Anthropic and OpenAI Copy Palantir, While the Emirates Want to Hand Power Over to AI

The Emirates Want to Hand Over Government Operations to AI

The United Arab Emirates aims to conduct half of its government operations using agentic AI within two years – systems that independently analyze information, make decisions, and act with minimal human oversight. While other nations are still debating AI ethics councils, the UAE is prioritizing speed: each ministry is evaluated based on its adoption rate, and a task force led by Mohammad Al Gergawi is driving implementation. In practical terms, this means building permits in minutes instead of weeks and automated government services that adapt to demand in real-time. The leadership is clear: AI is seen not as a tool, but as an operational partner. This is a trial run for state transformation that will either set new standards or spectacularly fail due to the complexity of government processes. → Techpresso

Synthszr Take: The UAE is treating artificial intelligence like Singapore once treated its port logistics: as a systemic upgrade, not an add-on. The country understands that in the AI era, speed is more important than perfection – a state that takes two years to form an ethics committee while others are already iterating will lose its connection to global data flows. The Gulf states are using their autocratic structure as a speed advantage: no federal blockades, no GDPR, no elected local politicians raising concerns. The model is reminiscent of Estonia's digital transformation, but with an oil state's budget and without democratic friction. Whether agentic AI works in administration depends less on the technology and more on how well the UAE can retrain its civil servants to become process designers. The real bet is that AI-driven efficiency will replace the legitimacy through wealth that was previously based on oil.

Palantir's Sales Pitch Becomes a Model for Anthropic

Anthropic is launching a $1.5 billion joint venture with Blackstone, Hellman & Friedman, Goldman Sachs, and General Atlantic to integrate Claude into the private equity giants' portfolios. The structure follows OpenAI's DeployCo model from last month but is more focused: Anthropic itself is contributing $300 million, as are Blackstone and Hellman & Friedman. Goldman Sachs is investing $150 million. The vehicle functions as a mix of a consulting arm and a deployment factory, compressing the enterprise sales cycle from years to months. OpenAI's DeployCo was larger with a $4 billion commitment, but it also guaranteed PE partners an annual return of 17.5 percent. Anthropic, instead, is focusing on prestige: fewer partners, more credibility. The timing is no coincidence – Anthropic is currently negotiating a $50 billion round at a valuation of $850–$900 billion, with a planned IPO in October 2026. → Techpresso

Synthszr Take: Private equity is turning AI models into what McDonald's did with burgers: a franchise system. The buyout firms own thousands of companies in healthcare, logistics, and manufacturing – each a potential Claude customer, but with no time for individual sales pitches. The joint venture isn't a technology partnership; it's pure sales infrastructure. Anthropic provides the product, and Wall Street provides access to captive customers. The real innovation is speed – what normally takes years happens in months because PE partners can 'encourage' their portfolio companies to use Claude (he who pays the piper calls the tune). Goldman Sachs has already had Anthropic engineers working internally on autonomous compliance agents for six months – that's the prototype now being industrialized. Anthropic projecting $30 billion in annual revenue by March 2026 shows that AI models are becoming a commodity, and sales channels are the new moat.

OpenAI Also Follows the Palantir Pattern: Forward Deployed Capital

OpenAI has finalized the most structurally unusual enterprise AI deal of 2026: a $10 billion vehicle with 19 investors and a guaranteed annual return of 17.5% over five years. The strategy turns private equity portfolios into a captive sales channel. The company confirmed The Deployment Company, a Delaware-based joint venture designed to bring OpenAI's enterprise products into the operations of some of the world's largest buyout firms. OpenAI's own commitment is up to $1.5 billion: $500 million in equity at closing, with an option to add another $1 billion later. The PE consortium is contributing about $4 billion over the same five-year period. The entity will be controlled by OpenAI through super-voting shares, allowing it to retain strategic control while financial sponsors take on the economics of a yield-oriented investment. → Techpresso

Synthszr Take: OpenAI is copying Anthropic's structured sales maneuver but putting the pedal to the metal: seven times larger and with a bolder omission – the usual system integrators (Accenture, Deloitte, Capgemini) are completely absent. They are normally the standard bridge for any major enterprise tech rollout. Here, they are being bypassed. OpenAI is building its own consulting vehicle directly with PE firms because system integrators sell hours, but private equity has shareholder pressure. When a Blackstone operating partner recommends deploying GPT-Enterprise to a portfolio CEO, it's not a recommendation. The 17.5% return guarantee is the price of admission for this access and, at the same time, a statement: model quality and compute scale are no longer the moat. Sales infrastructure is. Sam Altman is learning from Palantir, only faster: forward-deployed engineers were the move in 2010 — in 2026, it's forward-deployed capital.

Nvidia Loses China Completely, Huawei Takes Over

Nvidia CEO Jensen Huang confirms what US export controls have wrought: Nvidia's market share for AI accelerators in China has dropped to 0%. Just two years ago, Nvidia dominated the Chinese market with 66%. The H200 licenses, which were supposed to re-enable trade, have been stuck in a bureaucratic no-man's-land between US approvals and Chinese import restrictions. Commerce Secretary Howard Lutnick confirmed in April that no H200 chips were sold to Chinese companies. Huawei is exploiting the vacuum: the company expects to increase its AI chip revenue from $7.5 billion to $12 billion, with 750,000 units of its Ascend 950PR processor planned. China's AI chip market could grow to $196.2 billion by 2029. → Marcus Schuler

Synthszr Take: The zero in Nvidia's China business is a consequence of US export policy. Washington has provided a textbook example of how to lose a market while trying to control it: the H200 chips are stuck in a kind of Schrödinger's state, simultaneously approved and banned, while Huawei creates facts on the ground with the Ascend 950PR. This is reminiscent of Prohibition in the 1920s, where the attempt to curb alcohol consumption only strengthened the black markets. The Chinese AI chip market is reorganizing like an ecosystem after a forest fire: the dominant species (Nvidia) has disappeared, and local species (Huawei, SMIC) are spreading. Jensen Huang warns about losing control of the software stack, but the real problem lies deeper: the US has traded its most important currency in the AI age, technological interdependence, for an embargo that only one side respects.

Microsoft Copies the Old IBM Marketing Tactic: Fear, Uncertainty and Doubt

Microsoft is moving Agent 365 from preview to general availability – a management platform for AI agents costing $15 per user per month. The product is positioned as a central control plane through which IT and security teams can monitor, manage, and secure AI agents: within the Microsoft ecosystem, in third-party clouds like AWS Bedrock and Google Cloud, on employee endpoints, and increasingly across a sprawling ecosystem of SaaS agents. Microsoft refers to the phenomenon of local AI agents installed without IT knowledge as 'Shadow AI' – a new category of enterprise risk. David Weston, Corporate Vice President of AI Security at Microsoft, describes three specific security incidents that are already occurring: developers connecting agents to backend systems and inadvertently exposing sensitive infrastructure; attackers using cross-prompt injection via manipulated data sources; and agents accessing data sources not designed for agentic access patterns. Agent 365 acts as a central registry and policy engine for all agents in the company, regardless of whether they are developed with Copilot Studio, deployed on AWS Bedrock, or run as a local installation on Windows machines. → VentureBeat

Synthszr Take: Microsoft is playing a familiar game – only this time in the AI age. What's being sold here as a security platform structurally follows the classic FUD logic: first, a diffuse, hard-to-control risk space is established ('Shadow AI'), then the only credible solution is provided along with it. Microsoft is positioning Agent 365 not primarily as a product, but as a response to a threat that it helps define linguistically. The three attack scenarios described by David Weston are technically plausible – but above all, communicatively effective: they shift the perception of AI agents away from productivity and toward potentially uncontrollable security risks. This is precisely where the FUD pattern kicks in: uncertainty is not resolved but structurally reinforced. Anyone operating agents today without central control is implicitly moving in a gray area between data leaks, attacks, and governance failures. The narrative is strongly reminiscent of IBM's historical strategy: externalize complexity, emphasize risk, and establish your own platform as the necessary control instance. The only difference is that this time, the opponent is not a specific competitor but an entire paradigm – the decentralized, uncontrolled use of AI. Agent 365 thus becomes the institutional answer to a problem deliberately described as systemically unmanageable. The real bet is not technological but psychological: companies aren't paying for better agents, but for the reduction of fear, uncertainty, and loss of control. Or more precisely: Microsoft is not monetizing the solution – but the sense of threat that makes this solution seem indispensable.

Google Relaunches Its Gemini App

Google is currently rolling out a complete redesign of the Gemini app that goes far beyond a cosmetic update. The central element: a pulsating, colorful background animation with gradient effects that uses Google's 'Liquid Glass' design language on iOS. The new interface places the input box as a pill in the center, while the model selector moves back to the top left corner. The real innovation lies in the unified access to tools: Images, Videos, Music, Canvas, Deep Research, and Guided Learning appear as a descriptive list below the input. A plus button opens a bottom sheet with a carousel for photos, camera, and recently used images. The 'See thinking steps' option moves to the overflow menu and displays the thought processes as a bottom sheet. → 9to5google

Synthszr Take: Google is redesigning Gemini to be a membrane through which users interact with artificial intelligence. The pulsating background animation isn't a design gimmick but a signal: the interface itself is becoming a living organism that reacts to input. While OpenAI positions ChatGPT as a tool (text in, text out), Google is transforming Gemini into an environment. The unified tool access suggests deeper integration: instead of switching between different AI models, a meta-layer orchestrates various capabilities. This is reminiscent of biological systems where specialized cells work together to perform complex functions. Google is betting that the future of artificial intelligence lies not in ever-larger models, but in the elegant choreography of specialized components.

Unitree Opens First Store in Shanghai

Chinese robotics manufacturer Unitree is opening Asia's first 'Embodied Intelligence Experience Store' in Shanghai at the end of May. The 100-square-meter store in the Jing'an district showcases the company's entire product range and, according to Unitree Vice President Li Binjie, creates 'immersive consumption scenarios through embodied intelligence technology.' The opening is part of the Chinese Ministry of Commerce's 'First in Shanghai' initiative, which aims to develop new consumption drivers. Shanghai is expanding the spectrum of product premieres from beauty and fashion to sports, pet products, and cultural IP assets. In parallel, retail sales in China show an acceleration, with 2.8 percent growth in the first two months of 2026 compared to the previous year's month of December 2025. → Hello China Tech

Synthszr Take: Unitree is turning robots into lifestyle products, just as Apple once brought computers from the office into the living room. The 'Experience Store' follows the Tesla formula: don't sell technology as a B2B solution, but stage it as a consumer experience. China is testing a hypothesis here that Western markets haven't yet dared to try: if robots are to become the next smartphone category, they need to be tangible in shopping malls, not just on trade show floors. The state-run 'First in Shanghai' initiative shows how China intertwines industrial policy and retail. While Europe debates AI regulation, Shanghai is creating facts in physical space.

Banks Are Drowning in Loan Volume for Data Centers

The expansion of AI infrastructure in the US is increasingly straining the banking system. According to the Financial Times, major banks like JPMorgan Chase, Morgan Stanley, and SMBC are looking for ways to transfer risks from financing AI data centers to other investors. The loan volumes for new data centers have grown so large that individual institutions are hitting their internal risk concentration limits. One deal illustrates the scale: a $38 billion loan package is financing data centers in Texas and Wisconsin linked to Oracle. JPMorgan and MUFG have been trying for months to distribute parts of these loans more broadly in the market. Some banks have even tried to sell the loans at a discount to non-bank buyers. → Techpresso

Synthszr Take: Banks are currently experiencing what happens when infrastructure grows faster than the ability to finance it. We know this pattern from the 19th-century railway bubble: massive capital concentration in projects whose profitability only becomes apparent years later. Only this time, it's not about railway lines but about server farms with the electricity consumption of small cities. The $38 billion financing for Oracle shows that even major banks are reaching their limits (Matthew Moniot talks about 'choking' on the sums). The attempt to pass on risks to insurers and credit funds via 'significant risk transfers' is reminiscent of the securitization wave before 2008. Maine's Governor Mills has recognized the explosive nature of the situation: her veto against the data center moratorium shows the classic conflict between short-term jobs and long-term systemic risks. The banks are betting that AI data centers will become the new essential infrastructure – but if the bubble bursts, others will foot the bill.

The Recursive Trap: When AI Survives Its Own Optimization

Jack Clark of Import AI predicts with a 60% probability that by the end of 2028, AI systems will be able to develop their own successors without human intervention. His thesis is based on two key developments: SWE-Bench scores exploded from 2 percent (Claude 2, late 2023) to 93.9 percent (Claude Mythos Preview), marking the saturation of real-world software engineering tasks. In parallel, METR's time horizon plot shows a multiplication of task complexity that AI systems can reliably handle 50 percent of the time: from 30 seconds (GPT 3.5, 2022) to 40 minutes (o1, 2024) to 12 hours (Opus 4.6, 2026). Clark already sees initial proofs-of-concept for models training their successors within the next two years – initially with smaller models, as frontier models are still too expensive and complex. Most developers in Silicon Valley and major AI labs are already programming entirely with the help of AI systems, including tests and code reviews. → Import AI

Synthszr Take: Clark is describing a kind of evolutionary self-domestication, where AI systems become their own breeders. This is reminiscent of Stanislaw Lem's 'Summa Technologiae,' in which intelligence understands and optimizes itself as a material. The 12-hour mark on METR is not a random milestone: it corresponds roughly to a workday plus overtime – the time window in which humans solve complex creative problems. When AI crosses this threshold, it becomes the architect of its own cognitive architecture. Clark speaks of 'crossing the Rubicon,' but a better comparison might be the Cambrian explosion: as soon as evolution discovers itself, it accelerates exponentially. The irony is that we are building systems to help us build better systems, until they no longer need us.

The Proof: Language Models Must Dream

Researchers have mathematically proven what many in the AI industry fear: hallucinations in large language models are unavoidable. The paper formally defines hallucination as an inconsistency between a computable LLM and a computable ground-truth function. By applying learning theory, the authors show that LLMs can never learn all computable functions. Since the formal world is only a part of the more complex real world, hallucinations in real-world applications are all the more inevitable. The study identifies hallucination-prone areas, particularly in tasks with provable time complexity. The authors also discuss existing mitigation strategies and their limitations for the safe deployment of LLMs. → arxiv.org

Synthszr Take: The study hits a nerve: AI systems are like urban planners who must reconstruct a metropolis from satellite images but can never see all the backyards. The mathematical proof is reminiscent of Gödel's incompleteness theorems – some truths remain systemically unattainable. Silicon Valley is currently selling billion-dollar solutions for a problem that, according to this research, is fundamentally unsolvable. This isn't an implementation error; it's an epistemological boundary. Companies like Meta are pumping billions into 'hallucination reduction,' while mathematics says the best you can achieve is clever concealment. Perhaps we should stop trying to build perfect oracles and start working with creative liars.