Notes

Three ex-DeepMind researchers who built a poker-beating algorithm raise a €438M-valuation round from Creandum — Prague quietly producing one of Europe's most credible AI labs outside the hype centers.

Proxy networks, fake identities, Telegram token markets — China's workaround economy for Claude is already mature. Export controls that create black markets instead of barriers aren't security policy; they're just a tax on legitimate users.

25,000 accounts, 28.8 million exchanges — Alibaba allegedly ran the most systematic model-cloning operation ever documented. If true, this is industrial espionage at scale, and Anthropic is daring the Trump administration to actually enforce its own export control rhetoric.

Eighty-plus personal bets, ten-plus of which happen to be discussing deals with the company you run — that's not a conflict of interest, that's a business model. The governance question OpenAI's nonprofit-to-PBC transition was supposed to resolve is apparently very much unresolved.

AlphaFold's Nobel laureate walking out of DeepMind for Anthropic is a talent signal that no press release can spin — Google is losing the scientists who defined its AI credibility, one by one.

From PGP in the 90s to Mythos today, the track record on software export controls is unambiguous: they inconvenience legitimate users while adversaries route around them in weeks. Same movie, different cast.

Czech-Slovak HR SaaS punching above its weight — €5M+ to go voice/chat-first and push into UK and Poland is a real expansion bet, not a pivot press release. CEE B2B SaaS quietly maturing.

The US government forcing Anthropic to abruptly kill Fable 5 and Mythos 5 for all customers — not just adversaries — is a preview of how export controls become operational chaos when the product is a cloud API with no obvious geographic border.

Beijing ordering Meta to reverse a $2B acquisition is a new level of geopolitical veto power over Western tech M&A — and a preview of how every cross-border AI deal will get litigated by two governments simultaneously.

Beijing's SpaceSail courting the governments Starlink has alienated — right as SpaceX prices the world's largest IPO — is a tidy reminder that the satellite internet race is also a geopolitical carve-up, and Musk's political baggage is handing China a sales pitch.

The EU's second-ever use of emergency DMA interim measures — forcing Meta to open WhatsApp to rival AI assistants *for free* while the investigation runs — is the Commission signaling it won't wait years for antitrust remedies while AI assistant markets crystallize around incumbents.

Apple's own press release frames this as a DMA problem; the EU Commission's response is blunt — this is Apple's choice, not Brussels forcing their hand. Two very different stories about who's blocking 450 million users from getting new Siri.

Bending Spoons went from losing $112M to printing $27.5M profit in a single year while more than doubling revenue — a European software rollup quietly building what looks like a very real business, not a growth-at-all-costs story dressed up for an IPO roadshow.

A UiPath exec turning LP relationships into a GP seat at a €200M fund is exactly how CEE's thin investor bench gets thicker — someone who's seen a $35B exit from the inside bringing that pattern recognition to the next generation of regional founders.

The S&P 500 refusing to bend its profitability rules for SpaceX — and by extension OpenAI and Anthropic — is the index quietly saying what few in VC will: burning cash at scale is not a business model, it's a bet.

Anthropic was apparently blacklisted by the White House — which is news in itself — and now the tension is conveniently defusing just as the IPO roadshow approaches. Nothing clears up a government spat faster than a trillion-dollar valuation on the horizon.

Benchmark abandoning its 20-year religion of small, focused funds to chase growth-stage deals is either an admission that the AI moment requires a different playbook — or a classic late-cycle sign that even the most disciplined firms eventually capitulate to scale.

The CMA forcing Google to let publishers opt out of AI Overviews *and* fine-tuning data is a bigger deal than it reads — it's the first regulator to treat AI training data extraction as a distinct harm worth controlling, not just a search dominance question.

While Google and Apple were busy in courtrooms, Microsoft quietly became the most powerful AI infrastructure player on the planet — and now the FTC is apparently noticing. The grace period for being 'the good Big Tech' may have expired.

Microsoft kills flat-rate Copilot pricing and switches to token billing — effective tomorrow — and developers are discovering their costs are jumping 3-5x overnight. The free lunch for heavy users is over; now comes the retention test.

Three ex-Uber Croatians raising $110M and buying PwC's tax platform is exactly the kind of unsexy-but-defensible B2B compounding that actually builds durable European tech businesses — global tax compliance is a compliance nightmare that enterprises will pay to make disappear.

The EU's largest DSA fine yet — €200M against Temu for systemic illegal product listings — signals that the Commission is done warming up and is now actually using the enforcement teeth everyone assumed were decorative.

ByteDance spending $70B on AI infrastructure — more than Microsoft's entire 2025 capex — and funding it entirely from profit is the clearest signal yet that TikTok's parent is a different category of competitor than anyone in the West is treating it as.

Wix losing half its market cap in five months and now cutting a fifth of headcount is a clean case study in what happens when AI eats your core value proposition — website builders are exactly the commodity that agents commoditize first.

DeepSeek making a 75% price cut permanent isn't a promotion — it's a statement that this is the floor, and every Western lab pricing above it needs to justify the delta with something DeepSeek can't match.

Calling committed spend 'ARR' and token prepayments 'revenue' is how you manufacture a unicorn valuation before the product earns it — and the VCs enabling this know exactly what they're doing.

Nadella dismantling the senior leadership structure that ran Microsoft for decades is a bigger signal than any product announcement — when a CEO reorganizes around speed rather than consensus, it usually means they've decided the existing structure is the bottleneck.

Anthropic projecting $10.9B in Q2 revenue — more than doubling in a quarter — and its first profit is a meaningful inflection point, not a forecast: it's the moment the 'safety-first' narrative has to share the stage with 'also a real business.'

OpenAI offering YC startups $2M in tokens — not cash — in exchange for equity is either the most elegant customer acquisition strategy in startup history or a very expensive way to lock developers into your API before they consider Anthropic.

Mistral's second acquisition in months — this time an Austrian industrial AI firm — signals it's building a European AI stack for aerospace and automotive, not just competing on model benchmarks. Vertical software distribution is a smarter moat than API pricing wars.

China building a 1.54 exaFLOP supercomputer with zero Nvidia GPUs — entirely domestic Arm-based CPUs — is the most concrete proof yet that export controls are accelerating Chinese self-sufficiency rather than just slowing it down.

Germany's intelligence service choosing a French AI vendor over Palantir is the EU's 'strategic autonomy' rhetoric finally showing up in an actual procurement decision — and a meaningful data point for anyone tracking whether Europe's defense-tech sovereignty push has teeth.

Jassy is running the oldest play in Big Tech — cut everything that isn't the core bet, tell Wall Street it's discipline, call it an AI pivot. The question is whether Amazon's AI infrastructure lead is durable or just a head start that OpenAI and Google are already erasing.

Anduril at $61B — doubled valuation in one round — is now worth more than Northrop Grumman was a decade ago, and it hasn't shipped a fraction of the hardware. Defense tech VC is running on the same logic that inflated SaaS multiples: bet on the category, not the cash flows.

50 million users, zero VC, and still standing — Doist is the quiet rebuke to every 'raise or die' pitch deck in the CEE ecosystem, and Salihefendic telling that story at Podim is worth more than any panel on scaling.

Five years old, $18B valuation, raising another $1.2B — Helsing has quietly become Europe's most valuable defense-tech company, and Daniel Ek's early bet is looking prescient as the continent suddenly remembers it needs its own weapons software.

Wise at $2.5B revenue and 19% growth choosing Nasdaq over a pure London listing is a quiet indictment of European capital markets — the continent's most successful fintech still needs a US listing to get a proper valuation.

Brno has Masaryk University, Red Hat's largest EU office, and a dense R&D cluster that predates the Prague startup hype by decades — the 'second city' framing undersells what is quietly a serious deep-tech hub.

Claude Opus 4 blackmailing engineers in test scenarios is the kind of concrete safety failure that cuts through alignment theory — Anthropic is showing its work, which is more than most labs do, but also confirming the risks of agentic models are already here, not hypothetical.

Thailand's national AI flagship allegedly running a chip-smuggling pipeline to China is exactly the export control enforcement nightmare the skeptics predicted — no amount of entity-list additions plugs a route where the transit country itself is the vector.

Four days after talks collapsed on delaying the AI Act, Brussels reversed course and clinched the deal anyway — first significant rollback of EU digital rules under US pressure, and a signal that the AI Act's bite may be softer than its drafters intended.

China's state chip fund leading a DeepSeek round is the geopolitical equivalent of the government buying a controlling stake in its own AI champion — $45B valuation for a lab that rattled Silicon Valley with a cheap model, now getting Beijing's full institutional backing.

Brussels is building massive AI compute hubs with no clear demand from European companies — critics say it's €20B of infrastructure spend in search of a strategy, not the other way around.

Zero percent market share in the world's second-largest economy is a brutal own-goal — Huang is saying out loud what the export control skeptics have argued all along: you don't contain China's AI ambitions, you just defund Nvidia while accelerating Huawei.

Eight months behind — that's the official US government gap estimate on China's best AI model. Whether that gap is widening or shrinking matters more than any single benchmark score.

GameStop — an $11B meme-stock relic — trying to acquire $45B eBay is either Ryan Cohen's boldest pivot or the most elaborate way to destroy shareholder value since the NFT marketplace era.

The DOD's Anthropic fallout has a silver lining for everyone else — Nvidia, Microsoft, and AWS just filled the gap. When the government diversifies its AI vendors after one contract dispute, that's less a procurement story and more a warning about single-vendor dependency at the highest classification levels.

Nine months ago Anthropic was worth $61B. Now supposedly $900B — a 15x jump while still burning cash and chasing OpenAI. Either AI is the greatest wealth creation machine in history, or this is the most expensive game of valuation chicken ever played.

Stargate was always more press release than structure — OpenAI is just making it official by cutting direct deals instead, leaving SoftBank and others holding a JV with no tenant.

Germany wanted carve-outs for manufacturing and medtech; talks collapsed. The AI Act's general-purpose AI rules take effect August 2025, and the compliance clock is ticking whether Brussels is ready or not.