Protesters march through downtown Memphis, Tennessee, carrying signs opposing Elon Musk and xAI during the "Get Out of Memphis" demonstration on October 4, 2025. (Photo by AUSTIN JOHNSON/AFP via Getty Images)

Addressing Regulatory Arbitrage in the AI Supply Chain

Brazilian President Luiz Inacio Lula da Silva holds a cabinet meeting at Planalto presidential palace in Brasilia, Brazil, Wednesday, June 3, 2026. (AP Photo/Eraldo Peres)

Big Tech's Proxy War Against Brazil's New Platform Accountability Rules

Members of the European Parliament Arba Kokalari and Michael McNamra pictured during the AI omnibus interinstitutional negotiations. / Photo: EP

Members of the European Parliament and Arba Kokalari and Michael McNamra pictured during the AI omnibus interinstitutional negotiations. / Photo: EP

EU Lawmakers Move to Ban AI Nudifiers, but Enforcement Remains Unclear

An array of colorful, fossil-like data imprints representing the static nature of AI models, laden with outdated contexts and biases.

Luke Conroy and Anne Fehres & AI4Media / 

A Case Against Forever: Why Cloud and Social Media Need Inactivity-Based Deletions

Howard Lutnick, US Secretary of Commerce, testifies before the House Committee on Appropriations Subcommittee on Commerce, Justice, Science, and Related Agencies in the Rayburn House Office Building in Washington, DC on Thursday, June 5, 2025. (Photo by Aaron Schwartz/Sipa USA)

Did the US Government Just Set An AI Export Precedent by Blocking Mythos?

An abstract image based on the US Constitution

Earning a Social License for Transformative AI

BRUSSELS, BELGIUM—JUNE 3: EU Commissioner for Tech Sovereignty, Security and Democracy Henna Virkkunen introduces the European Technological Sovereignty Package, featuring the Chips Act 2.0, Cloud and AI Development Act, and related strategies to reduce the EU's dependence on non-European suppliers for critical digital technologies. (Photo by Thierry Monasse/Getty Images)

Europe’s AI Sovereignty Problem Runs Far Deeper Than Frontier Access

Rep. Lori Trahan (D-Mass.) speaks during an interview with POLITICO in her office on Capitol Hill on May 21, 2026. (Francis Chung/POLITICO via AP Images)

Unpacking the Great American Artificial Intelligence Act of 2026

Anthropic's Co-Founder and CEO, Dario Amodei on stage at the Code with Claude developer conference on Wednesday, May 6, 2026 in San Francisco. (Don Feria/AP Content Services for Anthropic)

Anthropic's Mythos Recall and the White House's Missing AI Safety Playbook

G7 Summit Set to Kick Off Amidst Allies' Widening Rift Over AI Sovereignty

A collage featuring a vintage illustration of a woman’s head mapped with labeled sections resembling a phrenology chart. The mapped sections are overlaid by a neutral network diagram– depicting crisscrossing black lines. Two anonymous hands extend from the left side, pulling on two wires from the diagram.  In the background is a panel of the Turing Machine with numerous knobs and switches, highlighting a connection between the history of computing, psychology, biology, and artificial intelligence. 

Personifying AI Harms People and Protects Companies

Helen McEntee, Ireland's Minister for Foreign Affairs and Trade, outlines priorities for Ireland's upcoming EU Council Presidency on June 10. Source: Irish Presidency of the Council of the EU 2026.

Ireland Targets AI Omnibus, Child Safety in EU Presidency Agenda

Northwest view up to the pediment, rotunda, and dome of the California State Capitol in Sacramento. (Radomianin / Wikimedia Commons)

Northwest view up to the pediment, rotunda, and dome of the California State Capitol in Sacramento. 

 States Step in Where Congress Stalls on AI Safeguards for Kids

A man showing mental distress from constant exposure to harmful content online. His family, in the background, progressively disappears.

How AI’s Labor Supply Chains Fail Workers

Image of map and computer language over binary codes on server racks in server room.

Global Digital Policy Roundup: May 2026

Abhijeet Dipke, center, founder of the Cockroach Janta Party addresses supporters during a protest demonstration in New Delhi, India, Saturday, June 6, 2026. (AP Photo/Manish Swarup)

After Test Data Scandal, India’s Youth Stake Their Claim on the Digital Square

A blur effect added to Leonardo da Vinci’s Mona Lisa

AI Can Rebuild Blurred Faces, So How Do We Protect People Now?

Close-up of Keir Starmer, wearing a dark suit, white shirt, and tie, looking directly at the camera.

Prime Minister Sir Keir Starmer delivers opening remarks ahead of a meeting with senior figures from TikTok, X, Meta and other social media giants at 10 Downing Street, Westminster. Picture date: Thursday April 16, 2026. 84281335 (Press Association via AP Images)

Could a UK Teen Social Media Ban Work Without Expanding Surveillance?

The image shows a hand holding a smartphone in front of a computer screen with war and other words projected across.  

Hidden Spyware Use in War May Be More Widespread Than We Realize

How OpenAI, Google, and Anthropic Plan to Handle the 2026 US Midterms

Tatiana Dias is a fellow at Tech Policy Press.

Just hours after the Brazilian government published decrees expanding the liability of internet providers in Brazil last month, a congressional offensive to strike down the new rules began. Within 24 hours, 26 legislative decree proposals 

The move was visibly coordinated. The proposals were submitted by lawmakers from right-wing and far-right parties — most from PL, the party of former president Jair Bolsonaro — and share nearly identical arguments. Some are word-for-word the same.

As João Brant and Nina Santos, Secretary and Deputy Secretary for Digital Policies at the Brazilian Presidency, 

, the government's decrees aim to regulate last year's Supreme Court ruling — known as the STF — 

 of Brazil's Internet Civil Rights Framework, the Marco Civil da Internet.

For ten years, Brazilian law held that platforms were not liable for user-generated content. Last year, the STF ruled that article partially unconstitutional.

In practice, the decision shifted the liability regime for platforms in Brazil, but it 

. The STF is currently reviewing appeals filed by Big Tech companies against last year's ruling. The reporting justice, Minister Dias Toffoli, included a 

. The final ruling is expected to be published on Wednesday, June 17.

But it fell to Congress — according to the 

 — to establish the rules for the new liability regime to take effect. That hasn't happened so far. So the executive branch stepped in with its own regulatory proposal.

One of the Lula government's decrees is directly tied to the Marco Civil: it would require companies to exercise a duty of care over criminal content, prevent online fraud, and give the National Data Protection Authority, the ANPD, oversight power over platforms. The second focuses on protecting women online, requiring platforms to remove non-consensual intimate images within two hours and to put in place mechanisms to protect women — especially public figures and journalists.

As soon as the government released the proposals, an intense congressional push to block them began. Unlike previous episodes — 

 which was defeated after intense Big Tech lobbying — this time the companies did not act directly.

Instead, lawmakers moved in with requests to suspend the executive decrees. The proposals share the same argument: first, they invoke the Constitution to claim the decrees exceed executive authority. Then they assert that the decrees create "unprecedented obligations"; they challenge terms like "systemic failure" and "duty of care"; they warn of private censorship and free speech violations; and they conclude that regulating the matter falls outside the ANPD's mandate. Some call for the decrees to be struck down entirely; others target specific provisions.

None of the proposals have been voted on yet, but tensions are running high.

, the Brazilian Chamber of the Digital Economy — which includes companies such as Meta, OpenAI, Google, Kwai, and TikTok — alongside the Latin American Internet Association and the Conselho Digital, a Big Tech lobbying group, criticize the executive decrees as an "unusual" regulatory path and warns of legal uncertainty and excessive content takedowns.

Public figures mirrored these arguments. Lawyer Ronaldo Lemos, a member of the Oversight Board (Meta's content moderation council), an advisor to Spotify, and director of the think tank ITS-Rio, 

 sharply criticizing the government's decrees, calling the executive's move an "attack on the Constitution" and claiming that, by designating the ANPD as the regulatory body, Brazil is "copying the Russian model of concentration."

Carlos Affonso Souza, law professor and also director of ITS-Rio, said that Brazil uses the the 

digital sovereignty narrative as a 'shield’

compared the decrees to a Pandora’s box

: “once opened, jars or boxes, certain evils don't easily go back inside: diffuse surveillance, lawless competition, the space to silence legitimate dissent”.

In response, other commentators — including law professor Conrado Hubner — 

's authority to regulate the STF's decision, particularly given Congress's inaction on the issue. "The weak regulation of Big Tech offers a clear example of how legislative silence is not always an authentic expression of popular sovereignty. In this case, there is documented evidence of congressional capture by the most insidious corporate power ever seen in the history of capitalism," he wrote.

"Regulating the administrative operation of these rights, within the boundaries set by the STF and while awaiting congressional action, is not mimicking authoritarian countries. It is exercising the digital sovereignty of a democratic state under the rule of law," wrote seven other researchers in law, journalism, and sociology, in a 

separate piece defending the government's decrees

Amid the dispute, Big Tech still holds the 'Trump card'

There is yet another external element adding pressure to the debate. On June 1, the USTR proposed new 25% tariffs on Brazilian goods. The US investigation, which began last year, accuses Brazil of adopting anti-competitive practices that harm American companies — particularly in digital markets, payment systems, ethanol, corruption, and deforestation.

The USTR singles out Brazil's PIX payment system as an unfair practice and also references "secret court orders" suspending accounts and platforms — a reference to 

decisions issued by Justice Alexandre de Moraes

The STF ruling that declared Article 19 of the Marco Civil partially unconstitutional is also directly cited in the 

. According to the USTR, the decision "compounds this situation of uncertainty and risk, and effectively requires companies to choose between potentially incurring substantial liability for user-generated content and preemptively taking down potentially lawful content."

The tariff announcement came on the same day Trump 

with Jair Bolsonaro's son, Flávio, the far-right presidential candidate in this year's elections. Although Flávio 

 the new tariffs and the visit, last year his brother Eduardo has actively lobbied for U.S. retaliation against Brazil and publicly celebrated Trump's support after the announcing of the tariffs.

That moment, when Trump first floated the possibility of tariffs of up to 50% on Brazilian goods, Brazilian Vice President Geraldo Alckmin took the lead in negotiations, meeting not only with US officials but 

directly with the tech companies themselves

. In a series of meetings with Meta, Google, Amazon, Apple, Visa and others in July, following the tariff announcement, Alckmin signaled openness to data center investments and argued that Brazil "should not rush" to regulate Big Tech. 

With the decrees, the picture that emerges is one of coordinated pressure on multiple fronts: in Congress, through legislative decrees; in the public debate; and now on the geopolitical stage, with Washington using tariffs as leverage to defend American corporate interests in Brazil's digital market.

For the Lula government, the question is beyond regulatory. It is whether Brazil has the autonomy to set the rules of its own digital environment, or whether that decision will be negotiated, as it was in 2025, in closed meetings with the tech giants themselves.

 that would prohibit AI systems designed to create nonconsensual intimate images, one of the most 

 measures in a broader package aimed at simplifying the bloc’s Artificial Intelligence Act.

 reached on May 7 on legislation known as the "AI omnibus," which includes a new prohibition on so-called "nudifiers" — AI systems that generate explicit images of real people without their consent. Yet while lawmakers have repeatedly described the measure as a ban, many national authorities still lack the legal powers needed to enforce it.

The agreement requires companies to implement safeguards against the creation of nonconsensual sexual deepfakes and could remove offending systems from the EU market by the end of the year. “This is not a strict prohibition,” said Sadia Berdaï, head of the AI Innovation and Technology Unit at Luxembourg’s Data Protection Authority. “It’s about the intended purpose, the way you use the systems, not the technology itself.”

The issue of sexualized deepfakes returned to headlines recently after 

AI images of Italian Prime Minister Giorgia Meloni

 in lingerie circulated online. The case is a reflection of the broader problem, given that 

 of AI pornographic videos online are non-consensual, and 

“We're talking about systems designed to strip clothes from photographs of real people, not synthetically created people, but real people, overwhelmingly women and children, to humiliate them, degrade them, and objectify them. To do so at scale and industrialize that for profit,” Michael McNamara, a co-rapporteur of the AI Omnibus and a member of the European Parliament, said during the debate on Monday ahead of the vote.

The agreement prohibits AI systems intended to generate nonconsensual intimate content involving identifiable individuals. Users would also be barred from intentionally creating such content without consent.

The measure would not prohibit all AI-generated sexual content. Instead, it would allow the creation of synthetic intimate material when all depicted individuals have provided consent.

To qualify for protection, the targeted individual must be identifiable under the General Data Protection Regulation (GDPR), including through a face, name, location, online identifier or other distinguishing characteristics.

Belén Luna Sanz, a women's and digital rights advocate, said that the requirement may create challenges for victims. "It creates a significantly higher threshold for protection of victims and enforcement at the level of product safety," she said.

The proposal also excludes fictional characters and certain artistic, cartoon and medical uses. Critics say questions remain about how different cultures understand intimacy. Silvia Semenzin, a researcher focused on digital violence, pointed to examples involving AI-generated images of Muslim women without hijabs, which some may experience as a serious violation of privacy despite not being sexually explicit.

"Challenges remain regarding the definition of what constitutes 'intimate' content, a concept that varies across cultures and even between individuals," Luna Sanz said.

Who bears responsibility for consent and safeguards?

, consent must be freely given, specific, informed, unambiguous and explicit. Companies offering systems that generate or manipulate intimate content would be required to establish reliable mechanisms to collect and demonstrate consent while complying with the GDPR.

For Semenzin, the burden of proof remains a central concern.

"The person creating or sharing the images should have to prove they had consent and not the other way around," she said. "Victims should not be forced to prove, after the image has been created and it's already circulating, that they did not authorize it."

The agreement would also place new obligations on developers. Regulators could scrutinize companies whose systems predictably and repeatedly generate nonconsensual intimate content if they fail to implement adequate safeguards.

Luna Sanz said the approach is significant because "it puts responsibility not only on the perpetrators of such abuse, but also on the companies."

According to the text, providers would be expected to deploy safeguards including data cleaning, refusal training, prompt-design protections, output controls, runtime guardrails, content filtering, usage restrictions, abuse-detection systems, and notice-and-action mechanisms. 

“Platforms will be required to implement robust safeguards, including prompt-level filters, building on existing practices in other regulated domains,” Member of the European Parliament and shadow rapporteur for the Omnibus, Sergey Lagodinsky, told Tech Policy Press. 

Markéta Gregorová, another member of the European Parliament and shadow rapporteur on the Omnibus, said many providers already employ similar protections. “Most AI providers don’t want these images to be generated regardless and already have measures in place,” she said. “The cost for these providers will be minimal since they already comply with the ban.” X chatbot, Grok, still 

As with other parts of the AI Act, enforcement is the real 

, requiring coordination between national AI 

, and the European Commission. At the moment, 

 have designated a market surveillance authority, and even those bodies have limited powers until national legislation is in place.

“Until national laws are voted, we are not competent to enforce the AI Act,” Berdaï said. According to Berdaï, Italy and Denmark are among the few countries that have approved national laws allowing them to enforce nudity prohibitions. In Luxembourg, regulators can act under the GDPR when personal data processing is involved, but not under the AI Act itself. “It’s just a matter of competence, and as long as the law is not voted on, we don’t have the competence, even though we are skilled,” said Berdaï.

Lithuania faces similar challenges. The country’s Communications Regulatory Authority told Tech Policy Press that “the practical enforcement arrangements are not yet fully defined at the EU level,” including how responsibilities will be divided between national market surveillance authorities and the European AI Office, and that it was still too early to know the exact procedures that will apply from December 2026.

Breaching any of the AI Act’s prohibitions can lead to fines of up to EUR 35 million or, for companies, up to 7% of worldwide annual turnover from the previous financial year, whichever is higher.

After Tuesday’s vote in Parliament, the agreement needs to be adopted by the Council, then published in the Journal of the European Union. Given that the AI Omnibus delays the high-risk provisions scheduled for August 2026, Brussels wants the file ready before that deadline. However, the ban on nudifiers is only due to start on Dec. 2.

 is published in partnership with the AI Now Institute.

This weekend, the United States administration 

 to cut off foreign access to two of its advanced AI models. In Europe, the news has been interpreted almost unanimously: a wake-up call that shows dependency means access can be cut off overnight.

, we make the case that Europe’s AI market is 

 with the ecosystems of dominant US players, in ways that boosting both supply and demand alone cannot disentangle. As a result, interventions that seek to secure sovereignty in AI risk leading to much more entrenched dependence. The first part of this series looks at the demand-side of things.

The limits of AI sovereignty at the application layer

 is lauded as one of Europe's biggest AI success stories — and understandably so.

Founded in Stockholm in 2023, it was quickly 

 as the fastest-growing software startup in history. Currently, the company is 

 at $6.6 billion after raising a total of $540 million and is frequently cited alongside Mistral as proof that Europe can build globally competitive AI companies. Lovable lets anyone build a website or an app (“something they love”) by chatting with an AI, and lots of people do.

But Lovable’s rise is dependent on the kindness of strangers. Instead of developing their own models, deep technological moats, or exploiting a local technology stack, the company builds on the American tech stack: prompts are routed to models made by Anthropic and Google and distributed in the existing hyperscaler ecosystems. For example, on June 4, Lovable 

 a deepened partnership with Google Cloud that is supposed to lead to a “fivefold increase” in Lovable's Google Cloud footprint, with recent 

 of the products into the Google Workspace.

Let us be clear. There’s nothing wrong with focusing on making AI useful, and building on top of existing models. It’s also hard, requires speed of execution, great UX, nurturing a community of users, and getting there first. As a long-term business strategy, however, it carries a risk.

 that Anthropic created a new in-chat app builder that lets users generate applications from simple prompts — essentially competing with their client Lovable’s entire product. Weeks later, Anthropic launched 

, a law-focused offering that put legal AI startups on the defensive — including Harvey, Legora, and Robin AI, all of which deploy Claude models at the core of their products. The supplier had become a competitor twice over within a matter of weeks. This pattern — an incumbent absorbing the feature a smaller company built on top of it — is known as getting “

,” is a well-known, yet often-forgotten trope in the tech industry, named after Apple introduced a new feature that rendered a popular third-party tool irrelevant. 

Not all European start-ups build on proprietary US foundation models: some train their own, some run smaller, domain-specific models on proprietary data and on-premise infrastructure, or focus on physical-AI. However, we begin this series by interrogating application-layer companies that primarily create end-user software and workflows on top of existing foundation models, rather than training frontier models or operating large-scale AI infrastructure. That is because the application layer is a specific bet European policy has chosen to place.

Lovable is an example of a broader trend in European tech strategy. It is based on an assumption that Europe's opportunity in AI lies in building tools that make AI useful at the application layer (rather than spending billions on building deep technology, compute infrastructure and parallel ecosystems). The application layer is where most of the profit will be made, so the argument goes, and focusing on applied AI helps Europe stay in the ‘AI race’ by essentially piggybacking on the investments of others (

 by SAP’s CEO in the Economist and these 

 by Skype founder-turned-venture capitalist are prime examples for this position).

This idea is genuinely tempting. The bet (call it “sovereignty-at-the-application-layer”) provides a profitable way forward without playing the prohibitively expensive game of competing at the frontier-model layer. Asset-light software companies, with low headcount and few hard, illiquid assets, are a far better fit for how venture capital is built to allocate money. It also pattern-matches to how the internet era seemed to play out. From travel platforms like Booking.com to streaming services like Netflix, major winners of the internet economy emerged at the application layer.

The idea is also alluring to policy-makers. “Applied AI” first entered the European policy world with von der Leyen’s 

Political Guidelines for the 2024-2029 Commission

, which set out the intention to build an Apply AI Strategy to boost new industrial uses of AI.

But the policy argument conflates financial efficiency with strategic autonomy. There are three serious arguments that speak against the long-term sustainability of the application-layer bet for Europe: creeping supplier competition, loss of control and value capture upstream

Suppliers as competitors and ultimate beneficiaries

Anthropic competing with their client, Lovable, isn’t an isolated incident. There’s been an industry-wide trend of model-makers entering the application market, driven by the need to capture value beyond the raw model provision. Claude Managed Agents (alongside Claude Cowork) are a direct answer to the self-hosted infrastructure for autonomous agents developed by OpenClaw (the free and open-source AI agent that can run on device). OpenAI tried to acquire coding assistant Windsurf (which was ultimately 

Not just model makers, but also dominant hyperscalers are both suppliers and potential competitors for application-level start-ups. Amazon, Google, and Microsoft all compete at the application layer directly, Microsoft on office and productivity software integrated into Office 365, Google across search and productivity, and AWS now also on industrial applications of AI. All three companies also sell AI as a service bundled into their cloud platforms, making it excruciatingly difficult to compete with these players.

This is what distinguishes AI from earlier software eras. Companies like Salesforce and Adobe could build lasting application-layer franchises because their suppliers sold commodity inputs and had no designs on their markets. AI applications rent their defining capability from a handful of firms that are expanding into applications themselves, and that set the marginal cost of every customer interaction.

If all of this sounds familiar to astute historians of Europe’s relationship with dominant tech companies, that’s because it is. Europe already has its application-layer success story, and it is a tenant. Spotify, its most celebrated consumer-tech champion, runs on someone else's cloud, hands most of its revenue to the rights holders it cannot do without, and has spent years fighting Apple over App Store fees, payment restrictions, and market competition. Succeeding at the application layer and achieving independence turn out to be different things.

Application-layer AI companies inherit an even harder version of this. Their marginal costs are rigid and set upstream by model providers. And while they could theoretically switch to open models, once these reach parity with proprietary ones, leaving is hard, often by design, and the financial and infrastructural dependencies that are being built are hard to disentangle. As a result, these companies bear far more resemblance with a different corner of the internet economy: the flight-comparison sites that ran on Google traffic until Google Flights displaced them; the publishers who reorganized around Facebook's algorithm and pivoted to video, until 

 the engagement figures had been vastly inflated by Facebook; and apps like Spotify that remain subject to Apple’s ever changing App Store terms, 

and a cut of up to 30% on many transactions

Today, we’re seeing this exact dynamic play out at a much larger scale. European AI companies building on American APIs risk becoming something like an outsourced R&D department for leading model providers: they take the risks of discovering which applications actually work, while the value of those discoveries flows back down the stack. The moment a European startup comes up with a seriously profitable idea, the model makers have every incentive to acquire the company or simply rebuild the product as a feature.

This is bad for Europe’s long-term independence, but for the typical venture capital cycle, selling early to a wealthy incumbent is an attractive way to reach exit liquidity — and resisting that pull is hard for a startup stacked with VC. Lovable currently has 54, mostly venture investors in its cap table, with Google’s growth venture arm CapitalG leading the latest funding round. Instead of battling for market share against stupefyingly wealthy incumbents, selling to these companies provides a successful exit for the early investors and a small team of founders. This points to the 

 between the strategic objective of technological sovereignty and the 

 of venture capital financing. While the European AI ecosystem has notable 

, this typical VC model still skews startups’ trajectory towards the US ecosystem.

It is often noted that European startup founders invest in a new generation of founders, 

. But this flywheel might never spin free of the structure of which it is part. It might lead to a cycle of serial exits to the balance sheets of US tech ecosystems, without generating enough escape velocity to create alternative gravity wells that could serve as a scalable platform anchoring European companies to pursue pathways. Against this gloomy picture, a counterargument is routinely offered: European companies can protect themselves through proprietary data, scale up fast enough to become dominant themselves, expand through mergers, or form a captured user base into long-term contracts. Moreover, the irreplaceability of context-specific data generated by proprietary know-how and workflows of vertical industries and incumbents 

 the insulating layer protecting the European companies from US competition, and creating staying power in the European ecosystem

But these mythical, vertical-specific data lakes in manufacturing or healthcare are not available for horizontal startups such as Lovable, which have to depend on customer traction, user experience, and speed as their differentiating factors. These are relatively weak moats compared to the financial dominance, ecosystem control, and integration advantages of existing platforms that control the engine on which these corporations depend. The first-mover advantages are real, as OpenAI testifies. However, the first to move is not necessarily the last to survive.

Infrastructure and model layers competing with application-layer start-ups that depend on them isn’t the only reason why focusing on the application layer is risky. When delegating a core part of their infrastructure, companies also delegate control over pricing models, and APIs can change overnight, while model behavior can be altered without notice. This leaves a fundamental vulnerability at the heart of an API-based business model. Recently, for instance, Anthropic 

 from a flat rate to usage-based pricing, destabilizing application builders downstream.

stake their hopes on staying “provider-agnostic”

: switching between models rather than tying their products to one. The bet is intuitive, but it leaves them exposed on two fronts at once. On prices, API costs are currently depressed sector-wide as providers still fight for market shares. Anyone building on top of these APIs is heavily exposed once prices reflect real costs, or when the model layer 

 starts to capture value and is able to extract rent. On performance, reaching the absolute competitive frontier for a given use case may require model-specific optimization. An agnostic start-up has to forgo these optimizations to prevent lock-in

A counter-argument is that orchestration layers that sit between the app and several model APIs could make switching between models more seamless — and many start-ups, such as the German 

, now offer exactly this. But that raises the question of who sets orchestration standards, and AI companies at the application layer would still be dependent on multiple, though now competing, oligopoly of model makers.

The most serious counterargument is open-weight model parity. Simply put, if open-weight models reach a good enough quality, on par with leading, proprietary frontier models, then application-level AI start-ups can self-host competitive models and build on top. But the largest open-weight models are still released by a small number of AI labs, which means dependence on these models being released under truly open licenses. Most open-weight models 

 in the sense of open-source, and even maximally open models do not by themselves guarantee competition, as the inputs needed to train and deploy large-scale AI remain concentrated in the same handful of corporations, who have a history of using the rhetoric of openness to entrench, rather than dilute their dominance. While open-weight models do shape the bargaining position, the ecosystem is an integrated whole – the distribution layer can shape model and infrastructure choices of application companies squeezed in between.

How these technical developments and advances will unfold is still uncertain. But crucially, we caution against betting that any of these technical changes will automatically solve the underlying political economy that takes into account the broader ecosystem in which those models live.

From an industrial policy perspective, the question of where value flows is perhaps the most important one, and the one that’s often sidelined. 

The Commission's report supporting the Apply AI strategy

, for instance, counts over 4,100 AI startups in the EU with a combined enterprise value of €161 billion. But it counts startups, not supply chains and tracks where AI companies are headquartered, not where value is created or captured.

The ambition is to enable the next Google, or at least the next OpenAI. But Google's power was never the product at the top of the stack, it was owning almost everything beneath it, from chips to cloud to models to data. And the leading proprietary models are each tied to one of the dominant hyperscalers, for instance, OpenAI with 

. As a result, any start-up that builds on leading, proprietary models incurs not just ongoing costs for tokens, but also for inference and infrastructure. For application-layer companies, those costs consume a substantial share of revenue, which flows back to the US model providers and cloud platforms. The access to distribution might also come with strings attached – for instance, Lovable's Google Cloud deal comes with expanded access to Google's own Gemini models and Anthropic's Claude models, where Google (via its investment arm Capital G) 

. To support their expansion of market share, they are motivated to direct token demand to these foundation models by using their investment partnership to determine model choices. As a result, control of margin sits upstream, not at the application layer, vulnerable to being squeezed at will by their core technology provider.

What this means for EU AI industrial policy

This dynamic changes a few core assumptions that are often implicit in EU AI industrial policy: Simply increasing the number of AI start-ups in Europe will not necessarily lead to more independence from US tech. European AI start-ups at high valuations don’t necessarily mean that Europe is building a sustainable AI sector that is here to stay. A thriving scene of European AI start-ups doesn’t automatically mean an economic upside for Europe. And more private investment, especially money bound by the logic of venture capital, does not automatically produce gains that stay in Europe.

The problem is not that European founders build on the best available models. The problem, for the long-term sustainability of Europe’s economy, is that countless rational and perfectly reasonable individual choices further entrench Europe’s dependency in ways that may be even harder to unwind than the status quo. The seeds of European dependency lie not only in the coercive force of the monopolies of the digital economy. They are also sown as a byproduct of ideas that appear self-evident enough to be left unexamined.

When Europe accepts that European companies should build on market-leading foreign partners for fear of being left behind; when hyperscaling venture capital is elevated as the sole viable financial driver of European tech; and when the spirit of Silicon Valley startup handbooks — built in California, by California, and for California — is carried as the absolute model to emulate, the path is largely set. 

The question is not whether Europe can and should build applications on top of existing models — it is rather to note that applications alone do not solve the question of technological sovereignty that Europe so desperately seeks.

, the Commerce Department’s Bureau of Industry and Security issued an export control directive to Anthropic to suspend access to its Fable 5 and Mythos 5 models for any foreign national, whether located abroad or inside the United States, citing national security concerns. Because the company could not reliably screen users by nationality, it 

In the immediate coverage of these events, so far, most of the attention has been paid to questions about the models, the reported jailbreak that prompted the action, the White House’s deliberations, and the company’s response. What’s needed now is policy analysis. And from that perspective, it’s clear that the novel development here isn’t another intrigue about a Anthropic and its large language models. It’s the legal instrument the White House deployed to force Anthropic’s hand, and the precedent it sets going forward.

Export controls have long reached more than physical goods. Software, source code, and electronic transmissions of non-public technical data all fall within the 

, and the policy debates post-encryption suits of the 1990s settled that intangible 

. So the surprise in this scenario isn’t that software can be export-controlled at all. It’s that these authorities, built around discrete transfers of an identifiable item from one party to another, are being applied to a continuously available frontier model reached by API. When the controlled thing is a service that anyone can call from anywhere at any time, the familiar questions get harder. What exactly is the export? The model weights, the act of inference, the capability itself? The regulations have answers for releasing a file or transmitting a dataset. A model that responds to a prompt in real time fits those categories awkwardly, and that friction is now being worked out through enforcement rather than rulemaking.

That's worth dwelling on, because there's a real difference between AI policy that takes shape through a public regulatory process, with notice and comment and the chance for affected parties to weigh in, and AI policy that takes shape through ad hoc actions over time. With a technology this new and moving this fast, the second path is often where things start, simply because events outrun the rulemaking process. But it has a quiet consequence: each enforcement decision becomes precedent, the accumulated precedents begin to function like a rule, and a framework can end up assembled before anyone has stepped back to define it as one. That's not unique to this case or this administration. It's a recurring feature of governing at the frontier, and it's worth recognizing for what it is.

Much remains unclear about this specific situation. Almost everything in the public record is secondhand, much of it from parties with something riding on how this is understood. The directive arrived by letter, which has not been made public, so the precise scope and rationale are known mostly secondhand, as well. The trigger itself is contested. The government acted after another company reported jailbreaking Mythos in a way that alarmed officials. Anthropic characterizes the episode as a 

and says the behavior at issue is widely available from other deployed models, and at least 

one security researcher who saw the underlying work

 disputes the "jailbreak" framing altogether, describing it instead as defensive research. Further, reporting has since suggested the controls were prompted partly by 

concern that a China-linked group had accessed Mythos

, with the worry being that an adversary could reverse-engineer or distill the model, though that account is single-sourced and Anthropic says the issue was never raised. That factual uncertainty isn't a side issue. It's central, because when the determination is made privately and the facts are this contested, it becomes genuinely hard to tell which theory of control the government is even applying.

And underneath all of it sits the question that actually determines how broad this gets: how does a government decide when a frontier model crosses the line into something that warrants control? There are two basic approaches, and they lead to very different worlds.

The first is an incremental-risk approach, where the question is whether a given model materially expands what an adversary could do beyond what's already available to them elsewhere. If similar capability is already widespread, controlling one more instance of it accomplishes little, so this approach tends to focus attention on genuinely novel capabilities and exceptional cases, and it keeps controls relatively rare.

The second is a capability-based approach, where the mere presence of a sensitive capability is the trigger, regardless of whether comparable capability already exists in the wild. This one is potentially far broader, because nearly every frontier model possesses some capability that could be characterized as sensitive, and once you adopt that framing the universe of models potentially subject to control expands dramatically.

Strip away the mechanics and a real question is whether the government is regulating net-new capabilities or regulating capabilities period, and that's the distinction that will determine how expansive this approach ultimately becomes. The current case sits right on that fault line. Anthropic's defense is essentially an 

: whatever Mythos can do, comparable models can already do, so controlling it denies little.This view seems to inform the perspective of a group of cybersecurity experts who 

 urging the government to reverse its position. A capability-based view would find that beside the point, because the sensitive capability is present regardless of what else exists.

Both have a serious argument behind them. The case for incremental risk is that controlling capability already available everywhere imposes real costs on US firms and their allies while doing little to deny anything to a determined adversary. The case for the capability-based view is harder to dismiss than its critics admit, because you can't un-export a capability once it has proliferated. Waiting for proof that a model is truly net-new may mean acting too late to matter. If the goal is preventing irreversible harm, a government may reasonably decide it can't afford to wait for certainty.

That tension doesn't resolve neatly, and pretending otherwise would oversimplify things. But the choice between the two framings is the actual policy decision, and right now it's being made implicitly, buried inside discrete enforcement actions that read like one-offs. This isn't for lack of institutional machinery. The Center for AI Standards and Innovation already runs pre-deployment evaluations of frontier models under agreements with the major labs, and has examined what those models can do with their safeguards stripped away. The capacity to assess capability exists. What's missing is a published standard linking those assessments to consequences: which findings trigger which restrictions.

Which raises a further problem policymakers haven't fully grappled with: do the agencies making these calls even have the technical capacity to make incremental-risk determinations at the frontier? Judging whether a model materially expands an adversary's options requires deep, current knowledge of what already exists across the global landscape, and that's a demanding analytical task. A capability-based approach is, among other things, far easier to administer, which creates a quiet institutional gravity pulling toward the broader framework whether or not it's the right one.

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe