Venture Geopolitics Issue 46
05 May 2026
Last week I wrote about the opportunity to out-coordinate rather than outspend. I did not have insider knowledge about SAP’s acquisition of Prior Labs (HQ’d in Freiburg), which was announced yesterday (with Prior remaining independent)! But it represents a perfect example of what that can look like in practice. It also reopens a question I keep coming back to - where in the AI stack does the value accrue?
There is no doubt that AI will create immense economic value, but the harder question is who captures it. The dominant assumption, reinforced by every fundraise and every gigawatt-scale compute deal, is that it will be captured by the frontier labs building the models.
But the model layer looks like something on a fast path to commoditisation, with rapid capability convergence (underscored again by Deepseek’s latest release), multiple credible providers, falling cost per token and customers who increasingly ‘mix and match’.
The value will accrue into the layers around the model where things are harder to replicate. Proprietary data, domain expertise, workflow integration, distribution, the customer relationships that make any of those things stick.
SAP has those things. It is not an AI-native company, but nor is it normal SaaS. It sits underneath the operations of much of the enterprise economy, and its systems are deeply embedded and highly customised, which makes it very difficult to leave. The problem this incumbent faces therefore is how it gets used in an AI-native world, what value gets eroded and what it can still capture.
Mercedes-Benz - a flagship SAP customer - reportedly cut its SAP seat count by around 40%, leaning on external AI tools to do work it once paid people to do inside the platform. Customers rarely move off SAP, they reduce usage, and SAP’s reaction was not to pretend the shift was not happening but to insist that whatever agent now sits on top of its systems, it will define the terms of access.
Prior Labs is the constructive half of the same instinct. Tabular foundation models are the unsexy cousins of LLMs, built to predict time series, fill gaps in structured datasets and categorise messy back-office data, which is the kind of work LLMs do not in fact do particularly well and also the kind of work that quietly drives a huge amount of economic output.
Prior Labs’ TabPFN-2.6 tops benchmarks in its category and has been downloaded >3m times. But left to itself, it is the sort of group that might have continued producing excellent papers in relative obscurity or been replicated by a better-capitalised (likely American) lab two years and several billion dollars later.
Inside SAP, with €1 billion committed over four years and Prior continuing to operate as a separate entity, it becomes something else, an applied capability deployed at scale into problems where SAP already owns the data, the customer and the workflow.
The two SAP announcements this week describe a posture rather than a transaction. Acquire uncommoditised capability before the market notices it, and control the agent layer over a platform too embedded for customers to leave. Both flow from the same recognition that the model layer is unlikely to be where the value finally settles, and the players best positioned to capture it are potentially those who own what is hardest to replicate.
The press has already framed the Prior Labs acquisition as the establishment of Europe’s frontier AI lab, but that framing flatters the labels and underplays the move. A frontier lab is a capability. What SAP is assembling is capability plus data plus workflow plus customer, the four ingredients value tends to settle into once a technology starts to commoditise, anchored inside the infrastructure the European economy already runs on. As I wrote a month ago, with specific reference to SAP, “Europe builds institutions”. This appears to be a move following that European hallmark - it isn’t making headlines (at least compared to Gamestops more vibes-driven $55B acquisition of ebay), but it is building deep and significant value.
IPOs / Publics
S&P Dow Jones Indices is considering accelerating mega-cap IPO inclusion from 12 to 6 mo & dropping profitability requirements for top 100 market cap companies, potentially benefiting SpaceX’s expected mid-June $75B raise at $1T+ valuation alongside future Anthropic & OpenAI listings (here)
Cerebras raised target IPO size to $4B at $40B valuation, up from initial $2B plan, with banks receiving $10B+ in early interest orders as semiconductor stocks surge 50% this year amid AI infrastructure investment boom (here)
Retail investors are driving massive premiums in closed-end funds offering exposure to SpaceX, OpenAI & other pre-IPO tech firms - Fundrise Innovation Fund hit a 3,000% premium to NAV in March before falling back to ~85 after short-seller criticism. The funds use complex structures like SPVs & forward contracts, prompting OpenAI to warn it may invalidate underlying shares acquired through such transactions (here)
SpaceX IPO filing reveals Musk can only be removed as CEO & chairman by his own consent through Class B super-voting shares he controls, giving him effective veto power over board removal attempts - an unusual governance structure even by dual-class standards (here)
IPO expert Jay Ritter questions SpaceX’s potential $2T valuation ahead of its historic public offering, citing scepticism over Starlink’s projected cost savings & margin expansion - noting IPOs with 40x+ price-to-sales ratios typically underperform markets over 3 years (here)
SoftBank is creating Roze AI, a robotics company focused on automating US data centre construction, & plans to IPO the business in the US this year at up to $100B valuation, incorporating ABB Robotics acquisition & data centre assets, as Son seeks to offset $30B+ OpenAI commitments whilst approaching leverage limits (here)
Google Cloud CEO Thomas Kurian says the company’s full-stack AI strategy - including proprietary TPU chips & DeepMind models - is helping it gain ground vs Amazon & Microsoft, with cloud revenue growing 48% to $70B annually. Google controls ~25% of global AI compute capacity & struck a $200B+ deal with Anthropic, while Kurian predicts a market shakeout as AI startups struggle with unsustainable economics (here)
AWS launched OpenAI products on Bedrock after Microsoft’s exclusive rights ended, including latest models, Codex, & new Bedrock Managed Agents service - marking deeper AWS-OpenAI collaboration as Microsoft-OpenAI relations deteriorate (here)
Microsoft CEO Satya Nadella said he plans to ‘fully exploit’ royalty-free access to OpenAI’s models through 2032, despite losing exclusivity as OpenAI partners with Amazon - Microsoft’s AI business hit $37B annual run rate, up 123% year-over-year (here)
Big Tech’s AI capex plans jumped to $725B this year as Google Cloud grew 63% (vs Amazon’s 22% & Microsoft’s 40%), helping Alphabet shares rise 6% to near $4.5T market cap whilst Meta fell 8% despite 33% revenue growth amid rising costs & vague AI timelines (here)
SAP published new policies threatening to block unauthorised AI agents like OpenClaw from accessing customer data, citing performance & IP protection concerns - a defensive move as customers like Mercedes-Benz reduce SAP usage by 40% whilst using external AI tools for data analysis (here)
SAP acquired Prior Labs (tabular foundation models) with €1B+ investment over 4 years to establish Europe’s leading frontier AI lab, targeting structured business data where LLMs struggle - Prior Labs’ TabPFN-2.6 tops benchmarks & has 3M+ downloads (here)
Atlassian shares jumped 25% after Q3 results showed 32% revenue growth to $1.8B, with CEO noting customers using its AI search tool Rovo generate twice the annual recurring revenue of non-users - a sign AI integration may be driving product uptake rather than threatening traditional software providers (here)
Big Dogs
OpenAI expanded its Amazon partnership to distribute AI models via AWS Bedrock after renegotiating Microsoft exclusivity terms, with Amazon having invested $15B of a committed $50B & securing $138B in data centre capacity deals (here)
OpenAI projects 122M subscribers this year driven by $8 ChatGPT Go tier, expecting 80% of $20 Plus users to downgrade as it shifts from subscription to advertising revenue (targeting $102B ads by 2030, 36% of total revenue) (here)
OpenAI has abandoned its original $500B Stargate joint venture model in favour of bilateral deals for computing capacity, halting data centre projects in the UK & Norway while securing over 8GW through partners like Oracle ($300B, 5-year deal) - now targeting $600B spending by 2030 despite missing revenue targets (here)
OpenAI dismissed reports of missed internal targets as ‘clickbait’, insisting consumer & enterprise businesses are ‘firing on all cylinders’ despite growing investor concerns over AI infrastructure spending & computing costs outpacing revenue growth (here)
Musk testified he was a ‘fool’ to provide $38M to OpenAI as a nonprofit, accusing Altman & Brockman of manipulation before converting to for-profit status now valued at $800B - seeks $180B damages & leadership removal (here). He also testified that xAI used distillation techniques to train Grok on OpenAI models, confirming widespread industry suspicions about frontier labs copying each other’s work - whilst OpenAI & Anthropic simultaneously campaign against Chinese firms using identical methods (here)
Anthropic collecting investor allocations within 48 hours for $50B round at potential $900B+ valuation, closing within 2 weeks - would eclipse OpenAI’s $852B valuation & likely be final private round before IPO, as revenue run rate hits $40B vs $30B disclosed (here)
Anthropic formed a $1.5B joint venture with Blackstone, Goldman Sachs & others to deploy its AI across their investment portfolios, with Blackstone & H&F investing $300M each whilst Goldman & General Atlantic commit $150M apiece - a revenue diversification play ahead of Anthropic’s expected IPO this year (here)
Anthropic in talks to buy AI chips from UK startup Fractile as part of broader effort to reduce Nvidia dependence & lower inference costs, with Fractile raising 100M+ at 1B+ valuation partly on strength of potential Anthropic deal (here)
Venture Capital
Startup bankruptcies exceeded exits for the first time in 2025, with insolvencies representing 51% of US portfolio exits & 54% in Europe - nearly 4,000 bankruptcies versus 3,600 exits as high 2020-21 valuations collide with ongoing liquidity constraints (here)
UK’s Pension Schemes Bill gained Royal Assent with mandation watered down to 10% cap (5% for UK assets), after Lords rejected full compulsion of pension funds into specific asset classes - VC CEO Duncan Johnson argues self-regulation via Mansion House Accord (17 pension providers pledging 10% to private markets by 2030) beats mandation’s blunt instrument approach (here)
Coatue launched Next Frontier to acquire land for AI data centres, with tens of billions expected to be deployed across projects targeting customers like Anthropic - the $70B investment firm is already developing a 430MW complex in Indiana through a JV that raised $5.7B in bonds (here)
Founders Fund raised $6B for its fourth growth-stage fund just 10 mo after closing a $4.6B predecessor, with $4.5B from LPs & $1.5B from employees including Peter Thiel (here)
137 Ventures raised $700M across two growth-stage funds, deploying $1B in the past year into defence, AI & industrial companies including SpaceX, Anduril & Hadrian - positioning ahead of SpaceX’s expected $1T+ IPO (here)
Earlybird VC closed €360M Fund VIII, its largest fund ever, focusing on deeptech & AI infrastructure with early investments in Black Forest Labs, Aleph Alpha & others (here)
BMW i Ventures raised $300M for its third fund targeting agentic AI & physical AI startups across North America & Europe, bringing total assets under management to $1.1B as the German automaker’s VC arm positions for AI-driven manufacturing transformation (here)
Ridgeline raised $180M through the SBA’s Critical Technologies Initiative, leveraging 1.25x debt against $81M private capital to back dual-use tech companies (here)
Kompas VC raised €160M for its second fund to back early-stage startups in manufacturing, supply chains & sustainability across Europe’s fragmented markets (here)
Illuminate Financial closed its $135M Early Growth Fund backed by BNP Paribas, Citi, Deutsche Börse, HSBC & others, targeting Series B+ AI & fintech companies serving financial services (here)
Regulation
EU moves closer to fining Meta up to 6% of global turnover for failing to prevent under-13s accessing Instagram & Facebook, escalating transatlantic tech tensions as Trump administration calls DSA enforcement attacks on American platforms (here)
UK tech ministers worry EU regulatory alignment could ‘smother’ British AI innovation & strain US ties, seeking opt-outs from Brussels’ stricter AI Act while maintaining laissez-faire approach that has attracted investment (here)
US District Judge Yvonne Gonzalez Rogers told Elon Musk & OpenAI leadership to stop social media commentary during their high-profile trial, after Musk called CEO Sam Altman ‘Scam Altman’ & amplified critical coverage on X - all parties agreed to a ‘clean slate’ (here)
Venture Geopolitics
New research warns Europe’s heavy reliance on Chinese green tech (90% of solar modules, 80% of wind turbines & batteries) creates national security risks including cyber attacks, supply disruptions & potential US pressure to remove Chinese components - echoing UK’s rejection of Chinese wind factory & National Grid component removal over cyber concerns (here)
US warns European allies including UK, Poland, Lithuania & Estonia to expect serious delays for missile systems as Iran war depletes American stockpiles, forcing Pentagon to reallocate weapons from Indo-Pacific amid concerns about deterring China (here)
Goldman Sachs stopped Hong Kong bankers from using Anthropic’s Claude AI models following strict contract interpretation, whilst maintaining access to other AI vendors like OpenAI - highlighting how US-China tensions are creating operational complexities for global financial services (here)
UK Technology Secretary Liz Kendall announced a new AI hardware plan to secure domestic semiconductor capability & supply chains, calling AI the ‘defining currency of the age’ & warning Britain risks falling behind without tech sovereignty (here)
Google agreed to let the Pentagon use its AI tools in classified settings, joining OpenAI & xAI after Anthropic refused over ethical concerns about mass surveillance & autonomous weapons - despite 600+ Google employees signing a letter opposing the deal (here)
Russia now imports >90% of its sanctioned military technology through China, up from 80% last year, despite EU sanctions targeting semiconductors, electronics & machinery needed for weapons production. Most EU countries avoid harsher penalties on China, fearing Beijing’s retaliation, while China maintains it doesn’t recognise international sanctions (here)
White House state dinner for King Charles III brought together tech leaders including Tim Cook, Jeff Bezos, Marc Andreessen & Marc Benioff alongside Trump administration officials, highlighting Silicon Valley’s continued access to power despite political shifts (here)
Pentagon signed classified AI deals with seven companies (xAI, OpenAI, Google, Nvidia, Microsoft, AWS, Reflection) for ‘any lawful operational use’ on secret networks, after Anthropic refused broad licensing & was subsequently banned as a supply chain threat - now suing for political retaliation (here)
Google CEO Sundar Pichai met Trump administration officials ostensibly about cybersecurity but primarily over worries about AI compute capacity for government defence systems, amid concerns that Anthropic’s restricted Claude Mythos model could throttle high-priority government access during crises (here)
Build American AI, a nonprofit linked to OpenAI & Andreessen Horowitz executives, is funding influencer campaigns to promote US AI development whilst stoking fears about Chinese AI capabilities, highlighting how tech giants are deploying dark-money tactics to shape public opinion on AI geopolitics (here)
UK defence officials warn the state pension triple lock could face suspension if Britain enters direct conflict with Russia. No main UK political party, except the Green Party, has made it official policy to axe the triple lock as it is seen as a politically toxic vote-loser, but the Office for Budget Responsibility last year forecast that the extra cost of the triple lock to state pension spending will be around £15.5bn by 2030. (here)
Banksy installed a new statue overnight in central London depicting a suited man blindfolded by a flag walking off a ledge, placed among historical monuments in Waterloo Place. London authorities plan to keep the provocative artwork in place despite growing crowds (here)
Notable deal:
Panthalassa raised $140M from Peter Thiel & others at nearly $1B valuation for wave-powered floating data centres, deploying 85m steel nodes that generate electricity from ocean waves to power AI chips via self-propelled vessels operating in international waters. Thiel, a co-founder of Palantir & PayPal, said: “The future demands more compute than we can imagine. Extraterrestrial solutions are no longer science fiction. Panthalassa has opened the ocean frontier.” (here)
Strategic Sectors
AI
DeepSeek released V4-Pro (1.6T parameters) & V4-Flash (284B parameters), both with 1M token context & MIT licensing, pricing 80-90% below frontier models whilst claiming performance within 3-6 mo of GPT-5.4 & Gemini 3.1-Pro - a reminder that Chinese labs continue advancing open-weight capabilities at aggressive price points (here)
Poolside launched two coding-focused AI models - flagship Laguna M.1 (225B parameters) & open-source Laguna XS.2 (33B parameters) - with the smaller model achieving 44.5% on SWE-bench Pro despite running locally on consumer hardware, positioning US-trained open weights against Chinese alternatives like DeepSeek (here)
NVIDIA B200 GPU spot prices surged 114% to $4.95/hour in 6 weeks, driven by frontier AI model releases like GPT-5.5 requiring advanced memory capacity - signalling inflationary demand outpacing chip improvements & creating pricing power for cloud providers (here)
OpenAI revealed GPT-5.5 developed an obsession with ‘goblin’ & fantasy creature metaphors during training - a byproduct of their personality selection feature (specifically “nerd” mode), which then required explicit system prompts to suppress the behaviour - highlighting how reinforcement learning rewards can create unintended biases that leak across model functions (here)
Cloudflare launched agent-first infrastructure provisioning via Stripe Projects, allowing AI agents to autonomously create accounts, purchase domains, deploy code & handle billing without human intervention - a protocol shift that could accelerate agent-driven development workflows (here)
Zuckerberg & Chan’s Biohub commits $500M over 5 years to build AI models of human cells, targeting datasets 10x larger than current billion-cell scope with partners including Nvidia & Allen Institute (here)
Academic study finds one-third of websites created since 2022 are AI-generated, making the internet more cheerful but less semantically diverse - contrary to expectations, AI content isn’t spreading more falsehoods or failing to cite sources (here)
Spotify launched ‘Verified by Spotify’ badges to distinguish human artists from AI-generated content, requiring sustained engagement, platform compliance & off-platform presence like concerts - response to synthetic tracks comprising 44% of daily uploads on rival Deezer (here)
Nebius agreed to acquire Eigen AI for $643M in cash & shares to strengthen its Token Factory inference platform, combining Eigen’s optimization stack with Nebius’s global GPU capacity - the MIT HAN Lab team will establish Nebius’s Bay Area presence as inference becomes two-thirds of AI compute demand (here)
Cybersecurity
White House blocked Anthropic’s proposal to expand Mythos AI model access from ~50 to ~120 entities over security concerns & computing power constraints, highlighting ongoing tensions despite recent reconciliation efforts following Pentagon disputes (here)
CrowdStrike & Palo Alto Networks shares up 20% this month as Anthropic’s Mythos AI model identifies thousands of new cybersecurity vulnerabilities, positioning incumbent security firms to benefit from expected surge in enterprise cyber spending after 2 years of sluggish growth (here)
NSA is using Anthropic’s Mythos model to find security vulnerabilities in Microsoft software, whilst Microsoft deploys the same AI tool to identify & patch flaws faster - highlighting how AI cyber capabilities are being shared selectively with governments & select companies to stay ahead of threat actors (here)
Defence
Defense Secretary Pete Hegseth defended the Iran war costing $25B so far, attacked Democratic critics as the US’s biggest adversary, & falsely inflated Ukraine aid figures during contentious congressional testimony on Operation Epic Fury (here)
Founders Fund increased its stake in Anduril as the defence startup expects revenue to double to $4.3B in 2026, though operating losses will hit $1.2B & won’t turn profitable until 2030 - reflecting the fund’s strategy of concentrated bets on perceived winners whilst Anduril burns venture capital to compete with traditional primes like Northrop Grumman (here)
Space
Starcloud seeks funding at $2.2B valuation for space-based data centres, having launched its first demo satellite (single Nvidia H100 GPU) in November & planning 88,000-satellite constellation - facing heavyweight competition from SpaceX’s 1M satellite plans & Blue Origin’s 50,000-satellite ambitions despite only 25,000 satellites ever launched in human history (here)
SpaceX plans to file for IPO this week in what would be the largest US public offering in years, pitching space launch, Starlink internet & orbital data centres whilst allocating 20%+ of shares to individual investors rather than the typical 10% - though recent $250B xAI acquisition may depress initial financials (here)
Energy
Ming Yang Smart Energy accused the UK of ‘politicisation’ after being banned from offshore wind projects over national security concerns, despite a planned £1.5B Scottish investment. The world’s 4th-largest turbine maker said it would continue European expansion plans, eyeing Spain & other markets as Chinese wind tech faces growing geopolitical resistance (here)
Belgium in talks to nationalise nuclear assets from Engie, suspending reactor dismantling plans as the country reconsiders extending reactor lifespans beyond 2025 amid Europe’s broader nuclear renaissance following energy shocks (here)
Notable deal:
CMBlu Energy (Germany) raised €50M at €1B as Europe’s latest unicorn, with Samsung Ventures backing its non-lithium long-duration storage tech targeting data centres & utilities - highlighting European energy storage competitiveness amid supply chain reshoring (here)
Crypto
Robinhood shares fell 6% after crypto trading revenue dropped 47% year-over-year, though total revenue rose 15% to $1.07B driven by record prediction market volumes of 8.8B event contracts; the firm plans to launch its own prediction market exchange this year via joint venture with Susquehanna (here)
Polymarket partnered with Chainalysis to deploy insider trading detection tools following high-profile cases including a US soldier’s $400K bet on Maduro’s capture & Israeli suspects using classified intel, as prediction markets face regulatory scrutiny despite Trump’s recent criticism of the sector (here)
EVs / AVs
Uber plans to equip millions of driver vehicles with sensors to create a massive data collection grid for autonomous vehicle companies, expanding beyond its current AV Labs fleet to become the data infrastructure layer for the entire self-driving industry (here)
BYD reported Q1 net profit of $600M, down 55% year-on-year, as domestic Chinese EV sales slowed despite government subsidies phasing out - though export strength helped offset domestic weakness with European sales more than doubling to 74,000 vehicles (here)
Robotics
Meta acquired Assured Robot Intelligence (ARI) for undisclosed terms, bringing co-founders Xiaolong Wang (ex-Nvidia) & Lerrel Pinto (ex-NYU) to its Superintelligence Labs to advance humanoid robotics foundation models for physical labour tasks (here)
Notable deal: ZaiNar raised funds at $1B valuation from Steve Jurvetson & Jerry Yang after 9 years developing 5G-based location tracking accurate to 4 inches indoors/outdoors, securing $500M in contracts & negotiating $5B more as robotics & physical AI drive demand for precise positioning beyond GPS capabilities (here)