Compute & Infrastructure

Both Samsung and SK Hynix have issued explicit warnings that high-bandwidth memory supply will remain severely constrained through 2027 and likely beyond, as reported by Tom's Hardware. The critical detail is customer behaviour: hyperscalers and AI labs are now reserving HBM supply years in advance, a market signal that demand visibility extends well past current order cycles. The spillover into standard DRAM is also tightening — indicating that HBM demand is consuming fab capacity that would otherwise serve general-purpose memory production.

HBM production is not simply a matter of scaling standard DRAM lines. It requires specialised through-silicon via stacking processes, advanced packaging, and dedicated fab investment with lead times of two to three years. This means the supply response to current demand signals will not materialise before 2028 at the earliest. For infrastructure planners, this is a confirmed physical constraint — not a forecast risk — that limits how quickly AI accelerator deployments can scale regardless of GPU availability.

Analysis from Semiconductor Engineering confirms what supply chain observers have suspected: access to TSMC's N3 and N2 nodes is now so capacity-constrained that it is functioning as a market entry barrier rather than simply a technical threshold. Companies without long-term wafer supply agreements — secured years ago — cannot reliably access leading-edge silicon at commercial volumes. This is structurally advantageous for NVIDIA, Apple, and AMD, all of which hold significant TSMC commitments, and disadvantageous for newer entrants attempting to compete on raw process node performance.

The practical consequence is a bifurcation in chip design strategy. Firms locked out of leading-edge monolithic die production are pivoting to chiplet and multi-die architectures on mature nodes — a path that introduces its own complexity around NoC coherency, thermal management, and packaging yield, as separately flagged by Semiconductor Engineering. This is not a clean workaround — it trades one set of constraints for another — but it does open design space for differentiated AI accelerator architectures that do not depend on TSMC N2 allocation.

A Bloomberg analysis Bloomberg documents a structural shift in how hyperscalers are financing AI infrastructure: Alphabet, Meta, and Amazon are now issuing debt at scale rather than funding data centre buildouts from operating cash flows. This represents a meaningful change in risk profile for companies that historically maintained fortress balance sheets. The shift reflects both the magnitude of required investment — which exceeds what free cash flow can absorb — and management confidence that AI infrastructure will generate returns justifying the leverage.

Goldman Sachs analyst Jim Covello's concurrent recommendation to favour hyperscalers over chipmakers — reported by Bloomberg — rests on the thesis that hyperscalers capture more durable value from infrastructure buildout than equipment suppliers subject to capex cycle volatility. However, Bloomberg Opinion's Dave Lee directly challenges Meta's position, arguing that unlike Amazon and Google, Meta generates no meaningful cloud revenue growth from its AI infrastructure spend, making its $145 billion capex commitment difficult to justify on hyperscaler logic. The distinction matters for infrastructure planning: Amazon and Google are building capacity that generates direct third-party revenue; Meta is building internal capability with no external monetisation mechanism.

PJM Interconnection — the largest US grid operator — has received 220GW of new connection requests under its reformed interconnection process, per Data Center Dynamics. This figure is striking: it represents a volume of power demand that cannot be served by the existing grid on any near-term horizon, regardless of regulatory process improvements. PJM is now deploying an agentic AI system developed by Google-backed Tapestry to manage the queue — an irony that underscores the self-reinforcing nature of AI infrastructure demand.

At the ground level, AI data centre buildout is already creating measurable labour market distortions. Texas construction projects are experiencing two-month delays as data centre developers offer electricians 75% salary premiums over residential construction rates, per Tom's Hardware. This is a confirmed, present-tense impact — not a forecast — and it signals that the skilled labour constraint is arriving ahead of the power constraint in the fastest-growing data centre markets.

Cambricon's Q1 2026 revenue of $423 million confirms that China's domestic AI chip ecosystem has reached commercial scale, per Tom's Hardware. The trajectory — driven by sustained policy support, captive demand from Chinese hyperscalers prohibited from buying NVIDIA's best chips, and a deepening domestic design talent base — represents a structural market shift, not a temporary substitution. US export controls designed to limit China's AI capabilities have functioned instead as an industrial policy forcing mechanism, accelerating indigenous capability development at a pace that would not have occurred in an open market.