Compute & Infrastructure

US authorities suspect that a company central to Thailand's national AI initiative facilitated the transhipment of Supermicro servers — containing advanced Nvidia GPUs subject to export restrictions — to Chinese end-users including Alibaba, according to Bloomberg. The scale is described as billions of dollars in hardware, making this one of the most significant alleged circumvention operations identified since the expanded October 2023 controls. The use of a sovereign AI programme as a routing mechanism is particularly alarming: it suggests adversarial actors are deliberately co-opting government-endorsed entities to exploit the political sensitivity around disrupting national tech initiatives.

The chokepoint implications are multi-directional. Nvidia faces reputational and regulatory exposure from distribution channel failures it does not directly control. Supermicro — already under prior scrutiny — is again a named vector. TSMC-manufactured silicon ends up in restricted territory regardless of the formal sales chain. The Bureau of Industry and Security will face intensified pressure to impose distributor-level compliance obligations, potentially slowing legitimate chip deployments in Southeast Asia as collateral damage. This incident will also accelerate scrutiny of other Southeast Asian nations — Malaysia, Vietnam, Indonesia — that have signed large AI infrastructure deals with US hyperscalers and chip vendors.

SK Hynix customers have moved beyond standard long-term supply agreements and are reportedly offering to directly purchase EUV lithography equipment and co-fund new fab lines, according to Tom's Hardware. This is a structurally significant shift: it means hyperscalers and AI hardware vendors are willing to take on balance-sheet exposure in memory manufacturing — a capital-intensive and technically specialised domain well outside their core competency — because the market cannot deliver supply fast enough. The offers are described as running into the hundreds of millions of dollars per commitment.

The bottleneck is compounded by the ASML dependency: EUV machines have lead times measured in years and ASML's production capacity is itself constrained. Even if customer financing unlocks SK Hynix's willingness to expand, the physical equipment pipeline limits how quickly new HBM capacity can come online. Samsung and Micron face similar constraints. For AI training clusters — where HBM bandwidth is the binding constraint on GPU utilisation — this shortage directly caps the effective deployment rate of new Nvidia Blackwell and successor hardware regardless of chip availability.

As of April 2026, 69 US jurisdictions have enacted bans or moratoriums on new data centre construction, with four classified as permanent, according to Tom's Hardware. The acceleration is notable: this number has grown rapidly from earlier counts tracking fewer than 20 active restrictions. The drivers are primarily local: power grid strain, water consumption for cooling, property tax base concerns, and community opposition to industrial-scale facilities in residential or agricultural zones. These are not federal interventions — they are distributed local zoning and utility decisions that collectively produce a fragmented and unpredictable permitting environment.

The geographic compression effect is underappreciated. As bans proliferate in established markets — Northern Virginia, the Pacific Northwest, parts of the Midwest — remaining viable locations command premium pricing, faster utility interconnection queues, and greater political leverage over hyperscaler operators. This creates a structural advantage for states that have proactively positioned themselves as data centre-friendly, including Texas, Georgia, and Ohio, at the cost of accelerating grid stress in those concentrated locations. The Three Mile Island restart — targeting mid-2027 — is a direct response to this dynamic, with nuclear offering a politically defensible power source in a context where new gas generation also faces permitting opposition, as reported by Bloomberg.

Anthropic has confirmed a $1.8 billion multi-year computing deal with Akamai, pulling a CDN and edge infrastructure provider into the frontier AI inference stack, according to Bloomberg. This is a deliberate diversification away from exclusive AWS dependency, which Anthropic has previously relied upon as its primary cloud backer. Akamai's distributed edge footprint offers latency advantages for inference at scale and geographic reach without the concentration risk of a single hyperscaler. The deal size — $1.8 billion — represents a meaningful commit that will reshape Akamai's infrastructure investment priorities.

Simultaneously, OpenAI and Broadcom are in reported discussions to finance an $18 billion custom ASIC project, with the structure reportedly tying Broadcom's investment commitment to Microsoft purchase guarantees, according to Data Centre Dynamics. This remains at the negotiation stage and should be treated as a speculative announcement rather than a confirmed programme. If executed, it would represent one of the largest custom silicon commitments in history and would deepen the Microsoft-OpenAI infrastructure co-dependency even as OpenAI publicly pursues independent data centre capacity. The tension between OpenAI's sovereign compute ambitions and Microsoft's structural leverage as a balance-sheet backstop is a defining dynamic in frontier AI infrastructure for the next three to five years.

CoreWeave shares fell following Q1 earnings after forward guidance triggered concerns about the sustainability of its hyperscale GPU rental growth trajectory, with CEO Michael Intrator defending the company's data centre buildout pace in a Bloomberg interview. Separately, Cerebras Systems is reported to be raising its IPO price range ahead of its public debut, according to Bloomberg, reflecting strong institutional demand for exposure to alternative AI chip architectures. The contrast is instructive: markets are rewarding differentiated hardware bets while scrutinising capital-intensive GPU rental intermediaries whose value proposition depends on sustained Nvidia supply advantages and hyperscaler pricing dynamics.

CoreWeave's situation highlights a structural vulnerability in the GPU cloud model: it is caught between Nvidia on the supply side — as a dependent customer for H100 and Blackwell allocation — and hyperscalers on the demand side, which are simultaneously customers and competitors building internal capacity. As hyperscalers complete their own GPU cluster buildouts, the addressable market for third-party GPU cloud shrinks toward enterprise and mid-market segments that are less profitable at scale.