AI Infrastructure Hits Hard Limits: Chips Smuggled, Memory Gone, Land Banned

Three distinct supply-side crises are converging to cap the effective deployment rate of AI infrastructure regardless of how much capital is committed. SK Hynix's HBM capacity has reached zero, with customers so desperate they are offering to purchase EUV machines and co-fund fab lines directly — a distortion of normal supplier-customer relationships that reveals how severely AI training workloads have exhausted advanced memory supply chains. Even with customer financing, ASML's constrained EUV production pipeline means new HBM capacity cannot come online quickly enough to relieve pressure on Nvidia Blackwell deployments. Memory, not logic chips, is now the binding hardware constraint.

Simultaneously, 69 US jurisdictions have enacted data centre construction bans or moratoriums — four permanent — compressing viable buildout geography and creating pricing premiums and grid stress in the remaining permissible states. The Three Mile Island nuclear restart targeting mid-2027 is a direct response: nuclear offers a politically defensible power source when both new gas generation and grid interconnection face opposition. Taken together, these physical constraints — memory, land, and power — represent a class of problem that multi-billion-dollar capex commitments and hyperscaler balance sheets cannot simply override. Strategic advantage will accrue to operators who secured permissible sites, long-term power agreements, and memory supply commitments earliest.

The suspected Thailand transhipment operation — routing billions in Nvidia-chip-laden servers through a government-endorsed national AI initiative to Chinese end-users including Alibaba — is the most consequential alleged breach of the US AI export control regime to date. The mechanism is deliberately sophisticated: co-opting a sovereign AI programme exploits both the political sensitivity around disrupting national tech initiatives and the legitimate-seeming end-user credentials such entities carry. Challenging a state-backed AI programme is diplomatically costly in ways that sanctioning a private company is not, and adversarial actors have clearly identified this asymmetry.

The enforcement response will almost certainly involve tighter end-use certification requirements across Southeast Asia and expanded entity list designations beyond direct Chinese recipients to third-country facilitators. This creates collateral damage for every legitimate data centre investor operating in ASEAN markets — Malaysia, Vietnam, Indonesia — that have attracted substantial hyperscaler and chip-vendor capital on the premise of serving as neutral compute hubs. The pattern also has implications for the US CHIPS Act bilateral partnership framework and the AI Diffusion Rule's Tier 2 country structure: both programmes have inadvertently created entities that are politically protected and technically capable of handling restricted hardware at scale. Capital allocators with exposure to Southeast Asian AI infrastructure should price this regulatory risk explicitly into underwriting assumptions.

Multiple signals this week collectively mark a transition from broad AI capital enthusiasm to discriminating conviction. TCI Fund Management's reduction of its Microsoft stake from 10% to 1% — an estimated $8 billion liquidation from a historically high-conviction, low-turnover holder — reflects institutional concern that AI disruption to Microsoft's core software revenue is a structural risk, not just a near-term transition cost. CoreWeave's forward guidance miss raised questions about whether GPU cloud intermediaries caught between Nvidia on supply and hyperscalers building internal capacity on demand can sustain the margins required to justify current valuations. SoftBank cutting its OpenAI margin loan target adds a further note of caution to the frontier lab funding stack.

Against this backdrop, Cerebras Systems reportedly raising its IPO price range signals strong institutional appetite for differentiated chip architectures perceived as having defensible moats — a direct contrast to the CoreWeave reception. Anthropic's Akamai deal and Isomorphic's $2 billion-plus raise confirm that capital is still flowing at scale to frontier labs and AI-for-science verticals with clear differentiation. The pattern is bifurcation rather than retreat: investors are rewarding verifiable technical moats — values-based alignment, production-validated security tooling, custom silicon — while scrutinising capital-intensive intermediaries and companies where AI efficiency gains are compressing headcount without yet accelerating top-line growth. Cloudflare eliminating 1,100 roles while revenue growth slowed is the clearest single illustration of this timing mismatch.