Compute & Infrastructure

Three converging developments this week reveal the full architecture of Anthropic's compute strategy. First, Bloomberg reported that Google has agreed to backstop lease payments at five data centers, enabling Anthropic to access what amounts to a $35 billion loan against chip infrastructure it does not own. This is a confirmed financial arrangement, not a speculative plan — Google's credit support is the mechanism that makes the deal financeable at this scale.

Simultaneously, Data Center Dynamics confirms that Broadcom, Apollo, and Blackstone have formally launched a 20GW XPU platform with an initial $35 billion tranche, with Anthropic as the anchor customer via infrastructure operator Fluidstack. The dollar figures align across both reports, suggesting this is the same underlying transaction viewed from different angles — Google providing demand-side credit backstop while Apollo and Blackstone provide supply-side capital for the physical build. Separately, Data Center Dynamics reports that Clive Chan, who led OpenAI's custom chip program working on a Broadcom ASIC expected later this year, has now joined Anthropic — directly relevant given Broadcom's centrality to the XPU platform. Anthropic is simultaneously securing third-party compute at massive scale while building the internal talent to reduce that dependency over time.

Taiwan is considering extending AI chip export restrictions beyond currently blacklisted Chinese entities to cover all customers in China, with criminal penalties for violations, according to Bloomberg and confirmed in additional detail by Tom's Hardware. This is a proposed measure under active consideration — not yet enacted — but the direction of travel is unambiguous. The motivation is twofold: aligning with US BIS controls to reduce diplomatic friction with Washington, and closing the grey-market channels through which advanced chips have continued to reach Chinese AI labs via third-country intermediaries.

The strategic significance is that Taiwan's involvement transforms export control from a US bilateral tool into a multilateral chokepoint. TSMC already restricts production of advanced nodes for Chinese customers under US pressure, but sales of existing inventory and lower-node chips have continued to flow. Criminal liability for any China sale would make Taiwanese fabless designers, distributors, and system integrators legally exposed — a qualitatively different deterrent. Beijing's likely response will be to accelerate SMIC capacity and domestic EDA tool development, but those programs remain years behind on leading-edge nodes.

Bloomberg reports TSMC's monthly revenue rose 30% year-on-year, driven explicitly by AI chip demand. This is confirmed revenue data, not a forecast. The sustained growth rate — following multiple consecutive strong quarters — validates that demand is not front-loaded inventory building but reflects genuine end-use consumption across training and inference deployments.

The concentration risk this represents cannot be overstated. TSMC manufactures the overwhelming majority of leading-edge AI accelerators including NVIDIA's Blackwell series, AMD's MI-series, and the custom ASICs being developed by hyperscalers. No meaningful alternative exists at 3nm and below. Intel Foundry Services remains subscale for AI workloads, and Samsung's advanced node yield rates have not reached competitive parity. Any disruption to TSMC — geopolitical, natural disaster, or supply chain — propagates immediately into the entire AI buildout. The proposed Taiwan export controls add a further policy variable to this concentration.

Meta has confirmed a partnership with Reliance Industries to build its first AI data center in India, per Bloomberg. The Reliance partnership gives Meta access to land, power procurement relationships, and regulatory navigation in a market where foreign infrastructure ownership faces friction — the same logic that drove Microsoft and Google to partner with local conglomerates in their India expansions. This is a confirmed partnership announcement; construction timeline and capacity have not been disclosed publicly.

In the UK, Data Center Dynamics reports that Civo and Era4 have launched a sovereign AI cloud product housed in Era4 facilities. While modest in scale, it reflects the growing regulatory and procurement demand from UK public sector and financial services clients who require data residency guarantees. Separately, IBM CEO Arvind Krishna confirmed a $1 billion federal commitment to build a chip fabrication facility in Albany, New York, per Bloomberg. This is a government-backed facility, but the timeline to meaningful advanced-node production capacity from a greenfield US fab is measured in years, not quarters.

SpaceX has unveiled an 11-million-square-foot Gigasat factory and announced targets of 1 GW of space-based AI compute by late 2027 and 100 GW per year by 2030, according to Tom's Hardware. The factory announcement is confirmed; the compute capacity targets are Elon Musk's stated projections and should be treated as aspirational. The 100 GW by 2030 figure is implausible on current satellite manufacturing and launch cadence trajectories and appears designed to signal strategic intent rather than reflect operational planning.

The strategic logic is real even if the numbers are inflated. Orbital compute circumvents the two most acute terrestrial bottlenecks — power grid interconnection queues and land availability near population centres. Solidion's announcement of a battery platform specifically designed for LEO-based AI data centers, reported by Data Center Dynamics, suggests a nascent supplier ecosystem forming around this use case. However, latency constraints, launch costs per kilowatt, thermal management in vacuum, and radiation hardening requirements for AI accelerators at scale remain unresolved engineering challenges.