Compute & Infrastructure

Nvidia's Kyber NVL144 rack system, designed to host the Rubin Ultra generation of accelerators, has been pushed to 2028 — a delay of more than 12 months from prior expectations. According to SemiAnalysis via Tom's Hardware and Data Center Dynamics, the root cause is a manufacturing defect in the PCB midplane — a high-complexity interconnect component central to the rack's power delivery and signal integrity architecture. A reported stopgap solution was also axed following customer pushback, leaving no confirmed bridge product.

The delay is strategically significant beyond a single product cycle. Hyperscalers building multi-year capacity plans around Nvidia's roadmap now face a gap between Blackwell deployments and the next meaningful performance uplift. It also reinforces a pattern: as rack-scale systems grow in complexity — integrating custom power, liquid cooling, and high-bandwidth interconnects into a single deliverable — manufacturing execution risk scales accordingly. The PCB midplane problem is not unique to Nvidia; it reflects an industry-wide challenge as compute density pushes against the limits of conventional board fabrication.

Despite high-profile claims of American-made AI chips, Tom's Hardware confirms that every Blackwell die produced at TSMC's Arizona fab is shipped to Taiwan for advanced packaging — specifically CoWoS (Chip-on-Wafer-on-Substrate), the process that integrates HBM memory stacks onto the GPU interposer. This step remains entirely offshore, with no domestic advanced packaging capacity at scale confirmed before 2028. Intel's Ohio and Arizona fabs are not yet a viable alternative for this process node.

This is the most precise articulation of the U.S. semiconductor policy gap currently available. The CHIPS Act investments have successfully stimulated front-end wafer fabrication, but advanced packaging — where TSMC, ASE, and Amkor's Taiwan operations hold decisive expertise — remains the unresolved chokepoint. For national security planners and infrastructure strategists, the implication is direct: a geopolitical disruption affecting Taiwan would halt U.S. AI chip supply even if Arizona fabs continue operating at full utilisation.

A new survey reported by Bloomberg shows Chinese enterprises are actively replacing Nvidia accelerators with domestic alternatives — primarily Huawei Ascend and emerging fabless vendors — driven by a combination of U.S. export controls restricting access to H100/H200/Blackwell-class chips and a deliberate government-backed push toward technology self-sufficiency. This is no longer a marginal or aspirational trend; it represents a structural reorientation of China's AI infrastructure procurement.

Simultaneously, Tom's Hardware reports Huawei is preparing to enter South Korea's market with the Atlas 950 SuperPod, deploying 8,192 Ascend 950 accelerators per cluster and claiming triple the inference throughput of Nvidia's H20 at one-quarter the cost. These performance claims are unverified by independent benchmarks, and the H20 is itself an export-controlled downgrade from Nvidia's full portfolio — but the pricing aggression is credible given Huawei's vertically integrated stack and state backing. South Korea is a strategically chosen market: home to Samsung and SK Hynix, deeply integrated into the U.S. semiconductor alliance, yet under commercial pressure to diversify AI infrastructure costs.

Anthropic has signed a confirmed 19 billion dollar, 20-year lease with TeraWulf for a Kentucky data centre, as reported by Data Center Dynamics. TeraWulf simultaneously sold its stake in a joint venture with Fluidstack, suggesting a strategic consolidation around the Anthropic relationship. At 19 billion dollars over 20 years, this is one of the largest single data centre commitments by an AI-native company, and it reflects a deliberate move away from pure hyperscaler dependency toward owned or long-leased dedicated infrastructure.

TeraWulf's background in Bitcoin mining gives it experience with high-density power procurement and grid co-location — skills directly transferable to AI data centre operations where power is the primary constraint. Kentucky's grid characteristics and land availability make it a logical choice for large-scale buildout. For the broader market, Anthropic's commitment signals that frontier AI labs at sufficient scale are now behaving like infrastructure operators, not just cloud tenants — a pattern that will intensify competition for power-rich sites and long-term grid capacity agreements.

Samsung Electronics reported a 19-fold surge in quarterly profit, driven by AI memory demand, yet shares declined as investors questioned whether the AI-driven capital expenditure cycle has peaked, per Bloomberg. The concurrent selloff in SK Hynix shares compounds the signal: memory is the bellwether for AI hardware demand, and markets are pricing in uncertainty about demand sustainability rather than celebrating current results. RBC's Janet Mui characterised the pullback as healthy consolidation given still-strong fundamental demand, per Bloomberg, but the gap between record earnings and investor reaction reflects a market recalibrating forward expectations.

For infrastructure analysts, the Samsung dynamic matters because HBM3E and future HBM4 supply — critical for next-generation AI accelerators — is concentrated between Samsung, SK Hynix, and Micron. If investor sentiment causes Samsung to moderate capex guidance, HBM supply constraints could tighten further into 2027, reinforcing bottlenecks at the memory layer even as logic fabrication capacity expands.