Compute & Infrastructure

Google is set to announce a new TPU generation this week explicitly optimised for inference — the serving of deployed models at scale — rather than training, where Nvidia's H100 and H200 dominance has been most entrenched. According to Bloomberg, leading AI developers including Google's direct competitors are already procuring existing TPU generations, a striking shift given the historical reluctance of hyperscalers to build on a rival's proprietary silicon. The inference market is structurally different from training: workloads are more latency-sensitive, more continuous, and ultimately larger in aggregate compute terms as model deployment scales. Custom inference silicon can achieve significant efficiency gains over general-purpose GPUs, and Google's tight integration of TPU design with its own model development pipeline — particularly Gemini — gives it an architectural feedback loop Nvidia cannot replicate.

The strategic implication is that Google is no longer simply hedging against Nvidia dependency for internal use; it is positioning TPUs as an external compute product competing in the same market as Nvidia's accelerators. If the new generation achieves competitive inference performance per watt and per dollar, it introduces genuine pricing pressure on Nvidia at the inference layer precisely as inference becomes the dominant workload category. The risk for Nvidia is structural: training clusters are built infrequently, but inference infrastructure is procured continuously.

AI-related silicon is approaching one-third of TSMC's total revenue, according to Next Platform. This is a confirmed directional shift in the foundry's revenue mix, not an analyst projection, and its implications are layered. On one side, it cements TSMC's indispensability to the AI economy and gives it leverage to maintain premium pricing on advanced node capacity — particularly CoWoS advanced packaging, which remains the binding constraint for high-bandwidth memory integration on AI accelerators. On the other side, it creates a meaningful concentration risk: a sustained pullback in AI capex, whether from regulatory disruption, a model capability plateau, or a macro correction, would now have direct and outsized impact on TSMC's financials.

The packaging bottleneck deserves particular emphasis. TSMC's CoWoS capacity — essential for stacking HBM on logic dies in GPUs and TPUs — has been the most acute supply constraint in the AI hardware chain for the past 18 months. Expansion is underway but capital-intensive and slow. Until alternative packaging suppliers such as ASE and Amkor can absorb meaningful share, TSMC remains the single point of failure for next-generation AI accelerator volume.

Two converging pressures documented this week confirm that the UK's AI infrastructure buildout is running into hard physical limits. First, The Register reports that AI datacenter capacity is migrating away from London as power shortages and planning constraints — combined with reduced latency dependency on London's financial sector — make regional sites more competitive. This is a structural geographic shift, not a cyclical adjustment: once grid connections are secured in secondary markets and operators prove that AI inference does not require sub-millisecond proximity to financial exchanges, the economic case for London's premium land and power costs weakens significantly.

Simultaneously, a UK parliamentary committee has launched a formal inquiry into low-energy chip architectures — neuromorphic, analogue, and sparse computing designs — as a potential mitigation for grid bottlenecks, according to The Register. This is an early-stage policy signal rather than a near-term market intervention, but it indicates that UK policymakers are treating energy intensity as a binding constraint requiring a hardware-level solution, not just a grid investment problem. The inquiry's findings could influence domestic R&D funding priorities and procurement criteria for public sector AI infrastructure.

Anthropic is actively negotiating datacenter leasing deals in Europe and Australia, according to Data Center Dynamics, at the same moment OpenAI is pulling back from its Stargate Europe commitments. The contrast is instructive: rather than building or co-investing in owned infrastructure as OpenAI's Stargate model envisioned, Anthropic is pursuing a leased colocation model that preserves capital flexibility while establishing jurisdictional presence. For European and Australian governments prioritising data sovereignty and local AI capacity, Anthropic's demand represents an opportunity to anchor international AI operators to domestic infrastructure — particularly in Australia, where government investment in sovereign AI compute has been explicit policy.

The leasing approach also reflects a different risk calculus than hyperscale build-to-own models. Anthropic does not operate a general-purpose cloud business; its infrastructure needs are inference-dominated and tied to commercial API demand that remains forecast-dependent. Leasing rather than building is rational under uncertainty, but it means Anthropic's geographic expansion is contingent on colocation operators having available AI-grade capacity — dense power, liquid cooling readiness, and sufficient fibre connectivity — in the target markets.

Victory Giant Technology, a Huizhou-based printed circuit board manufacturer and Nvidia supplier, raised $2.6 billion in Hong Kong's largest IPO in seven months and surged 60% on debut, according to Bloomberg. PCBs are a foundational but often overlooked layer of the AI hardware stack — every GPU, switch, and server requires high-layer-count, high-speed boards capable of handling the signal integrity demands of modern AI accelerators. Victory Giant's founder has publicly projected high-speed growth for five years, a timeline consistent with the multi-year capex commitments hyperscalers have made to AI infrastructure.

The listing's scale and market reception confirm that capital markets are pricing AI infrastructure exposure well below the silicon layer — into substrates, packaging, and interconnect. For supply chain analysts, this is a useful leading indicator: a PCB manufacturer at this valuation implies sustained conviction in AI server build rates, not a near-term inventory correction. It also highlights a geopolitical dimension — a Chinese supplier deeply embedded in Nvidia's supply chain raises questions about how export controls and supply chain diversification pressures interact with component-level dependencies that are harder to monitor than chip shipments.