Featured

#AnthropicDiscloses$84.5BComputeDealWithSpaceX

The AI race is starting to look less like a software competition and more like an infrastructure competition.

Anthropic’s IPO filing gives a remarkable view into just how much computing capacity the next generation of AI companies believe they will need. According to Reuters, Anthropic has disclosed a potential computing-capacity arrangement worth as much as $84.5 billion with xAI, while its broader long-term infrastructure commitments are expected to reach at least $518 billion over the next decade.

Those numbers are difficult to process at first. But the more important question is not simply how much money Anthropic is committing.

It is why a company building some of the most advanced AI models in the world needs this much computing power in the first place.

The answer is becoming increasingly clear.

Training a frontier model is only one part of the equation. Once a model becomes commercially useful, computing demand does not disappear. Every user request requires inference. Every AI agent running continuously consumes resources. Longer context windows require more memory. More capable models require larger and more sophisticated infrastructure. And as AI moves from occasional chatbot interactions toward systems that can perform tasks continuously, the amount of computation required can increase dramatically.

That changes the economics of the AI industry.

The scarce resource is no longer simply access to talented researchers or good algorithms. Reliable computing capacity itself is becoming a strategic asset.

Anthropic’s infrastructure commitments show how companies are trying to secure that asset before demand becomes even harder to satisfy.

The $84.5 billion arrangement is particularly interesting because it represents a different approach from simply building everything internally. Leasing computing capacity can give an AI company access to large-scale GPU infrastructure without having to construct every data center, secure every power connection, purchase every server, and manage the entire physical deployment process itself.

That flexibility matters.

AI demand is growing quickly, but it is also difficult to forecast several years ahead. Leasing allows companies to scale capacity more rapidly and potentially avoid owning enormous amounts of hardware that could become underutilized later.

But there is another side to this strategy.

Long-term infrastructure commitments can become enormous fixed obligations. Reuters reported that roughly 80% of Anthropic’s disclosed infrastructure commitments are binding regardless of actual usage. That means the company is effectively making a massive bet that future AI demand, revenue and model usage will justify today's infrastructure decisions.

This is where the AI infrastructure story becomes much more interesting.

Anthropic is not relying on a single source of computing power. Its disclosed commitments involve multiple infrastructure and technology partners, including very large agreements connected to Google, Amazon, Microsoft and Broadcom. The strategy appears to be about securing enough capacity from multiple sources while continuing to expand the underlying AI business.

In other words, the AI race is creating an entire infrastructure economy behind the models.

And this brings us directly to memory.

GPUs get most of the attention when people talk about AI computing, but GPUs cannot operate in isolation. High-bandwidth memory, or HBM, is a critical component of modern AI accelerators because large-scale training and inference require extremely high memory bandwidth.

The latest TrendForce research makes this part of the story even more important. It expects HBM supply to remain constrained into 2027 and has raised its forecast for the blended average HBM selling price in 2027 to a 121% year-over-year increase, driven by tight supply, the growing mix of HBM4 and continued AI server demand.

That means the AI infrastructure race does not stop at data centers.

More AI computing demand creates pressure across GPUs, HBM, networking equipment, power infrastructure, cooling systems, data-center construction and electricity supply.

This is why I think the most interesting part of Anthropic’s filing isn't the headline number itself.

It is the confirmation of a structural shift.

For years, the AI conversation was dominated by model architecture, benchmark scores and product features. Now the bottleneck is increasingly physical. You can have brilliant researchers and sophisticated algorithms, but without enough chips, memory, electricity and data-center capacity, those models cannot be trained or served at the scale the market expects.

At the same time, simply throwing more hardware at the problem cannot be the industry's permanent solution.

If computing costs continue rising, AI companies will eventually need better efficiency as well as more capacity. Smaller models, quantization, sparsity, inference optimization, improved architectures and specialized accelerators can all reduce the amount of computation required for a given task.

That creates an interesting balance for the next phase of AI.

More computing power expands what models can do. Better efficiency determines how economically those capabilities can be delivered.

Anthropic’s enormous infrastructure commitments therefore tell us something much bigger than the spending plans of one AI company.

They show that the AI industry is moving into a phase where securing compute can be just as important as developing the model itself.

The real competition may not be simply about who builds the smartest AI.

It may increasingly be about who can secure enough computing power, memory, energy and infrastructure — while still making the economics work.

And that is a much bigger story than an $84.5 billion contract headline.