Compute & Infrastructure

Anthropic has entered a compute agreement with SpaceX to access the full capacity of the Colossus 1 data centre — originally built by xAI for Grok model training — and has signalled interest in multi-gigawatt orbital compute capacity, according to Bloomberg and Data Centre Dynamics. The arrangement is confirmed as a signed agreement. The orbital compute reference remains in the 'expressed interest' category — no contracts or hardware have been confirmed for that component.

The strategic implication is significant: Anthropic, backed by Amazon and Google, is deliberately diversifying compute supply away from its own investors' infrastructure. Colossus 1's scale — originally reported at 100,000 Hopper-class GPUs — gives Anthropic a large, immediately available inference and training block. The orbital compute angle, while speculative, points toward a longer-term thesis that SpaceX's Starlink constellation could serve as a latency-tolerant distributed compute fabric. For the broader market, this deal signals that frontier AI labs are willing to pay a premium for capacity not controlled by their cloud equity partners, which has direct implications for AWS and Google Cloud's strategic leverage over Anthropic.

SpaceX has filed for permits for a semiconductor fabrication facility in rural Texas under the Terafab brand, with capital figures in the filing reaching $55 billion and total potential investment cited at $119 billion — versus the $20 billion announced by Musk in March, according to Tom's Hardware. This is a permit filing, not a confirmed construction contract or funded commitment. The gap between the March announcement figure and the filing figures is large enough to warrant scepticism about whether the higher numbers represent firm plans or regulatory envelope-maximising.

If even a fraction of the stated investment materialises, Terafab would represent one of the largest single semiconductor infrastructure commitments in US history, rivalling TSMC's Arizona and Intel's Ohio investments combined. The strategic logic is clear: SpaceX's Starshield, Starlink, and xAI GPU clusters all require chips at scale, and domestic production insulates those programs from export control exposure. However, the absence of a disclosed process node partner, equipment supplier agreements, or technology licensing arrangements means this project is firmly in the speculative column. Permitting timelines for projects of this scale typically run 2-4 years before ground breaks.

Denmark's Energinet grid operator has suspended new data centre connection requests after accumulated demand applications reached 60 GW against a national grid capacity base that cannot absorb that volume, according to Tom's Hardware. Denmark joins Ireland, the Netherlands, and parts of the UK in imposing hard grid connection freezes, a pattern that is now systemic across Northern European markets that had previously been preferred for data centre buildout due to cool climates and renewable energy availability.

The 60 GW figure demands context: Denmark's total installed generation capacity is approximately 20 GW, meaning the pipeline of data centre requests alone represents three times the country's entire grid. This is not a near-term capacity problem — it reflects speculative land-banking by operators filing connections to secure optionality. Nevertheless, the freeze creates real delays for operators with genuine build plans. The University of Southern Denmark simultaneously brought an AI supercomputer online in Sønderborg with waste heat recovery feeding district heating — a model that demonstrates what compliant, grid-integrated AI infrastructure looks like, and one likely to receive regulatory preference during connection rationing.

NVIDIA has committed $300 million to Corning to fund construction of three new US-based optical fibre manufacturing plants, a deal that would increase domestic fibre production capacity by over 50%, according to Tom's Hardware. This is a confirmed investment, not a letter of intent. The strategic framing is explicit: NVIDIA wants its US-based deployment partners — hyperscalers and co-lo operators building NVL-scale clusters — to have reliable domestic fibre supply without exposure to import disruption.

This move complements NVIDIA's Spectrum-X Ethernet MRC protocol, a custom RDMA transport layer already deployed in gigascale AI clusters, which ServeTheHome reports is designed specifically for the latency and bandwidth requirements of multi-thousand GPU training runs. Together, the fibre investment and proprietary network protocol mean NVIDIA is progressively owning the physical and logical interconnect stack surrounding its GPUs — silicon, networking ASICs, transport protocol, and now the cable plant itself. This creates switching costs well beyond the GPU purchase decision.

AMD posted record Q1 results driven by data centre CPU demand, with the company already developing Zen 7 architecture and planning a more specialised EPYC portfolio targeting distinct AI and cloud workloads, per Tom's Hardware. AMD simultaneously warned that consumer and gaming revenue will decline in Q2, a divergence that confirms AI infrastructure is now the only structurally growing segment of the semiconductor market. The global semiconductor industry hit $298.5 billion in Q1 sales, according to Semiconductor Industry Association data cited by Tom's Hardware, putting the full year on track to exceed $1 trillion.

Apple's quiet discontinuation of 128GB Mac Studio and earlier 512GB models due to supply constraints — confirmed by Tom's Hardware — illustrates the downstream effect: high-bandwidth memory and advanced packaging capacity is being allocated toward AI accelerators at the expense of consumer products. Arm's guidance simultaneously flagged smartphone market weakness while projecting AI data centre growth as the compensating driver. The pattern is consistent across the supply chain: memory, packaging, advanced nodes, and interconnects are all flowing toward AI infrastructure at the cost of consumer and mobile segments.