📨 AI Capital Concentrating Fast

Read time: 7 min

This week, we cover:

Watch the 2-min video overview:

Global VC doubled to $330.9B in Q1'26. Over 60% of this value came from just ten $2B+ rounds, primarily in US AI platforms like OpenAI, Anthropic, and xAI. The Americas secured roughly 80% of global VC (US: $267.2B), while Europe ($25.7B) and Asia ($31.8B) reached multi-year highs but remained significantly smaller. Software, largely driven by AI, attracted a record $225.2B. However, deal count fell, indicating a shift towards fewer, larger late-stage investments and increased valuation risk. Exit value surged to $413.5B, skewed by a few mega-M&A transactions, while IPO volumes remained subdued [KPMG].

Chinese humanoid startups ship more robots than their US counterparts but command much lower valuations. US firms are valued as broad AI platforms, whereas Chinese players are seen as industrial hardware bets. Geopolitical factors are diverting US pension capital from China, opening opportunities for Middle Eastern funds to invest in Chinese "hard tech" and gain diversified exposure to humanoids in manufacturing and real-world applications [CNBC].

Big Tech’s planned $600 billion-plus AI spending is unsettling investors: Amazon, Alphabet and Meta fell, while Nvidia, Microsoft and Tesla rose on demand hopes. Software and data analytics firms are under pressure as markets question returns and disruption risk [Reuters].

A leaked memo from Sergey Brin to Gemini engineers, urging them to "urgently bridge the gap" with Claude, reveals Google's serious concern about Anthropic's lead. Brin is personally embedded in a coding strike team, and every Gemini engineer is now ranked by internal AI tool usage. A key metric shows Anthropic uses AI for nearly 100% of its own coding, while Google is at roughly 50%. Watch for accelerated Gemini releases in Q2 and Q3, and the implications for enterprise AI procurement decisions [Sherwood News].

AI reshapes math research by speeding up proof discovery: AI is now assisting mathematicians in discovering and verifying proofs more quickly, blurring the lines between human and machine authorship. Experts note that AI accelerates research without replacing human judgment, as results still require line-by-line verification [Quanta Magazine].

NSA uses Anthropic’s Mythos despite Pentagon blacklist: The NSA is reportedly using Anthropic's Mythos, despite the Pentagon labeling the company a security risk. This highlights how urgent cyber needs are overriding internal restrictions. A White House meeting between Anthropic's CEO and senior administration officials suggests a potential thaw, possibly broadening Mythos access to federal agencies beyond Defense [Axios].

Adobe launches AI tools for corporate marketing: Adobe introduced CX Enterprise, an AI suite for automating and personalizing marketing. This comes as the company faces pressure from cheaper autonomous tools and broader concerns that AI could reduce software demand. Shares remain down roughly 30% year-to-date, but the product launch suggests incumbents are adapting faster than the selloff implies. This is a bellwether for how legacy software companies will monetize AI rather than be displaced by it [Reuters].

AI demand keeps memory shortages tight, with relief unlikely before 2028: AI demand is straining broader memory supply, not just RAM, keeping prices high and delaying relief until at least 2027 or 2028. Global smartphone shipments are projected to drop 13% in 2026 (a loss of 160M units), with average selling prices hitting a record $523, up 14% year-on-year. Sub-$100 phones are now considered permanently uneconomical [Gizmodo].

Anthropic’s Mythos reaches unauthorized users, raising security concerns: A small group reportedly accessed Anthropic's restricted Mythos model through a third party within days of its launch. This underscores the risk that powerful AI tools can spread beyond intended controls. The first alleged unauthorized access came not from a rival nation-state but from a private Discord group, a reminder that the hardest security perimeter to defend is often the closest [Bloomberg].

Vercel confirms breach after hackers stole customer keys via third‑party AI tool: Hackers breached Vercel via a compromised third-party AI app, then sold exposed customer credentials online. Vercel states a limited number of accounts were affected and urges users to rotate non-sensitive secrets. The root cause — a contractor downloading malicious scripts — highlights how AI tooling rapidly expands the enterprise attack surface in ways traditional security frameworks were not built to handle [TechCrunch, Vercel].

AI chipmaker Cerebras files for US IPO after turning profitable in 2025: Cerebras Systems has filed for a US IPO, reporting $510M revenue and $87.9M net income in 2025, a significant improvement from a $484.8M loss the prior year. With a $20B compute contract from OpenAI and a target valuation exceeding $35B, this marks the first meaningful public-market test for a credible Nvidia alternative [Bloomberg].

Cursor draws $2B at $50B+ valuation as SpaceX locks $60B acquisition option: AI coding startup Cursor is raising fresh capital from a16z, Nvidia, and Thrive, while securing a strategic partnership with SpaceX. SpaceX holds an option to acquire Cursor for $60B before year-end. This deal provides Cursor with the necessary compute runway and gives Musk a shortcut into frontier coding, an area where xAI has struggled to compete with Claude Code and Codex [CNBC, X].

Canva targets 2027 IPO amid shift to AI credit model: Canva plans a 2027 IPO after transitioning from subscriptions to an AI credit model, aiming to solidify its business and compete with leading AI platforms. This will be an early indicator of how public markets will value AI-native design tools relative to their enterprise software predecessors [Capital Brief].

VW unveils "Agentic AI for All" roadmap, deploying onboard AI agents with personality in China EVs from late 2026 for intuitive vehicle control [AI Business].

Meta is deploying software on US work computers to capture mouse movements, clicks, keystrokes, and screenshots. This data will train AI agents on real human-computer interactions, making Meta the first hyperscaler to openly state this. The new training data moat is not the open web, but the workforce being automated [Fortune].

AI firms, including Google DeepMind, OpenAI, and Anthropic, are developing autonomous "co-scientists." These agent teams hypothesize, plan experiments, and execute them via robots, accelerating discoveries like protein synthesis while raising concerns about narrowing scientific diversity [MIT Tech Rev].

📅 9-10 Sep 2026 | 🌐 TBC

This premier UK forum connects government, industry, finance, and infrastructure leaders to address AI deployment challenges, from workforce reskilling and policy to public-private partnerships.

"The measure of intelligence is the ability to change." — Albert Einstein

That’s it for this week.

The CLNM Capital