Compute & Infrastructure

Anthropic's $11.6 billion compute contract with Akamai — confirmed as the largest deal in the CDN-turned-cloud-infrastructure firm's history — is analytically significant for two reasons beyond its headline size. First, it demonstrates that frontier AI labs are actively diversifying away from the AWS-Google-Azure hyperscaler triumvirate to secure guaranteed capacity at negotiated rates. Akamai's distributed edge infrastructure gives Anthropic leverage on inference latency and geographic distribution that pure hyperscaler arrangements do not readily offer. Second, the deal validates Akamai's multi-year pivot toward AI compute as a core product, not a value-add — a strategic repositioning that now has a credible anchor customer. Bloomberg

The deal structure and timeline have not been publicly disclosed, so it remains unclear how much capacity is pre-committed versus on-demand, and over what contract period the $11.6 billion is spread. Analysts should watch whether this prompts competing labs — particularly Google DeepMind and Meta AI — to pursue similar non-hyperscaler arrangements, which would materially alter the revenue assumptions underpinning AWS and Azure's AI infrastructure growth forecasts.

Elon Musk has confirmed that Colossus 2 will receive 220,000 NVIDIA GB300 GPUs by next week, with two further tranches of identical size expected before year-end — bringing the site to over one million AI GPUs and finally hitting a target Musk set two years ago. Tom's Hardware The total GPU count at full build-out — 1.44 million — would represent one of the highest concentrations of AI compute under single-operator control globally.

The critical qualifier is power: xAI is constructing a 1.2-gigawatt power plant to bring systems fully online. Until that plant is commissioned, the operational GPU count is constrained by available grid and on-site generation capacity, not hardware supply. This is the defining infrastructure bottleneck pattern of 2025-2026: hardware delivery timelines have compressed significantly, but power interconnection and on-site generation remain multi-year processes. The GB300's power profile — each GPU drawing substantially more than its predecessor — means that 1.44 million units at full utilisation represents a power demand that cannot be met from existing grid connections at most sites.

Tower Semiconductor's $4 billion investment in Japanese operations, targeting a 40x increase in optical connectivity output by 2029 versus 2025 baseline levels, is a direct play on silicon photonics becoming the dominant interconnect technology for AI accelerator clusters and data centre networking. Tom's Hardware As GPU cluster sizes scale — illustrated by xAI's million-plus GPU buildout — copper interconnects face fundamental bandwidth-distance limitations, making co-packaged optics and silicon photonics a near-mandatory upgrade path for next-generation AI infrastructure.

Japan is a deliberate geographic and geopolitical choice: it positions Tower away from US-China export control crossfire, aligns with Japan's national semiconductor strategy, and accesses a mature precision manufacturing ecosystem. The 40x output target by 2029 is an announced plan, not confirmed capacity; the investment commitment and site selection decisions carry more weight than the output projection at this stage. This investment also reinforces the emerging supply chain reality that optical connectivity — not just logic chips or HBM — is becoming a strategic bottleneck requiring deliberate sovereign and corporate capacity planning.

New Jersey's $1.07 million fine and 45-day shutdown deadline issued to the DataOne facility — linked to Microsoft — for operating 62 unpermitted portable turbine generators marks a meaningful escalation in how local regulators are engaging with data centre power infrastructure. Tom's Hardware The record fine quantum is less significant than the shutdown threat: a 45-day deadline to comply or cease operations represents the kind of operational risk that infrastructure operators and their cloud customers take seriously.

The pattern here is instructive. Data centres have been deploying diesel and gas turbine generators at pace to bridge the gap between grid interconnection delays and hardware delivery timelines — a workaround that regulators in multiple US states and European jurisdictions are now scrutinising. The combination of noise pollution complaints, air quality permitting requirements, and community opposition is creating a compliance surface area that operators systematically underestimated during the 2024-2026 buildout acceleration. This is a structural risk for the sector: the fastest path to operational capacity increasingly runs through regulatory friction that cannot be solved with capital alone.

Google is launching an experimental satellite on October 1 carrying four tensor processing units and approximately 1,000W of solar power capacity. Tom's Hardware At four TPUs, this is explicitly a proof-of-concept for operating AI accelerators in the space environment — thermal management, radiation tolerance, and power cycling under orbital conditions — rather than a capacity play. The 1,000W solar budget is negligible relative to any production AI workload.

The strategic significance is the research agenda it signals. Space-based compute eliminates terrestrial power and land constraints and could serve latency-sensitive inference at global scale in a configuration that bypasses sovereign jurisdiction in novel ways. These are decade-horizon considerations, but Google's willingness to commit hardware to orbit for experimentation indicates the company is stress-testing infrastructure assumptions well beyond the 2030 planning horizon that governs most current data centre investment decisions.