Sovereign Compute, Price Wars, and Governance Without Teeth

Three distinct capital mobilisations this week illustrate the fracturing of what was briefly a unified global AI infrastructure market. Alibaba is building a vertically integrated stack — proprietary silicon plus 20 GW of global real estate — explicitly designed to reduce exposure to US-controlled hardware. Anthropic is negotiating direct data centre tenancy funded by private credit to escape hyperscaler pricing leverage. SoftBank is raising junk debt at record yields to finance positions that span both US and non-US infrastructure. These are not competing bids for the same asset; they are the construction of parallel compute ecosystems with different ownership, regulatory exposure, and customer bases.

The sovereign dimension extends beyond the corporate level. China's Inner Mongolia buildout is proceeding at state-backed pace; India's domestic data centre IPO market is activating; BDx is breaking ground in Indonesia with a timeline aligned to near-term frontier lab demand. Meanwhile, US export controls — designed to slow Chinese frontier training — are successfully redirecting Chinese engineering capital into adjacent domains: Hygon's embedded robotics processors target industrial AI markets where Western controls are thinner and Chinese manufacturing proximity is an inherent advantage. The controls are not failing, but they are producing displacement effects that Western policy architecture has not fully anticipated.

This week produced a striking concentration of governance activity with negligible enforcement substance. Trump's 'AI Force' exists as executive intent without statute, budget, or interagency framework. The Trump-Xi AI safety dialogue lacks verification mechanisms. The UN tech envoy candidly described multilateral AI safety cooperation as a 'scorpions in a bottle' dynamic. OpenAI's push for US-led international evaluation standards is a call for voluntary benchmarks, not binding treaty obligations. The UK House of Lords is building an evidence record that may eventually produce binding recommendations — but has not yet. Citizens Advice's 'right to human' demand is the outlier: it targets existing statutory hooks under the FCA's Consumer Duty, making it the most operationally proximate governance development of the week.

The pattern across jurisdictions is consistent: executive and political actors are racing to claim governance leadership while avoiding the institutional costs of genuine enforcement architecture. Australia's proposed copyright liberalisation for AI training is the clearest example of the inverse dynamic — a government actively dismantling existing enforcement of rights to attract AI investment. Taken together, these signals confirm that the binding regulatory layer is consolidating at the national and regional level — EU AI Act, FCA Consumer Duty, potential US cloud compute restrictions — while the international layer remains aspirational, creating a variable-geometry world where compliance obligations depend entirely on the jurisdiction of operation.

The simultaneous release of cheaper GPT-6 variants and Claude Opus 5.5 at 40% lower cost than its predecessor is not a promotional event — it reflects a structural dynamic in which frontier labs are competing for enterprise API wallet share at the expense of established SaaS revenues. Procurement data showing AI providers rising from 1.4% to 8% of enterprise software spend in twelve months confirms this is budget displacement, not incremental allocation. The downstream effect is visible in incumbent responses: Microsoft authorising 30–50% Copilot discounts for large seat commitments, and Amazon, Figma, and Workday all cutting prices to defend against churn. The entire enterprise software pricing stack is in compression.

The infrastructure implications compound the commercial ones. OpenAI and Anthropic are simultaneously cutting model prices and scrambling for 20–30 MW interim data centre capacity because gigawatt-scale campuses are years from completion. This bifurcation — aggressive price competition on the revenue side, constrained and expensive capacity acquisition on the cost side — creates structural pressure on unit economics that capital markets are beginning to price in. CoreWeave's $4.2 billion bond upsizing and Snorkel AI's tripled valuation reflect continued institutional confidence in the infrastructure and data layer; the stalling of multiple AI IPOs reflects growing scrutiny of concentrated revenue and unproven margin structures at the application layer.