Venture Geopolitics Issue 56
14 July 2026
Governments and companies spent the week trying to contain AI capability. Apple is suing OpenAI, alleging that former employees carried trade secrets into its new hardware business (and complaining that the company poached more than 400 Apple employees). Anthropic accused Alibaba’s Qwen team of mounting a distillation attack using ~25,000 fake accounts. Washington already restricted access to Anthropic’s most sensitive models; Beijing announced it too is considering similar limits on its own.
Each story begins with the same instinct: identify valuable knowledge, draw a legal boundary around it and assume the boundary will hold. But AI is built from mobile people and reproducible software, and the incentives to move both are enormous.
Washington’s restrictions prevented people from accessing Anthropic’s models for a while, but they also redirected demand for those models elsewhere. Chinese open-source models quickly looked not only cheaper but less risky. Chinese models moved from 11% of US OpenRouter usage last year to more than 30% every week since February, peaking at 46%. The effort to preserve an American advantage strengthened the market for its substitutes. Now Beijing is preparing to repeat the exercise itself.
Putting a hand over a fire hydrant changes the direction of the spray. It does not stop the flow. It may even send the water back into your face.
This does not mean legislation is powerless. China’s helium export ban, announced this week, will probably achieve its narrow purpose for two reasons. First, helium is physical: shipments can be identified and stopped at the border. Second, the ambition is modest - this is conservation, not leverage. China imports ~85% of the helium it consumes (and produces <2% of global supply). Beijing is not using a domestic monopoly to pressure others; it is preventing a scarce imported resource from leaking back out*.
Model controls fail on both counts. Models are not physical: they travel in a person's head, through an API or inside open-source weights, so there is no border at which to stop the shipment. And the ambition is anything but modest: the aim is to deny a rival the defining technology of the era, at the very moment AI has become synonymous with power - which makes the incentives to circumvent the controls overwhelming. This is the arithmetic of the war on drugs: harsh penalties suppress some activity, but where the prize is big enough, the trade stays alive.
Law is therefore a thumb on the scale, not an absolute. It can hold scarce matter inside a border; it cannot hold knowledge that wants to move.
That distinction - between theft and movement - brings us back to Apple, and to non-competes. Trade-secret theft is inexcusable. Changing jobs is not. Apple is not seeking an order stopping those 400 employees from working at OpenAI - it couldn’t; California has famously never enforced non-competes (and its innovation ecosystem has thrived).
Of course, non-competes exist for a good reason - to protect companies’ IP and so incentivise its creation. But they do not reliably protect IP once created, while discouraging the mobility and competition that create it in the first place. The UK and European ecosystems have been debating whether to follow California’s lead (see here and here). Research links the clauses to weaker mobility, firm creation, innovation, wages and productivity. So let’s be pragmatic here too: lose the law, retain talent by giving people reasons to stay, and find other ways to promote the creation of valuable IP that cannot be so easily circumvented - like building infrastructure, which is just like helium but MUCH, MUCH HEAVIER.
*The irony is not lost on me. Helium, which due to physics we cannot prevent from leaking off the planet, may somehow not leak out of China. But data? Can’t be done.
IPOs / Publics
SpaceX joined the Nasdaq-100 at $2T with 1.34% weighting, drawing an estimated $4.3B in passive inflows as index funds automatically absorb the stock - though shares fell 5.4% on inclusion day (here)
Apple hit an all-time high, outperforming Mag 7 peers as investors reward its light-capex AI strategy - monetising its 2.5B device install base through hardware upgrades & services rather than data-centre buildouts that have pushed Amazon & Meta cashflows negative (here)
IBM issued a profit warning that sent its shares plunging more than 20%, as customers raced to redirect spending to servers & storage ahead of price increases driven by the AI boom (here)
ASML reports Q226 earnings on July 15, with TSMC following on July 16. Both ones to watch, given the firms represent the foundational infra layer of the entire AI boom. TSMC already posted NT$442.68B in June revenue, a 68% increase YoY (here)
Apple sued OpenAI in California federal court alleging the ChatGPT maker stole hardware designs & trade secrets via former employees as it prepares to launch AI devices. Over 400 ex-Apple staff now work there. The lawsuit names Tang Tan (OpenAI’s chief hardware officer, formerly 24 yr at Apple) & Chang Liu (electrical engineer who accessed files after departure). Apple is seeking injunction & damages; the suit marks collapse of their partnership after OpenAI acquired Jony Ive’s io for $6.4B in May 2025 to compete in consumer hardware (here)
PrismML compressed Alibaba's open-source Qwen 3.6 (27B parameters) to under 4GB to run on Apple's iPhone 17 Pro, enabling advanced on-device AI without cloud inference - a shift that could dampen hyperscale capex if edge models rival data-centre performance (here)
Mark Zuckerberg (Meta) told staff AI agent technology had progressed slower than expected over the past 4 mo, acknowledging the company’s 10% headcount cut & restructuring in May had not been “clean” & bets on agentic AI “haven’t come to fruition yet”. Meta expects to spend up to $145B on AI infrastructure this year & anticipates benefits within 3-6 mo (here)
Slack integrated Slackbot with Salesforce’s Headless 360 platform using Model Context Protocol servers, allowing employees to query CRM data, generate Tableau charts & trigger DocuSigns via chat - democratising Salesforce access beyond the 20-30% of staff who typically hold licences. Salesforce expects to spend $300M on Anthropic tokens this year, yet Anthropic’s Claude Tag (also in Slack) raises internal concern about competitive overlap (here)
Big Dogs
Anthropic alleges that Alibaba’s Qwen AI lab orchestrated a coordinated distillation attack using roughly 25,000 fake accounts that generated over 28.8 million interactions with Claude between April 22 & June 5, 2026. The operation’s goal, according to Anthropic, was to extract Claude’s advanced coding capabilities & use the outputs to train Alibaba’s own Qwen models (here)
Anthropic launched Claude Cowork on mobile & web, enabling tasks to run autonomously across devices & in the background. Usage data from 1.2M sessions shows 33% of Cowork activity is business operations (reports, checklists, reconciliation), 16% content creation, whilst software development accounts for just 9% - a deliberate push beyond the developer seat into the far larger knowledge-worker market (here)
Anthropic hired Tom Blomfield (Monzo cofounder, Y Combinator GP) to join its compute team, continuing a high-profile hiring spree that includes Andrej Karpathy (OpenAI cofounder) & John Jumper (Google DeepMind, Nobel laureate for AlphaFold) (here)
xAI launched Grok 4.5, its strongest model yet, with improved speed, lower cost & token efficiency compared to Anthropic’s Opus; excels at autonomous tasks, coding, finance & legal work, & can build complex Excel models & PowerPoint decks without human intervention (here)
OpenAI publicly released GPT-5.6 models (Sol, Terra, Luna) after a 2-week limited rollout requested by the US government, which had sought pre-release assessments under Trump’s June AI executive order. The company also launched GPT-Live voice models that can listen & speak simultaneously, rolling out globally. The release follows Anthropic’s restoration of Claude Fable 5 & Mythos 5 access after a clash over export controls (here)
The power developer behind OpenAI’s Stargate data centre is in talks to sell a stake, signalling early investor liquidity interest in infrastructure tied to AI compute buildout (here)
Venture Capital / Finance
CVC closed its European mid-market PE fund Catalyst III at €3.0B, nearly double the €1.75B target, reflecting LP appetite for established platforms with local presence; the fund has already deployed into cybersecurity (WithSecure) & medical devices (WillowWood) (here)
Eurazeo closed its fifth-generation secondaries programme at €2.3B, above the €2B target & more than double its 2021 predecessor (€1B). The firm’s Secondaries & Mandates platform now manages €6.4B across GP-led & LP secondaries, focused on European mid-market buyouts, with ESF V already ~50% deployed across 22 transactions (here)
Bregal Milestone closed Fund III at €915M (oversubscribed hard cap) in 8 mo, targeting European B2B software, AI & cybersecurity with embedded AI deployment capability through its Amplify programme; the London-based firm now manages c.€2.3B across three funds (here)
Expeditions closed €197M for Fund II (2x initial target) backed by BAE Systems, European Investment Fund & Keysight Technologies, alongside NATO Innovation Fund & Polish Development Fund. The Warsaw/London firm focuses on autonomous systems & critical capability gaps (here)
Nest, the UK’s largest pension fund by membership (£68B AUM, 14M members), is allocating up to £1B to VC by 2030, starting with £20M via Schroders Capital into late-stage companies including Synthesia & Wayve. The move follows UK government efforts to funnel pension capital into private markets - signatories to the Mansion House Compact doubled unlisted equity allocations from £800M to £1.6B in 2025, with a further £50B expected from the 2025 accord requiring 10% private market exposure (here)
Value of takeover bids for UK companies is 27 times higher than the value of new listings this year. Only seven companies have listed on the LSE so far in 2026, raising a small amount compared to the large acquisitions (here)
Regulation
France will finally publish a decree banning advertising for imported fossil fuels by end-2025, closing a 5-year gap since the 2021 Climate & Resilience law. Environmental groups call it too narrow - the ban excludes broader corporate communications from hydrocarbon firms & fossil-fuelled products (SUVs, air travel) - & question why enforcement was delayed when direct fuel advertising has already largely disappeared (here)
The New York Times & Daily News claim OpenAI hid evidence in their copyright lawsuit, alleging the company secretly searched chat logs & training data for copyrighted journalism while telling courts such searches were technically infeasible (here)
The European Commission will propose age restrictions for children accessing social media after summer, with phased access by age group, following pressure from member states concerned about child safety. Australia enacted a ban for under-16s last year; the EU is also investigating Meta platforms under the Digital Services Act for potential breaches related to children’s exposure (here)
European Commission ordered Meta to redesign Instagram & Facebook to remove addictive features - including infinite scroll & autoplay - or face fines up to 6% of global revenue under the Digital Services Act, part of a broader EU push on social media design following similar action against TikTok in February (here)
Apple lost its challenge at the EU General Court against Digital Markets Act designations covering its App Stores & iOS as gatekeepers subject to interoperability & competition obligations, with judges rejecting claims the rules are disproportionate or threaten privacy protections. Apple can appeal to the CJEU on points of law (here)
Venture Geopolitics
China imposed an immediate helium export ban to secure domestic supply for chipmaking as Middle East conflict threatens shortages; China imports ~85% of its helium (half from Qatar), but typically re-exports surplus to Asia - Beijing now blocking that arbitrage to prevent local price spikes amid AI & semiconductor demand (here)
AI researchers & former OpenAI staff (the same people behind AI 2027) published their next piece: “Plan A”, a detailed policy proposal calling for an international deal to delay superintelligence until 2040, mandate full research transparency & enable multiple nations to scale safely together rather than race (here)
Beijing held meetings with Alibaba, ByteDance & Z.ai over potentially restricting overseas access to China's most advanced AI models, including unreleased systems, according to three sources. Officials discussed classifying AI theft under national security law & limiting foreign funding of domestic AI startups, mirroring U.S. moves to restrict Anthropic's Mythos model. Chinese authorities reportedly fear Mythos could be weaponised against Chinese interests. The talks follow Beijing's April order forcing Meta to unwind its $2B Manus acquisition & June rules tightening control of outbound tech deals (here)
Brad Smith (Microsoft) criticised Washington’s AI policy as “regulation without transparent or complete rules”, arguing the government lacks proper tools beyond export controls. He said the Anthropic Fable 5 shutdown was justified on cybersecurity grounds but executed with a blunt instrument designed for hardware, not API-delivered models (here)
Fed chair Kevin Warsh tapped Marc Andreessen to co-lead a task force advising the central bank on AI’s impact on productivity & labour markets, giving the a16z co-founder influence over how the US central bank models structural employment & output shifts in the AI era (here)
Von der Leyen & Rutte called for NATO to become “more European” ahead of the Ankara summit, with Rutte arguing “to stay transatlantic, we have to become more European” as allies showcase increased defence spending in response to US drawdown (here)
Trump praised NATO’s ‘tremendous unity’ at the Ankara summit despite earlier lashing major European allies for not joining the US-Iran war & calling Spaniards ‘hopeless’ while threatening to cut trade. The summit delivered a win for Ukraine: Trump said he would grant Kyiv a license to manufacture Patriot air defence systems domestically & NATO members pledged $80B in military aid over 2 years & reaffirmed mutual defence commitments (here)
Germany is struggling to retain skilled migrant workers who leave due to bureaucratic delays (naturalisation, permits, qualification recognition), language barriers & jobs below qualifications. Around 60% return home, 40% move to other EU countries (here)
Chinese AI models are gaining ground in the U.S. companies because they work well & cost less than American models. Share of tokens used by U.S. companies on Chinese AI models via OpenRouter - a platform that enables developers to access a range of AI models - has sat above 30% each week since Feb. 8, with that figure rising as high at 46%. The average across the previous 12 months was just 11%, falling to 4.5% in the first half of 2025 (here)
The OBR warned UK debt could rise from ~95% of GDP (2030-31) to 300% by 2075-76 under baseline assumptions, driven by ageing population pressures (health, care, pensions), defence, & net zero investment - though it stressed governments would act before such paths materialise. The UK has seen one of the largest debt increases among advanced economies over 20 yrs (here)
Trump’s naval blockade & 20% Hormuz transit fee (later dropped) sent Brent crude up 10% to $83/barrel, yet refined product shortages remain acute - diesel, jet fuel & petrol prices are 35-70% above pre-war levels as global refining output sits 8M b/d below normal. Gulf refining capacity down 1.4M b/d from Iranian strikes, China’s export ban still constraining supply, & Ukrainian drones knocked out 1.5M b/d of Russian refining capacity. Diesel queues stretch for days in Russia; Turkey & Brazil are redirecting flows, tightening Atlantic markets. Markets bet on further price rises unless Hormuz reopens & Gulf/Chinese/Russian production recovers before winter demand (here)
Strategic Sectors
AI
SoftBank’s Son told executives at SoftBank World that AI would account for 20% of global GDP ($46T) by 2040, requiring 3 terawatts of data centre capacity - equivalent to a third of global installed power today - & urged Japanese business leaders to prioritise AI or “play the role of spouse” (here)
A workforce study projects the US semiconductor industry will face a deficit of up to 157,000 skilled workers by 2030, threatening domestic fab expansion including Intel’s Ohio facility - a supply-side constraint on reshoring ambitions driven by CHIPS Act subsidies (here)
DeepSeek is developing its own AI inference chip to reduce reliance on Nvidia & Huawei silicon, according to three sources - a strategic shift that follows OpenAI’s Jalapeno reveal & reflects both global hyperscaler vertical integration & China’s semiconductor self-sufficiency imperative (here)
Nscale’s £2B Nvidia-backed AI data centre in Essex faces delay after being told its 90 MW grid connection won’t be ready for the planned 2027 opening (here)
Notable deals:
Mercor (AI data labeling) is discussing raising funds at $20B, doubling its $10B valuation from under a year ago, despite a March security breach that led Meta to pause work indefinitely (here)
Positron (AI inference chips) is in talks to raise $750M across two tranches at $3.5B then $5B, tripling its February valuation in 5 mo amid fierce competition for Nvidia alternatives - part of a broader wave that saw SambaNova raise $1B at $11B & Etched secure $800M with $1B in sales contracts, though Cerebras’ post-IPO slide to below issue price highlights execution risk (here)
Swedish vibe-coding startup Lovable is in talks to raise $300m at a $13.2B post-money valuation (here)
Cybersecurity
VentureBeat survey of 107 enterprises found 69% share API keys across AI agents, creating lateral compromise risk after a single breach. Only 30% sandbox high-risk agents, & isolation drops to 20% at firms over 1,000 employees - where incident rates hit 63%, up from 49% at smaller firms. Most enterprises (82%) rely on provider-native guardrails (OpenAI, Google, Microsoft) rather than dedicated agent identity platforms (Entra Agent ID at 13%, CrowdStrike at 6%), though 59% plan to adopt or replace tooling within 12 mo (here)
Iran-linked actors reportedly conducted a phone-tracking campaign targeting US military personnel, exploiting mobile device vulnerabilities to geolocate service members during heightened regional tensions (here)
Energy / Climate
Notable deal:
Proxima Fusion raised €411M at €2.4B, Europe’s largest fusion investment on record, led by XTX Ventures & East X Ventures with RWE & Google as strategic investors. The Munich-based stellarator company will build Alpha, a net-energy demonstrator near Munich by the early 2030s, ahead of Stellaris, the first commercial stellarator power plant later that decade (here)
Defence
Anduril CEO Brian Schimpf said the defence tech firm (valued at $61B in May) is not rushing to IPO amid what he calls a “hype cycle” with “crazy high valuations” in the sector, warning that irrational pricing could backfire for investors. SpaceX shares down ~25% from their post-IPO peak offer a cautionary example (here)
Pentagon seeks cheap hunter-killer drones after losing 30+ Reapers worth over $1B to Iranian air defences, drawing on Ukraine’s mass-drone playbook to overwhelm layered defences with expendable platforms rather than $30-50M aircraft (here)
Notable deals:
Helsing raised $1.8B at $18B from Goldman Sachs Alternatives, Dragoneer, CPPIB & others, cementing its position as Europe’s largest defence startup. The Munich-based firm has expanded from AI-powered weapons software into drones & maritime vessels, securing €493M in German government orders this year (here)
Kraken Technology Group (autonomous maritime vehicles) raised $175M at $1B led by DTCP, with backing from NATO Innovation Fund, Rheinmetall & British Business Bank. The London-based unicorn builds uncrewed naval vessels for surveillance & coastal combat, with contracts across UK MoD, NATO partners & US Special Operations Command, plus manufacturing partnerships with Rheinmetall, Anduril & Davie Shipbuilding (here)
Space
Eight NATO allies (Canada, Denmark, Finland, Germany, Netherlands, Norway, Sweden, Turkey) are building HALO, a mega-constellation linking national military satellites for comms, intelligence & missile tracking - echoing the US Space Force’s Starshield-backed data transport model (here)
Notable deal:
Blue Origin is raising $10B at $130B pre-money (with $2B from Bezos), marking its first external funding since founding in 2000. Led by Coatue, the round comes as Blue races to close the gap with SpaceX, which raised $86B at a $2T+ valuation in June & has completed hundreds of launches versus Blue's handful of orbital tests (here)