers than attracting funding.
If you notice an error, select it with the mouse and press CTRL+ENTER.
