Venture Geopolitics Issue 57
20 July 2026
IBM built a massive and successful company on great execution. A deep understanding of its customers and relentless delivery built a pre-eminent brand, while its systems became so deeply embedded that leaving became almost unthinkable: “No one gets fired for buying IBM”.
That execution has faltered - over recent months by the CEO’s own admission, but over years if you include the long and painful retreat from Watson. The brand is now one the world associates with an earlier technological era, yet switching costs kept customers locked in. That moat - “Switching Costs” in Helmer’s 7 Powers - has served IBM well, but is now under pressure. Earlier this year, Anthropic suggested Claude Code could modernise the COBOL running on IBM mainframes, and last week the stock fell 25%.
The interesting part is not just IBMs predicament, but the fact that an entire class of technology moat - switching costs - has been “demoted” (ha!) by AI.
At the same time, another moat is being “promoted” (doh.). Ben Thompson, whose Aggregation Theory defined a generation of technology thinking, used Monday’s Stratechery to explain why its assumptions apply less cleanly in the AI era. His theory described a world in which software and distribution had near-zero marginal costs: products could be reproduced and delivered at almost no additional expense, creating extraordinary returns to scale and allowing power to accrue to whoever controlled demand.
“Scale” - another of Helmer’s 7 Powers - still matters, but token costs mean it is no longer absolute. In contrast to traditional SaaS, AI products carry real COGS: compute, chips, electricity. As Thompson puts it, “marginal costs are back in a big way”.
If you haven’t read Aggregation Theory, you should. As those marginal costs return, power is beginning to shift towards those who dominate supply-side efficiency matters over pure demand – cheaper electricity, better chips, models that spend fewer tokens reaching the same answer. This is the promoted moat: Helmer’s “Process Power” moat - deeply intrinsic business culture and processes that allow a company to operate with lower unit costs.
Ultimately, the relative importance of different moats is being rebalanced. Switching costs are weakening; culture and processes are strengthening.
For you and me at the junction of Venture and Geopolitics, this is one of the most significant shifts we have seen in our industry. The fundamental inputs shaping AI economics - energy, raw materials, labour and trade - are also what constitute Geopolitics, and returns to scale define Venture Capital. I expect we are not done exploring the impacts of this change!
IPOs / Publics
Shein secured approval from Hong Kong’s listing committee for an IPO, targeting $40-50B valuation (down from $100B in 2022) with public filing expected week of 27 July & potential late-August debut (here)
Data center operator Csquare (NYSE: CSQR) priced below the range at a steep discount to peers and fell -2% (here) while early-stage nuclear fuel maker Standard Nuclear (NYSE: STDN) slashed its offering by more than half, then plunged -18% (here)
SpaceX shares fell below $125, down $1T from their June peak & below the $135 IPO price, after an aborted Starship launch & ahead of employee lockup expiry in August (here)
IBM’s stock experienced the worst day in its history (the company has been a public stock for 115 years). Mainframe sales & the software that runs on them are faltering; management blamed AI spend, but per Ben Thompson, the real concern for IBM is that AI’s ability to port the essential backend programs that run on archaic technology will mean those missed sales never come back (here)
Apple briefly overtook Nvidia as the world’s most valuable company, closing Friday at $4.902T vs Nvidia’s $4.908T, as investors view it as a safer AI play that avoids hyperscaler capex risk (here)
Stripe & Advent International offered to acquire PayPal for $60.50/share ($53B+), backed by $50B in committed financing - a 28% premium to market but still well below PayPal’s 2021 peak of $360B. The deal would combine the two largest merchant payment platforms (~$3.7T annual volume), giving Stripe direct consumer relationships via PayPal’s 430M accounts & Venmo, plus scale to bypass Visa/Mastercard fees & bolster stablecoin ambitions via Bridge (here)
UK government has convened private equity firms including Hg Capital, CVC, EQT & Elliott to encourage London IPOs of portfolio companies as only 7 floats have occurred in 2025, with takeover value outstripping new listings 27:1 (here)
London Stock Exchange will launch LSE 24, a 24/5 trading venue operating 5pm-7:50am, starting with ETPs in H1 2027 - the platform will support algorithmic & agentic trading, integrate with LSEG's Digital Securities Depository for digitised settlement, & expand into equities, positioning London for extended-hours liquidity as trading becomes more automated & globally connected (here)
Big Dogs
Microsoft CEO Nadella criticised Anthropic’s Fable model for being ‘editorially controlled’, questioning why a creation tool would block user requests so aggressively - a public rebuke of a partner in which Microsoft invested $5B & which agreed to spend $30B on Azure. Nadella also told engineers that economic dependence on a few token-capital providers ‘makes no sense’, signalling Microsoft’s push toward cost-efficient custom models that keep client data in-house (here)
Anthropic is in talks to buy $10B of compute from Meta, following its $1.25B/mo deal with SpaceX for Nvidia GPUs at Colossus 1. The move reflects acute capacity constraints across frontier AI labs - Anthropic has imposed usage limits despite a near-$1T valuation, whilst Google pays SpaceX $920M/mo & has capped Gemini access to Meta. For Meta, selling compute to a direct rival marks a strategic pivot into cloud services, helping justify capex & recent headcount cuts to shareholders (here)
Anthropic has 32 open roles focused on preventing AI misuse for radiological, nuclear, chemical, explosive & cybercrime harms, paying mid- to upper-$200K & requiring domain expertise to stress-test models before deployment. OpenAI is hiring similar roles at $295K-$445K as labs bolster safety teams amid absent regulation, with Anthropic breaking with the Pentagon this year over surveillance & autonomous weapon concerns (here)
Anthropic is scheduling investor meetings ahead of a potential October IPO at its $965B valuation, moving ahead of OpenAI (now targeting 2027) & China’s DeepSeek in the race to public markets (here)
Apple sent legal letters to ~40 former employees now at OpenAI, demanding document preservation & meetings as part of its trade secrets lawsuit alleging OpenAI stole secret hardware plans (here)
OpenAI said it is ‘not aware of any evidence’ supporting Apple’s lawsuit alleging trade-secret theft tied to OpenAI’s device ambitions (here)
Regulation
Burnham will scrap Labour’s digital ID scheme, redirecting the £600M annual cost to cost-of-living support (here)
Demis Hassabis published a framework proposing a US-led Frontier AI Standards Body modelled on FINRA, funded largely by industry, to assess AGI-class models before deployment - labs meeting capability thresholds would submit models for voluntary (then mandatory) 30-day reviews covering cyber, bio & agentic risks, with quarterly benchmark updates. He frames AGI as 2-3 years away, comparable to electricity or fire, requiring a technical testing regime that can adapt fast enough to match acceleration while incentivising responsible behaviour & forming the basis for international coordination (here)
Treasury Secretary Scott Bessent said the US will examine Chinese open-source AI models for IP theft, warning that watermarks from American LLMs have been found in Chinese models & sanctions may follow. The administration is also considering requiring companies to disclose use of Chinese AI models, following concerns that distillation techniques allowed rivals like Moonshot AI's Kimi K3 to develop competitive models at lower cost (here)
Venture Capital
ICONIQ survey of 300+ software executives shows AI products now comprise 42% of revenue (up from 32% in 2025, projected 53% in 2027), with gross margins expanding from 45% to 53% as consumption-based & outcome-based pricing models gain ground. Vertical applications (43% of builders) dominate, especially in financial services & healthcare; Anthropic adoption jumped from 51% to 81% in 6 mo as multi-model strategies (avg 3.3 models) became default. High-growth firms run flatter orgs, deploy AI tools 1 mo faster, write 59% of code with AI assistance (vs 47% peers), & see 48% productivity gains vs 32% - a compounding structural advantage in the Builder’s Economy (here)
Venture Geopolitics
Andy Burnham became Labour leader & was sworn in as UK prime minister on Monday, promising greater public control of essential services, devolution to local government & a rejection of 1980s privatisation policies (here)
The Trump administration has repeatedly considered measures to restrict Chinese AI models - including Entity List additions, security advisories & executive orders requiring liability guarantees - that would effectively lock in OpenAI & Anthropic's duopoly. Pro-innovation voices killed earlier attempts, but with personnel shifts & Kimi's recent rise, national security hawks are gaining ground. The approach now favours procurement rules, public pressure & security warnings rather than outright bans, though the effect on U.S. companies using cheaper Chinese models would be similar (here)
Xi Jinping outlined China’s ambition to rival the US as a global AI leader at a Shanghai conference, announcing 5,000 training opportunities for developing countries over 5 years, signalling Beijing’s push to position China as a governance convenor whilst Western export controls tighten. Nearly 30 countries including Russia, Brazil & Indonesia joined China’s new World Artificial Intelligence Cooperation Organization, giving Beijing influence over international AI standards as Chinese models (including Moonshot’s latest release) narrow the gap with US labs & attract cost-conscious western customers like DoorDash, Siemens & Airbnb (here). Jinping also called for “open source, open collaboration” (here)
Israel is spending $50M on AI-generated texts, conservative media payments & Trump-linked advisers to shift US public opinion after favourability dropped to 40% (Pew, March) amid wars in Gaza & Iran - a state-funded influence campaign that raises questions about digital sovereignty & foreign information operations targeting domestic audiences (here)
China’s GDP grew 4.3% year-on-year in Q2 2026, the slowest in 3 years & below the 4.5-5% target, as a property crisis & weak consumer spending offset AI-driven export strength (June exports surged 27%, led by chips, batteries & cars). The economy split between manufacturers supplying global AI demand & 14M construction workers who lost jobs since the property crash, with retail sales choppy & households focused on saving amid higher fuel prices from the Iran war (here)
German Chancellor Friedrich Merz is pitching Germany as “Europe’s bedrock of stability” ahead of an October investment summit in Berlin, modelled on Macron’s “Choose France” events - but investors meeting investment tsar Martin Blessing have asked for concrete evidence of reform impact & clearer industrial strategy. Germany leads Europe in patents but struggles to commercialise innovation, lacks deep capital markets, & faces slower data centre buildout vs France (which secured €93B pledges in June, mostly from SoftBank) due to multi-tiered approvals & higher energy costs after abandoning nuclear (here)
Jingye (China’s owner of British Steel since 2020) has threatened international arbitration after the UK nationalised British Steel last week, calling it “blatant extortion” & demanding compensation. The UK considers the commercial value nil; China’s foreign ministry warned the dispute will affect perceptions of the UK investment climate (here)
EU unity on Russia sanctions is fracturing as Greece blocks measures to protect Prokopiou’s Dynagas (which has shipped $24B of Russian LNG since 2022), whilst Germany & Portugal demand fish-import carve-outs & Austria seeks €2B unfrozen for Raiffeisen. Diplomats warn that national capitals - seeing Russia as less of a direct threat & unwilling to tolerate corporate costs - are prioritising commercial interests over solidarity (here)
US to impose 50% tariffs on wide range of Canadian goods (excluding energy, potash, critical minerals, steel & aluminium) in 30 days, citing discriminatory trade practices by Canadian provinces; move applies even to goods compliant with USMCA, marking escalation after earlier Supreme Court ruling curbed Trump's global emergency tariffs. Canada-US trade totalled $716B in 2025 (here)
China is considering export controls on AI model weights & training data, targeting Alibaba, ByteDance & Zhipu downloads to foreign users, alongside restrictions on overseas chipmakers fabricating Chinese-designed advanced semiconductors. The move follows Moonshot's Kimi K3 exceeding Anthropic's Opus 4.8 on most benchmarks & aims to protect China's narrowing gap with US frontier AI, though domestic firms warn tighter rules could slow development (here)
Strategic Sectors
AI
Moonshot launched Kimi K3, China’s largest AI model at 2.8T parameters, matching or beating Anthropic’s Claude Opus 4.8 & OpenAI’s GPT 5.5 on most coding & agent benchmarks while being available as an open-weight model at a third of Anthropic’s cost - raising at $31.5B (vs Anthropic’s $965B) as the US-China performance gap narrows faster than the consensus 8-12 mo timeline (here)
Kimi K3 matches Claude Opus on coding tasks at 70% lower API cost ($3/$15 per M tokens vs $10/$50) & offers uncapped $39/mo plans while Anthropic quietly downgrades its $20 tier when economics fail (here)
…Kimi paused new subscriptions for its K3 model after 48 hrs of demand hit GPU capacity limits, splitting membership into general & code-focused tiers to manage compute allocation - a tactical response to unexpected adoption in China’s LLM market (here)
SpaceX is in talks to provide the Pentagon with billions of dollars’ worth of data-centre capacity for running AI models, deepening Musk’s ties with the US military as defence agencies race to deploy AI at scale (here)
Kalshi launched a forward curve for AI compute pricing, plotting expected GPU rental costs up to 1 yr ahead using prediction markets, as CME & ICE also race to build derivatives markets around the infrastructure underpinning the AI build-out (here)
Local opposition to data centres is intensifying across the US, with residents in Oklahoma & other states protesting construction on concerns over water use, community impact, & lack of transparency - a rare bipartisan flash point as hyperscalers expand infrastructure for AI workloads (here)
ONS data shows UK businesses using AI more widely (35% of firms, up from 12% in 2023) but not more deeply - average tools per firm rose only 1.4 to 1.6, just 10% use AI extensively, & 60% focus on cost-cutting rather than new products or markets. Job impact remains limited outside creative & admin roles, raising questions about productivity optimism (here)
Thinking Machines (founded by former OpenAI CTO Mira Murati) released Inkling, a 975B-parameter open-weight multimodal model trained on 45T tokens, positioning it as a customisable alternative to closed frontier models (here)
Notable deals:
Databricks raising $3B at $188B (led by Coatue), following a $5B round earlier this year, as data analytics & AI application infrastructure attracts continued late-stage capital ahead of widely anticipated IPO (here)
CuspAI (AI for materials discovery) raised $450M at $2.6B co-led by Kleiner Perkins & NEA, with participation from Bezos Expeditions & the UK’s new Sovereign AI Fund. The Cambridge startup - advised by LeCun & Hinton - is working with ASML, Meta, Nvidia & Samsung on materials for semiconductors, carbon capture & cleantech via its AI Materials Foundry network (here)
OpenEvidence (AI medical chatbot) considering $200M round at $20B but unlikely to proceed, having doubled revenue to ~$300M annualised in 7 mo via pharma ads at 90% gross margin; also in acquisition talks with large tech co (here
Cybersecurity
UK AI Security Institute finds Chinese open-weight models now trail US frontier systems by just 4 mo on cyber capabilities, down from 6-10 mo in 2025. Beijing-based Z.ai’s GLM-5.2 performed comparably to Anthropic’s Opus 4.6 & OpenAI’s GPT-5.2-Codex released 4 mo earlier, while running far cheaper. AISI warns the narrowing gap threatens to compress patching windows globally, as open models cannot be withdrawn or access-controlled once released - a risk amplified as models approach the autonomous vulnerability-chaining capability Anthropic saw in Mythos (here)
Apple warned iPhone & iPad users of a social engineering campaign using FaceTime calls, with scammers impersonating trusted organisations via caller ID spoofing to harvest credentials & financial data - a reminder that video calls now carry enough perceived legitimacy to become a scalable fraud vector (here)
Cursor (AI code editor) disclosed a critical zero-day on Windows - opening any repository with a file named git.exe in the root triggers automatic execution with full user privileges, no warning or interaction required. The flaw creates a supply chain risk: attackers can embed malicious executables in repositories that run silently when developers open projects, granting access to SSH keys, credentials & source code. No patch yet released (here)
Notable deal:
Neo (agentic software security) raised $100M from Andreessen Horowitz & Bessemer to secure AI agents & AI-enabled applications across enterprises (here)
Energy
Andy Burnham will maintain Labour’s ban on new North Sea exploration licences despite industry pressure (here)
Nscale’s £2B Essex AI data centre, backed by Nvidia with Microsoft as anchor customer, is exploring alternative power after National Grid delays threatened timelines despite securing 90MW connection approval 6 days prior - underscoring UK grid bottlenecks as AI demand collides with Labour’s clean power goals. LCP Delta forecasts Britain will miss its 2030 clean electricity target (reaching 83% vs 95% goal) unless grid upgrades accelerate, while Capgemini reports 68% of electricity executives expect AI-driven power shortages. Over 100 UK projects now considering on-site gas generation to bypass grid queues stretching close to a decade (here)
Notable deal:
Valar Atomics (small modular reactors) is in talks to raise $1B at $5B pre-money from Sequoia, weeks after achieving criticality at its Utah test site. The 3-year-old startup - founded by 27-year-old Isaiah Taylor - aims to mass-produce reactors for data centres & industrial facilities, signing Nvidia as an early customer for a pilot nuclear-powered data centre (here)
Defence
France & Germany pledged to develop a European alternative to Palantir’s military software, citing France’s Arcadia command-and-control platform & comparable German solutions, after both countries recently dropped Palantir contracts in favour of European providers (here)
Ukraine’s ousted defence minister Mykhailo Fedorov, 35, publicly attacked General Oleksandr Syrskyi, accusing him of blocking tech-driven initiatives & sabotaging reforms, exposing a rift between Ukraine’s tech-focused reformers & Soviet-trained military leadership 4+ years into the war. Fedorov, credited with scaling drone production & deep-strike campaigns into Russia, said Syrskyi presided over dysfunction & intrigue rather than asymmetric warfare strategy - prompting spontaneous protests in Kyiv & resignations by air force officials (here)
UK Ministry of Defence awarded a £2B, 15-year contract to Omnia Training (Raytheon UK, Capita, Cervus, Rheinmetall UK & Skyral) to deliver an AI-enabled training system for the British Army (here)
UK MoD awarded £3.16M to three SMEs (Frankenberg Technologies, Greenjets, Cambridge Aerospace) to develop low-cost anti-drone interceptors under LCADE, a five-nation European programme responding to mass drone threats - Russia launched over 200 drones/day into Ukraine in March 2026. UK is first of five partners (UK, Poland, France, Italy, Germany) to award contracts, with focus on affordable solutions that can be manufactured at scale across allied nations (here)
Nokia & NestAI are deploying 5G, AI & sensing to maintain battlefield connectivity, addressing the military requirement for resilient comms in contested environments. The partnership reflects commercial telco hardware being adapted for tactical networks, a pattern accelerating across NATO as operators target dual-use scale (here)
Lockheed Martin announced the PAC-3 ACE, a new Patriot interceptor costing under $2M (less than half its current $4M PAC-3 MSE), to be co-produced with US & European partners. The move reflects pressure from nimble startups like Ukraine’s Fire Point, which tested a low-cost anti-missile system in June, & European demand for indigenous air defence as the US pulls back under Trump (here)
Crypto / Stablecoins
Citadel Securities has now committed $600M across two rival crypto exchanges - $400M into Crypto.com (July 2026) & $200M into Kraken (Nov 2025), both at $20B valuations - as each pursues tokenised securities, derivatives & on-chain financial products; Citadel gains exposure to the infrastructure layer without backing a single winner, though only Kraken has disclosed a formal liquidity & market-structure collaboration (here)
DTCC is running a trial with ~40 firms including JPMorgan, Goldman Sachs, BlackRock, Vanguard & NYSE to tokenise shares & Treasurys on blockchain infrastructure, marking a step toward digitising settlement rails at the core of U.S. capital markets (here)
Robotics
Notable deal:
Humanoid (SKL Robotics) raised $150M at $1.2B in a Series A led by founder Artem Sokolov’s $100M seed, with Schaeffler investing & Bosch in talks for a September extension. The London-based firm has delivered 15 robots & secured conditional orders for 10,000+ units, including a $300M deal with Schaeffler for up to 1,850 robots over 5 yrs; Bosch will manufacture 500 units next year (here)
Biotech
Isomorphic Labs unveiled its Drug Design Engine (IsoDDE), which more than doubles AlphaFold 3’s accuracy on novel protein-ligand structures, matches physics-based binding affinity methods at a fraction of the cost, & identifies cryptic binding pockets from sequence alone - moving from structure prediction to computational drug design at scale (here)