Compute & Infrastructure

Taiwanese prosecutors have indicted nine individuals, including a senior Nvidia manager, for orchestrating the illegal export of Nvidia B300 GPU servers to China using a structured five-point strategy to exploit and circumvent customs controls. According to Bloomberg and Tom's Hardware, this is Taiwan's first known prosecution of AI chip black-market activity. The involvement of an Nvidia employee raises immediate questions about internal compliance controls at the company, and the sophistication of the five-point evasion methodology suggests this was not an opportunistic operation but a structured supply chain circumvention ring.

The strategic implication is significant: despite layered US export controls and TSMC's own customer screening processes, advanced Blackwell-generation silicon was reaching Chinese end-users through a Taiwan-based smuggling route. This case will likely accelerate pressure on semiconductor companies to implement more granular post-sale tracking and channel monitoring. It also raises the question of how much B300-equivalent compute has already reached Chinese AI labs through informal channels, complicating assessments of the US-China compute gap.

Two converging pressures are now materialising simultaneously in AI memory supply. First, Nvidia customers are being warned of approximately 15% price increases on Blackwell and Rubin systems as DRAM costs surge, per Bloomberg citing Raymond James analyst Simon Leopold. Second, SK Hynix disclosed at Hot Chips 2026 that HBM cubes are constrained to a maximum 775-micron total thickness — the standard thickness of a 300mm logic wafer — meaning hybrid bonding cannot be introduced at HBM4E and will be pushed to HBM5, as reported by Tom's Hardware. SK Hynix confirmed it will extend the existing MR-MUF process through Nvidia's Rubin generation.

Marvell's response at FMS 2026 — a three-tier AI memory infrastructure that incorporates CXL memory built on recycled DDR4 — reflects an emerging market adaptation to scarcity. Recycling older DRAM into CXL-attached tiered memory pools does not resolve the HBM supply bottleneck for high-bandwidth training workloads, but it addresses inference and memory-expansion use cases at lower cost, potentially relieving some capacity pressure. The combination of price inflation and a hard physical constraint on HBM stacking density means bandwidth-per-dollar for frontier AI systems will worsen before it improves.

Hot Chips 2026 served as a major disclosure event for the next generation of AI hardware infrastructure. AMD detailed its MI400 GPU and Helios rack-scale system, presenting a full rack architecture designed to compete with Nvidia's NVLink-based scaling approach, per ServeTheHome. Nvidia in turn detailed the Vera Rubin NVL72 rack, its next-generation AI factory building block, per ServeTheHome. Intel disclosed Crescent Island, a GPU offering 160GB to 480GB of LPDDR5X capacity, positioning itself on memory capacity rather than raw compute throughput, according to ServeTheHome.

These disclosures are architectural announcements, not shipping product confirmations. AMD's Helios and Intel's Crescent Island represent planned capabilities; actual volume availability and datacenter adoption timelines remain unconfirmed. However, the sophistication of AMD's rack-level design and Intel's pivot to memory-capacity differentiation suggest both companies have identified specific niches — scale-out interconnect competition for AMD, large-context inference for Intel — rather than attempting direct GPU-to-GPU benchmark parity with Nvidia. IBM's Hot Chips disclosure of a dual-ISA 2nm mainframe AI processor with z/Architecture and ARM support in the same core is a separate signal: enterprise infrastructure is beginning to demand architectural flexibility rather than single-vendor lock-in.

TerraPower CEO Chris Levesque confirmed at Bloomberg that the company will announce a deal with a data centre customer for a new nuclear power plant this year, explicitly naming hyperscalers as the primary demand driver for new nuclear capacity, per Bloomberg. This follows a broader pattern of AI infrastructure operators moving from power purchase agreements on existing grid capacity to direct commissioning of new generation assets. Levesque also indicated TerraPower's strategic focus on reducing construction timelines and consumer costs — acknowledging that current economics require improvement for nuclear to scale as a primary AI energy source.

The parallel signal from the data centre development market is consistent: Sphere 3D's announced 50MW facility in Kentucky and Greenidge Generation's rebrand as Vulcan Infrastructure targeting AI and HPC both reflect the conversion of stranded or legacy power assets — Bitcoin mining capacity in both cases — into AI workload infrastructure. These are announced plans, not confirmed commissioned capacity. However, the asset conversion trend is structurally significant because it brings existing grid connections and cooling infrastructure online faster than greenfield builds, even if the compute density requirements of AI impose substantial retrofit costs.