Anthropic, a global leader in AI large models and applications, has signed an agreement to pay up to $84.5 billion for access to AI compute resources from SpaceX (SPCX.US), the AI and space exploration pioneer founded and led by Elon Musk, through 2029, according to people familiar with the matter cited by media reports. This means the compute fees Anthropic has agreed to pay SpaceX are nearly double earlier estimates, underscoring that the faster model commercialization progresses, the more strategically valuable it becomes to have compute clusters that can be delivered on time and run reliably.
This amount is far higher than the scale of the agreement disclosed in May of this year. In its IPO filing, SpaceX revealed that Anthropic agreed to pay $1.25 billion per month for AI compute for a maximum term of three years. The agreement can be cancelled with 90 days' notice. If the agreement is fully executed over three years, the total amount would be approximately $45 billion. In addition, according to media reports, after factoring in other large AI compute infrastructure partners such as Google (GOOGL.US), Amazon (AMZN.US) and Microsoft (MSFT.US), the AI lab led by Dario Amodei is expected to spend at least $518 billion over the next decade to support its workloads. The scale of compute procurement disclosed by Anthropic as it pushes toward an IPO is reinforcing SpaceX's market positioning as a major AI infrastructure supplier.
The agreement between Anthropic and SpaceX's xAI could bring its compute procurement spending through 2029 to as much as $84.5 billion, about 87.8% higher than the earlier estimate of $45 billion based on full three-year performance. It should be noted, however, that the disclosed figure reflects the latest expansion in the scale of cooperation and cannot yet be used to determine how much unit compute rental prices have risen.
The $84.5 Billion Compute Agreement Emerges, Giving SpaceX's AI Growth Engine Fresh Support
The financial data shown in Anthropic's prospectus provides a more concrete window into this round of expansion: revenue in 2025 grew to about 12 times the prior year's level, approaching $4.6 billion, while operating losses exceeded $8 billion. Compute and infrastructure spending reached $7.33 billion, roughly triple the 2024 level and accounting for about 58% of total operating expenses of $12.65 billion. The data in this latest IPO prospectus shows that frontier model commercialization is simultaneously expanding both revenue and computing resource consumption, with AI agent and AI large model training, inference and infrastructure investment still significantly ahead of profit realization.
In addition, Anthropic's infrastructure spending arrangements of at least $518 billion over roughly the next decade mean that about 80% of large AI infrastructure project expenditures are non-cancellable or require payment regardless of actual usage. SpaceX's external customer base is also expanding. Its filing with the SEC shows that Google has contracted for about 110,000 Nvidia GPUs along with CPUs, memory and other supporting compute, agreeing to pay $920 million per month from October 2026 through June 2029. Calculated mechanically over the full 33 months, that corresponds to $30.36 billion, plus earlier ramp-up period costs; the contract also includes delivery requirements and early termination clauses.
On the AI compute resource supply side, according to Musk's latest disclosure on September 25, the two major super AI data centers Colossus 1 and Colossus 2 together are equipped with about 780,000 AI GPUs, with plans to add GB300, Nvidia's frontier AI GPU compute cluster, in batches. If all proposed batches are fulfilled, the total scale could reach 1.44 million units, though the batch at the end of December still carries uncertain conditions.
"Compute Time Premium" — Whoever Delivers AI Compute Resources First Can Capture Commercialization First
As Anthropic seeks to claim the global IPO crown with a $2 trillion valuation and an IPO fundraising scale of roughly $100 billion, Wall Street analysts have begun focusing on the "compute time premium" — complete clusters that can be put into use earlier and stably may have greater commercial value than simply owning chips. For leaders in AI large models, waiting for power connections, network debugging and server deliveries means delaying expansion of user capacity and taking on enterprise orders.
Anthropic has previously made clear that the SpaceX partnership and other new compute arrangements enable it to raise usage limits for Claude Code and its API, and improve the capacity available to paying subscribers. This forms a clear transmission chain: more compute delivery, expanded product and service capability, and in turn conditions for growth in paid usage. From the perspective of the underlying AI inference architecture, a single task for the most advanced AI agents may continuously trigger planning, retrieval, tool calling, code execution and result verification. GPUs handle model computation, CPUs handle task orchestration and tool execution, while memory and storage carry context, caching and task state; slow CPU links can also increase pressure from cache eviction and recomputation. Therefore, what the spread of agents drives procurement toward is a system in which GPUs, CPUs, memory, networking and power and cooling operate in coordination.
SpaceX's earlier disclosed compute configuration for Anthropic included about 325,000 Nvidia AI GPUs, along with high-performance CPUs, exabyte-scale storage and high-speed optical interconnect systems within the data center, enough to show that large contracts correspond to complete AI compute cluster services. Nvidia's AI GPU compute clusters are undoubtedly still the clearest supply bottleneck at the AI compute industry chain level. TrendForce estimates show that in 2027, NVL72 rack shipments covering the Blackwell and Vera Rubin platforms are expected to grow more than 50% year over year; market research graphics show that related system output value is expected to rise from about $226 billion in 2026 to $711 billion in 2027, a sharp year-over-year increase of 214%.
TrendForce expects the average selling price of Vera Rubin systems to be about twice that of GB300, and rising prices for key components such as wafers and HBM will also increase system value; GB300 demand is expected to continue into the first half of 2027, after which VR200 will ramp up and become the platform with the highest shipment share in 2027. For SpaceX, model competition and infrastructure cooperation can coexist: its own models compete for the application market, while providing compute externally expands infrastructure revenue sources. By extension, its investment value should be assessed around delivered capacity, actual utilization and continuity of customer payments.
Investors should note, however, that most of the latest $84.5 billion arrangement can still be cancelled with 90 days' notice, so it should be viewed as a signal of potential revenue scale and customer demand, not directly as $84.5 billion in locked-in revenue.
Comments