Compute & Infrastructure

DeepSeek released a preview of its V4 model — described as a 1.6 trillion parameter mixture-of-experts architecture — trained on Huawei Ascend chips rather than NVIDIA H100s or H800s, the latter of which are now subject to US export controls. Tom's Hardware reports that the US government is simultaneously escalating allegations of IP theft against DeepSeek and other Chinese AI firms, adding a sanctions-enforcement dimension to what is already a technology competition story.

Benchmark assessments reported by Bloomberg suggest V4 does not close the gap with leading US frontier models, and DeepSeek is competing aggressively on price — slashing API fees in what Bloomberg characterises as a Chinese domestic price war. Bloomberg The combination of Huawei-native training, a 1.6T parameter scale, and aggressive inference pricing represents a meaningful proof point that China can sustain frontier-adjacent AI development under chip embargo conditions, even if performance parity with US leaders remains elusive.

A Taiwanese court handed a 10-year prison sentence to a former Tokyo Electron employee convicted of stealing TSMC proprietary process data, one of the heaviest penalties Taiwan has imposed for semiconductor IP theft. Bloomberg The case illustrates a structural vulnerability that goes beyond any single actor: TSMC's process technology is so uniquely valuable, and so concentrated in a single geography, that it has become a persistent target for state-adjacent industrial espionage at multiple points in its supply chain.

The Tokyo Electron link is notable — TEL is one of the world's largest semiconductor equipment suppliers and a critical node in the wafer fabrication supply chain, meaning espionage risk is not confined to TSMC employees but extends across its entire vendor ecosystem. Taiwan's stock market surpassing the UK's in total value — with TSMC alone exceeding 40% of that market cap per Tom's Hardware — quantifies exactly how much value is concentrated in this single institution, and why the espionage threat vector will intensify rather than diminish.

What began as localised opposition to individual data centre projects has consolidated into a recognisable political pattern: community resistance is now sufficiently organised and widespread that it is stalling builds across the US and becoming a campaign issue ahead of the 2026 midterms. The Verge The concerns are threefold — grid stress and electricity cost pass-through to residential ratepayers, water consumption for cooling, and labour displacement — and they are converging in communities where data centre density is highest.

This is a material infrastructure planning risk. Hyperscalers and co-location developers have been operating on the assumption that permitting friction is manageable at the project level; organised political opposition that influences zoning law, utility commission policy, and state-level legislation changes that calculus. The pipeline of announced capacity commitments — across AWS, Microsoft, Google, and Meta — depends on permitting and grid interconnection timelines that are now subject to a new layer of democratic accountability that was not priced into original buildout projections.

SoftBank is seeking a $10 billion margin loan collateralised against its OpenAI equity stake, according to people familiar with the matter cited by Bloomberg. This follows SoftBank's anchor role in the $500 billion Stargate initiative and its aggressive commitments to AI infrastructure investment globally. The margin loan structure suggests SoftBank is deploying capital faster than it can liquidate legacy assets, using illiquid private equity as collateral to maintain investment velocity.

The risk profile here is asymmetric: if AI infrastructure valuations correct — or if OpenAI's primary market valuation compresses from its current levels — the collateral supporting the loan deteriorates simultaneously with SoftBank's broader portfolio. Given that secondary market valuations for AI companies have reached levels that most primary market analysts consider speculative (Anthropic reportedly trading at $1 trillion on secondary markets per Tom's Hardware), the use of these valuations as loan collateral introduces systemic fragility into AI infrastructure financing.