Compute & Infrastructure

Anthropic has executed an $11.6 billion, seven-year contract with Akamai Technologies for computing power, one of the largest infrastructure commitments by a frontier AI developer outside the major hyperscalers. The deal adds to a growing portfolio of Anthropic data centre arrangements and suggests the company is deliberately distributing its compute dependencies rather than consolidating on a single cloud provider. Akamai, historically a content delivery and edge network operator, has been aggressively expanding into cloud compute infrastructure, and this contract validates that strategic pivot with a marquee AI customer. Bloomberg

The strategic logic for Anthropic is clear: concentration risk on any single hyperscaler creates both pricing leverage and potential supply constraints during peak demand periods. A seven-year term also locks in capacity at current rates, hedging against the continued surge in GPU cluster pricing. For the broader market, this deal reinforces the emerging pattern of non-hyperscaler infrastructure providers winning meaningful AI workloads — a structural shift that challenges the assumption that AWS, Azure, and Google Cloud will capture the entirety of frontier AI compute spend.

Google is launching a satellite equipped with its Tensor AI processors next week to evaluate in-space performance under radiation, thermal cycling, and vacuum conditions. This is a confirmed launch under Project Suncatcher, Google's initiative exploring orbital AI infrastructure. The test is explicitly preliminary — assessing whether consumer-grade AI silicon can survive the space environment — but it constitutes real hardware in orbit, not a conceptual proposal. The Verge

The strategic rationale for orbital data centres includes freedom from terrestrial power grid constraints, potential access to continuous solar energy, and elimination of land permitting and water cooling dependencies that are increasingly blocking ground-based buildout. However, the technical and economic barriers remain substantial: radiation hardening, latency to ground stations, launch costs, and maintenance access are all unresolved at commercial scale. This launch should be read as a long-horizon R&D investment, not near-term infrastructure capacity. The gap between a single test satellite and a functioning orbital compute cluster is measured in decades and hundreds of billions of dollars.

A leading semiconductor analyst has publicly called for AMD to be investigated for treason following evidence that export-controlled RFSoC adaptive radio platform chips — valued at approximately $36,000 per unit — were being quoted at around $1,000 for a Chinese crowdfunding project. AMD has attributed the availability to supply chain diversion rather than direct sales, a distinction with significant legal and geopolitical weight. Tom's Hardware

This episode is analytically important beyond the AMD-specific allegation. It illustrates that export controls on advanced semiconductors are only as effective as enforcement mechanisms downstream of the initial sale — and that grey-market diversion through third-country intermediaries remains a systematic vulnerability in the US technology restriction regime. The dramatic price differential (36x markdown) suggests a well-organised diversion channel rather than opportunistic single-unit resale. For infrastructure professionals, this matters because it raises questions about whether controls on AI-relevant chips — including high-bandwidth memory and advanced logic — are actually constraining Chinese AI compute buildout at the pace Washington assumes.

Micron is discontinuing its 2GB GDDR7 memory chips used in consumer gaming GPUs and redirecting production capacity toward 3GB high-density GDDR7 dies targeting AI and professional GPU applications. The higher-margin AI-focused memory commands significantly better economics, making the pivot financially rational. Tom's Hardware

This is a concrete example of AI demand structurally reshaping semiconductor production priorities rather than simply adding incremental capacity. Memory is a critical constraint in AI infrastructure — HBM for training accelerators and high-density GDDR for inference — and the willingness of Micron to sacrifice consumer GPU market share for AI margins confirms that memory manufacturers view AI as the dominant long-cycle demand driver. The secondary effect is potential supply tightening for consumer graphics cards, but the primary infrastructure signal is that AI compute's appetite for high-bandwidth, high-density memory is now large enough to redirect a major memory manufacturer's product roadmap.

Tower Semiconductor is establishing a large-scale optical chip manufacturing hub in Japan as part of a $4 billion investment pledge, with the CEO providing expanded details on plans announced earlier this year. Silicon photonics — the technology underpinning optical chip interconnects — is increasingly recognised as a structural bottleneck for scaling AI data centres beyond current GPU cluster densities. Electrical interconnects face fundamental bandwidth and power efficiency limits at the rack and pod scale that optical interconnects can circumvent. Data Centre Dynamics

The Japan location is strategically significant. It aligns with Japan's sovereign semiconductor strategy — which includes the Rapidus advanced logic fab and substantial government subsidies for foreign chipmakers establishing domestic capacity — while also positioning Tower outside both US and Taiwan geopolitical risk concentrations. For AI infrastructure planners, expanded silicon photonics capacity matters because co-packaged optics and optical interconnect fabrics are on the critical path for next-generation GPU clusters operating at 100kW-plus per rack densities. The investment is confirmed as a pledge with CEO-level elaboration, but the buildout timeline and production ramp details remain to be specified.