Compute & Infrastructure

A filing reviewed by Tom's Hardware reveals that UK-based Nscale signed a data center agreement with Spring (SG) Pte Ltd — a Singaporean subsidiary of ByteDance — through which the TikTok parent obtained access to more than 2,000 Nvidia B200 chips hosted in Norway. Nscale's US IPO filings referenced Spring only as a 'significant customer,' omitting any mention of ByteDance, which contributed 73% of the neocloud's 2025 revenue. The B200 is among the most restricted chips under current US export control rules, and its availability to a Chinese entity via a European data center and a shell subsidiary represents a significant structural failure in enforcement.

This case illustrates the core vulnerability in the current control architecture: restrictions apply to chip sales and exports, but access-as-a-service through third-country intermediaries creates a legal grey area that regulators have not yet closed. Norway is not subject to US export controls, and Nscale as a UK company is not automatically bound by US Entity List restrictions. The Commerce Department and BIS will face pressure to extend cloud access rules — analogous to the Know Your Customer provisions floated in prior rulemaking — to foreign-hosted GPU infrastructure. The strategic implication is clear: geography-of-hardware is no longer a sufficient proxy for geography-of-access.

Alibaba's semiconductor division T-Head has unveiled the Zhenwu V900 AI accelerator, claiming three times the performance of its predecessor M890 with 216GB of on-chip memory and native support for superclusters scaling to 500,000 chips, according to Tom's Hardware. Alibaba is positioning the V900 as the foundation for training a 10-trillion-parameter Qwen model — a scale that would exceed current publicly disclosed frontier model sizes. The chip is described as supporting NVLink-equivalent interconnect densities, suggesting serious investment in the cluster networking layer that has historically been a bottleneck for non-NVIDIA hardware.

Independent performance verification of Chinese-domestic chips remains difficult, and Alibaba's claims should be treated as marketing benchmarks until third-party validation is available. However, the architectural ambition — 500,000-chip superclusters — is strategically significant regardless of current performance gaps. China's strategy appears to be achieving sufficiency for domestic frontier training rather than achieving parity with NVIDIA on individual chip metrics. The V900 announcement also arrives as Alibaba simultaneously releases efficient inference-optimised models like Qwen 2.1 Image that run on consumer hardware, suggesting a two-track approach: domestic superscale training infrastructure plus globally distributable open-weight models that reduce dependence on high-end inference hardware.

Governor Newsom signed a package of bills mandating that data centers disclose their electricity and water consumption to California regulators and affected communities, according to The Verge. This is confirmed legislation — not a proposal — and it creates binding reporting requirements for facilities operating in California. The law responds directly to community protests over grid strain and water depletion in regions where large-scale AI infrastructure has been sited without adequate public consultation.

The strategic consequence for infrastructure planners is immediate: California-based data centers will now generate public consumption records that can inform permitting challenges, utility negotiations, and environmental litigation. Hyperscalers and colocation providers with California footprints — including major campuses in Santa Clara County and the broader Bay Area — will need to build disclosure compliance into their operational reporting. More significantly, this creates a legislative template that other large states with data center buildout activity (Texas, Virginia, Georgia) may follow. The combination of mandatory disclosure and community standing to challenge projects materially changes the permitting risk calculus for new AI infrastructure in high-value markets.

Two converging signals indicate that AI inference infrastructure is entering a structurally supply-constrained phase. First, Data Center Dynamics reports that OpenAI and Anthropic are actively procuring 20–30MW data center deployments — a scale that positions them as anchor tenants for purpose-built inference facilities rather than simply consuming hyperscaler capacity. This shift suggests both companies are experiencing cloud capacity constraints or seeking cost structures unavailable through AWS, Azure, or Google Cloud at their current inference volumes. Second, Bloomberg reports that Modal Labs is in talks to raise at a $15 billion valuation — triple its valuation from four months ago — while Baseten is simultaneously raising at a significantly higher valuation, both driven by enterprise demand for managed GPU inference platforms.

The two developments are structurally linked. As frontier labs consume available hyperscaler capacity and begin building or leasing dedicated inference infrastructure, mid-market enterprises are pushed toward GPU cloud intermediaries like Modal and Baseten that aggregate and abstract remaining capacity. The rapid valuation inflation at Modal — tripling in four months — reflects genuine supply scarcity in the inference GPU market rather than speculative enthusiasm alone. The 20–30MW range being targeted by OpenAI and Anthropic is large enough to require purpose-built facilities but small enough to be served by colocation providers or neoclouds, creating opportunities for operators outside the hyperscaler tier.