Compute & Infrastructure

Envision Group, one of China's largest wind and energy storage companies, has commissioned the first phase of a gigawatt-scale data center in Inner Mongolia, according to Bloomberg. The facility represents an unusual but strategically coherent move: a company that controls renewable generation assets building compute infrastructure adjacent to its own power supply, effectively short-circuiting the grid interconnection bottleneck that is throttling data center expansion globally. Inner Mongolia offers abundant wind resources, relatively low land costs, and existing high-voltage transmission infrastructure, making it one of China's preferred locations for power-intensive AI workloads.

The confirmed commissioning of the initial phase distinguishes this from the large volume of announced-but-unbuilt capacity in the market. That said, 'gigawatt-scale' describes the full build target — the commissioned phase is almost certainly a fraction of that figure. The strategic model here — energy company as data center operator — is one to watch as a template for bypassing utility constraints, particularly in markets where grid connection queues run to years.

Frore Systems has publicly claimed that its LiquidJet direct-die coldplate technology can lower NVIDIA Rubin GPU junction temperatures by 10°C and translate that thermal headroom into a 15% improvement in token-generation throughput, per Tom's Hardware. Crucially, hyperscalers are reportedly evaluating delidded GPU configurations in production — removing the integrated heat spreader to allow direct contact between the coldplate and the die. This is aggressive thermal engineering that trades hardware warranty and handling complexity for sustained frequency and power delivery gains.

The 15% token-generation efficiency claim requires scrutiny: it likely reflects a best-case scenario under sustained thermal throttling conditions rather than a universal baseline improvement. Nevertheless, the direction is analytically important. As GPU TDP continues to climb — Rubin-class hardware operates in the 1,000–1,200W range per GPU — the delta between thermally-constrained and thermally-unconstrained performance widens, and cooling vendors increasingly determine effective compute yield per rack. This is shifting cooling from a facilities afterthought to a first-order design parameter in AI factory architecture.

Elon Musk confirmed via social media that both SpaceX and xAI will exclusively use NVIDIA AI accelerators for training and inference workloads, citing the Vera Rubin architecture as the best available compute platform, per Tom's Hardware. Musk further stated that a space-optimised variant of the Vera Rubin NVL72 is planned for launch in 2027, implying NVIDIA is developing or has already developed radiation-tolerant or thermally-adapted versions of its flagship AI accelerator for orbital deployment.

The exclusive commitment from xAI reinforces NVIDIA's demand position at the frontier, though it also carries strategic concentration risk for both parties. For NVIDIA, dependence on a small set of very large customers creates revenue volatility; for xAI, exclusive platform lock-in limits negotiating leverage as AMD, Intel, and custom silicon alternatives mature. The orbital deployment angle is analytically distinct: space-based AI compute is nascent but directionally significant for defence, Earth observation, and autonomous satellite operations. If NVIDIA's silicon is qualifying for radiation environments, it expands the addressable market and creates a new certification moat.

A Semiconductor Engineering analysis confirms that multi-physics co-simulation across 2.5D and 3D chiplet stacks has become the primary constraint on heterogeneous design iteration — not toolchain capability gaps, per Semiconductor Engineering. Simulating thermal, electrical, mechanical, and signal-integrity interactions simultaneously across a multi-die package is computationally expensive to the point where full-fidelity simulation runs take days or weeks, forcing engineers to work with degraded models or serialise what should be parallel design loops. This is a direct tax on the development velocity of next-generation AI accelerators, which almost universally rely on chiplet-based packaging — HBM stacking, I/O chiplets, compute tiles — to circumvent monolithic die scaling limits.

The implication for the AI infrastructure supply chain is that simulation compute infrastructure — high-core-count workstations, EDA-optimised cloud instances, and simulation acceleration hardware — is itself a bottleneck in the path from architecture to production silicon. As AI accelerator design cycles compress under competitive pressure, the firms that invest in simulation infrastructure gain a compounding velocity advantage.

The US National Science Foundation has announced a $100 million AI Infrastructure Hubs programme designed to extend AI compute access to research institutions in underserved regions of the US, per Data Center Dynamics. The programme is confirmed as announced — it is not yet confirmed as funded and deployed. At $100 million, the investment is modest relative to the scale of private hyperscaler buildout (Microsoft, Google, and Amazon each announced over $50 billion in 2026 capex commitments), but the policy objective is distinct: distributing compute access geographically rather than maximising aggregate capacity. This addresses a real structural problem — AI R&D capability is concentrating in a handful of coastal metro areas, limiting the diversity of institutions that can participate in frontier research.