Compute & Infrastructure

Oracle has agreed to purchase up to 2.8 GW of fuel-cell power from Bloom Energy to supply AI data centres, according to Bloomberg. This is a confirmed commercial agreement, not a letter of intent. At 2.8 GW, the contract is among the largest single power procurement deals in the data centre industry's history and reflects a structural shift in how hyperscalers are approaching energy: the grid is no longer the default solution, it is the fallback.

Bloom Energy's solid-oxide fuel cells run on natural gas and can be sited directly adjacent to data centres, bypassing utility interconnection queues that now routinely run three to seven years in primary markets. The deal validates the emerging playbook — pair on-site generation with long-term offtake agreements to guarantee power availability independent of grid capacity. The environmental tradeoff is real: fuel cells are cleaner than diesel but are not zero-carbon, which creates tension with hyperscaler sustainability commitments. OpenAI's concurrent policy push for accelerated grid development, noted by Bloomberg, implicitly acknowledges that grid modernisation alone will not move fast enough to meet near-term compute demand.

The US Commerce Department's Bureau of Industry and Security has lost approximately 20% of its export licensing staff over the past year, according to Tom's Hardware. The direct consequence is a processing bottleneck on export licence applications for NVIDIA and AMD AI accelerators destined for or potentially rerouted through China. Export controls are only as effective as the enforcement apparatus behind them, and a depleted BIS creates two distinct risks: approved shipments slow down, frustrating legitimate customers in allied markets; and the reduced oversight capacity increases the probability that restricted chips reach prohibited end-users through third-country transshipment.

This development cuts against the strategic logic of the October 2022 and subsequent export control frameworks, which were premised on a functioning, well-resourced licensing and enforcement agency. The timing is particularly sensitive given ongoing US-China tensions over semiconductor access and China's accelerating domestic chip development programs. If BIS cannot process licence applications at operational speed, the de facto effect is a generalised slowdown in AI chip exports globally — including to allied nations — while bad actors face a reduced likelihood of detection.

Anthropic is exploring the design of proprietary AI chips, though the effort remains pre-organisational — no dedicated team has been assembled as of reporting by Data Centre Dynamics. The strategic logic is straightforward: Anthropic's rapid growth, detailed by The Next Platform, has made it a major source of revenue for both Broadcom — which supplies custom AI accelerator silicon through its XPU program via Google's TPU infrastructure — and Google Cloud. As training and inference costs become central to competitive differentiation, dependency on third-party silicon at this scale creates both cost and strategic exposure.

The Broadcom-Google dynamic is worth examining separately. Anthropic's growth is generating significant ASIC revenue for Broadcom through the Google TPU ecosystem, which means Anthropic is effectively subsidising a supply chain it does not control. If Anthropic moves toward in-house silicon design — likely fabbed at TSMC, given no credible alternative for leading-edge process nodes — it would join a pattern established by Google (TPUs), Amazon (Trainium/Inferentia), and Microsoft (Maia). The critical distinction is that Anthropic lacks the cloud platform revenues that funded those programs. This makes the financial sustainability of an in-house silicon program the central uncertainty.

US hyperscale data centre investment is shifting materially toward Texas and Midwest markets as primary markets — Northern Virginia, the Bay Area, Phoenix — face binding energy constraints, according to Data Centre Dynamics. Microsoft's confirmed multi-billion-dollar commitment to Ontario, Canada, reported by Data Centre Dynamics, is part of a broader $19 billion Canadian program and reflects the same dynamic operating at a cross-border level: power availability and interconnection timelines now drive site selection more than latency or proximity to customers.

The Festus, Missouri council ouster documented by Tom's Hardware is directly relevant here. As hyperscalers move into secondary and tertiary markets that have the grid capacity but lack the planning infrastructure and political normalisation around large industrial projects, community opposition becomes a material siting risk. The Festus situation involved a $6 billion project and resulted in democratic backlash severe enough to remove elected officials. This dynamic will recur as the industry pushes into markets where residents were not anticipating industrial-scale power draw.