Anthropic pays SpaceX $15B/yr for compute; Alibaba's Qwen3.7-Max runs 35h

Anthropic has agreed to pay SpaceX $1.25 billion per month through May 2029 for access to cloud computing infrastructure, totaling roughly $15 billion annually, according to details revealed in SpaceX's IPO filing. The deal grants Anthropic access to graphics processing units at SpaceX's Colossus and Colossus II data centers, which house more than one gigawatt of computing power. SpaceX originally built those facilities for xAI, Elon Musk's AI venture, but did not need all the capacity. Anthropic will pay a reduced fee for May and June 2026 before the full monthly rate kicks in. Separately, Alibaba's Qwen Team released Qwen3.7-Max, a proprietary model capable of approximately 35 hours of continuous autonomous execution. The model marks a strategic departure from Alibaba's prior open-source releases, driven in part by the departure of key Qwen Team leaders earlier this year. Qwen3.7-Max is only accessible from Chinese-based endpoints, limiting its appeal for US and European enterprises. These two developments underscore a fundamental shift in the AI industry: the cost of compute is skyrocketing, and the open-source ethos that defined the early frontier model era is giving way to proprietary walled gardens. Why this matters now: the infrastructure arms race is reshaping competitive dynamics, with hyperscale compute deals and closed-source strategies becoming the new baseline for AI leadership.

Anthropic's $15B annual compute payment dominates its P&L

Anthropic's $1.25 billion monthly payment to SpaceX represents a staggering fixed cost that will dominate the company's income statement for the next three years. To put this in perspective, Anthropic's total revenue in 2025 was estimated at roughly $3.5 billion, meaning the SpaceX deal alone consumes more than four times that figure annually. The company is clearly betting that future revenue growth from products like Claude Code and Claude Opus 4.7 will eventually cover this expense. Anthropic charges $30.00 per million tokens for Claude Opus 4.7, positioning it at a premium to OpenAI's GPT-5.4, which costs $17.50 per million tokens. The math is brutal: even if Anthropic captures a significant share of the enterprise AI market, the company will need to sustain hypergrowth for years to achieve profitability. The deal structure also reveals a strategic hedge: by locking in compute capacity through May 2029, Anthropic insulates itself from the GPU supply constraints that have plagued competitors. However, the fixed monthly payment creates a massive operating leverage problem: if demand softens or a competitor releases a superior model, Anthropic still owes SpaceX $1.25 billion every month. The reduced fee for May and June suggests a ramp-up period, giving Anthropic a brief window to optimize utilization before the full cost burden hits. The company's revenue in 2025 was $3.5 billion, so the deal is more than four times that figure annually, underscoring the scale of the bet on future growth. Anthropic's burn rate now exceeds $1 billion per month before any other operating costs, a figure that will force the company to either achieve profitability by 2028 or seek additional capital from investors.

Compute bottleneck drives pricing power across the AI stack

The Anthropic-SpaceX deal crystallizes a fundamental truth about the AI industry: compute is the new oil, and those who control it wield enormous pricing power. SpaceX built the Colossus facilities for xAI but ended up with excess capacity, creating a natural arbitrage opportunity. By leasing that capacity to Anthropic at $15 billion per year, SpaceX effectively monetizes its infrastructure advantage without bearing the full risk of xAI's model development. This dynamic mirrors what is happening across the hyperscaler ecosystem. Google, which operates its own TPU infrastructure, has a structural cost advantage over competitors who must rent compute from third parties. OpenAI relies on Microsoft's Azure infrastructure, but the terms of that arrangement are not publicly disclosed. The pricing power extends downstream: as compute costs rise, model providers pass those costs to customers. OpenAI's GPT-5.4 at $17.50 per million tokens and Anthropic's Claude Opus 4.7 at $30.00 per million tokens reflect this pass-through. The gap between the two, roughly 71% premium for Anthropic, suggests that Anthropic is either pricing to cover its higher compute costs or betting that its model quality justifies the premium. Either way, the compute bottleneck creates a natural floor for API pricing, benefiting infrastructure providers like SpaceX, Microsoft, and Google while squeezing margins for model developers. The Colossus facilities consume more than one gigawatt of power, equivalent to a small nuclear reactor, and that power draw will strain local grids in the regions where Colossus is located, potentially triggering regulatory scrutiny.

Competitive reshuffle: xAI, OpenAI, and Alibaba reposition around compute constraints

The Anthropic-SpaceX deal reshapes the competitive landscape by giving Anthropic access to infrastructure originally built for xAI. This creates an awkward dynamic: Elon Musk sits on SpaceX's board, which is stacked with Musk empire loyalists, yet SpaceX is now enabling a direct competitor to xAI's Grok models. The three AI infrastructure lease agreements disclosed in SpaceX's IPO filing suggest that xAI itself is scaling back its capacity ambitions, freeing up compute for third parties. For OpenAI, the deal represents a threat and an opportunity. OpenAI benefits from Microsoft's Azure infrastructure, but the Anthropic-SpaceX arrangement gives Anthropic access to a dedicated, gigawatt-scale facility that is not shared with other tenants. That exclusivity translates into faster training cycles and lower latency for inference. Meanwhile, Alibaba's shift to proprietary models with Qwen3.7-Max signals a broader industry realignment. By restricting access to Chinese-based endpoints, Alibaba effectively cedes the US and European enterprise market to OpenAI, Anthropic, and Google. The departure of key Qwen Team leaders earlier this year suggests internal turmoil at Alibaba's AI division, which accelerated the move to a closed-source model. Alibaba now joins OpenAI and Google in offering its latest models exclusively via paid APIs, eroding the open-source advantage that once defined the Qwen family. The 35-hour autonomous execution capability of Qwen3.7-Max is impressive, but its geographic restrictions limit its commercial impact to the Chinese domestic market.

Downstream effects on hyperscalers, enterprise buyers, and GPU supply chains

The $15 billion annual compute deal has second-order effects that ripple across the entire AI supply chain. For hyperscalers like Google and Microsoft, the deal validates the thesis that dedicated AI infrastructure commands premium pricing. Google's TPU v5 clusters and Microsoft's Azure AI supercomputers now have a clear pricing benchmark. Enterprise buyers face a stark choice: pay $17.50 to $30.00 per million tokens for frontier models or accept lower-quality outputs from smaller, cheaper alternatives. This pricing dynamic will accelerate the adoption of model distillation and fine-tuning as enterprises seek to reduce their API bills. On the supply chain side, the deal locks up a significant portion of global GPU capacity for the next three years. SpaceX's Colossus facilities consume more than one gigawatt of power, equivalent to a small nuclear reactor. That power draw will strain local grids in the regions where Colossus is located, potentially triggering regulatory scrutiny. For GPU manufacturers like NVIDIA, the deal represents a guaranteed demand signal that supports their production planning. However, the exclusivity of the arrangement means that other AI startups will find it harder to secure large-scale compute, potentially consolidating the frontier model market around a handful of well-capitalized players. The three AI infrastructure lease agreements in SpaceX's IPO filing suggest that this model of dedicated, multi-year compute deals will become the norm, squeezing out smaller competitors who cannot commit to billion-dollar contracts. The deal locks up a significant portion of global GPU capacity for the next three years, and the exclusivity of the arrangement means that other AI startups will find it harder to secure large-scale compute.

Policy and strategy signal: the end of open-source AI and the rise of infrastructure monopolies

The convergence of Anthropic's compute deal and Alibaba's proprietary pivot sends a clear signal about the direction of AI policy and strategy. Open-source AI, which dominated the early frontier model era with releases like Meta's Llama and the original Qwen models, is giving way to a closed-source, infrastructure-driven paradigm. Alibaba's decision to make Qwen3.7-Max proprietary, influenced by the departure of key Qwen Team leaders, reflects a broader recognition that the economic value of frontier models lies in exclusivity, not openness. This shift has policy implications: regulators in the US and Europe have expressed concern about AI concentration, but the compute economics make concentration inevitable. The Anthropic-SpaceX deal effectively creates a vertically integrated compute monopoly, with SpaceX controlling the infrastructure and Anthropic controlling the model. Elon Musk's dual role as SpaceX board member and xAI founder raises conflict-of-interest questions that regulators will examine during SpaceX's IPO review. The deal also signals that AI companies are willing to pay extraordinary premiums for compute access, which will drive further investment in data center construction. For policymakers, the key question is whether this infrastructure arms race creates systemic risk: if a single data center failure or geopolitical disruption cuts off compute for a major AI player, the downstream effects on enterprise customers will be severe. The era of cheap, abundant AI compute is over, replaced by a world where access to silicon determines who wins.

The trajectory of AI development is now inextricably tied to infrastructure finance. Anthropic's $15 billion annual commitment to SpaceX is not an anomaly: it is a template. Every frontier model developer will need to secure similar compute arrangements or risk falling behind. The IPO-bound SpaceX will likely disclose additional AI infrastructure lease agreements as it courts public investors, revealing the full scale of this emerging asset class. For Alibaba, the shift to proprietary models with Qwen3.7-Max represents a strategic retreat from the global market, ceding ground to US-based competitors while focusing on the Chinese domestic market. The 35-hour autonomous execution capability of Qwen3.7-Max is impressive, but its geographic restrictions limit its commercial impact. The real action is in the compute deals, where billion-dollar monthly payments are becoming the cost of admission for AI leadership. Investors should watch for similar arrangements from OpenAI, Google, and xAI in the coming months, as each company races to lock down the infrastructure needed to train and deploy the next generation of models. The AI industry is entering a phase where balance sheet strength, not algorithmic innovation, will determine the winners.