The Great Decoupling: China vs. USA vs. Europe

Decoupling (1): DeepSeek-V4 No Longer Needs NVIDIA Chips

After months of delay, DeepSeek has released its long-awaited V4 model, competing with Moonshot's Kimi K2.6 and Xiaomi's freshly launched Mimo 2.5. The 1.6 trillion parameter Mixture-of-Experts model was trained on 32 trillion tokens and achieves a context of 1 million tokens using new Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) techniques. DeepSeek is providing both Base and Instruct versions—a rare move that may herald a future “DeepSeek R2.” The model positions itself on par with Gemini 3.1, GPT 5.4, and Opus 4.6. Remarkably, it runs entirely on Huawei Ascend chips instead of NVIDIA hardware. → AINews

Synthszr Take: DeepSeek demonstrates what happens when technology embargoes meet ingenuity: you build your own hardware stack. Running the model on Huawei Ascend instead of NVIDIA GPUs is like switching from imported wheat to homegrown rice—less glamorous, but vital for survival. The simultaneous release of Base and Instruct versions suggests a more open development strategy, while Western providers increasingly close off their models. With 32 trillion training tokens and a 1 million token context, DeepSeek is playing in the same league as the major US models, just on completely different silicon. China is no longer copying architecture; it's building its own AI ecosystem in parallel—including its own chips.

Decoupling (2): China No Longer Wants Silicon Valley Dollars

China plans to cut off tech companies from accepting American investments without government permission. According to Bloomberg, the National Development and Reform Commission (NDRC) has told several private companies in recent weeks to reject US financing in their funding rounds. Among those affected are AI startups Moonshot AI and Stepfun, as well as TikTok parent company ByteDance. The trigger was Meta's $2 billion acquisition of AI startup Manus in late 2025, which led to investigations in Beijing for possible illegal foreign investments and technology exports. Although Manus was registered in Singapore, its founders were from China. Critics accuse the deal of transferring valuable AI technology to a geopolitical rival. The new rules could further cut off China's tech sector from Western venture capital. → Techpresso

Synthszr Take: China is activating a capital immune system against Western investors, reminiscent of biological rejection responses. The NDRC acts like a blood-brain barrier, selectively filtering which foreign bodies (dollars) are allowed into the system. The Manus trauma acts like a vaccine: a single case sensitizes the entire ecosystem. While the US uses entity lists and chip embargoes, China relies on prophylactic capital controls—both sides are building their own closed-loop systems. The irony: Chinese startups are losing access to the world's most liquid capital markets at the very moment they need it most for the global AI race. Beijing is betting that technological sovereignty is more important than Silicon Valley money.

Decoupling (3): Europe Is Building Its Own Infrastructure

The Finnish company Verda (formerly Datacrunch) has raised €100 million in a funding round led by Lifeline Ventures, with participation from byFounders, Tesi, Varma, and other investors. The capital comes from equity and debt financing from Nordic financial institutions. Verda already operates data centers in Finland and Iceland and plans to expand to Sweden, the US, and Asia. CEO Ruben Bryon positions his company as a European alternative to the American hyperscalers AWS, Microsoft, and Google, with the explicit promise not to move the company headquarters to the US. The company is cash-flow positive, according to its own statements, and expects to double its revenue to over $60 million in the first quarter of 2026. The strong demand for computing resources for AI training and inference meant that the €55 million raised in September 2024 was already used up by January. → StrictlyVC

Synthszr Take: Verda isn't primarily selling computing power; it's selling European independence as a premium product. The business model is reminiscent of Swiss private banks: neutrality as a differentiating feature in a market shaped by geopolitical tensions. The irony: to compete with AWS, Azure, and Google, Verda is leasing data center capacity itself instead of building its own—a sort of franchise model for cloud infrastructure. Demand is apparently exploding so much that €55 million was burned through in four months, which points to either extreme inefficiency or genuine market need. Bryon's bet is clear: in a world where data is the new oil, European companies will gladly pay a premium for the guarantee that their data is not subject to the CLOUD Act.

Claude Agents Negotiate Better Than Humans: eBay Stock Drops 5%

Anthropic had 69 employees trade for a week via Claude agents on an internal marketplace. The experiment, “Project Deal,” operated entirely through Slack: after a brief interview, the AI agents independently conducted all negotiations, from posting ads to finalizing prices. Humans only intervened at the end to exchange the physical items—from snowboards to ping-pong balls. The real test ran in parallel: in two of the four marketplaces, participants were randomly assigned either Claude Opus 4.5 (the most powerful model) or Claude Haiku 4.5 (the weakest model). Opus agents closed an average of two more deals and achieved prices that were $3.64 higher for identical products. A used folding bike sold for $65 with Opus, but only $38 with Haiku. → Techpresso

Synthszr Take: Anthropic has accidentally recreated the digital equivalent of Plato's allegory of the cave. People are sitting in front of differently focused shadows of the same reality, believing their version is fair, while measurable differences impact their wallets. The experiment not only shows that stronger AI makes better deals—it reveals the coming two-tier society of algorithms, where model strength becomes an invisible tax. The 11 out of 28 participants who preferred their Haiku experience are reminiscent of studies on cognitive dissonance: people rationalize poor outcomes as personal preference. In a world where 46% of participants would pay for such a service, access to better models becomes a new form of social inequality—invisible, measurable, and accepted by those affected.

Musk vs. Altman in Court — A Founder Dispute Becomes Industrial Policy

Starting Monday, a California court will hear the lawsuit by Elon Musk against Sam Altman and OpenAI. Musk, a co-founder and early funder of OpenAI, accuses Altman and President Greg Brockman of breaking their promise to run OpenAI as a non-profit organization. He is seeking over $100 billion in damages, the removal of Altman and Brockman, and the reversal of the recent conversion to a for-profit company. OpenAI counters that Musk's allegations are “baseless” and motivated by jealousy—he regrets his 2018 departure and now wants to sabotage a competitor. At the time, Musk had tried unsuccessfully to integrate OpenAI into Tesla; in 2023, he founded his own AI company, xAI. Witnesses include Musk and Altman, as well as Microsoft CEO Satya Nadella and several OpenAI insiders. → Morning Brew

Synthszr Take: This trial is a proxy war over who dictates the terms of the AI revolution. Musk portrays himself as a betrayed idealist decrying the commercialization of an originally non-profit mission—all while running his own for-profit AI company, xAI. The irony: both sides invoke the same founding vision of “safe AI” but interpret it in diametrically opposite ways. OpenAI argues that only with billions in investment can it compete with Google and China; Musk sees this as a betrayal of the founding promise. What's being litigated here is less a legal question than one of industrial policy: can AI development even happen outside the venture capital cycle anymore? The outcome could decide whether the next generation of AI is developed by non-profit foundations or publicly traded corporations.

Claude Becomes a Mega-Aggregator

This week, Anthropic announced 15 new app integrations for Claude, including Tripadvisor, Booking.com, Instacart, Spotify, TurboTax, Uber, and Thumbtack. The common denominator: each of these companies built its business on aggregating supply. Thumbtack gives Claude access to over 300,000 service providers; Intuit is connecting TurboTax directly to the AI assistant. These are not simple API partnerships but transfers of customer relationships. The aggregators of the last decade are being aggregated themselves, and they are lining up for it. → Aakash Gupta

Synthszr Take: Claude is becoming a meta-aggregator, turning the winners of the platform economy into interchangeable data sources. The pattern is reminiscent of biological food chains: primary producers (hotels, drivers) are aggregated by secondary consumers (Booking, Uber), which are in turn swallowed by an apex predator. The irony: venture capital has invested billions in building these aggregators, only to watch them voluntarily cede their hard-won market position to AI assistants. What is being sold as a “partnership” is the gradual devaluation of their core competency: the direct customer relationship. Claude doesn't need $300 million for marketing like Booking.com if users are already talking to the AI first anyway.

Grok Voice Think Fast 1.0 — The Voice Agent with Thinking Time

xAI has released Grok Voice Think Fast 1.0, a voice agent that, for the first time, integrates reasoning capabilities into real-time conversations without increasing response latency. The model leads the τ-voice-bench rankings, outperforming Google's Gemini 3.1 Flash Live and OpenAI's GPT Realtime 1.5 in realistic scenarios with background noise, accents, and interruptions. At Starlink, the agent already handles all telephone support with 28 different tools, achieving a 20% conversion rate in sales calls and a 70% autonomous resolution rate for support requests. The unique feature: while the model is speaking, it thinks in parallel about complex queries without users noticing any delay. xAI demonstrates this with a simple test: when asked, “Which months contain the letter X?” Grok correctly answers “none,” while other models confidently name “February.” → Techpresso

Synthszr Take: xAI is solving the fundamental problem of voice agents in a surprisingly elegant way: instead of having to choose between a quick answer and deep thought, Grok does both simultaneously. This is reminiscent of Daniel Kahneman's System 1 and System 2 thinking, except here both systems run in parallel rather than sequentially. The Starlink numbers are the real eye-opener: a single model replaces an entire call center infrastructure and converts every fifth caller into a customer. While OpenAI and Anthropic optimize their models for developers, xAI is building directly for end-user telephony. Musk is once again betting on vertical integration: Starlink as an in-house test lab, Tesla data for training, X for distribution. The irony: the very man who warns of AI risks is now democratizing its most dangerous application—persuasive voices that never get tired.

Token Spending Is Getting Out of Control: Companies Are Just Now Discovering That Inference Is More Expensive Than Training

A survey of 15 tech companies reveals a pattern that should keep CFOs up at night: token spending has increased tenfold in the last six months, with no signs of slowing down. An engineering director at a publicly traded company with 5,000 employees reports multiple budget increases in April alone after teams switched to Claude's new “high-effort” mode. The cost per pull request exploded as developers used the most expensive models even for trivial tasks. At another company with 10,000+ employees, Sonnet was set as the default model to reduce costs, yet developers manually select the more expensive Opus at every start. A Director of Engineering warns: “We are spending hundreds of dollars per day per active developer”—budget discipline is giving way to the fear of being left behind technologically. → The Pragmatic Engineer

Synthszr Take: Token spending is following the same pattern as city highways in the 1960s: the more lanes that were built, the more traffic emerged. It's the Jevon's Paradox, stupid. Companies are currently creating their own induced demand by spending unlimited token budgets as a sign of their innovative spirit. The irony: while everyone is talking about “AI-native” workflows, no one is optimizing for cost—a 10x increase in six months is not an adoption curve, it's a spending explosion without ROI measurement. This is reminiscent of the early cloud migrations when AWS bills suddenly became seven-figure sums and no one knew why. The difference: cloud costs could be optimized through reserved instances and spot usage, but with token spending, there's only one direction. The industry is just learning that inference costs aren't a one-time expense like training costs; they scale like a permanent lease.

SpaceX Admits: AI Data Centers in Space Could Be a Really Bad Idea

SpaceX CEO Elon Musk is touting orbital AI data centers as a “no-brainer.” At the World Economic Forum in Davos, he claimed that the unlimited solar energy beyond Earth's atmosphere would make space the “most cost-effective place for AI” within two to three years. The vision: up to a million satellites, each larger than the ISS, would orbit as data centers. But in the pre-IPO documents obtained by Reuters, SpaceX itself sounds far more skeptical. The company admits that the “initiatives to develop orbital AI computation are in early stages, involve significant technical complexity and unproven technologies, and may not achieve commercial viability.” The sensitive AI chips could wear out faster in the “harsh and unpredictable environment of space.” Moreover, the entire concept hinges on the success of the Starship rocket, which has yet to complete a full test flight without exploding. As SpaceX heads toward a $1.75 trillion valuation in its upcoming IPO, experts warn of the ecological consequences: millions of burning-up satellites could damage the ozone layer. → futurism.com

Synthszr Take: Musk is turning SpaceX into a cosmic server farm while the Mars mission remains on hold. The pattern is familiar from biology: parasites often hijack the nervous systems of their hosts, forcing them into self-destructive behavior. Here, the AI hype is playing the role of the parasite, diverting SpaceX from its original mission. The physics of space are brutal for electronics: cosmic radiation, extreme temperatures, no possibility of maintenance. Every chip engineer knows that semiconductors age faster in space than on Earth. His own engineers seem to know this too, while the CEO dreams of solar energy utopias. SpaceX risks becoming a victim of its own scale: the space pioneer is turning into an oversized cloud provider with a space gimmick.

The AI Absorption Problem in Organizations Remains Stubborn

The claim sounds provocative, but Will Schenk of TheFocus.AI just signed a $300,000 consulting contract that supports this very thesis. In his analysis for the Turing Post series “The Org Age of AI,” he shows: companies are diligently implementing AI technology into their products (AI in the business), but the actual corporate organization remains untouched (AI on the business). A bank can introduce AI-powered fraud detection and still coordinate its quarterly planning using PowerPoint slides passed between assistants. A manufacturer automates its warehousing with computer vision but continues to budget in rigid FTE units. The problem runs deeper: large companies are not unitary actors but internal economies with their own physics. Information becomes currency, departments become tribes defending their territory. Every organization reaches an equilibrium that resists change—because the people inside are optimized for exactly that. → Turing Post

Synthszr Take: Schenk describes a phenomenon reminiscent of the double agents of biology: enzymes that can be both catalysts and inhibitors depending on the environment. AI in companies works the same way—it accelerates product processes while simultaneously petrifying organizational structures because no one wants to touch the power distribution. Andrew Bosworth's “Career Cold Start” algorithm (30 minutes with each team member, three targeted questions) shows the way: the true organization lies in the shadow structures, not the org chart. But making precisely these structures machine-readable would undermine their power. We are not experiencing an AI transformation, but an AI layering on top of old power structures.