Compute & Infrastructure

Executives connected to Supermicro allegedly constructed a diversion network using a Thai government entity to export restricted Nvidia AI accelerators to Chinese customers, with Alibaba named as a recipient, according to reporting by Tom's Hardware. This is the most operationally sophisticated circumvention route disclosed to date, going beyond individual grey-market resellers to implicate a state-adjacent institution as a transshipment node.

The exposure of this network has several compounding implications. First, it validates long-standing concerns among export control analysts that entity-list restrictions are being systematically gamed through Southeast Asian intermediaries — a pattern US Commerce has been attempting to close via the diffusion rule and additional country-level licensing requirements. Second, Supermicro's involvement, even if limited to executive-level actors rather than corporate policy, adds to the company's already strained compliance record. Third, if Alibaba received restricted H100-class hardware, Chinese frontier model training capacity may be materially higher than US intelligence estimates based on sanctioned trade data. Enforcement action and likely additional regulatory pressure on SE Asian transshipment hubs should be expected.

CoreWeave posted wider losses and issued a disappointing Q2 2026 revenue forecast, according to Bloomberg, despite continuing to spend aggressively on data centre capacity. The company's model — borrow capital, lease Nvidia GPUs at scale, resell compute on flexible contracts — is sensitive to two variables that are now moving adversely: the cost of capital has not returned to 2021 levels, and customers are increasingly negotiating harder on GPU rental rates as the H100 spot market has softened.

This is a structurally important data point because CoreWeave was the canonical proof-of-concept for the independent GPU cloud model. If its growth is decelerating while capex remains elevated, it raises questions about whether the addressable market for non-hyperscaler GPU compute is as deep as venture and debt markets assumed at IPO. The contrast with ByteDance's 25% capex increase is instructive: hyperscale and large model labs are consuming compute at scale, but they are increasingly doing so through direct infrastructure ownership or bespoke deals — not through CoreWeave's spot and reserved capacity products.

Maryland has filed a formal complaint with federal energy regulators against PJM's proposal to allocate $2 billion in grid upgrade costs to the state, arguing that the transmission infrastructure primarily serves data centres located in neighbouring Virginia and other jurisdictions, not Maryland ratepayers, according to Tom's Hardware. This is not an isolated grievance — it is the first major inter-state regulatory confrontation over how to allocate the transmission costs triggered by the data centre load boom concentrated in Northern Virginia.

The structural problem is that PJM's cost allocation methodology was designed for a geographically diffuse load growth model, not for the hyperconcentrated demand spikes created by multi-gigawatt data centre campuses. States that host the data centres capture property taxes and economic activity; states that sit downstream on the grid absorb transmission costs. Without a federal-level resolution of this allocation framework, similar disputes are likely across every major grid operator territory where data centre demand is growing — MISO, ERCOT, and SPP included. This regulatory friction will not stop buildout, but it will slow transmission permitting and add cost uncertainty to grid interconnection timelines.

Argyll and SambaNova have launched a sovereign AI cloud in the UK, fulfilling a partnership first announced in October 2025, according to Data Centre Dynamics. SambaNova's architecture — purpose-built inference hardware rather than Nvidia GPUs — is a deliberate sovereign compute differentiator, allowing UK government and enterprise customers to process sensitive data on infrastructure not subject to US hardware export dependencies or hyperscaler terms of service.

Simultaneously, ByteDance's confirmed 25% capex increase to 200 billion yuan for 2026, reported by Bloomberg citing the South China Morning Post, reflects Beijing's implicit policy of encouraging domestic tech giants to build out China's internal AI compute base. Rising memory chip costs are noted as a driver of the budget increase, suggesting that even well-resourced Chinese operators are absorbing commodity component inflation. Together, these two developments — a mid-tier sovereign cloud in the UK and a hyperscale domestic buildout in China — illustrate the bifurcating global compute investment landscape, where geopolitical alignment increasingly determines infrastructure architecture choices.

On the hardware supply side, the emergence of hybrid 3.5D packaging — combining 2.5D interposer approaches with vertical 3D stacking — reflects the semiconductor industry's attempt to extract more performance per die without waiting for next-node lithography advances, as analysed by Semiconductor Engineering. For AI accelerator roadmaps, this matters because CoWoS capacity at TSMC remains a binding constraint on HBM-to-GPU bandwidth; 3.5D architectures could relieve some of that pressure by enabling denser integration without requiring entirely new packaging lines.

On the demand side, an Ipsos survey cited by Tom's Hardware found that 47% of Americans oppose data centre construction near their communities — a figure that directly translates into zoning opposition, extended permitting timelines, and site selection constraints. Community opposition is not a new phenomenon for energy infrastructure, but at this scale and with this level of political salience, it creates meaningful friction for operators trying to site gigawatt-scale campuses in suburban and semi-rural areas with existing grid access.