#Anthropic与SpaceX签署845亿美元算力协议 Spending $84.5 billion to lock in computing power! Anthropic bets on AI foundational infrastructure as the computing power race heats up

Anthropic’s IPO filing has revealed a computing power lease agreement with SpaceX worth up to $84.5 billion, with the agreement running through 2029. The company behind the Claude large language model is expected to invest at least $518 billion in AI infrastructure over the next decade. The computing power arms race is intensifying, with HBM memory prices expected to rise.

The competition in the AI industry appears on the surface to be a contest between model capabilities and product experiences, but at its foundation lies a never-ending battle for computing power.
Anthropic’s latest IPO filing has dropped a bombshell: the company has signed a computing power lease agreement with SpaceX worth up to $84.5 billion, with the agreement remaining in effect through 2029.
Reuters reported, citing the prospectus, that this ultra-expensive computing power contract is far above the $45 billion previously estimated by the market. The filing also shows that Anthropic’s total investment in AI infrastructure over the next decade is expected to be no less than $518 billion.
These astonishing figures offer a direct view of the massive investment that leading large-model companies are making in computing resources.

What exactly is being purchased in this $84.5 billion computing power deal?
The agreement is essentially a lease of computing resources: Anthropic will lease NVIDIA GPU clusters at data centers owned by SpaceX for Claude-series large-model training, inference, and AI agent operations.
The agreement includes a flexible provision under which either party can terminate the partnership by giving 90 days’ advance notice.

Why doesn’t Anthropic build its own data centers and instead choose to lease computing power from SpaceX?
Building an ultra-large-scale computing cluster from scratch involves land, power supply, data center construction, and hardware procurement, making the process lengthy and financially demanding. Computing power leasing enables rapid access to massive GPU resources and quick expansion to match the explosive growth of large-model businesses. Amid highly volatile AI demand, flexible leasing can mitigate the risk of idle hardware.
However, this model also has drawbacks. The total cost of long-term leasing can ultimately far exceed that of building independently; control over computing resources is not in the company’s own hands, and any change in the partnership could directly affect the stability of model services. Leading companies are increasingly building their own facilities while also leasing extensively from external providers, making hybrid deployment the industry’s mainstream approach. 📌 📈 Expectations of higher HBM memory prices bring changes to the hardware supply chainThe computing power arms race continues to heat up, directly driving demand for upstream hardware. TrendForce predicts that the average price of HBM high-bandwidth memory will rise substantially in 2027. HBM is a core supporting component of GPUs, and both large-model training and concurrent inference by AI agents depend heavily on it. Major large-model companies are rushing to buy computing power, while GPU and HBM supply remains tight relative to demand. In the past, everyone focused on the software capabilities of large models; now, more and more people realize that without sufficient and stable computing power, even the best model algorithms cannot be deployed and operated. Computing power has become a strategic factor of production for AI companies. 📌 ⚖️ Computing power arms race: advantages and concerns coexistMassive investment in computing power brings highly visible benefits. Sufficient computing power can support models with larger parameter counts and longer context windows, while running large numbers of AI agents in parallel and accelerating model iteration. The stronger a company’s computing power reserves, the more room it has to continuously refine model capabilities and respond quickly to market demand.
However, massive computing power investment also creates hidden risks for the industry. Sky-high hardware costs raise the barrier to entry, while resources continue to concentrate among a small number of leading companies. Once commercialization revenue falls short of expectations, companies will face enormous financial pressure after large amounts of capital are spent on hardware procurement. At the same time, the electricity consumption and carbon emissions of operating large-scale GPU clusters are challenges that the global AI industry must address together.

Industry status and future development direction
Today, leading AI companies worldwide are securing computing power resources through multiple channels.
Anthropic’s list of partners includes multiple computing power providers, such as SpaceX, Google, Amazon, and Microsoft, reducing the risk of relying on a single provider.
The future computing power market will become more diversified, with multiple models coexisting, including self-built data centers, leasing from cloud providers, and third-party computing power services. Computing power will not expand indefinitely. In the long term, the industry cannot rely solely on piling up hardware to improve AI capabilities. Model lightweighting, inference optimization, and sparsification technologies can reduce computing power consumption. The simultaneous evolution of hardware and algorithms is the path to healthy development.
Computing power is the foundation, but the ultimate value of AI still depends on whether real-world applications can create genuine value.

Some say that AI competition is fundamentally a competition for computing power, while others believe algorithms and application scenarios are the core. Which view do you agree with more? Feel free to share your thoughts in the comments. $SPCX