Good morning. I’m on my way to Legal Tech Talk in London this morning, speaking on the new shape of in-house legal.
Part of that new shape has been on display this week. Anthropic was made to switch off its two newest models at just 90 minutes’ notice after a US export-control order. Like many legal and governance stories, the most interesting question sat below the headlines. Leadership teams will look past the drama and towards how to figure their own underlying dependencies. This week we unpack practical steps to take and why it matters. 🎯
— Philip
BRIEFING ROOM
Rent control
Last week, at almost immediate notice, Anthropic was forced to suspend access to its new Fable 5 and Mythos 5 models after receiving a US export-control directive. The order cited national security, barring any foreign national, whether inside or outside the USA, from accessing specified AI models. The suspension made headlines because it was dramatic, visible and immediate. It raised legitimate questions about sovereignty, geopolitics and dependence on US technology providers.
For legal teams and organisations navigating day-to-day commercial reality – and likely underutilising AI capability, far away from frontier models – such topics are abstract. Yet the story does underline the exposure of UK and European companies building business models on top of mostly rented American tech.
Bristows compared the ban to “Crypto Wars” examples from the 1990s as evidence that such restrictions are unlikely to hold. Given a more combative, competitive international landscape, boards may find little comfort in a history lesson. Even momentary disruption has operational and financial consequences. But the more persistent risk sits underneath it. Model providers already influence business economics regardless of political interventions. Kill-switches are likely the rarest, most visible and most reversible version of the risk. The suspension made the headlines; supplier leverage exists every day.
Plug and pray
Vendor-dependence has most often been discussed in terms of pricing power. Anthropic's shift earlier this year from flat-rate to usage-based pricing destabilised the companies building on top of it, and sector-wide API costs are currently held low by providers fighting for share — meaning anyone built on those APIs is exposed the day the model layer starts to charge what it costs.
The deeper threat to a balance sheet is the risk that grows every day the model is not switched off: the supplier choosing to become the competitor. Within weeks this spring, an in-chat app builder appeared that implicitly could compete with Anthropic's own client Lovable; and a Claude for Legal offering put model-dependent legal-AI firms Harvey and Legora on the defensive. The industry word for a supplier absorbing the feature you built on top of it is sherlocking, and the model-makers have every incentive to do it.
Capabilities that once supported independent products increasingly emerge as native features within model ecosystems themselves. What begins as a supplier relationship can quickly become a competitive one. The question for leadership teams is whether the organisation has a moat beyond access to the model itself.
Finally, there is contractual power. Most standard AI terms already anticipate disruption. Providers typically reserve broad rights to modify services, suspend access for legal or regulatory reasons, or replace functionality. What matters is whether customers have negotiated meaningful remedies when those changes affect business-critical operations.
Fabled sovereignty
The structural answer to all of this is sovereign AI. Governments in both the EU and UK are actively pursuing it, but for most businesses it remains too far away to be commercially meaningful. Brussels published its Technological Sovereignty Package this month, with a Cloud and AI Development Act that will grade providers across four assurance levels. At London Tech Week, the UK government announced £1.1bn for AI hardware alongside its £500m Sovereign AI Fund — but industry voices warned the money risks buying “British-branded infrastructure on somebody else's silicon” with no domestic model competing anywhere near Fable or Mythos today. That reality was reflected in Europe 2031, an 18,000-word manifesto published this week, which argues the continent must make “ugly trade-offs" with US providers to avoid a “slide into irrelevance”, because the frontier race has already been lost.
Why it matters
As with any AI debate, futurology can obscure practical next steps. GCs and leadership teams will look for opportunities to mitigate their exposure.
🔹 Do we know where our AI concentration risk sits? Most boards can identify their major cloud providers. Far fewer can identify the foundation models underpinning critical business processes, or understand how much operational dependency exists on a small number of suppliers. Mapping exposure and diversification is a sensible first step. Adding a fallback or provider-agnostic architecture could keep a critical process running if a model vanishes at 48 hours' notice, noting that trade-offs would include loss of model-specific tuning.
🔹 Have we modelled the economics of dependency? Many organisations are focused on AI capability, at the expense of leverage. The useful step is identifying how pricing works, who controls it, what happens if costs increase materially, and how exposed key workflows are to changes in commercial terms. Simple steps like identifying AI cost per transaction and the impact on gross margins if that cost doubles creates a quantifiable input to plan against.
🔹 Understand what is negotiable. Terms may be flexible once you learn the right questions to ask. Getting the underlying LLM named in the contract, with a duty to notify you before it changes or is withdrawn, lets you map concentration and catch a model swap that degrades a workflow, rather than discovering it when output quality drops and no one can say why. Negotiating to exclude regulatory change from the definition of force majeure, that a fallback to a weaker model triggers service credits or a fee cut, and that material suspension allows termination without penalty, converts an uncontrollable event into a priced, bounded one with a remedy attached.
Fable will likely be restored within weeks. The dependency underneath it, and adjacent LLMs, isn't going anywhere. Forward-thinking GCs will use this week's story to identify where leverage sits before they find themselves scrambling contracts at 5:21pm on a Friday afternoon.
FROM THE SIDEBAR
Quick signals worth clocking
🇬🇧 UK Government Office for Science war games its AI Scenarios 2030, from “Slow Burn” to “Take-off”. One finding for leadership teams: “The opportunity for UK businesses to benefit from AI deployment is substantial, and AI-enabled gains are expected to become the main source of the UK’s continued productivity growth”
💼 Anthropic is recruiting a Policy Counsel EMEA in London at up to £240k, a well-paid seat that, after this week, comes with a unique context the advert doesn't quite capture.
🥸 As the UK announces its social media ban for under 16s to be enforced in part via face age-estimation tech, consider the risk of adversarial attacks from children tricking age-verification tools with... drawn on moustaches.
Enjoying the signal?
If this edition would help a colleague thinking through AI governance, enterprise risk or legal operating change, feel free to forward it on.
💬 Forward to a colleague
🧠 Was this forwarded to you? Subscribe here to get Profiles in Legal every Wednesday.
Here’s how I can help
I advise technology businesses and leadership teams on AI, product and regulatory strategy and enterprise readiness to move from informal AI adoption to scalable governance.
If your organisation is navigating AI deployment, maturity or commercial negotiation challenges, feel free to reply directly.
- Philip
Profiles in Legal examines how AI, governance and technology are reshaping modern businesses and legal teams.
This publication is for general information only and does not constitute legal advice. Seek professional advice for specific situations.