Compute & Infrastructure

The HBM and high-bandwidth memory shortage has moved from a frequently cited risk to a confirmed earnings differentiator. Bloomberg reports that the worsening shortage is producing measurable divergence in corporate results, with suppliers exposed to AI memory demand posting sharply different outcomes from those serving commodity DRAM markets. This reflects a supply chain structure where HBM3E production remains concentrated at SK Hynix, with Samsung still ramping yield on competitive product and Micron at an earlier stage of volume — meaning three firms globally control the supply of the component that most constrains GPU utilisation at scale.

The market dynamics are compounded by the findings from ServeTheHome, which documented AI memory pricing as sufficiently distorted that operators are now deploying AI tooling to identify procurement arbitrage. When infrastructure teams are using ML to navigate memory markets, it signals that pricing opacity and supply concentration have reached a level that introduces meaningful operational risk for data centre operators planning capacity at 12-24 month horizons.

Microsoft's Kenya project collapse is the most instructive case study in current infrastructure economics. Tom's Hardware reports the Kenyan government assessed the planned Olkaria geothermal-powered facility would require switching off roughly half the national grid to meet its power draw — a capacity disagreement that has stalled a confirmed $1 billion commitment. This is not a speculative risk; it is a project that has already stopped. The dynamic illustrates the fundamental mismatch between data centre power density ambitions and the grid headroom available in emerging markets that tech companies are increasingly courting for power-advantaged sites.

At the physical infrastructure level, IEEE Spectrum frames the problem at the rack level: GPU clusters operating beyond 100 kW per rack generate high-frequency, synchronised spike loads that stress power chain resilience in ways that conventional UPS and PDU architectures were not designed to handle. SoftBank's response — announced as a confirmed partnership with South Korean startups Cosmos Lab and DeltaX to manufacture zinc-halogen batteries at gigawatt-hour scale by 2028 — represents a vertically integrated bet on solving the energy storage layer internally rather than relying on incumbent vendors. Tom's Hardware notes the 2028 target for GWh-scale production, which should be treated as an announced plan at this stage, not confirmed capacity.

Bloomberg confirms that China's domestic AI hardware suppliers are failing to meet domestic demand, with component shortages across the supply chain constraining output. This is structurally significant: China has been forced into accelerated domestic substitution by US export controls, but the supply chains required to support that substitution — particularly for advanced packaging, memory, and specialised components — remain immature or dependent on inputs that are themselves subject to restriction.

The Loongson disclosure provides a useful benchmark. Tom's Hardware reports that Loongson's 3B6600 CPU and 9A1000 GPU — its next-generation products targeted for 2027 — are expected to reach parity with Intel's 12th-generation architecture and AMD's RX 550, both of which are multiple generations old by current standards. This performance delta is not closing at a rate that would allow Chinese sovereign compute to support frontier AI training workloads within a credible planning horizon, reinforcing that export controls are achieving their stated objective of denying China access to leading-edge compute for AI development.

CME Group's partnership with Silicon Data to create a futures market for AI computing power — reported by Bloomberg — is a structural inflection point in how compute capacity is priced and allocated. A functioning futures market would create transparent price discovery for GPU-hours, enable operators to hedge compute costs, and potentially shift procurement away from the current model of bilateral long-term contracts with hyperscalers. This is announced as a plan, not a live market, but CME's institutional credibility means this is more than speculative.

On the network fabric layer, Next Platform covers the OpenAI and Microsoft-led initiative to develop a more scalable Ethernet standard for AI cluster interconnects. This matters because NVIDIA's NVLink and InfiniBand have dominated high-performance cluster networking, giving NVIDIA a systems-level lock-in that extends beyond GPU silicon. An industry-standard Ethernet fabric capable of matching InfiniBand performance at scale would erode that advantage and potentially open the market to disaggregated accelerator deployments from AMD, Intel, and custom silicon vendors.

SiPearl and Semidynamics have confirmed a partnership to develop a European sovereign rack-scale compute platform targeting AI inference workloads, per Data Centre Dynamics. This is a confirmed development agreement, not yet a production commitment, but it represents the EU's most concrete rack-level initiative to reduce inference dependency on US hyperscaler hardware. The platform is designed to serve both public and private European initiatives, aligning with the EU's broader AI Act compliance infrastructure requirements.

On the US side, Jensen Huang's exclusion from Trump's China state visit — documented by Tom's Hardware — is being interpreted by at least one cited expert as a deliberate diplomatic signal that chip export controls remain non-negotiable. Given that NVIDIA's H20 and downgrade-spec products for China represent a significant revenue line, and that further export tightening remains a live policy risk, the exclusion has direct read-through to NVIDIA's China revenue outlook.