Anthropic locks $10b compute deals as AI cloud wars heat up

Anthropic is aggressively broadening its AI compute partnerships well beyond its initial collaboration with SpaceX, securing monumental deals with industry giants such as Meta and Volta. The proposed $10 billion agreement with Meta underscores Mark Zuckerberg’s ambitions to challenge AWS and Google in cloud AI infrastructure, while the $10 billion deal with Volta—a neocloud renting colocation from bitcoin miner Bitdeer—highlights Anthropic’s strategic diversification into emerging and unconventional AI infrastructure providers. These moves reflect Anthropic’s intent to cement its footprint across a variety of compute ecosystems amid intensifying competition.

Despite ideological differences and past public criticisms between Anthropic and Elon Musk’s ventures, pragmatism prevails as Anthropic commits approximately $1.25 billion per month to SpaceX for AI compute resources, leveraging Musk’s vast pool of 220,000 idle GPUs. This partnership exemplifies how critical infrastructure access trumps rivalry in the fiercely competitive AI landscape, where compute scarcity drives unlikely alliances. Anthropic’s deals with SpaceX, alongside other major players like Google and Amazon, contribute to a staggering $47.5 billion backlog in AI compute contracts, signaling a consolidation of power among a select few compute 'haves' poised to dominate the market.

Anthropic’s supply chain intricately intertwines with the bitcoin mining sector, as evidenced by its $15 billion data-center debt being offloaded by banks with Google’s backstop activating only post-construction. This crossover between mining and AI infrastructure not only reflects a broader industry pivot but also underscores the growing complexity and scale of Anthropic’s compute procurement strategies. The convergence of these sectors illustrates how AI compute scarcity is reshaping traditional infrastructure financing and operational models.

The AI compute market’s increasing concentration is creating a stark divide between the resource-rich and the compute-starved, with Anthropic exemplifying the necessity of securing access to hundreds of thousands of GPUs to remain competitive. Smaller AI firms lacking such scale face the risk of being squeezed out as the compute 'haves' pull further ahead, intensifying the market’s winner-takes-all dynamics. Anthropic’s multi-billion dollar deals and diverse partnerships position it firmly within this elite tier, highlighting the critical role of massive infrastructure access in the future of AI development.

SpaceX is aggressively positioning itself as a neocloud AI infrastructure provider by uniquely integrating its extensive satellite constellation, terrestrial data centers, and launch capabilities. Leveraging the proven reliability and profitability of its 10,000 Starlink satellites—which already incorporate compute, laser-linked data transmission, and power systems—SpaceX is scaling this model to develop orbital AI data centers in partnership with NVIDIA, aiming to deploy solar-powered AI compute satellites like the upcoming Starmind network with up to a million GPUs in orbit. This vertical integration, combining rocket manufacturing, satellite operations, and AI compute layers, sets SpaceX apart from traditional cloud and aerospace companies, enabling it to process data closer to its source in space and reduce latency and transmission inefficiencies.

SpaceX’s terrestrial AI infrastructure, anchored by its rapidly expanding Colossus data centers in Memphis and Texas, complements its orbital ambitions and fuels a booming AI compute rental business. With contracts granting Anthropic access to roughly 325,000 Nvidia GPUs for $1.25 billion per month and Google leasing 110,000 GPUs at $920 million monthly, SpaceX is generating a staggering $26 billion annually from AI compute rentals alone. CFO Brett Johnson highlights a less than one-year payback on these capital deployments, underscoring strong demand amid a supply-constrained market, while the company’s talks with the Pentagon for ground data center capacity further cement its foothold in terrestrial AI compute markets.

Despite the promising growth and strong backlog, SpaceX’s ambitious AI infrastructure expansion demands massive capital expenditures, with spending more than doubling from $7.7 billion in Q1 to over $15 billion in Q2 2026, heavily weighted toward AI compute. While this aggressive investment strategy has driven a 247% increase in AI segment revenue to $2.56 billion, it has also raised investor concerns about cash runway and execution risks, especially given the complexities of scaling orbital data centers and the critical reliance on the Starship program as the transport monopoly for deploying these costly payloads. Notably, Jamie Dimon’s rare endorsement of SpaceX’s orbital AI concept lends institutional credibility, even as regulatory and capital deployment challenges loom.

SpaceX’s strategic vision extends beyond conventional cloud computing by pioneering orbital AI data centers that represent the next evolution of edge computing, processing data in orbit to drastically reduce the inefficiencies of transmitting vast satellite data back to Earth. Collaborating exclusively with NVIDIA on the Vera Rubin architecture, SpaceX plans to begin launching prototype AI compute satellites as early as next year, aiming to scale gigawatt-level orbital compute capacity. This approach not only leverages SpaceX’s unique assets—reusable rockets, satellite manufacturing, and the Starlink network—but also positions the company to outpace hyperscalers in speed and scale, exemplified by xAI’s rapid GPU cluster deployments that dwarf traditional timelines.

Premium AI data center leases such as SpaceX's Colossus command extraordinarily high rates—around $50 billion per gigawatt—due to the scarcity and immediacy of large-scale, coherent GPU capacity optimized for heavy AI workloads. These massive, tightly interconnected clusters reduce communication overhead and enable higher utilization compared to smaller, fragmented, or multi-tenant neo-cloud blocks, justifying their substantial premium in the market.

Colossus-style leases typically clear at a roughly 32% premium over one-year H100 contracts, reflecting their short-term availability and the urgent demand for large coherent clusters, whereas neo-cloud mega-deals often span five years or more with prepayments and predictable offtake, resulting in lower effective annual rates. This dynamic underscores how immediacy and scale drive pricing disparities in AI compute leasing.

The high cost structure of AI-optimized data centers, with shell and power infrastructure alone ranging from $10 to $20 million per megawatt due to liquid cooling and high-density IT fit-outs, further elevates lease rates and sustains scarcity rents in the market. Despite cancellation flexibilities such as 90-day notices, the constrained supply of premium AI compute capacity ensures that spot and short-term rates frequently exceed longer-term contracted prices.