Compute & Infrastructure
Top Line
Nvidia restarts H200 manufacturing for China after receiving US licenses and customer orders, marking the first resumption of its China supply chain in over a year despite ongoing export restrictions.
DayOne Data Centers seeks to double its borrowing to a record $7 billion for Asian data centre expansion, while Microsoft powers on the first Nvidia Vera Rubin NVL72 system and begins second-phase construction in Denmark.
Strait of Hormuz blockade from US-Iran conflict threatens Taiwan's semiconductor industry within days due to critical LNG and helium import dependencies required for TSMC fabrication and grid operations.
Alibaba raises AI computing and storage prices by up to 34% citing surging demand and infrastructure costs, while Samsung considers multi-year memory contracts to stabilise supply amid shortage concerns.
Nvidia integrates acquired Groq's LPU architecture into its Vera Rubin platform for low-latency inference workloads, removing previously planned CPX accelerators from its roadmap in favour of specialised inference processors.
Key Developments
China GPU Supply Chain Restarts as US Eases H200 Export Stance
Nvidia CEO Jensen Huang announced the company is restarting manufacturing of H200 AI accelerators for Chinese customers after receiving export licenses from the US government and purchase orders from multiple Chinese clients, according to Bloomberg. This marks the first time Nvidia's China supply chain has been operational since export restrictions froze shipments over a year ago, as reported by Tom's Hardware. The H200, while not Nvidia's latest generation, represents a significant capability restoration for Chinese AI development amid ongoing technology restrictions. The licensing appears selective rather than a blanket policy reversal, with Huang specifying licenses have been granted for 'many customers' rather than unrestricted access.
The timing coincides with Alibaba raising AI computing service prices by up to 34% due to demand surges and infrastructure costs, per Bloomberg, suggesting Chinese cloud providers face constrained capacity that H200 availability could partially address. However, the H200 remains multiple generations behind Nvidia's current Vera Rubin platform, creating a widening technological gap between Chinese and Western AI infrastructure capabilities.
Strait of Hormuz Blockade Creates Immediate Semiconductor Supply Chain Risk
Taiwan's semiconductor industry faces imminent crisis as the US-Iran conflict enters its third week with a Strait of Hormuz blockade threatening critical energy and industrial gas imports, reports Tom's Hardware. Taiwan imports nearly all its energy requirements including substantial LNG volumes to sustain the electrical grid powering TSMC and other fabrication facilities. Additionally, helium imports essential for semiconductor manufacturing processes are disrupted by the same chokepoint. The situation represents a days-not-weeks timeline for grid stress and fabrication impacts, given limited strategic reserves.
TSMC manufactures the majority of the world's advanced logic chips including GPUs from Nvidia and AMD, making this supply disruption systemically significant for global AI infrastructure buildout. Unlike typical supply chain vulnerabilities that unfold over months, the energy dependency creates an acute crisis with limited mitigation options beyond diplomatic resolution or military intervention to reopen shipping lanes.
Data Centre Financing and Deployment Accelerate Amid Capacity Constraints
DayOne Data Centers is seeking to double an existing loan to as much as $7 billion, according to Bloomberg, which would represent the largest borrowing for the sector by any firm in Asia. The scale of debt financing signals both aggressive expansion plans and investor confidence in sustained demand for Asian data centre capacity. Meanwhile, Microsoft commenced second-phase construction at its Køge, Denmark facility per Data Center Dynamics and invested $52 million across two sites in Medina, Texas according to Data Center Dynamics. Google was confirmed as the developer behind a $500 million data centre in Lima, Ohio, previously known as Project BOSC, reports Data Center Dynamics.
Counterbalancing expansion, Applied Digital effectively withdrew from its planned data centre near Toronto, South Dakota, stating the project is 'unlikely to go forward' per Data Center Dynamics, demonstrating that not all announced capacity materialises. Microsoft also claimed to be the first hyperscaler to power on Nvidia's Vera Rubin NVL72 system according to Data Center Dynamics, indicating leading cloud providers are securing early access to next-generation infrastructure. Samsung is considering a shift toward multi-year memory contracts, reports Bloomberg, a departure from typical shorter timeframes that could stabilise supply but reduce pricing flexibility.
Nvidia Restructures Product Portfolio with Groq Integration and Architecture Diversification
Nvidia revealed substantial changes to its data centre roadmap at GTC 2026, removing the previously planned Rubin CPX accelerators and instead integrating Groq LPU architecture for low-latency inference workloads, as reported by Tom's Hardware and analysed by Next Platform. The $20 billion Groq acquisition now appears strategically focused on specialised inference acceleration rather than general-purpose compute expansion. Nvidia demonstrated the Rubin Ultra tray featuring the industry's first AI GPU with 1TB of HBM4E memory per package, reports Tom's Hardware, representing a four-fold memory capacity increase over current generation products.
The updated roadmap includes the Rosa CPU and stacked Feynman GPUs with optical NVLink interconnects, per Tom's Hardware. Nvidia also launched the DGX Station featuring the GB300 Grace Blackwell Superchip with 784GB of combined memory and 1,600-watt power draw, now available for order with shipping expected in coming months according to Tom's Hardware. The BlueField-4 STX storage architecture for agentic AI was introduced to address data access bottlenecks in inference workloads, reports Tom's Hardware.
Energy and Cooling Innovation Attempts to Address Infrastructure Constraints
Microsoft expects to commercialise MicroLED datacenter cables by late 2027, reports Tom's Hardware, with the MOSAIC technology replacing lasers with inexpensive MicroLEDs and promising 47% faster data transmission with approximately 33% lower latency. The company is also expanding Hollow Core Fiber deployment for inter-rack connectivity. Meanwhile, researchers demonstrated a zeolite-based thermal battery system that could enable data centres to use factory waste heat for cooling, though Data Center Dynamics notes the research remains at an early stage.
Emerald AI and InfraPartners unveiled a 'Flex-Ready Data Center' design using AI to adjust load according to grid conditions, per Data Center Dynamics, claiming greater flexibility over power needs. These developments address different constraint vectors: MicroLED targets interconnect bandwidth and latency, thermal batteries attempt cooling efficiency gains, and flex-ready designs address grid integration challenges. However, none are in production deployment, and commercialisation timelines extend into 2027 or beyond.
Signals & Trends
Memory Supply Shifts Toward Long-Term Commitments Amid AI-Driven Tightness
Samsung's consideration of multi-year memory contracts represents a fundamental change in semiconductor supply dynamics, moving from spot pricing and short-term agreements to capacity reservation models more typical of utility infrastructure. This reflects hyperscalers and AI infrastructure operators prioritising supply certainty over pricing flexibility as memory becomes a strategic constraint. Alibaba's 34% price increase for AI computing services demonstrates that cloud providers face margin pressure from infrastructure costs and are passing increases to customers, suggesting the AI compute market is supply-constrained enough to absorb significant price escalation. The combination of long-term contracts and price increases indicates the memory market is transitioning from cyclical oversupply-shortage patterns to sustained tightness driven by AI workload growth.
Architectural Specialisation Displaces General-Purpose Compute Consolidation
Nvidia's removal of CPX accelerators in favour of Groq LPU integration signals a strategic reversal from its previous trajectory toward unified GPU-based computing. The company now acknowledges that inference workloads have fundamentally different requirements than training, requiring purpose-built silicon rather than scaled-down training accelerators. This mirrors broader industry movement toward heterogeneous computing with specialised processors for vision, natural language, recommendation systems, and other specific tasks. The trend challenges the economic model of general-purpose accelerators amortising development costs across multiple workload types, potentially fragmenting the market and increasing customer integration complexity. Companies building AI infrastructure must now evaluate whether to deploy homogeneous GPU clusters or hybrid systems combining training and inference-optimised processors.
Geographic Diversification Rhetoric Outpaces Actual Capacity Shifts
Despite ongoing discussion of supply chain resilience and geographic diversification, announced projects remain heavily concentrated in established markets while new geographies see limited actual deployment. Google's $500 million Lima, Ohio facility and Microsoft's Denmark expansion represent investment in traditional Western markets, while Applied Digital's South Dakota project cancellation demonstrates that secondary locations face execution challenges. DayOne's record Asian borrowing confirms the region's continued dominance despite Taiwan risk concerns exposed by the Strait of Hormuz crisis. The gap between diversification intentions and deployment reality suggests that despite acknowledged concentration risk, economics and technical ecosystem advantages keep driving investment toward existing clusters. True geographic diversification appears contingent on either government subsidies reaching levels that overcome inherent location disadvantages, or actual supply disruption forcing redistribution.
Explore Other Categories
Read detailed analysis in other strategic domains