Capital & Industrial Strategy

The White House Office of Science and Technology Policy, under Michael Kratsios, issued a memo accusing Chinese firms — led by DeepSeek — of systematically distilling US AI models at industrial scale, effectively extracting capability without the underlying compute investment. Separately, the US State Department ordered a global diplomatic warning to allies, per Reuters, signalling that IP enforcement is now a coordinated foreign policy instrument, not merely a domestic concern.

The strategic logic here is significant: if model distillation is cheap and effective — and DeepSeek's R1 demonstrated it can be — then US export controls on chips may be insufficient to preserve frontier model advantage. The administration appears to be moving toward a doctrine that treats model weights and training methodologies as controlled assets, analogous to dual-use hardware. The implications for open-source AI policy are profound: any US push to restrict weight distribution would directly conflict with the open-source advocacy from players like Meta and the academic community. Separately, Anthropic is probing possible unauthorised access to its restricted Mythos model, reportedly accessed by users who guessed its API endpoint location — a security failure that reinforces White House concerns about model leakage.

Cohere's acquisition of German AI startup Aleph Alpha, with financial backing from Schwarz Group (the Lidl and Kaufland parent with roughly €140bn in annual revenue), is the most strategically coherent European AI consolidation move to date. Per TechCrunch, both companies' governments have blessed the deal, and the explicit strategic intent is to offer European enterprises a credible non-US alternative for AI deployment — addressing data sovereignty concerns that have historically slowed hyperscaler adoption in regulated European sectors.

The deal is notable for what it combines: Cohere's enterprise API and model infrastructure with Aleph Alpha's European regulatory relationships, data residency credentials, and government contracts — particularly in German public sector and defence-adjacent clients. Schwarz Group's backing provides both capital and an immediate enterprise customer base at scale. This is not a distressed acquisition; it is a deliberate industrial strategy move, likely encouraged by Berlin and Ottawa, to create a viable European AI champion before US and Chinese players lock in the enterprise market. The deal signals that the 'sovereign AI' investment thesis is consolidating from multiple small national bets into fewer, better-capitalised regional platforms.

ServiceNow's Q1 results — which beat earnings and revenue expectations — nevertheless sent the stock down over 16%, with Salesforce, Workday, and Oracle falling in sympathy, per CNBC. The trigger was a subscription revenue miss attributed partly to geopolitical uncertainty from the Iran war, but the market reaction was disproportionate because it confirmed a structural fear: that AI-driven automation is beginning to reduce enterprise seat counts and per-user licence revenue. ServiceNow's own CEO acknowledged the company would not backfill open roles, relying instead on AI productivity gains — a statement that simultaneously validates the AI investment thesis and undermines the software-as-headcount-multiplier model that has driven SaaS valuations for a decade.

J.P. Morgan flagged elevated hedge fund short interest in ServiceNow, per Reuters, and the EQT warning that AI concerns will stall PE exits from software stakes — reported by the Financial Times — suggests the repricing is not confined to public markets. Private equity software portfolios are now facing a valuation compression cycle driven by AI disruption risk, creating a potential overhang on PE exits across the sector. The talent dynamic is compounding this: CNBC reports that senior enterprise software executives are defecting to OpenAI, accelerating capability drain from incumbents precisely as they need AI talent most.

Meta's 8,000-person layoff — framed explicitly by the company as relieving AI spending pressure, per Fortune — combined with Microsoft's concurrent buyout programme, totals up to 23,000 roles. This is a structurally different dynamic from prior tech layoffs: the capital freed is not being returned to shareholders or used for debt reduction, but is being redirected into AI infrastructure capex. Meta's parallel initiative to track employee keystrokes across Google, LinkedIn, and Wikipedia as AI training data — reported by CNBC — suggests the company is simultaneously cutting labour costs and harvesting remaining workforce behaviour as training signal.

Google CEO Sundar Pichai's disclosure that 75% of Google's new code is now AI-generated, per Semafor, with Meta targeting similar levels by mid-year, provides the productivity justification for these headcount reductions. The operational logic is now explicit across hyperscalers: AI-generated code reduces the marginal return on software engineering headcount, making labour a source of AI funding rather than a growth input.

DeepSeek's V4 preview release was the most anticipated Chinese AI model drop since R1 in January, but the market and analyst response was notably subdued. Bloomberg concluded the model fails to narrow the US lead, and TechCrunch noted the company itself only claimed to have 'almost closed the gap' on reasoning benchmarks — a significant walk-back from R1's shock performance. Crucially, V4 is specifically optimised for Huawei Ascend chips, per Reuters, which is the strategically significant architectural choice: it demonstrates that China's AI development stack is explicitly designed around domestic silicon, reducing dependence on NVIDIA and validating Huawei's position as the hardware backbone of Chinese AI.

The combination of a muted benchmark performance and the White House IP theft allegations creates a more complex picture than either a pure 'China catching up' or 'China falling behind' narrative. The Huawei chip optimisation suggests DeepSeek is prioritising deployability within China's constrained hardware environment over competing head-to-head with GPT-5.5 or Gemini on international benchmarks — a rational strategic pivot that serves the domestic market and government infrastructure deployment goals regardless of frontier performance.