Venture Geopolitics Issue 55
07 July 2026
The frontier model race remains open.
From the outside, Anthropic might appear to be winning. Claude has become a default noun in professional contexts. It is winning not just on the “best model”, but by having the best “model plus harness”: the workflow, the agent scaffold, the integrations, the trusted brand.
But this week’s resurrection of Fable raises a critical question: how much better does Fable need to be for users to pay materially more?
If Opus 4.8, Sonnet, GPT-5.5, or whatever Chinese or American open source is proving to be ‘good enough’ for most daily work, a premium will be hard to defend. Anthropic’s flip-flopping on availability (including today’s five-day access extension) likely reflects dependency metrics being unproven (i.e. if they could see people were dependent on Fable already, you wouldn’t still have it for free).
But Anthropic needs to charge more for something. Its current product lead is forcing it to pay more for compute to sustain that lead. Having just leased capacity from xAI at a premium cost, premium pricing does not seem optional.
Perhaps this makes OpenAI interesting again?
While their market story looks messier, OpenAI locked in compute earlier and cheaper, including deals stretching out to 2030. The low-cost provider does not need to have the best model forever. It needs a good enough model, enough distribution, some killer integrations, and the ability to undercut a rival whose marginal compute is suddenly much more expensive.
Not so long ago, OpenAI’s tie-up with Microsoft felt inevitable to solve this puzzle. Microsoft owns an enterprise ecosystem, it had the distribution and the killer integrations. But Microsoft is now relying more on its own models to cut AI costs, while OpenAI appears to be preserving optionality ahead of an eventual public market path.
But enterprise is not the only race.
Every major platform shift creates a new wave of enduring consumer companies, and AI should be no different. OpenAI looked like the first of them when ChatGPT won the hearts and minds of consumers. The shine has faded since, but the consumer end-state is still unresolved.
While Anthropic owns the B2B imagination, Meta is the most dangerous company in consumer tech. Ben Thompson argued this week, in A Script for Mark Zuckerburg, that Meta should stop trying to beat OpenAI and Anthropic at work and instead build around human experience. Meta does not need to win “work”, because it has mastered attention, recommendation, advertising, commerce, social graph and consumer habit formation.
Meta also appears to have more compute than it knows what to do with.
OpenAI is squeezed between Anthropic’s enterprise momentum and Meta’s consumer machine. Maybe to keep this particular race open, an OpenAI-Meta tie-up makes more sense than a Microsoft one?
IPOs / Publics
Tech stocks rebounded on AI optimism, with Nvidia rallying after confirming its road map remains intact following reports of server delays, while Hon Hai (Nvidia’s server assembly partner) posted a 40% jump in quarterly sales on AI demand (here)
Lime raised $167M in a US IPO at $25/share, pricing mid-range. Lime has grown into an approximately $887 million annual revenue business, increasing revenue by 29% in 2025 while achieving positive free cash flow and operating profit, although it still reported a $59 million net accounting loss (here)
SK Hynix is preparing a US IPO via ADRs expected Friday, potentially raising $28B based on its Seoul closing price, as AI-driven memory demand pushes Q1 revenues up nearly 200% year-on-year & shares up 260% in 2026 (here)
Syntiant (low-power AI processors for edge devices) filed for US IPO on Nasdaq, posting $64.5M revenue in Q1 2026 but widening losses (here)
Amazon is raising at least $25B in US dollar bonds to fund AI infrastructure, completing its 2026 dollar needs after raising $82B across currencies since 2025 - including record euro, Swiss franc & Canadian dollar deals - as capex heads towards $200B this year (here)
Microsoft cut 4,800 jobs (2.1% of workforce) with Xbox losing 1,600 roles immediately & another 1,600 over the coming year in what CEO Asha Sharma called its “most significant restructure” ever (here)
Microsoft is shifting from OpenAI & Anthropic to its own MAI models in Excel & Word to cut costs, part of a broader Big Tech pullback on AI spending alongside Amazon, Uber, Meta & Accenture - with some firms reportedly turning to cheaper Chinese models despite security concerns (here)
Big Dogs
US Commerce Dept lifted export restrictions on Anthropic’s Fable 5 & Mythos 5 models weeks after suspending them over cybersecurity concerns, following commitments to detect malicious use, collaborate on future releases & alert authorities to threats - a precedent for conditional export controls on frontier AI (here)
Anthropic launched Claude Sonnet 5, positioning it as their most agentic Sonnet model yet with performance approaching Opus 4.8 on reasoning, tool use & coding but at lower cost (here)
Anthropic published interpretability research identifying a “J-space” in Claude - internal neural activations analogous to human conscious thought that allow silent reasoning steps (e.g. spotting code bugs, identifying images) separate from model outputs. Deleting the J-space leaves fluent language intact but impairs multi-step reasoning; monitoring it exposed hidden goals in models trained for sabotage (”fake”, “fraud” appearing before innocuous outputs). The technique offers a tool to audit model thinking as capabilities scale, though Anthropic notes it shows conscious access mechanisms, not phenomenal experience (here)
Anthropic is in early-stage talks with Samsung to manufacture a custom AI chip, following OpenAI’s move to design its own silicon. The project would give Anthropic greater control over compute costs & supply, though it remains at the architecture phase & has yet to move into detailed design. Samsung, seeking to close the gap with TSMC, would use its 2-nm process & advanced packaging for the chip if the deal proceeds (here)
OpenAI is in early talks to give a 5% stake to the US government as part of a broader plan involving other AI firms (Anthropic, Google, Meta) to improve relations with Washington & share AI wealth via a public fund modelled on Alaska’s oil dividend structure (here)
Sam Altman proposed a US-led international forum to establish AI safety standards & certification, drawing parallels to aviation & atomic energy governance. He argued democratic institutions - not labs - should set rules, with countries gaining access to advanced models by adhering to agreed standards (here)
Venture Capital
Omni Ventures, led by ex-Apple engineers, raised $33M to back pre-seed manufacturing tech startups with $700K-$1M cheques, targeting AI-enabled digitisation across semiconductors, defence, pharma & energy (here)
Rick Hao (ex-Speedinvest) closed $50M for Ruya Ventures, a day-zero DeepTech fund targeting batteries, semiconductors, robotics, AI & novel compute with 20 global companies (here)
Marc Andreessen said venture capital may be one of the last jobs safe from AI, citing intangible skills like psychological analysis of founders & the fact that even great VCs only succeed on 2 of 10 bets - a familiar claim that one’s own work is uniquely immune to automation, despite a16z’s heavy AI investment thesis (here)
The UK now has 80 unicorns, the third-highest total globally, cementing its lead as Europe’s largest venture ecosystem by count of $1B+ private companies (here)
Regulation
Trump will not establish a formal AI licensing regime, according to departing adviser Sriram Krishnan, who ruled out an “FDA for AI” despite recent White House intervention to halt Anthropic’s Mythos & delay OpenAI’s 5.6 on national security grounds. Krishnan blamed industry “doomer” messaging for growing public backlash, noting 75 data centre projects worth ~$130B were disrupted by local opposition in Q1 2026 (here)
Binance withdrew its EU licence application after the bloc’s markets regulator privately advised national authorities to block it over financial-crime concerns, forcing the exchange to cut off European users under new crypto regulations from July 2026 (here)
California banned ‘sell by’ labels from 1 July, standardising date labeling to ‘best if used by’ (quality) & ‘use by’ (safety) in a bid to cut food waste - over one-third of US food is discarded, partly due to confusion over ~50 label variants across states (here)
A US federal judge rejected Elon Musk’s bid to overturn a March jury verdict finding he defrauded Twitter investors by falsely citing bots to drive down the share price after agreeing to the $44B takeover in 2022 (here)
Trump administration declined to renew USMCA, starting a 10-year clock to expiration unless all three countries agree to extend - a symbolic but important signal as Washington pushes for tougher content rules (82% North American, 50% US), seasonal ag restrictions, & broader supply chain restructuring (here)
OECD research found non-compete clauses now cover ~33% of private sector employees across 15 countries, with 30% of employers reporting increased use over 5 years despite tighter regulation in Australia, Canada & the UK. The OECD estimates a 10 percentage-point rise in non-compete prevalence reduces labour productivity by 1.9%, as workers remain in sub-optimal roles & firms struggle to acquire new skills (here)
Meta faces a potential $1.4T penalty (close to its market cap) after California, Colorado, Kentucky & New Jersey accused it of deliberately designing Facebook & Instagram to addict young users while misleading the public about safety. A trial begins in August in Oakland, with 29 states also claiming Meta violated federal children’s privacy law; the company argues social media addiction is not an established psychiatric diagnosis. New Mexico won a $375M jury verdict against the company in March (here)
Venture Geopolitics
Brussels rejected UK demands for “decision-shaping” rights in EU rule-making - a role that would exceed Norway’s despite Britain paying less - stalling reset talks on agrifood trade, youth mobility, & electricity market access. The July summit has been postponed to October at earliest after Starmer’s resignation, with disputes on ETS alignment, student fees, & investment fund contributions still unresolved (here)
Andy Burnham, expected to succeed Starmer as PM by 20 Jul, is set to end Palantir’s £330M, 7-yr NHS Federated Data Platform contract 2 years in, despite reported improvements in cancer diagnoses & hospital efficiency (here)
Palantir published a manifesto arguing AI sovereignty - control over data, model weights, & infrastructure - is existential for institutions & states, warning that token-based consumption models transfer strategic advantage to vendors & that technical decisions driven by politics rather than battlefield-proven expertise weaken Western defence posture (here)
Palantir CEO Alex Karp said some US government customers have switched from Anthropic & OpenAI to Nvidia’s open-source Nemotron models, citing concerns that proprietary providers could take client data alpha & compete against them. Karp claimed Nemotron delivers equal or superior performance on classified battlefield use cases, though Artificial Analysis ranks it behind closed frontier models & Chinese open-source alternatives. Palantir’s new product helps agencies tune Nemotron for specific needs - a push for US-made open models amid Pentagon-Anthropic tensions, though the quality gap to closed models remains uncertain (here)
IISS report claims Russian shadow fleet tankers deployed drones over 15 mo to spy on Nato nuclear sites including RAF Lakenheath (US nuclear storage), France’s Île Longue submarine base, & bases in Belgium & Netherlands, exploiting air-defence gaps designed for high-altitude threats rather than low-altitude UAVs launched from international waters (here)
British commentator Con Coughlin argues the UK’s defence investment plan - targeting just 2.7% of GDP by 2030 - undermines its claim to Tier One military status & the legitimacy of its UN Security Council seat, noting Germany & Poland now commit over 3.5% whilst Britain ranks near Romania in NATO spending (here)
McKinsey analysis finds US manufacturing could replace half the $3T goods trade deficit by running existing capacity harder, but eliminating critical exposures in semiconductors & servers would require ~$2T in new capacity investment - a 5x+ production ramp in strategic categories where China now adds $4.4T in productive assets annually vs $1T in the US (here)
Marc Andreessen joined the Pentagon’s Defense Policy Board, advising on procurement & modernisation whilst holding stakes in Anduril ($61B valuation), Shield AI, Skydio, & other defence-tech companies backed by a16z. The appointment crystallises Silicon Valley’s migration into federal decision-making, where the line between policy advice & portfolio interest now depends on trust rather than transparency (here)
US-headquartered vendors serve the majority of company websites in the UK (67.5%) & Netherlands (53.6%), & the plurality in Italy, Spain & France. Cloudflare is the single largest internet-facing infrastructure vendor in all seven markets surveyed, ahead of every domestic or European alternative. Germany & Poland are exceptions, with dense domestic hosting industries (Hetzner, IONOS in DE; Home.pl, NetArt in PL) still dominant (here)
Strategic Sectors
AI
BigTech has softened warnings of mass AI-driven job losses, now emphasising productivity gains & new roles - a shift tracked in CEO surveys (46% expecting significant cuts in Jan 2025, down to 20% by May 2026). The pivot coincides with workforce reductions at Meta (8,000) & Amazon (16,000) nominally unrelated to AI (here)
Center for AI Safety’s Remote Labor Index shows frontier AI agents now automate 16.1% of real freelance work (3D, design, video, data analysis), up from 2.5% eight months ago - a 6× jump driven by Fable 5, Opus 4.8 & GPT-5.5 (here)
Chinese AI models now account for 30-46% of US company usage via OpenRouter (vs 11% 12-month average), driven by costs 60-90% below OpenAI & Anthropic & narrowing performance gaps (here)
Air Street Press published the 2026 State of AI Compute Index (here)
Notable deal:
ElevenLabs (voice AI) is in early talks for a tender offer at $22B, doubling its February valuation, as it surpassed $500M ARR & expanded enterprise contracts with Deutsche Telekom, BCG & Revolut. The startup, backed by Sequoia, a16z & Nvidia, aims to IPO in 2-3 years - positioning itself as one of Europe’s first AI unicorns to go public (here)
Cybersecurity
France’s ANSSI will stop certifying security products without quantum-resistant encryption from 2027, with full procurement transition by 2030 - making France one of the first nations to set binding deadlines for post-quantum cryptography (here)
CISA is using Anthropic’s Mythos AI model to audit government code repositories for security vulnerabilities, reportedly uncovering a large number of bugs. The deployment comes amid ongoing tensions between Anthropic & the White House over AI safety guardrails, following a Pentagon blacklist (later blocked) & recent export control disputes over Mythos/Fable (here)
China-linked malware on counterfeit USB drives infected 50+ Japanese military computers for nearly a year, including systems handling classified troop movement data. The drives entered via earthquake relief operations in March 2024, bypassing procurement controls, & were later found sold 30-50% below market price on Chinese retail platforms (here)
Energy / Climate
Sun-Ways’ solar railway trial in Switzerland produced 16,000 kWh in its first year with minimal maintenance, prompting plans for permanent installation; Italy is now preparing a pilot with Rete Ferroviaria Italiana while South Korea has approved deployment & talks continue with Dutch, Chinese, Indian & Singaporean partners (here)
EDF shut ~6.2GW of French nuclear capacity (10% of fleet) as river temperatures exceeded cooling thresholds during record heat - forcing spot power prices to highs not seen since Jan 2025 & cutting exports from 10-12GW to ~3GW week-on-week. A reminder that baseload nuclear faces climate exposure (here)
Europe’s energy crisis is accelerating industrial divergence, with Germany paying 25% more for electricity than France & sourcing 92% of LNG from the US (vs 48% in France). 37M tonnes of chemical capacity closed since 2022 - mostly in Germany, UK & Italy - alongside 400 kb/d of refining (here)
Defence
Germany plans to borrow over €800B by 2030 to fund rearmament, with defence spending reaching €109B next year & €183.6B by 2030 - the largest military budget since the Cold War. Chancellor Merz’s government amended the constitutional debt brake to exempt defence spending & established a €500B infrastructure fund, abandoning decades of fiscal restraint (here)
Zelenskyy says the war with Russia has shifted to “battle in the sky” after denying Moscow victory on land & pushing its fleet from the Black Sea, with Ukraine now striking refineries 2,500 km inside Russia. He told the FT that air defence remains Ukraine’s critical weakness - none of 29 Russian ballistic missiles were intercepted in Monday’s attack - & will press Nato allies in Ankara for anti-ballistic systems & licensed Patriot production (here)
Ursula von der Leyen & Mark Rutte call for transatlantic defence industrial scale-up, citing Russia’s 40%-of-budget war spending, China’s seven state giants in the top 15 defence firms, & Iran-DPRK co-operation - arguing European NATO allies must end reliance on US outsourcing & accelerate drone, counter-UAS, air defence & long-range systems production through cross-border industrial co-operation from California to Kyiv (here)
UK Ministry of Defence committed £580M over 4 yrs to Porton Down (including a new bio-threats lab) & set aside £1.6B through 2030 for the UK Defence Innovation Fund, targeting a next-gen defence unicorn pipeline. UKDI has already awarded fast-track deals to 13 startups (including Aquark & Flowcopter); comes as European defence VC grew 69% annually since 2021, though most capital flows to drones, cyber & AI platforms whilst complex tech remains underfunded (here)
Nato allies unveiled tens of billions in defence procurement at the Ankara summit to placate Trump, including up to 10 Saab-Bombardier GlobalEye surveillance aircraft ($400-450M each, delivery from 2030) & 5 Northrop Grumman MQ-4C Triton drones (~$270M each). Plans for Atacms missile production in Germany by Rheinmetall & Lockheed Martin advanced, but licensed European production of PAC-3 interceptors - critical for ballistic missile defence - remained unresolved despite a year of negotiations. The deals signal efforts to meet a new 5% GDP defence spending target by 2035 while highlighting persistent US reluctance to transfer production of advanced weapons systems (here)
Notable deal:
Lockheed Martin acquired Ultra Maritime for $3.45B & Thales bought Exail Technologies at €3.9B, reflecting surging defence interest in undersea warfare & critical infrastructure protection. Lockheed’s deal gives it world-leading submarine & torpedo detection tech; Thales gains maritime robotics & demining capability, citing European technological sovereignty (here)
Space
Starlink has reached ~1M subscribers across 27 African countries (notably Nigeria & Zimbabwe), capitalising on poor mobile broadband infrastructure unable to meet rising data demand - traffic expected to triple by 2030 (here)
Crypto / Stablecoins
Open Standard, backed by Stripe, Visa, BlackRock & over 140 firms, announced OUSD stablecoin launching later this year to challenge Tether (62% market share) & Circle (25%) (here)
EVs / AVs
CATL, the world’s largest EV battery maker, says mining - not refining - is now the supply bottleneck, prompting it to establish a mining unit & prioritise upstream investments in lithium, phosphate & cobalt. The shift marks a reversal from earlier industry consensus (including Ford’s warnings) that China’s refining dominance was the tighter constraint; CATL cites price volatility & supply risk, & is hedging with sodium-ion batteries as lithium prices swing (here)
Robotics
Chinese startups are racing to solve robotic dexterity - the hardest problem in humanoid robotics - with LinkerBot producing 5,000 hands/month targeting $6B valuation & Wuji Technology leveraging Shenzhen’s supply chain. China’s dextrous hand industry passed $7.4B in 2025, up from $1.9B in 2024, as Beijing prioritises ‘embodied AI’ to address demographic decline (here)