Is AI finally going local?

What happened

On June 14th, the US Commerce Department ordered Anthropic to immediately disable access to its newly launched Fable 5 and Mythos 5 models for all foreign nationals. The trigger was a convergence of security findings: UK researchers had developed jailbreak exploits within hours of launch, and Amazon reportedly alerted Treasury that Fable 5 had been used to extract cyberattack-applicable information. Anthropic's CEO Dario Amodei resisted, characterising the vulnerabilities as minor and comparable to those in OpenAI's GPT-5.5. The ban was imposed regardless, resulting in an abrupt access blackout for a material portion of Anthropic's user base, and a public demonstration that federal regulators have both the legal authority and political will to shut down frontier models without warning or appeal.

Less than 24 hours later, AMD CEO Lisa Su held up a lunchbox-sized device running a 235-billion-parameter model on stage, with no cloud, or data centre connection. The AMD Ryzen AI Halo, powered by the Ryzen AI Max+ 395 chip with 128GB of unified memory, is open for pre-orders at $3,999 (third-party units as low as $2,000). For a developer spending $440/month on cloud inference subscriptions, the hardware pays for itself in under a year.

Impact on our Investment Case

A new category of enterprise AI risk

Until June 14th, enterprise AI risk discussions centred on familiar territory: GDPR, data sovereignty, EU AI Act classification: real but slow-moving risks, largely navigable through contracts and data-residency choices. Export control risk is structurally different: immediate, binary, and extraterritorial. Affected enterprises did not receive a compliance window; they received an outage. This introduces a second, independent risk axis beyond vendor concentration risk, namely the risk that a model becomes legally inaccessible. The two compound rather than offset: the most capable model is also the most likely to attract regulatory scrutiny.

From single event to structural shift

The assumption underpinning most enterprise AI strategies since 2022, that US-hosted frontier models were a utility, reliably accessible and beyond the reach of domestic policy, has been falsified. What replaces it is a fragmented deployment landscape, with three paths opening simultaneously.

Sovereign and private cloud are the most accessible near-term options. Hyperscalers with dedicated on-premise offerings (Azure Stack, Google Distributed Cloud) benefit from stickier enterprise contracts, existing government compliance infrastructure, and meaningfully higher political friction around cutting off private deployments versus disabling a public API endpoint.

On-device inference is the more radical escape. AMD's Ryzen AI Halo sits alongside Nvidia's DGX Spark and Apple's Mac Studio in a category of unified-memory devices capable of running large models locally. AMD claims the chip outperforms the RTX 5080 by over 3x on DeepSeek R1 inference, though this applies specifically to models exceeding the 5080's 16GB VRAM limit and has not been independently verified at scale. The category has a real ceiling: these are prototyping and developer environments, not production infrastructure. At high concurrency, the throughput gap versus a data-centre rack is not a rounding error.

Open-weight models are the thread connecting all three paths: DeepSeek, Qwen, and Llama are the only models deployable across sovereign cloud, on-device, and on-premise infrastructure alike. Every time a closed frontier model becomes inaccessible, their competitive moat widens directly.

The geographic dimension

The impact of June 14th is not symmetrical across geographies.

In Europe, it validates a policy direction already in motion: Mistral and other domestic sovereign-AI initiatives across France and Germany are pointing to it as Exhibit A, and EU procurement shifts toward local providers will accelerate.

The more structurally significant dynamic is in China. Of the major technology markets, it is the most advanced along the AI sovereignty path and by a substantial margin. Domestically anchored open-weight models have achieved near-frontier performance at a fraction of comparable training costs. With roughly 75% of Chinese technology sector revenues generated domestically, these companies are insulated from precisely the kind of access disruption June 14th demonstrated. The government's 90% AI adoption target for state-owned enterprises by 2030 is a procurement schedule, not merely an aspiration. Yet, as we discussed in detail, Chinese AI and technology continues to trade at a significant discount to US peers, a valuation gap that reflects a market still anchored to the pre-2024 risk narrative rather than the new structural reality.

Our Takeaway

These events sharpen rather than change our view on where durable value will be captured within AI. We are exposed to this transition through cloud infrastructure with sovereign deployment optionality, on-device inference hardware, and the open-weight model ecosystem. The geographic dimension, with China technology as a distinct asset class with different return drivers, domestic revenue anchoring, and a structural valuation discount, adds diversification with its own logic beyond a single macro call, and one that June 14th has made more, not less, relevant.

June 14th is the first exercise of Washington's authority over frontier model access; it will likely not be the last. What matters for positioning is whether regulatory action broadens to other providers or model categories; either way, the direction of travel is the same. We are monitoring closely how quickly the structural shift from centralised cloud AI to a multi-polar deployment landscape translates into measurable revenue. We are positioned to capture that transition across all three vectors.