Compute & Infrastructure

Anthropic has signed a confirmed $1.8bn cloud services contract with Akamai, according to reporting by Data Center Dynamics. This is notable not for its scale alone but for its counterparty: Akamai is an edge and content delivery network provider, not a hyperscaler. The deal implies Anthropic is deliberately constructing a multi-vendor compute strategy, reducing concentration risk on AWS, Google Cloud, and Azure while gaining distributed inference capacity closer to end users.

For the infrastructure market, this validates edge-cloud providers as a credible tier in AI deployment architecture — not merely for latency optimisation but as a strategic hedge. Expect other frontier AI labs to explore similar arrangements as hyperscaler capacity remains constrained and pricing power stays elevated. The contract value also implies significant committed GPU or accelerator capacity being provisioned through Akamai's network, the specifics of which are not yet public.

SoftBank founder Masayoshi Son is in active talks with French President Macron about a major AI data centre project in France, according to Bloomberg. This remains unconfirmed — sources describe ongoing negotiations ahead of a potential joint announcement in coming weeks. Simultaneously, SoftBank has confirmed the launch of a Japanese battery storage business tied to a production plant that will house an AI data centre, with production expected to begin in 2028, per Data Center Dynamics.

The dual moves reveal SoftBank's infrastructure thesis: co-locate energy storage with compute to solve power reliability constraints, and pursue sovereign partnerships in Europe to access both capital and regulatory goodwill. The France play is clearly a response to European governments' appetite for domestic AI infrastructure after years of dependence on US hyperscalers. However, SoftBank's historical pattern of announcing ambitious infrastructure deals that take years to materialise — or do not — means the France project should be treated as speculative until contracts are signed.

A confirmed structural trend is now well documented: AI data centre developers are systematically targeting unincorporated rural land to avoid municipal permitting, rezoning votes, and land-use reviews, according to Tom's Hardware. Rural parcels also offer the physical scale needed for gigawatt-class campus designs that urban footprints cannot accommodate.

The infrastructure risk calculus here is not straightforward. While regulatory speed increases, rural sites introduce transmission interconnection challenges — rural grids are frequently under-built for the load profiles of hyperscale AI facilities. Developers are effectively trading regulatory friction for grid infrastructure gaps, and in many cases are being required to fund transmission upgrades themselves. This is compressing margins and extending actual power-on timelines even as groundbreaking accelerates.

Cerebras has increased its IPO target by one-third to $4.8bn, according to Bloomberg. This is a confirmed upward revision to a live offering, not a speculative valuation — it reflects real institutional demand. Simultaneously, analyst data cited by Tom's Hardware reveals that Arm's AGI CPU order book has doubled to $2bn over the next two years in just 1.5 months — yet this still represents below 5% of overall data centre CPU market share, underlining how entrenched x86 and NVIDIA GPU architectures remain.

These two data points together indicate a market in early-stage architectural fragmentation: investors and hyperscalers are placing meaningful but not dominant bets on alternative compute architectures. Cerebras' wafer-scale chip design targets training and inference workloads where memory bandwidth is the constraint; Arm's AGI CPUs target orchestration and inference at the chip level. Neither displaces NVIDIA in the near term, but the combined capital flowing into alternatives is now large enough to sustain a parallel ecosystem.

Goldman Sachs has flagged that South Korea and Taiwan's AI-driven chip export booms are generating K-shaped economic divergence within each economy — benefiting semiconductor sectors while pressuring central banks to raise rates as trade surpluses widen, according to Bloomberg. JPMorgan has separately raised its bull-case Kospi target to 10,000, explicitly citing the memory semiconductor cycle and AI demand, per Bloomberg.

From a supply chain perspective, this concentration of AI hardware production in two jurisdictions — TSMC in Taiwan for leading-edge logic, Samsung and SK Hynix in Korea for HBM memory — represents a systemic geographic risk that the AI infrastructure buildout is actively deepening rather than diversifying. China's export rebound, partly driven by AI investment-related trade flows per Bloomberg, adds a further dimension: the global AI hardware supply chain remains deeply intertwined across geopolitically contested jurisdictions.