Venture Geopolitics Issue 45

28 April 2026

Is AI heading toward a two-superpower carve-up, or toward something messier and more interesting?

If the future of AI is determined solely by two superpowers, power and monopoly rents will concentrate quickly. But this week’s headlines suggested something more interesting may be emerging: a messier, more plural world in which others still have room to compete.

In China, the most underplayed line in DeepSeek-V4’s release was that it validated fine-grained Expert Parallelism on Huawei Ascend NPUs with a 1.50–1.73x speedup - a technical way of saying advanced AI workloads are becoming more viable on domestic Chinese chips rather than relying entirely on Nvidia’s stack. The announcement landed just as Washington escalated accusations that Chinese actors were distilling US frontier models and engaging in industrial-scale AI theft. China’s strategy appears straightforward: absorb capability while lowering dependence.

In the United States, the week showed a different version of the same contest. Google prepared to deepen its backing of Anthropic. Cirrascale’s new air-gapped Gemini appliance offered governments, banks and regulated industries a way to run frontier AI on their own infrastructure rather than inside a public cloud. All while SpaceX/xAI secured the right to acquire Cursor for $60 billion, and Nvidia surged to an over $5 trillion market capitalisation as investors continued to reward the incumbent provider of the world’s AI compute backbone.

But beyond Washington and Beijing, a third camp may be taking shape. Ineffable Intelligence’s record UK seed round was a reminder that Britain continues to generate frontier talent and model-layer ambition. The Canadian-German tie-up between Cohere and Aleph Alpha - a roughly $20 billion combination backed by Schwarz Group’s €500 million lead investment - suggested scale may increasingly come through alliance rather than isolation. Middle-power coalitions need not outspend the giants if they can out-coordinate them.

The US and China may still steal the headlines, but they do not have to own the whole map. Put DeepSeek on Huawei chips next to Google’s sovereign deployment push and Cohere/Aleph Alpha’s transatlantic merger, and a different picture emerges: lower-cost models, trusted deployments and cross-border coalitions are creating more routes to relevance than a simple duopoly would allow. A more complex technological order - more ecosystem than empire - is a healthier one.

In an open ocean, only sharks are happy. On a coral reef, there is room for many more winners.

IPOs / Publics

• US IPO activity surged with $17.3B targeted this month as companies rush to debut before SpaceX’s record-breaking June IPO plan, with recent deals delivering 21% weighted returns vs S&P 500’s 4.2% (here)

• 5 x IPOs planned this week - 3 biotechs (Seaport, Avalyn, Hemab) targeting neuropsychiatric disorders, pulmonary fibrosis & coagulation disorders plan to raise $200M each; Silver Bow Mining (exploration-stage Montana miner) seeks $50M; Bill Ackman’s Pershing Square vehicles also set to debut - expected to raise $5B, down from $10B target (here & here)

• X-Energy (Amazon-backed small modular reactor developer) raised $1.0B at $9.5B in the first traditional IPO for advanced nuclear tech, pricing well above range & closing up 27%; critical materials supplier Elmet Group raised $120M at $420M, up 21%, highlighting non-China supply chains amid defence & aerospace demand (here & here)

• Nvidia hit $5.06T market cap as shares reached $208.27, breaking out of a $170-190 trading range held since summer 2025 on renewed optimism around enterprise AI adoption where it dominates chip supply (here)

• BoE deputy governor warned global stock markets are overvalued & set to fall, citing AI bubble risks, $2.5T private credit market vulnerabilities & macroeconomic shocks - unusual public market call from BoE leadership (here)

• Google Cloud unveiled TPU 8t & 8i chips designed for AI training & inference respectively, claiming 117-124% better performance per watt than predecessors whilst announcing a $750M fund for corporate AI adoption (here)

• Google’s Gemini now runs on fully air-gapped servers via Cirrascale partnership, targeting banks & governments unwilling to send sensitive data to public cloud APIs - the model self-destructs if tampered with & vanishes when unplugged, marking a shift from cloud-first to private AI deployment for regulated industries (here)

• Google Cloud CEO Thomas Kurian argues the company’s integrated AI strategy - owning chips, models & infrastructure - positions it to close ground on AWS & Microsoft Azure, as Google invests $185B in capex whilst predicting a market shakeout for loss-making AI startups dependent on private funding (here)

• Hundreds of Google employees petitioned CEO Sundar Pichai to block Pentagon access to Google AI for classified military projects, echoing concerns that led to Anthropic being dropped from Defence Department programmes 2 mo earlier (here)

• Meta plans to cut 10% of its workforce (8,000 employees) on 20 May & freeze 6,000 open roles to boost efficiency & offset heavy AI spending, with Zuckerberg’s aggressive push into large language models & infrastructure driving record capex projections (here)

• Meta signed a multi-billion dollar deal with Amazon to use Graviton5 CPUs for AI reasoning & code generation, adding to $115B-135B capex plans that include chip deals with Nvidia, AMD & Google as Meta builds AI infrastructure whilst developing internal chips (here)

• Meta is installing keystroke & mouse tracking software on US employee computers to capture work interactions for training AI agents designed to autonomously perform office tasks, as the company plans 10% workforce cuts starting May 20th - a surveillance approach that would likely violate European data protection laws (here)

• Oracle’s $300B OpenAI deal faces financing bottlenecks as banks hit exposure limits syndicating multibillion-dollar data centre construction loans, with JPMorgan & others struggling to distribute risk across institutions - highlighting capital constraints that could slow AI infrastructure build-out alongside power grid & public opposition challenges (here)

Big Dogs

• SpaceX secured the right to acquire AI coding start-up Cursor for $60B as Musk attempts to catch up with OpenAI & Anthropic ahead of SpaceX’s record $1.75T IPO - if SpaceX doesn’t exercise the option, it pays a $10B termination fee (here)

• OpenAI released GPT-5.5, which the company calls its smartest model yet & a step towards creating a unified ‘super app’ combining ChatGPT, Codex & AI browser - intensifying competition with Anthropic & Musk’s X ambitions (here)

• Microsoft dropped its exclusive rights to sell OpenAI’s models, allowing the ChatGPT maker to pursue deals with Amazon & other cloud rivals whilst ending revenue-sharing arrangements - a shift that simplifies their complex partnership as OpenAI seeks broader distribution (here)

• OpenAI is committing up to $1.5B to a joint venture with PE firms TPG, Bain Capital, Advent & others to deploy AI across portfolio companies, guaranteeing investors 17.5% annual returns whilst competing with Anthropic’s enterprise push (here)

• OpenAI reportedly developing smartphone with MediaTek & Qualcomm for 2028 launch, replacing traditional apps with AI agents to bypass iOS/Android restrictions & capture deeper user data from ChatGPT’s 1B weekly users (here)

• OpenAI missed internal user & revenue targets, prompting CFO Sarah Friar to question $600B in data centre spending commitments as the company eyes an IPO by year-end amid slowing ChatGPT growth & competition from Google’s Gemini (here)

• Google to invest up to $40B in Anthropic ($10B initial at $350B, remainder tied to performance targets) plus 5 gigawatt compute deal starting 2027, following Amazon’s $25B investment & 5GW commitment - highlighting Big Tech’s escalating AI infrastructure arms race (here)

• DeepSeek-V4 launched with near state-of-the-art AI performance at roughly 1/6th the cost of GPT-5.5 & Claude Opus 4.7, featuring a 1M token context window & open-source MIT licensing - while notably training on Huawei Ascend NPUs rather than Nvidia hardware, marking China’s push for semiconductor independence in frontier AI development (here)

• DeepSeek is raising its first external funding round targeting a $20B valuation, primarily to retain AI researchers being poached by rivals offering lucrative stock options - the Chinese AI startup has lost key talent to ByteDance & Tencent despite competitive cash compensation (here)

Venture Capital / Finance

• Thoma Bravo is nearing a deal to hand over software firm Medallia to creditors including Blackstone, KKR & Apollo, wiping out $5.1B in equity after the 2021 $6.4B buyout - highlighting private equity’s software sector stress as AI disruption & high debt burdens collide (here)

Regulation

• China warned the EU it will take “countermeasures” if its companies are hurt by the proposed Industrial Accelerator Act, which would restrict foreign investments over €100M from countries with >40% global production in strategic sectors like batteries & solar panels, requiring 50% EU workers & tech transfer to European partners (here)

• California’s proposed 5% billionaire wealth tax gathered 1.5M signatures for November’s ballot, prompting Silicon Valley exodus including Google cofounders Page & Brin plus Zuckerberg relocating primary residences to avoid the levy (here)

• Maine Governor Janet Mills vetoed what would have been America’s first statewide data centre moratorium, citing failure to exempt a project in struggling mill town Jay - highlighting tension between AI infrastructure growth & local economic needs as states grapple with regulating data centres amid Trump administration threats to sue states over ‘cumbersome regulation’ (here)

• Musk vs Altman trial begins this week, with Musk seeking $134B damages & demanding OpenAI revert to full nonprofit status whilst removing Altman & Brockman from leadership - posing existential threat to the $852B-valued company ahead of its planned mega-IPO (here)

• China blocked Meta’s $2B acquisition of Manus (AI agent), ordering termination despite Butterfly Effect’s attempt to distance itself by moving headquarters to Singapore after raising $75M from Benchmark (here)

• China told AI startups Moonshot & StepFun to reject US capital without government approval, part of broader restrictions triggered by Meta’s $2B acquisition of Chinese startup Manus - a move that risks isolating China’s tech sector from two decades of American venture backing (here)

• UK ministers resist EU alignment on AI regulation & lab-grown meat rules, fearing it could undermine Britain’s innovation edge & US tech partnerships - negotiations continue as Starmer pursues broader EU reset whilst seeking sector-specific opt-outs (here)

• FCA led first operation to disrupt illegal P2P crypto trading, targeting 8 London locations with HMRC & SWROCU - no FCA-registered P2P crypto traders currently operate legally in UK (here)

Venture Geopolitics

• China’s R&D spending has reached parity with (and by purchasing power measures surpassed) the US, with both nations crossing $1T in research investment - marking China’s ascent from among the world’s lowest research spenders in 1980 to leading in scientific publications, highly-cited papers & patent filings whilst US federal R&D spending has declined from 1.86% of GDP in 1964 to 0.66% in 2021 (here)

• Palantir employees express internal turmoil over the company’s role as technological backbone of Trump’s immigration enforcement, with WIRED obtaining internal Slack messages showing worker concerns about civil liberties implications & “descent into fascism” (here & here)

• Palantir’s £330M NHS contract shows mixed results 15 mo in - some hospitals report benefits like 1,000+ patient transfers between sites, but critics call it ‘expensive database’ delivering similar capabilities to existing systems at higher cost (here)

• Iran war has cost Europe €22B in energy premiums compared to pre-war levels, forcing European manufacturers into 20-year US LNG contracts at premium rates whilst China’s electrification pace (30-32% electricity share vs Europe’s 24-26%) leaves the continent increasingly dependent on American fossil fuel exports as a strategic lever (here)

• White House accuses China of ‘industrial-scale’ AI theft via distillation attacks using tens of thousands of proxy accounts to extract IP from US labs, as Trump administration prepares countermeasures ahead of Xi Jinping meeting (here)

• King Charles addresses US Congress to repair US-UK ‘special relationship’ strained by PM Starmer’s initial reluctance to support US-led Iran war operations, with Trump calling the partnership ‘not good at all’ (here)

• India’s $1.3B ‘sovereign AI’ initiative faces contradictions as infrastructure relies on Nvidia chips whilst US legal frameworks could still compel American cloud providers to access India-hosted data, highlighting undefined sovereignty standards in AI governance (here)

• UAE announces plan to run 50% of government operations on autonomous AI systems within 2 years, positioning itself as the first nation to implement agentic AI at this scale across public services, with Sheikh Mansour overseeing execution (here)

Strategic Sectors

AI

• Cohere (Canada) & Aleph Alpha (Germany) agreed $20B merger to create transatlantic ‘sovereign AI’ alternative to US Big Tech, backed by German & Canadian governments plus $600M from Schwarz Group (Lidl owner) - early consolidation signal as middle powers seek AI independence from US-China duopoly (here)

• Martin Lueck, co-founder of £9B hedge fund Aspect Capital & quant pioneer, warns against fully delegating trading decisions to AI, arguing firms need to understand why models make trades rather than operating ‘black boxes’ - a counterpoint to AQR’s Cliff Asness who admitted to ‘surrendering more to the machines’ (here)

• Atlassian, HubSpot & 79 of 500 tracked software firms shifted to usage-based AI pricing by end-2025, double the 2024 figure, as flat-fee models proved unsustainable amid rising compute costs & concerns AI agents will reduce seat-based subscriptions (here)

• Jeff Bezos’s AI lab Project Prometheus is in talks to lease 38,000 sq ft at London’s King’s Cross, following its $10B raise at $38B valuation - part of an AI office space surge that could reach 4M sq ft in London by 2033, up from 1.5M today (here)

• Ex-OpenAI VP Jerry Tworek’s new startup Core Automation has recruited top researchers from Anthropic & Google DeepMind, including Rohan Anil & Anmol Gulati, aiming to build automated AI research systems beyond current scaling paradigms (here)

• Notable deals:

  • Cognition (creator of AI coding assistant Devin) in talks to raise hundreds of millions at $25B valuation, more than doubling from its $10.2B September round as demand for AI-generated code accelerates (here)

  • Ineffable Intelligence, founded by ex-DeepMind researcher David Silver, raised $1.1B at $5.1B to develop reinforcement learning AI that discovers knowledge without human training data - joining London’s growing AI hub alongside similar ‘pentacorn’ ventures from star researchers (here & here)

Cybersecurity

• Anthropic’s Mythos vulnerability detection model leaked to unauthorised users via third-party vendor breach linked to Mercor supply chain attack, but security researchers suggest the hype exceeds reality - the model hasn’t found zero-days that human experts couldn’t discover & claimed ‘thousands’ of vulnerabilities may only be around 40 confirmed (here)

• Meanwhile, UK government in talks with Anthropic over expanding access to Claude Mythos, the powerful AI model that can rapidly detect cyber vulnerabilities, as British banks seek expedited access to strengthen defences following warnings from Jamie Dimon & German central bank chief (here)

• Per the FT, Anthropic’s Claude Mythos AI model is driving a surge in software patches after identifying vulnerabilities, with Microsoft alone rolling out 150 updates for one bank since release - prompting cybersecurity chiefs to warn of risks to critical infrastructure & call for coordinated government-business response as the technology proliferates beyond US guardrails (here)

• Survey of UK companies with £100M+ revenue finds 61% lack full understanding of how their data is handled when processed by AI systems overseas, with three-quarters reporting weekly cross-border data transfers - highlighting compliance risks as AI adoption accelerates without adequate governance frameworks (here

• QBE & Beazley are introducing sublimits on AI-related cyber claims, capping LLMjacking losses at ~10% of total policy limits as insurers rush to limit exposure to emerging AI risks including hallucinations & regulatory breaches (here)

• Health records from 500,000 UK Biobank participants were found for sale on Alibaba before being swiftly removed by Chinese authorities, prompting data access suspension & highlighting ongoing security gaps in Britain’s premier genomic research programme (here). UK Biobank CEO blamed ‘a few bad apples’ at three academic institutions for posting the de-identified medical data (here)

Defence

• US defence stocks fell 5-10% despite massive munitions expenditure in Iran conflict as production bottlenecks & uncertainty over Trump’s $1.5T budget proposal weigh on investor sentiment - classic ‘buy tension, sell war’ dynamic as capacity constraints limit near-term profit realisation (here)

• US Army soldier Gannon Ken Van Dyke charged with making $400K by betting on Polymarket using classified intelligence from the January operation that captured Venezuelan President Maduro, allegedly placing his final wager hours before the raid (here)

Energy

• Trump cancelled Iran peace talks in Pakistan, driving oil prices up 2% (Brent crude to $107/barrel) whilst global markets remained mixed as shipping disruptions through the Strait of Hormuz continue (here)

• UK government corrected its AI data centre emissions forecast by 136x after routine review, now projecting 34 MtCO₂ over the decade to 2035 vs previous 0.25 MtCO₂ estimate, raising questions about net-zero commitments amid data centre expansion (here)

• US electricity demand will hit consecutive record highs in 2026-27, driven by AI data centres & crypto mining, with consumption rising to 4,381B kWh by 2027 as renewables reach 27% of generation mix (here)

• Notable deals

  • X-energy raised $1B in IPO at $23 per share (above $16-19 range), driven by data centre electricity demand & deals with Dow for industrial heat & Amazon for 5GW nuclear power by 2039 (here)

Robotics

• Notable deals:

  • Physical Intelligence (robotics AI) in talks to raise ~$1B at $11B valuation, nearly doubling from its $5.6B November round led by Alphabet’s CapitalG - Founders Fund participating with Lightspeed, Thrive & Lux in discussions (here)

  • Sereact raised $110M Series B led by Headline to develop Cortex 2.0, an AI model enabling robots to predict consequences & avoid errors before they occur - already powering systems for BMW & Daimler Truck with plans to expand into US retail applications (here)

Quantum

• Goldman Sachs disbanded its quantum computing team after discovering algorithms would require millions of years & 8M qubits to solve portfolio optimisation, whilst JPMorgan maintains 50+ quantum researchers across trading, risk & cryptography - reflecting Wall Street’s split on whether the technology remains too early despite McKinsey forecasting quantum revenue to reach $72B by 2035 (here)

EVs

• BYD’s Q1 net profit dropped 55% to $600M as Chinese EV sales slowed despite phased-out subsidies, though the Iran war’s energy shock boosted global demand & European sales doubled to 74,000 vehicles (here)