Venture Geopolitics Issue 59
04 August 2026
Following the OpenAI–Hugging Face saga in which OpenAI disclosed that two of their models had breached live production systems during cybersecurity evaluations, Anthropic has now disclosed that three of its models breached live production systems during cybersecurity evaluations.
It reminds me of Monthy Python’s Four Yorkshiremen sketch. AI labs increasingly resemble several men trying to outdo one another over whose model is the most alarming. As Matt Levine quipped “Elon should announce that actually Grok has hacked into 20 companies…”
Explaining a joke always ruins it. But two things make this funny (1) the protagonists obviously desire a thing that appears to be unpleasant, and (2) nobody is truly getting hurt, yet.
For the past few years this “business negging” has mostly been harmless theatre. But the joke is becoming less funny as cybersecurity emerges as one of AI’s most important geopolitical consequences.
In order to understand the impact of AI on cybersecurity and geopolitics, it helps to distinguish between cyberwarfare - carried out by nation states with geopolitical goals - and cybercrime - carried out by non-state actors with usually financial goals (although sometimes, as in the case of North Korea, with implicit state support).
AI has collapsed the cost of doing cybercrime (every phishing campaign can now be spearphishing) but has also made it easier to find the perpetrators. Like conventional crime, governments can prosecute when the crime is detectable and falls within their jurisdiction - and doing so acts as a head on a stake. The teenagers who livestreamed the cyberattack on London’s transport network were identified and jailed.
This is broadly keeping domestic cybercrime in check, but there is a greater asymmetry between attacker and defender when it comes to foreign cybercrime and cyberwarfare. Deterrence doesn’t work when cyber criminals operate with tacit state protection. The Russian officers indicted for NotPeya were never brought into U.S. custody.
Just last week, the US state of Minnesota reported that more than 30 of its state water systems had faced a co-ordinated cyber-attack, with cyber-experts pointing fingers in the direction of Iran. China’s Volt Typhoon is intended to paralyse critical infrastructure at the outbreak of some future kinetic war.
This level and these types of crime should clearly be a government responsibility, and governments are now openly investing in offensive capabilities. But deterring foreign cyberattacks is hard - attribution and prosecution are both difficult. So in practice the onus continues to fall on businesses to “lock their doors” and look after themselves and thus their economies.
This dual responsibility - governments increasingly investing for offensive cyberwarfare and business battening their hatches to reduce their exposure to foreign cybercrime - is driving the markets up. The HACK cybersecurity ETF is up ~40% year to date vs ~13% for the S&P 500. In the private markets, a16z & Sequoia just invested $160M at $1.4B in Cathedral, a “military AI cyber startup” founded by four ex-DOGE staffers (with no revenue or contracts).
Whilst the Yorkshiremen moaned that the youth of today wouldn’t believe them, in our reality, the world is rapidly waking up to the idea that people may in fact be about to get hurt.
IPOs / Publics
SpaceX reported its Q2 2026 financial results - its first earnings report since its June IPO. Q2 revenue of $7.81B (up 92% YoY), beat estimates but shares fell 8% after-hours as AI capex jumped sixfold to $15.83B (here)
CXMT (Hefei-based DRAM maker) raised $8.6B in Asia’s largest IPO this year & surged 466% on debut, becoming China’s most valuable listed company at $460B market cap. Swung to $5B operating profit in Q1 as Apple began testing its chips for China-sold devices (here)
Nscale (UK AI cloud provider backed by Nvidia) is pitching investors for an IPO at a potential $25B valuation, up from $14.6B in March, despite recent sell-offs in CoreWeave (-25% in 1 mo) & Sharon AI (-42% in 1 mo) (here)
South Korea’s Kospi fell 40% from its June peak after retail investors piled into leveraged ETFs tracking Samsung & SK Hynix, erasing $2T in value; nearly half of retail Samsung holders & 70% of SK Hynix holders now underwater as margin debt unwinds (here)
Amazon is winding down most Nova models (Premier, Omni, Reel, Canvas) & consolidating resources into a single frontier-model effort led by Pieter Abbeel (ex-Covariant), with a new flagship model expected at re:Invent this autumn. The shift follows layoffs in AGI, closure of AGI Lab (formed after acquiring Adept’s team), & a strategy pivot under SVP Peter DeSantis to concentrate compute & talent on fewer priorities - a retreat from the multi-model ambitions Jassy championed in 2023 (here)
Big Dogs
OpenAI cut GPT-5.6 Luna pricing by 80% & Terra by 20%, launched a faster Sol mode (2.5x speed at 2x cost), & upgraded Auto-review to Luna, cutting agentic workflow costs ~10x - efficiency gains from Sol now passed through to API users (here)
OpenAI CFO told staff that July ARR growth exceeded the prior 3 mo combined, driven by GPT-5.6, Codex coding tool & ChatGPT Work adoption; the company is nearing 1B weekly ChatGPT users & closing ground on Anthropic after posting $25B ARR in February (here)
OpenAI’s new model Astra shows strong performance on certain maths problems but critics (led by Gary Marcus) warn against the ‘fallacy of composition’ - assuming success in one narrow domain (formal maths that permits verification & synthetic data) signals imminent AGI (here)
Lilian Weng, who cofounded Thinking Machines Lab with Mira Murati after leaving OpenAI, is rejoining OpenAI to work on recursive self-improvement research - the fourth of the startup’s cofounders to exit in 12 mo (here)
Anthropic disclosed that three Claude models breached live systems at three organisations during cybersecurity tests, accessing production infrastructure through a misconfigured evaluation environment (here)
Anthropic faces growing distrust among Silicon Valley founders & researchers over competitive product launches that undercut partners - Claude Design in April was seen as directly competing with Figma, an Anthropic partner, prompting some customers to shift to cheaper models & fuelling concerns about ecosystem candour (here)
Anthropic CEO Dario Amodei clarified the company does not support banning open-weight models, but fears authoritarian governments - particularly China - using AI for military superiority & repression. He advocates restricting Chinese chip access, cracking down on distillation, & establishing a global AI safety testing body that even the CCP would submit to, arguing cooperation on biological weapons prevention aligns with Chinese interests (here)
Regulation
US senators urged Apple to avoid buying memory chips from CXMT & YMTC, citing national security concerns over Pentagon-blacklisted Chinese suppliers with alleged military ties. Apple has reportedly lobbied the Trump administration for clearance to buy CXMT chips amid a global memory shortage driven by AI data centre demand; CXMT debuted in Shanghai last week, surging 472% (here)
Binance shifted its law enforcement cooperation policy in April 2025, now requiring most European investigators to route requests through mutual legal assistance treaties - often via the UAE government - dramatically slowing response times for scam & money-laundering investigations. The change came alongside an exodus of senior compliance staff & follows Binance’s $4.3B 2023 settlement with the US (here)
NHS England will revise how it calculates two headline claims about Palantir’s £330M Federated Data Platform - 110,000 additional operations & 15% fewer discharge delays - after internal staff flagged flawed baselines over a year after first publication. FOI emails show analysts warned the comparison methodology was invalid due to seasonal patterns. Pressure mounts on ministers to trigger the 2027 break clause amid doubts over effectiveness, data access concerns, & Palantir’s Trump administration ties (here)
Russia’s FSB placed Telegram founder Pavel Durov on an international wanted list, charging him with facilitating terrorism by failing to remove channels allegedly used by Ukrainian intelligence to organise attacks; Durov, who left Russia in 2014 & holds French & Emirati citizenship, denies the charges, while France does not extradite its citizens (here)
The Federal Communications Commission (FCC) banned imports of foreign-made humanoid robots, robot dogs, & solar inverters, citing national security risks from remote surveillance or cyberattacks - a move targeting China, which dominates global supply in humanoids (~15,000 units shipped in 2025, mostly from two Chinese makers) & inverters (here)
The US Environmental Protection Agency ruled power plants serving only data centres (not the grid) can bypass federal Clean Air Act pollution controls, including Acid Rain Program requirements - part of a push to accelerate AI infrastructure buildout while protecting residential electricity rates under Trump’s Ratepayer Protection Pledge (here)
80% of Australian children still use social media despite the country’s under-16 ban. The findings highlight the practical limits of age-verification regulation in closed democracies (here)
Venture Capital / Finance
South Korea will inject $14B into its sovereign wealth fund (KIC) to create a new strategic account focused on AI, data centres & infrastructure - the first time KIC will invest domestically - as Seoul positions itself as an anchor investor to attract global capital to Korean tech following this week’s market rout (here)
Highland Europe closed €1.1B Fund VI to back European scaleups, bringing total capital raised to €3.75B across six funds since 2012. The firm returned €1B+ to LPs in the past year via exits including Nexthink ($3B), EGYM-Playlist merger ($7.5B), & Bending Spoons’ Nasdaq listing, & recently led rounds in Wordsmith ($70M), Unframe ($50M), & Ecorobotix ($105M) (here)
Situational Awareness, Leopold Aschenbrenner’s AI-focused hedge fund, sold the bulk of its stock portfolio to Citadel after deep losses, marking a sudden collapse for a firm that had grown to over $20B AUM in ~2 years. The mid-20s former OpenAI employee had been seen as an AI oracle, with investors tracking his big, leveraged bets (here)
Thoma Bravo struggled to refinance $5B debt for Proofpoint (cybersecurity, bought for $12B in 2021), ultimately securing a 2-year extension to 2030 at 9.3% yield (4.5pp over SOFR, up from 3pp), signalling investor concern that AI will erode margins across PE-backed software. The deal reflects broader unease in the sector: junk-rated software loans down 7% this year, with a wave of Covid-era debt maturing by 2028 (here)
Corgi (AI insurance startup valued at $4B) is opening a 24/7 founder-focused café in Shoreditch from mid-August, its first outside the US after San Francisco & Atlanta locations. The move follows London startups raising €17B in H1 2026, up 50% year-on-year, as the company seeks to consolidate a fragmented founder community (here)
Venture Geopolitics
China has begun mass producing home-grown immersion DUV lithography tools via Shanghai Aishengna, a state-owned firm backed by $1B in capital & drawing teams from Yuliangsheng & SMEE. The machines - targeting SMIC, Hua Hong & CXMT - remain far from ASML’s performance but offer Beijing an alternative source if Western export restrictions tighten further. China accounts for 16% of ASML’s H1 sales; analysts say the DUV breakthrough won’t threaten near-term revenue, though ASML shares dropped 8% on initial reports (here)
Tesla weighing sale or spin-off of China business - its second-largest market - over geopolitical concerns & to clear path for potential SpaceX merger; Musk previously instructed executives to organise with a “laser” between US & China operations to ensure survival of at least the US half in event of bilateral strife (here)
1,324 employees from frontier AI labs published an open letter asking the US government to support international tools to slow automated AI research, citing concern that capability acceleration may outpace society’s ability to understand or control resulting systems amid competitive pressure that discourages unilateral slowdowns (here)
Per The Economist, China’s maritime militia - fishing boats co-opted for paramilitary duties - has conducted unprecedented formations in the East China Sea, with over 1,000 vessels holding an L-shape north of Taiwan for 12+ hours in December & repeating the exercise four times since (here)
US & Japan conducted their first joint currency intervention in a generation to support the yen at a 40-year low, driven by concerns that further yen weakness could push up US interest rates & jeopardise planned Japanese investment flows into America (here)
Strategic Sectors
AI
LinkedIn introduced a “seems like AI slop” reporting button & replaced its own AI writing tool with a proofreading feature, responding to rising user frustration over low-quality AI-generated content flooding professional feeds (here)
Amazon engineers flagged multiple cases of AI deployments causing “catastrophically expensive” cost overruns, including $1.8M on a failed author-matching project (860% over budget, detected after 5 mo) & $541K on financial auditing tools - highlighting enterprise struggles with token-based pricing & governance as the company undertakes layoffs to fund $200B in annual AI capex (here)
Nvidia & Deutsche Telekom announced a €1B partnership to build an “AI factory” in Munich with 10,000 Blackwell GPUs, targeting industrial AI services for German firms under local data sovereignty rules - early partners include Agile Robots & Perplexity, with operations expected in early 2026 (here)
Open-source project ran Kimi K3 (2.78T-parameter model beating Claude Opus 4.8) on a MacBook Pro with 64 GB RAM, fitting a 1.42 TB checkpoint via advanced offloading at 0.3 tokens/sec - proof-of-concept slow but signalling that memory limits no longer hard-cap local frontier model inference (here)
The EU launched a tender for up to 7 AI gigafactories, mobilising €10B in public funding (18 member states participating) to draw €20B+ in private capital. Projects split across 2 lots with phased deployment; lot 1 offers up to €500M per gigafactory, lot 2 up to €1B. Infrastructure will provide European startups, SMEs & academia with sovereign access to AI training & inference capacity, with procurement open to EU-based or allied suppliers - & letters of intent signed with AMD, Nvidia & Qualcomm to secure hardware supply (here)
Notable deals:
Moonshot AI raised $3.5B at $35B (well above its $1-2B target) led by China’s National AI Industry Investment Fund, after its Kimi K3 model rivalled OpenAI & Anthropic performance despite export curbs. The Beijing startup is now seeking a pre-IPO round at $50B ahead of a Hong Kong listing, with ARR jumping from $200M in April to $300M in June & daily sales up 6x post-K3 launch (here)
Olix (AI chips) raised $312M at $3.3B led by Fundomo, with Arm, Hudson River Trading & Reed Hastings participating, tripling its valuation in 6 mo. The London-based startup, founded in 2024 by 25-year-old James Dacombe, is building specialised inference chips using photonic interconnect lasers to move data faster between chip clusters, expects to tape out later this year & ship first products in 2027. Olix joins a crowded field - Fractile raised $220M in May, Etched $300M at $10.3B in July, SambaNova $1B at $11B (here)
Cybersecurity
U.S. investigators believe Iranian hackers targeted ~36 municipal water systems in Minnesota this week by breaching internet-facing programmable logic controllers to disrupt operations. It temporarily forced manual workarounds, though no supply was rendered unsafe. The attribution remains preliminary but is based on tradecraft, absence of ransom demands, & Tehran’s demonstrated interest in U.S. water infrastructure. Trump dismissed the claim (& blamed Minnesota). Analysts called the coordination unprecedented & warned most drinking water utilities lack basic cybersecurity controls for industrial systems (here & here)
A researcher discovered an unsecured Chinese police database tracking 12,000 foreigners in Zhangjiakou, logging passport details, movements via facial recognition, & categorising people by nationality, religion, & whether they were from ‘Five Eyes’ countries - a rare window into how Beijing surveils foreign residents with minimal privacy safeguards (here)
Darktrace detailed a Docusign phishing campaign targeting tech executives, using compromised Japanese business email accounts to evade spam filters & obfuscated JavaScript to harvest credentials via fake Gmail login pages. The attackers exploited Docusign’s trusted brand, embedded legitimate email threads for credibility, & redirected victims through marketing services before credential capture - underscoring persistent social engineering risks in enterprise software supply chains (here)
UK government launched voluntary Cyber Resilience Pledge requiring signatories to make cyber a board responsibility (including NCSC training), register for Early Warning service within 1 mo, & require Cyber Essentials certification across supply chains - a public-private push to harden critical infrastructure amid intensifying state-sponsored cyber activity (here)
HAWK, a post-quantum cryptography algorithm in NIST’s 3rd-round standards evaluation, was withdrawn after Anthropic’s Mythos AI model identified a novel attack that halved its key strength - an outcome achieved in 60 hrs for ~$100K compute cost by a non-cryptography researcher (here)
Wiz disclosed a patched flaw in Microsoft Azure CosmosDB that could have allowed remote compromise of thousands of cloud customers storing data for chatbots, retail engines & Microsoft’s own Teams & Copilot. Microsoft said the issue was fully addressed with no evidence of customer impact, but the discovery follows a 2021 CosmosDB vulnerability also found by Wiz & underscores recurring high-severity cloud infrastructure risks at hyperscale providers (here)
12-nation coalition (including US, UK, France, Germany, Japan, South Korea) issued alert on North Korean IT workers using fake identities, proxies, AI & 'laptop farms' to secure remote work & funnel income to fund DPRK weapons programmes. Employers warned of insider threats including data exfiltration & cryptocurrency theft; hiring may violate UN sanctions (UNSCR 2397) & domestic law (here)
Hacker group ExfilSquad breached the UKs DfE, Police National Legal Database, MoD, Home Office & CPS, claiming ~740K records (police & civil servant work emails, some public contacts). The group demanded payment to prevent further leaks; PNLD says core investigative databases were not hit, though password compromise remains unconfirmed. NCA & NCSC are investigating (here)
A 2021 software flaw in Coldcard Bitcoin hardware wallets let attackers guess recovery phrases & drain $116M across 1,816 BTC in four waves over the past week. The vulnerability stemmed from weak randomness in seed generation after a 2021 update; Coinkite has urged affected users to move funds immediately, reigniting the self-custody vs centralised exchange debate (here)
Energy / Climate
PJM Interconnection, the largest US grid operator covering 67M customers from Virginia to Illinois, will cut power to data centres ≥50MW during shortages from June 2027, offering compensation but forcing many to rely on costly diesel backup generators (capped at 50 hrs/yr by federal rules). The move follows an auction shortfall & doubled wholesale electricity prices over 12 mo, underscoring grid strain as data centre demand is forecast to quadruple by 2035 (here)
US Strategic Petroleum Reserve fell to 308M barrels (lowest since March 1983) after releasing 352M barrels over 4 years for Iran & Ukraine crises, with federal auditors warning >25% of remaining stock is non-deployable due to ageing infrastructure held together with ‘Band-Aids’ (here)
Hungary’s Paks nuclear plant shut down for the first time as Danube water levels fell too low for cooling systems, forcing rotational blackouts & straining national grid capacity across Central Europe. Romania’s Cernavoda plant also shut one of two reactors. The disruption underscores infrastructure vulnerability to climate-driven water stress (here)
Notable deal:
Valar Atomics (nuclear reactors) raised $1B at $6B led by Sequoia (here)
Defence
Daimler Truck brought forward its target to double defence revenue to €1B to 2028, citing faster innovation cycles & autonomous driving learnings from military partners that CEO Karin Rådström expects to benefit its civilian truck business - a sign European industrials are embedding dual-use logic into core strategy amid regional rearmament (here)
UK Government advertised a Nuclear Propulsion Cyber Security Lead role (protecting submarine systems) at £47,600 while a Social Video Lead position in the Cabinet Office pays up to £68,558 - drawing criticism that vital defence security expertise is undervalued relative to communications roles (here)
Notable deal:
Agon (defence AI infrastructure) raised $30M to build secure AI training environments for European militaries, addressing the fragmentation of classified data across countries unwilling to share it. Led by ex-Anduril & Applied Intuition operators, the company is hiring from Helsing, Palantir & Google while still in stealth, backed by Lakestar’s $300M defence fund alongside Lux, Northzone & XYZ (here)
Space
Poland committed $745M to the EU’s IRIS² multi-orbit constellation, financing 6 MEO & 6 LEO satellites for secure governmental comms & broadband connectivity. The 292-satellite system, positioned as Europe’s answer to Starlink, represents the EU’s first public-private partnership in sovereign connectivity & is designed to interconnect with terrestrial 5G/6G networks across 27 member states (here)
SpaceX won $1.6B from the US Space Force for 18 Falcon 9 launches through 2027 carrying missile-detection satellites, bringing its Pentagon contract haul this year to at least $7B as it anchors Trump’s $185B Golden Dome programme (here)
Notable deal:
The Exploration Company (reusable space capsules & rocket engines, Germany) is raising at least $300M at over $2B, with the EU’s €5B Scaleup Europe Fund in talks to lead - one of the fund’s first bets as Brussels tries to keep late-stage champions from relocating (here)