Special Analysis
Two things happened this week that appear to point in opposite directions but in fact lead to the same conclusion. On capability, OpenAI announced that a model not yet released had produced ten research advances in mathematics and theoretical computer science, each result accompanied by a formal, machine-checkable certificate, at approximately 2,000 dollars in tokens used to generate the solutions, calculated at API-equivalent prices. On constraints, the EU AI Act's transparency obligations began to apply across the Union, Palantir posted a quarterly profit of 1.06 billion dollars while its chief executive argued on the earnings call that model providers are not the right custodians of enterprise data, and the AI talent market showed that even one of the world's best-funded AI startups cannot hold on to its founding team.
The common thread is more useful than any individual story. Frontier capability is advancing quickly and deploying it may be getting cheaper, but access, control, compliance, specialised talent and the ability to retain value all remain constrained. Model capability is rising fast. The right to deploy those models in line with EU rules, the power to keep the value they create inside the enterprise, and the ability to hire the people who built them are not rising at the same pace.
Ten results, and a certificate for each
On 1 August OpenAI announced that an internal version of Astra, its next major model, had produced ten new results in group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography and extremal combinatorics. According to OpenAI, the results "solve or make significant progress on" problems that had been open in each of these areas. The headline result is a construction establishing the existence of non-sofic groups, addressing a question in group theory relating to Gromov's notion of soficity.
How the work was carried out matters. Researchers used the model to prepare arguments in manuscript form, after which the model formalised each argument in Lean 4, a proof assistant whose output a program can verify line by line. OpenAI published a 249-page manuscript and the accompanying Lean certificates in a GitHub repository. The company estimates that the tokens used to generate the solutions would cost approximately 2,000 dollars at Sol API prices, which is an API-equivalent price rather than the actual cost of compute, and does not include the human work of preparing the manuscript. University of Manchester mathematician Thomas Bloom, who publicly challenged an inaccurate OpenAI claim from October 2025 about a catalogue of Erdős problems, called the results "big news" on X and judged them more significant than May's counterexample to the Erdős unit distance problem.
What the certificates make possible is real, but what they do not make possible matters just as much. Lean can verify that each formal proof follows from the encoded assumptions. It cannot independently establish whether the formalisation faithfully represents the original mathematical question, whether the assumptions are appropriate, whether the result is genuinely novel, or whether its mathematical interpretation is important. Those judgements still have to be made by experts. The burden of verification, however, shifts. The strongest earlier objection to AI-generated proofs, that checking them consumes scarce expert capacity, is now considerably eased.
The practical significance lies precisely in the cost figure at API prices. Approximately 2,000 dollars in tokens for research-level work changes the realistic ceiling of what a small technical team can attempt on a well-defined problem. A regional AI team, whether in Belgrade, Ljubljana or Zagreb, can now apply state-of-the-art reasoning to a genuine research question at a marginal cost that no longer looks prohibitive. Astra has not yet been released, and expertise, workflow and verification review are not included in the quoted token price. Even so, for applied work of this kind the increasingly scarce resource is problem definition and the expertise needed to interpret it, not computing power.
Article 50 transparency obligations under the EU AI Act
A significant EU AI Act compliance deadline fell on 2 August 2026, and the obligations of providers and deployers differ more than most enterprise-facing analysis acknowledges.
Article 50 covers four scenarios, divided by role. Providers of AI systems that interact directly with natural persons must as a rule ensure those persons are informed that they are interacting with an AI system, except where this is obvious. Providers of generative AI systems must ensure that synthetic image, audio, video and text content is marked in a machine-readable format and detectable as artificially generated or manipulated. Public disclosure obligations for deployers are narrower and apply primarily to emotion recognition and biometric categorisation, deepfake content, and AI-generated or altered text published to inform the public on matters of public interest.
The exceptions matter as much as the obligations. For text of public interest, disclosure is generally not required where the content has undergone human review or editorial control and where a natural person or organisation holds editorial responsibility. Where deepfake content forms part of an evidently artistic, creative, satirical or fictional work, the manner of disclosure may be adapted so as not to impair the work. A limited transitional period for the machine-readable marking obligation under Article 50(2) applies to certain generative systems already on the market before 2 August 2026: providers have until 2 December 2026 to comply. Non-compliance can lead to fines of up to 15 million euros or 3 percent of total worldwide annual turnover, whichever is higher, for entities that are not SMEs; for SMEs, Article 99 applies the lower of those two ceilings.
Enforcement is distributed rather than concentrated. National market surveillance authorities enforce the rules at member state level. The AI Office supervises certain systems. The European Data Protection Supervisor is competent for EU institutions. The territorial scope covers systems placed on the EU market, systems used in the EU, and situations where the output a system produces is used within the EU. Mere accessibility over the internet from the EU is not sufficient. Article 50 applies at the level of the AI system and differs from the separate regime for general-purpose AI models under Chapter V, which has its own obligations and enforcement framework.
For Slovenian companies, and for Serbian companies whose systems or outputs fall within the Act's territorial scope in the EU, the immediate task is to establish whether they act as a provider, a deployer, or both. Customer-facing chatbots, voice agents and generative features require a documented assessment of the applicable disclosure and marking obligations, including whether the transitional period for certain existing systems applies. For anyone touching EU users or content intended for the EU market, some part of the Act applies; the question is in which role, under which of the four scenarios, and what evidence exists in the documentation.
Palantir's 1.06 billion dollar quarter and the sovereignty question
On 3 August Palantir reported total second-quarter 2026 revenue of 1.94 billion dollars, up 93 percent year on year. US commercial revenue reached 764 million dollars, up 149 percent. GAAP net income was approximately 1.06 billion dollars. Adjusted earnings per share came in at 0.41 dollars, above the Bloomberg consensus estimate of 0.34 dollars. Full-year 2026 revenue guidance was raised to approximately 8.15 to 8.16 billion dollars. The shares rose around 12 percent in after-hours trading following the beat and the raised guidance.
The quarter matters in its own right, but how chief executive Alex Karp interpreted it is more interesting still. In a letter to shareholders and an interview with CNBC, Karp sharpened a critique of the leading AI laboratories that he has been developing for several quarters. He wrote that the business has "Marxist overtones and subtext" and argued that model providers "intend, knowingly or otherwise, to seize the means of production of their supposed partners". He framed customer control over data, models and institutional knowledge as "AI sovereignty", arguing that enterprises should not hand their operational data and intellectual property to model providers who might one day become their competitors.
Two things need separating. The financial results show rapidly growing demand for Palantir's products, particularly in the US commercial and government markets. They are consistent with rising demand for controlled AI deployment, but they do not prove that sovereignty or governance is the cause of that growth. Operating performance, government contracts, broader growth in enterprise AI capital spending and the acceleration of the US commercial business all contribute as well. Karp presents sovereignty as Palantir's strategic explanation for demand, but the quarter itself cannot confirm that interpretation. Moreover, "sovereignty" in Karp's sense is not the same thing as local, on-premises deployment. The underlying architectural pattern, that the customer retains control over data, models and institutional knowledge, can also be achieved in controlled cloud or hybrid environments.
For a regional enterprise weighing whether to send operational data to a leading American AI laboratory, the quarterly results and the chief executive's messaging join a growing set of signals that demand for controlled deployment is real and rising, even though there is as yet no conclusive answer as to why. The question on the table at a Slovenian bank, a Serbian state-owned enterprise or a regional insurer is not whether Karp is right about the motives of model providers. The question is whether the organisation's operational data and internal processes should sit in a controlled environment where the enterprise retains control, or in a model hosted by a large cloud infrastructure provider, where it does not. Palantir's quarter does not answer that, but it makes the question harder to keep postponing.
The AI talent market becomes a retention problem
On 27 July Lilian Weng announced she was leaving Thinking Machines Lab, saying she could no longer sustain the demands of startup work and that "constant stress and workload had pushed her past what her health can physically take". On 29 July an OpenAI spokesperson told TechCrunch that Weng would return to OpenAI, where she was previously VP of AI safety research, to lead a top-level team focused on internal research, including work on recursive self-improvement.
Weng is the fourth of Thinking Machines' six original co-founders to leave, after Barret Zoph, Luke Metz and Andrew Tulloch. Of the original six, according to available reports, only Mira Murati and John Schulman remain. Thinking Machines was valued at approximately 12 billion dollars following an initial two billion dollar funding round earlier this year, one of the largest early-stage rounds in the industry's history. The substance of the story is that a company of that size and with that financial backing lost two thirds of its founding team in under a year.
An Axios analysis of 3 August uses Weng's move as the occasion for a broader claim: the AI talent market is now primarily a retention problem. Top researchers move quickly between a small number of exceptionally well-funded organisations, among them OpenAI, Anthropic, Google DeepMind, Meta, Thinking Machines and xAI. Their decisions appear to be shaped by pre-IPO equity, access to compute, research autonomy, working conditions and organisational mission. A source cited by Axios said that Anthropic chief executive Dario Amodei had expressed concern that new hires may be arriving for the money rather than the mission; this is an indirect report rather than a direct statement. Weng's own stated reasons matter more than the tempting simplification: she left because the pace and volume of work were damaging her health, not because OpenAI outbid Thinking Machines for her.
For regional enterprises and national programmes, the implication is not defeat but a clearer picture of where they can compete. The concentration is real, but the market has not narrowed to four addresses: Thinking Machines, xAI and others remain genuine players. Regional teams can compete for individual researchers, missions and applied domains, although most will struggle to match the compensation and compute of the leading laboratories in general-purpose model research. The competitive space for a team in Belgrade or Ljubljana lies in fit of application, depth of integration, sector-specific knowledge and compliance-aligned engineering, in areas where proximity to the customer and familiarity with European rules are a genuine advantage.
Take away
Seen separately, these four stories point in different directions. Seen together, they give a clear picture. Frontier capability is advancing quickly and the expected cost of deploying it is falling; everything else is becoming more constrained or more concentrated. If the current trend continues, capability itself will not be the scarce resource. What will be scarce is compliance capacity, depth of integration, retained intellectual property, and the specific human expertise that turns a leading model into a working enterprise system.
For enterprises outside the European Union, the message is not that they should build a frontier model or try to poach researchers from Anthropic. It is that they should be ready when advanced capability becomes cheap enough for broad deployment, and should sit within the compliance framework that EU customers now require by law. Over the next twelve months, the winners will be the companies that control the application layer, the data layer and the compliance-aligned engineering in their sector. Everything else will increasingly be available to rent.