Anthropic has agreed to spend $1.25 billion a month on compute, and the infrastructure belongs to a direct competitor.
The counterparty is xAI, with the arrangement disclosed through IPO-related filings tied to SpaceX rather than a conventional joint announcement. Over the life of the contract, which runs through May 2029, the total value approaches $45 billion.
At face value, it looks like a capacity deal. In practice, it’s a workaround for a system that can’t build fast enough.
The Terms: A High-Cost, High-Flexibility Compute Lock-In
The agreement formalizes something closer to a rolling infrastructure hedge than a standard cloud contract.
- Monthly run rate: $1.25 billion
- Duration: Through May 2029 (~$45B total ceiling)
- Infrastructure scope: Capacity spanning both Colossus 1 and Colossus 2
- Exit clause: 90-day mutual termination window
- Ramp: Gradual utilization through May–June 2026 before full pricing applies
That last detail matters. This isn’t spare capacity quietly being offloaded. Its capacity is being activated at scale, fast enough to justify one of the largest compute commitments ever disclosed.
Why xAI Had the Capacity to Sell
The simple version — “unused compute” — doesn’t hold up under scrutiny.
xAI didn’t accidentally end up with idle infrastructure. It reallocated it.
Training workloads for newer Grok models have been shifting toward Colossus 2, a more advanced cluster designed for next-generation runs. That transition left significant capacity at Colossus 1 available for external use — not because demand disappeared, but because it moved.
From a financial perspective, the decision is straightforward. Infrastructure at this scale carries enormous fixed costs. Leaving it partially idle ahead of a public offering isn’t just inefficient — it’s indefensible.
Leasing it converts stranded capacity into predictable revenue.
Why Anthropic Was Willing to Pay a Premium
On the other side is Anthropic — and the constraint here isn’t demand. It’s a supply.
Inference demand has shifted from periodic spikes to continuous load, driven by enterprise usage and paid tiers. That creates a different kind of bottleneck: not just training capacity, but always-on serving infrastructure.
And crucially, this deal didn’t sit idle after signing.
Anthropic immediately deployed the capacity:
- Peak-hour usage caps were removed
- Claude Code rate limits were increased for Pro and Max users
- API throughput expanded for enterprise clients
This is the missing link most coverage ignores: the compute wasn’t warehoused — it was consumed almost instantly.
Which tells you something about the state of the market.
The Real Constraint: Power, Not GPUs
At $1.25 billion a month, Anthropic is almost certainly paying above what equivalent capacity would cost under normal conditions.
But normal conditions don’t exist.
The limiting factor isn’t chip supply alone — it’s power:
- Grid interconnection delays now stretch years in key US regions
- Permitting for new hyperscale facilities is increasingly constrained
- Multi-hundred-megawatt deployments face regulatory bottlenecks
Buying existing, already-powered infrastructure is often the only way to scale quickly.
Even at a premium, it’s faster than building.
The Neocloud Model, Defined Properly
This deal makes more sense when viewed as part of a broader structural shift.
Neocloud model:
AI companies build infrastructure for internal use first, then sell excess capacity opportunistically to external buyers when utilization falls below peak assumptions.
That’s a break from traditional cloud models led by Amazon Web Services and Google Cloud, where infrastructure is explicitly designed for third-party customers from day one.
Neoclouds invert that logic:
- Internal demand comes first
- External customers stabilize utilization
- Revenue smooths out overbuild risk
It’s less efficient than a hyperscale cloud in theory.
Right now, it’s more realistic.
A Temporary Truce Between Competitors
There’s an obvious tension in the arrangement.
Anthropic’s models are now running, at least in part, on infrastructure controlled by a company led by Elon Musk — someone who has publicly criticized competing AI labs, including Anthropic.
Under normal circumstances, that relationship would be untenable.
Under current constraints, it’s pragmatic.
The 90-day termination clause reinforces that reality. This isn’t a long-term alliance built on alignment. It’s a high-value, revocable agreement between two parties solving different problems:
- xAI needs utilization
- Anthropic needs immediate capacity
The overlap is temporary by design.
The Longer-Term Signal: Beyond Earth-Based Compute
Buried in the broader strategic context is a more speculative, but increasingly discussed, direction.
As terrestrial power constraints tighten, companies tied to infrastructure and launch capabilities — particularly SpaceX — have begun exploring orbital or space-based compute as a long-term solution.
The logic is straightforward:
- Fewer grid constraints
- Direct access to solar energy at scale
- Potential for multi-gigawatt deployments unconstrained by local regulation
It’s not imminent. But it’s no longer fringe.
And deals like this one make clear why the idea is getting attention.
The Bottom Line
This isn’t just a large contract.
It’s a snapshot of a system under pressure:
- Companies are overbuilding because they have to
- Underutilizing because demand is uneven
- And renting capacity to competitors because there’s no faster alternative
The neocloud model isn’t theoretical anymore.
It’s what happens when AI demand outruns the physical world’s ability to support it.
Related: An Anthropic’s 80x Growth Crisis: The Inside Story Behind the SpaceX Compute Deal
