AI Computing Power Race Heats Up! Anthropic Discloses $84.5 Billion Deal, Far Exceeding May Estimate; SpaceX (SPCX.US) Accelerates AI Growth Engine Again

According to English Finance APP, media reports citing informed sources say that global AI model and AI application leader Anthropic has signed an agreement to pay up to $84.5 billion to utilize AI computing power resources from SpaceX (SPCX.US), founded and led by Musk, the “AI + Space Exploration” pioneer, until 2029. This means that Anthropic has agreed to pay nearly double the previously estimated computing power fees to SpaceX, highlighting the strategic value of computing clusters that can be delivered quickly and operate stably as the commercial rollout of AI models accelerates.

This amount far exceeds the agreement disclosed in May this year. SpaceX revealed in its IPO filing that Anthropic agreed to pay $1.25 billion per month over a period of up to three years to access AI computing power. The agreement can be canceled with 90 days' advance notice. If the agreement is fully executed over three years, the total sum will be about $45 billion.

Additionally, according to media reports, when including other large AI computing 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 computing power procurement disclosed during Anthropic's IPO dash is reinforcing SpaceX’s market position as a major AI infrastructure provider. The agreement between Anthropic and SpaceX’s subsidiary xAI could bring Anthropic’s computing power expenditure up to $84.5 billion by 2029, which is about 87.8% higher than the estimated $45 billion for three full years. However, it should be noted that these figures reflect the latest disclosure of an expanded partnership and do not indicate how much the unit price of computing power has increased.

Computing Power Deal Worth Up to $84.5 Billion Emerges, SpaceX’s AI Growth Engine Gains Another Boost

Financial data disclosed in Anthropic’s IPO prospectus provides a more concrete observation window into this wave of expansion: revenue for 2025 is set to increase nearly 12 times compared to last year, approaching $4.6 billion; operating losses exceed $8 billion; spending on computing power and infrastructure hits $7.33 billion, about three times that of 2024, accounting for about 58% of total operating expenses of $12.65 billion. The latest IPO prospectus shows that the commercialization of cutting-edge models is simultaneously expanding both revenue and computational resource consumption, with AI agent/model training, inference, and infrastructure investments still significantly ahead of profit realization.

Furthermore, Anthropic’s infrastructure spending plan for roughly the next decade is at least $518 billion, of which about 80% of major AI infrastructure project expenditures are non-cancelable or must be paid regardless of actual usage.

SpaceX’s external customer base is also expanding. Documents submitted to the SEC show that Google has contracted to acquire about 110,000 Nvidia GPUs as well as CPUs, memory, and related computing power, agreeing to pay $920 million monthly from October 2026 to June 2029. That comes to $30.36 billion over the full 33-month term, with additional ramp-up costs from earlier phases; the contract also includes delivery and early termination clauses.

On the AI computing supply side, according to Musk’s latest disclosure on September 25, the Colossus 1 and Colossus 2 super AI data centers will be equipped with about 780,000 AI GPUs, with plans to incrementally deploy the GB300—the latest Nvidia advanced AI GPU cluster. If all proposed batches come online, total scale could reach 1.44 million units; however, some batches slated for late December are still subject to uncertainty.

“Computing Power Time Premium”—Whoever Delivers AI Computing Resources First Captures Commercialization First

With Anthropic aiming for a $2 trillion valuation and about $100 billion in IPO fundraising to become the world’s largest IPO, Wall Street analysts are focusing on the “computing power time premium”—that a complete cluster delivered earlier and operating stably may be more commercially valuable than simply owning chips.

For leading AI model developers, waiting for power access, network setup, and server delivery means postponing capacity expansion and enterprise order fulfillment. Anthropic has previously stated that the SpaceX deal and other new arrangements for computing power have allowed it to raise usage limits for Claude Code and API, as well as improve capacity available to paid subscribers. This forms a clear transmission chain: increased computing delivery expands product and service capabilities, which creates conditions for paid usage growth.

From the perspective of underlying AI inference architecture, cutting-edge AI agents may trigger a sequence of planning, retrieval, tool invocation, code execution, and result validation in a single task. GPUs handle model computation, CPUs orchestrate task scheduling and tool execution, while memory and storage provide context, caching, and task status; delays in the CPU process can increase cache eviction and recomputation. Therefore, the mainstream procurement to support AI agents includes systems coordinating GPUs, CPUs, memory, networks, and power/heat management. Early disclosures of Anthropic’s SpaceX computing power deal included about 325,000 Nvidia AI GPUs, plus high-performance CPUs, EB-scale storage, and high-speed optical interconnects within the data center—indicating that these massive contracts correspond to comprehensive AI computing cluster services.

The Nvidia AI GPU computing cluster remains the clearest supply bottleneck in the AI computing industry chain. TrendForce estimates show that shipments of NVL72 racks, covering both the Blackwell and Vera Rubin platforms, are expected to increase by more than 50% year-on-year in 2027; market research illustrations predict that related system market value will rise from about $226 billion in 2026 to $711 billion in 2027, a sharp 214% increase year-on-year. TrendForce also expects the average selling price of Vera Rubin systems to be about twice that of GB300, and the price increase of wafers and critical components such as HBM will further raise system value; GB300 demand is expected to continue into the first half of 2027, after which VR200 will ramp up and become the largest share of shipments for 2027.

For SpaceX, model competition and infrastructure collaboration can coexist: proprietary model efforts target the application market, while external computing power services expand infrastructure revenue sources. Thus, investment value should focus on delivered capacity, actual utilization rate, and payment consistency from customers. However, investors should note that most of the newly arranged $84.5 billion can still be canceled with 90 days’ notice, so it should be interpreted as an indicator of potential revenue and customer demand, rather than $84.5 billion in revenue already locked in.