Silicon Valley continues to "burn money" to ramp up computing power: OpenAI's budget has risen to 750 billion, and Anthropic and AMD are adopting the "equity-for-orders" approach.
A new wave of infrastructure investment is kicking off in Silicon Valley.
On July 22, OpenAI announced the launch of its first self-led $20 billion data center project, and raised its projected computing power expenditure by 2030 to $750 billion; on the same day, AMD announced it would invest up to $5 billion in Anthropic, while Anthropic committed to purchasing up to 2GW of AMD's latest-generation AI chips.
The two deals indicate that the competition among tech giants is no longer limited to simple chip procurement, but has fully extended to self-built data centers, energy security, and direct control over underlying infrastructure.
01
OpenAI's Computing Power Budget Surges to $750 Billion
To secure the massive computing power required to develop and run AI models, OpenAI has further expanded its data center plans and budget.
People familiar with the matter revealed that driven by new agreements with cloud service providers, OpenAI has raised its projected computing power expenditure by 2030 from approximately $600 billion at the start of this year to $750 billion.
However, as OpenAI advances its IPO process, balancing computing costs and financial health has become a key focus of management discussions. The computing power expansion plan strongly pushed by CEO Sam Altman has led CFO Sarah Friar to maintain a prudent attitude toward the company's capital chain status.
It is reported that Friar expressed concerns to the management that if revenue growth falls short of expectations, the company may face a funding gap when fulfilling its massive computing power procurement agreements in the future.
In terms of specific project advancement, OpenAI announced a $20 billion investment to launch a data center project codenamed "Project Camellia" in Effingham County, Georgia.
Regarding power security, OpenAI has reached an agreement with Georgia Power to obtain 3.2GW of power supply in phases between 2028 and 2032.
As understood, this project is the first data center that OpenAI has independently advanced as the lead designer and developer. Prior to this, OpenAI mainly leased computing power from cloud service providers such as Oracle and Amazon AWS.
To this end, OpenAI hired Brent Mayo, who previously led the construction of the Memphis supercomputer data center at Elon Musk's xAI, to serve as Head of Data Center Construction and Delivery. Mayo will report to Uday Ruddarraju, who was promoted to Chief Technology Officer of Computing Power this month.
To mitigate growing public reflection and resistance to data center construction, OpenAI has committed to providing $80 million in community benefit funds over the project's lifecycle for areas including schools, public safety, and healthcare. It will also offer up to $71 million in Codex vouchers to college students in Georgia. The project will adopt closed-loop water circulation cooling technology to minimize water usage, and voluntarily reduce power consumption during peak electricity periods to ensure local residents' electricity bills are not driven up.
In addition to the Georgia project, OpenAI is also advancing multiple computing power agreements: it previously signed a 6GW data center capacity agreement with Oracle; at the end of February, it expanded its cloud computing agreement with Amazon AWS to an 8-year, $138 billion deal (including 2GW Trainium chip computing power); last year, it also committed to an additional $250 billion in cloud spending on Microsoft Azure.
02
AMD and Anthropic Strike "Equity-for-Orders" Deal
Like OpenAI, Anthropic is also facing extremely high pressure from computing power demand. Previously, surging demand forced Anthropic to limit user access frequency, and some customers reported service outages.
To alleviate computing power bottlenecks and diversify its supply chain, AMD and Anthropic announced a major strategic cooperation on the same day. Under the terms of the agreement, Anthropic will deploy up to 2GW of AMD's latest-generation Instinct MI450 series GPUs starting in the first half of 2027. These chips will be integrated into AMD's Helios rack-level solution, which includes EPYC "Venice" CPUs, Pensando networking, and ROCm software.
At the same time, as specific deployment milestones are met, AMD will make a strategic equity investment of up to $5 billion in Anthropic.
The two sides also reached an engineering cooperation agreement, which will use Anthropic's Claude model to optimize workloads for AMD chips and accelerate the development of AMD ROCm software. AMD's own engineering and product development teams will also fully adopt Claude.
Lisa Su, CEO of AMD, said that the engineering teams of the two companies have been collaborating for some time, and the deployment of large-scale AI computing power requires long-term planning 12, 18, or 24 months in advance. According to people familiar with the matter, AMD is currently in discussions to provide financial guarantees for the lease debt of Anthropic's future data centers.
According to a report by The Wall Street Journal, Anthropic's computing power strategy in this transaction covers both direct purchases and leases: it will directly procure some AMD chips for its own facilities, and also lease some computing power through other or new cloud service providers.
Previously, Anthropic has signed relevant computing power or data center capacity agreements with Google, Amazon, and SpaceX, and its workloads run simultaneously on a variety of chips such as Google TPUs, Amazon Trainium, and NVIDIA GPUs.
This article is from the WeChat Official Account "Tencent Tech", author: Su Yang, editor: Xu Qingyang, published with authorization from 36Kr.