Compute & Infrastructure

ASML raised its full-year 2026 revenue guidance on the back of sustained AI infrastructure investment, according to Bloomberg. The revision is analytically significant not just as a financial update but as a real-time demand signal: ASML's EUV and high-NA EUV systems are the irreplaceable bottleneck in producing leading-edge logic at 3nm and below, and their order book is a direct proxy for how seriously the industry is treating AI compute buildout. No alternative supplier exists for high-NA EUV at any price.

The forecast upgrade comes despite ongoing US export controls that restrict ASML's ability to ship advanced tools to China. That those restrictions have not suppressed ASML's growth trajectory confirms that demand from TSMC, Samsung, Intel Foundry, and their hyperscaler customers is more than compensating. The strategic risk here is concentration: any disruption to ASML's Veldhoven manufacturing operations — whether from geopolitical pressure, supply of specialised components, or physical capacity — propagates immediately across every leading-edge AI chip programme globally.

Meta and Broadcom announced a deepened multibillion-dollar custom silicon partnership, with Broadcom leading the design and manufacturing coordination of Meta's next-generation AI accelerators, per Bloomberg. The departure of Broadcom CEO Hock Tan from Meta's board — likely to resolve conflict-of-interest concerns as the commercial relationship intensifies — underscores that this is a serious, long-term infrastructure commitment rather than a hedge.

The strategic logic is straightforward: at Meta's inference scale, a purpose-built ASIC optimised for its specific model architectures delivers meaningfully better performance-per-watt and total cost of ownership than NVIDIA GPUs. Google (TPUs), Amazon (Trainium/Inferentia), and Microsoft (Maia) have pursued the same path. The cumulative effect is that NVIDIA's addressable hyperscaler market for inference workloads is structurally shrinking, even as overall AI compute spend grows. NVIDIA retains dominance in training and in enterprises lacking the engineering scale to build custom silicon, but the high-volume, high-predictability inference workloads — the ones that justify NVIDIA's premium pricing — are migrating.

Three distinct sovereign compute stories landed simultaneously. Italy awarded a confirmed €211 million grant to 2D Photonics, a startup targeting photonic interconnects for AI data centres, per Bloomberg. This is a notable strategic choice: rather than funding another generic cloud facility, Italy is backing a technology layer — silicon photonics — that addresses the interconnect bandwidth bottleneck constraining dense GPU clusters. Japan's Rapidus, meanwhile, remains on track for 2nm production in 2027, supported by substantial government backing, with TSMC simultaneously expanding its own Japanese manufacturing footprint, per The Register. The Rapidus timeline is still classified as an announced plan, not confirmed production capacity.

The China subsidy data provides the starkest framing: $142 billion in chipmaking industrial policy between 2014 and 2023, versus $39 billion committed under the US CHIPS Act, per Tom's Hardware. Raw subsidy volume does not translate directly into leading-edge capability — China remains blocked from EUV access and has not demonstrated 5nm-class volume production — but at mature nodes, the capital deployed is enabling rapid capacity expansion that reshapes the economics of legacy chip supply chains. The gap also raises questions about CHIPS Act adequacy if a second tranche of US funding is not authorised.

Microsoft has agreed to take over data centre capacity at a Norway facility originally contracted for OpenAI and marketed under the Stargate brand, per Bloomberg. The Stargate initiative — the $500 billion US AI infrastructure programme announced with considerable fanfare in early 2025 — has always bundled confirmed near-term investments with speculative long-term commitments. This transaction illustrates the gap: capacity provisioned in OpenAI's name is being quietly redistributed to Microsoft, its primary commercial partner and investor, suggesting OpenAI's own infrastructure absorption is either slower than marketed or being consolidated elsewhere.

For infrastructure planners, the more important signal is that hyperscale data centre deals are being structured with enough flexibility to reassign capacity across related parties. This reflects the difficulty of accurately forecasting AI workload growth 18-36 months out. The Norway location also carries specific strategic weight: Nordic data centres benefit from cold climates reducing cooling energy costs and access to renewable hydroelectric power, making them economically attractive for energy-intensive AI training runs.