Compute & Infrastructure

Anthropic has agreed to a $35 billion cloud computing deal with Lambda, an Nvidia-backed cloud provider, according to Bloomberg. The scale is without precedent in AI infrastructure procurement and reflects a deliberate choice by Anthropic to expand capacity outside the dominant hyperscaler tier — AWS, Azure, and GCP — rather than deepen dependence on any single cloud partner. Lambda's Nvidia backing is directly relevant: it ensures priority access to GPU allocation and positions the deal as part of a vertically coordinated compute stack running from chip manufacturer through cloud provider to AI lab.

This deal has compounding strategic implications. First, it establishes a new benchmark for what frontier AI labs require in infrastructure commitments to sustain training and inference at scale. Second, it signals that Nvidia's ecosystem strategy extends beyond hardware sales into cloud financing and capacity provisioning — Lambda functions as an Nvidia-aligned distribution channel for compute. Third, with Anthropic reportedly preparing for a public filing in coming weeks, locking in multi-year compute capacity at this scale is a balance-sheet move as much as an operational one, providing investors with visibility on infrastructure cost and supply security.

Nvidia has made its largest-ever direct investment outside the United States in Taiwan's MediaTek, according to Bloomberg. MediaTek's shares rose approximately 10% on the news. The investment is explicitly framed as a deepening of collaboration between the two firms, and should be read in the context of mounting pressure on Nvidia's market position from hyperscaler-developed custom silicon — Google's TPUs, Amazon's Trainium, Microsoft's Maia, and Meta's MTIA all represent direct substitution risk.

MediaTek brings specific capabilities that complement Nvidia's position: strong SoC design expertise, deep relationships with TSMC, and a growing footprint in edge AI and automotive silicon. By elevating MediaTek into what Bloomberg describes as an exclusive AI chipmaker group, Nvidia is constructing a broader design ecosystem that can address market segments where its own products are either too expensive, too power-hungry, or export-restricted. This investment also has a geopolitical dimension — anchoring a major Taiwanese chipmaker more firmly within the Nvidia orbit strengthens supply chain alignment and creates a political-economy argument for Taiwan's continued role in Nvidia's supply chain against US legislative pressure to onshore production.

Taiwanese prosecutors have raided Unimicron Technology, one of the world's largest PCB and advanced substrate manufacturers, over allegations that it relabelled China-manufactured printed circuit boards as Taiwanese-origin products to avoid US tariffs, according to Tom's Hardware. Unimicron is a direct supplier to Nvidia, Intel, Google, and Amazon. The alleged practice — commonly called origin washing — would expose the company to a 40% US tariff penalty if confirmed.

This development matters beyond the legal and compliance dimensions. Advanced chip substrates are already a recognised chokepoint in AI hardware supply chains — substrate capacity constraints contributed to GPU supply delays in 2023 and 2024. Unimicron is not a peripheral supplier; it makes the ABF substrates and PCBs that sit directly beneath leading-edge processors. Any disruption to its operations — whether through penalties, loss of export privileges, or reputational damage causing customer diversification — would propagate rapidly into GPU and server board supply timelines. The probe also raises the question of how widespread origin-washing practices are across Taiwan's electronics supply chain under current tariff pressure, and whether other substrate or component makers face similar exposure.

SK Hynix is assessing a joint venture to manufacture memory chips in Japan, according to Bloomberg, framed as one of several options to meet AI-driven HBM demand while managing production costs. This is a studied feasibility, not a confirmed investment. Simultaneously, two separate academic research groups — one from Huawei, ETH Zürich, and HUST, the other from the University of Oxford — have published papers on High-Bandwidth Flash architectures for LLM inference, both available via Semiconductor Engineering.

The research convergence is significant. Both papers address the same fundamental constraint: HBM capacity per accelerator is becoming the binding limit on LLM inference, not compute throughput. High-Bandwidth Flash offers approximately 16x more capacity per stack at comparable bandwidth, and the Oxford research demonstrates that a hybrid HBM-HBF architecture can address capacity bottlenecks while preserving bandwidth for latency-sensitive operations. If this architecture matures toward productisation — particularly if Huawei is developing it for use in its own AI accelerators — it would reduce dependence on HBM, alleviate pressure on SK Hynix and Samsung as the near-monopoly HBM suppliers, and potentially reshape the memory market that currently commands the highest margins in the AI hardware stack.

A detailed analysis in IEEE Spectrum documents the structural contradiction at the heart of EU technology policy: the bloc's accelerating investment in AI factories, gigafactories, and data centres is increasing demand for advanced semiconductors that Europe produces at less than 10% of global output and cannot manufacture domestically at leading-edge nodes. The EU Chips Act targeted 20% of global chip production by 2030, a goal that analysts now widely regard as unachievable on current trajectories.

The practical consequence is that EU AI sovereignty initiatives are, in the near term, functioning as demand subsidies for TSMC, Samsung, and US chip designers — the precise suppliers that the sovereignty agenda is meant to reduce reliance upon. Intel's planned German fab has faced delays and cost escalations. TSMC's Dresden facility is under construction but will produce 28nm nodes, not the sub-3nm process required for frontier AI accelerators. The EU's position is structurally similar to its energy dependency situation pre-2022: policy ambition is real, but the industrial base to support it is a decade behind.