A useful image to think about AI is that of a big company CEO organising her layers of staff to construct a well-oiled machine. She can’t control or predict exactly what any individual component of the organisation will do, but she can design the incentives and systems that guide the machine. The miracle of AI is in democratising this ability: ordinary people are now running their own systems of agents mimicking the companies of billionaire CEOs. This lens is useful for us to understand some of the commercial realities of the future of AI.
This week, people are talking a lot about Jev, a new and incredibly cheap type of model. Jev is designed for decisions where you don’t actually need an LLM to generate text: classify this, approve that, pick A or B. It is much cheaper and simpler than using a frontier model for the same job.
Many have been questioning whether this will curb the appetite for leading-edge models from Anthropic and OpenAI. This is essentially the same question that commentators have been asking about the impact of improving open source models: does introducing a much lower cost solution to simpler problems decrease the market value of the more complex models?
Seen through the lens of our CEO allegory the answer is obvious. If you’re an extremely smart, talented CEO, does introducing lots of low wage labour to your society make your talent more or less valuable?
The introduction of Jev to LLMs is like the introduction of outsourcing to domestic labour forces in the early 2000s. The net result of offshoring was to increase high-skilled wages. Jev may do something similar for machine labour: it isn’t competing with white collar Opus and Fable, but it does provide a whole new tool set a CEO can use to construct her company.
Expect heterogeneous layers of models to evolve inside companies - cheap models doing simple work, more capable models handling harder tasks, and frontier models playing a management role.
The interesting question, then, is not whether cheaper models reduce the value of expensive ones. But what things that aren’t currently done by LLMs are going to become economical to do with LLMs once they become more capable with their new armies of cheap labour?
IPOs / Publics
Oura postponed its Nasdaq IPO 8 days after launch, citing market uncertainty despite strong demand & 90% YoY revenue growth expected for FY26 (here)
Anthropic filed its IPO prospectus warning of ‘existential risks to humanity’ from AI, reporting $4.6B revenue & $8B operating loss in 2025; backers target a $2T+ valuation, but safety concerns (including recent hacking incidents by AI agents) & extreme customer concentration (25% from two clients) complicate the autumn Nasdaq debut (here)
Nscale, the Nvidia-backed AI cloud provider targeting a $35B IPO, buried ByteDance as its largest customer in obscure SEC filings despite the Chinese tech giant accounting for 73% of its $33M revenue in 2025. ByteDance used Nscale’s Norwegian facility to access advanced Nvidia B200 chips blocked by US export controls - a legal but geopolitically fraught strategy that helped Nscale secure anchor revenue before landing $44B & $45B deals with Microsoft & Anthropic (here)
Firmus (data centre operator) expects a $77M H1 loss ahead of a planned $5B Australian IPO on 22 Oct, which would value the firm at up to $60B - the country’s second-largest listing ever (here)
Blockchain.com, founded in 2011, is targeting a $500M IPO at $4B-6B valuation this year after filing confidentially with the SEC, hoping to capitalise on Bitcoin’s 33% rally since mid-August. The crypto services firm raised $110M in 2023 at sub-$7B (down from $14B in 2022) & has been EBITDA-positive for 3 years, though prior crypto IPOs - Gemini, BitGo, Etoro - are down 50-80% since listing (here)
Sensofusion (anti-drone software) is seeking €300M in a Helsinki IPO at €1.3B pre-money, backed by €170M in anchor commitments from Finnish pensions. The Vantaa-based group holds a U.S. Marine Corps contract, doubled H1 2026 revenue to €26M, & is capitalising on European defence market momentum - though only 7% of 2025 revenue came from Ukraine (here)
Nvidia authorised a $150B buyback increase (largest in history), bringing total authorisation to $235B, to be executed by January 2028 - a signal management views shares as undervalued despite the stock trailing AMD’s 175% rally this year (up 23% vs S&P’s 12%) (here)
Nvidia is in early talks with insurers to underwrite credit risk on chip-backed loans to smaller cloud providers, exploring residual value protection if borrowers default & resold hardware doesn’t cover the debt - part of CEO Jensen Huang’s push to treat chips as an ‘investable asset class’ & expand demand beyond Big Tech balance sheets, alongside its $500B Wall Street facility & $105B OpenAI guarantee (here)
SoftBank drew $20B+ demand for an $11B junk bond sale to fund its OpenAI investment, pricing 7.5-yr notes at up to 9.875% - record highs for the company & above typical BB+ yields. The firm has sold nearly $15B in bonds this year, the largest speculative-grade issuer, as it pursues a $65B OpenAI commitment alongside recent margin loan expansions backed by Arm shares (here)
Meta launched Meta Enterprise Platform to sell its AI stack (including Muse assistant, agents & APIs) to businesses, hiring MongoDB CEO CJ Desai to lead the unit - a move to monetise $100B+ in AI capex while MongoDB shares fell 17% on his departure (here)
Big Dogs
Anthropic is seeking shareholder approval for a Palantir-style corporate structure that would grant its 7 co-founders collective 50.1% voting control ahead of IPO, despite CEO Dario Amodei owning only ~2% equity (here)
Regulation
20+ UK tech founders - including AI labs ElevenLabs, Synthesia, & several DeepMind spin-outs - called on PM Andy Burnham to limit non-competes & extended notice periods, arguing they hamper British competitiveness vs US rivals. CMA estimates 25%+ of UK employees face non-competes (here)
Venture Capital
1789 Capital, where Donald Trump Jr. is a partner, is in talks to raise $3B for a second growth fund (already secured $2B from existing LPs) to make larger bets in AI & defence (here)
Bessemer raised $5.75B ($1.75B early-stage, $4B growth) to back AI across the stack, having deployed $3B into 260+ AI-native companies since 2022 including Anthropic, Perplexity & Cognition (here)
Sofinnova Partners closed MD Start IV at €82M to build 6-8 medtech companies from inception across Europe & the US (here)
Energy Revolution Ventures closed the first tranche of a $50M Fund II targeting seed & Series A in energy storage, grid tech & advanced materials (here)
Venture Geopolitics
Trump & Xi met in Washington for Xi’s first DC visit in 11 years. US & China agreed to cut tariffs on $60B of goods under a ‘30-for-30’ framework, covering US coal, agricultural products, & Chinese toys, car seats, & holiday goods. US goods trade deficit with China has fallen 40% since Trump took office to a projected $140B in 2026 (here)
China & the US also agreed to establish an AI safety incident channel alongside commitments to military crisis communications. Trump signalled limits on AI information-sharing, citing a 1-2 year US lead (here)
Trump hosted Anthropic CEO Dario Amodei at the White House, signalling a thaw after the president called the firm “leftwing nut jobs” & the Pentagon banned its products over warfare-use restrictions. The dinner precedes a Tuesday meeting on AI regulation, where Amodei’s push for oversight clashes with Trump’s anti-regulation stance (here)
White House has asked OpenAI & Anthropic to let U.S. agencies review new models before sharing with UK’s AI Safety Institute, with Anthropic complying (Claude Mythos 5.1 now U.S.-only). The move underscores transatlantic friction over AI testing protocols, even as UK’s AISI - one of the world’s best-equipped government AI testers - retains access to frontier models including GPT-6 Astra. U.S. counterpart CAISI remains understaffed with no permanent director (here)
Brent crude rose above $107/barrel after Trump rejected Iran’s proposal to reopen the Strait of Hormuz within 7 days in exchange for sanctions relief & naval blockade withdrawal. Ship transits remain at ~130/week vs ~130/day pre-war, with one-fifth of global oil supply at stake - a sharp reminder that geopolitical choke points still set energy prices despite production diversification (here)
A US military analyst used AI to assess intelligence on a Chinese ship in spring 2026, producing a false report claiming it carried nuclear weapons components. Armed troops prepared to board the vessel before the error was caught - an incident sources said “almost started a war” with China (here)
China expanded overseas travel restrictions on top private-sector AI talent to include spouses & children, requiring families to obtain approval from Beijing before trips abroad - a broadening of curbs introduced in May covering executives at Alibaba, DeepSeek & other strategic firms (here)
ASML sells zero chipmaking machines in Europe today because no new fabs are under construction, its global public affairs chief warned, as the US, China & India court expansion while the EU Chips Act (2023) fails to deliver production capacity ahead of its 2030 doubling target (here)
Revolut launched a 3-day facial recognition payment pilot at London coffee shops, letting customers authenticate via face scan at checkout terminals with zero processing fees for merchants - opt-in only, with encrypted biometric data stored separately from merchant systems (here)
Monzo (consumer digital bank, 16M customers) is in early talks about a sale to NYSE-listed Nubank (Brazil, $65B market cap, 140M customers) at £8-10B (here)
Strategic Sectors
AI
Per the FT, corporate America is rapidly adopting open-weight AI models - primarily Chinese - as costs spiral; open models now account for 56% of tokens processed through Vercel’s AI Gateway (up from 7% in Dec), driven by pricing that’s a fraction of OpenAI & Anthropic’s closed systems. AT&T runs 40% of AI workloads on open models & aims for 70% within a year (here)
Nvidia released Open Agent Safety Platform, a software layer to contain AI agents after models from OpenAI, Anthropic, Meta & Google escaped sandboxes & attempted external breaches (including OpenAI’s July incident at Hugging Face). Huang framed safety as an engineering problem solvable through product design, countering Amodei’s recent call to slow AI advancement; partners include Microsoft, Oracle, Cisco & Anthropic (here)
China’s Ministry of Industry & Information Technology has asked ByteDance & Alibaba to report on plans to buy Nvidia’s RTX Pro 5500 chips (a workstation GPU not its top AI silicon), signalling possible approval as domestic AI firms strain for compute & Huawei can’t keep pace (here)
Manus (AI agents, China-founded, unwinding its Meta acquisition) launched Cue, a consumer personal-agent app offering multiple agents per user with email, phone, wallet & virtual computer access - positioning it against Meta’s Muse (here)
OpenAI pulled GPT-6.1 Astra after it scored worse than GPT-6 on alignment evaluations, unable to stay within instruction bounds (here)
Oracle invoked force majeure on Project Jupiter, its 2.45 GW New Mexico data centre, to defer rent if delayed beyond 2028 - the facility, part of Trump’s Stargate AI infrastructure programme, faces permit denials & pipeline delays, whilst $18B of project debt trades at distressed levels below 90 cents on the dollar (here)
A Hacker News thread traced an AI feature decision through multiple LLM layers - Claude-generated code from AI-written tickets based on AI-generated docs from Claude strategy memos - revealing no single human decision-maker, only layers of review responding to perceived market expectations. AI now amplifies the opacity & speed of “gradual disempowerment” (here)
Notable deals:
Energy
Britain’s electricity prices have risen 80% since 2000 despite renewables reaching 50% of generation, undermining electrification & leaving 85% of homes reliant on gas heating. The UK’s fragmented privatised system imposes costs through consumer-funded subsidies, foreign-owned infrastructure, imported equipment & complex regulation - creating a ‘zombie system’ more expensive than either pure market (Texas) or state-owned (France, Norway, China) models (here)
US Strategic Petroleum Reserve fell to 283.8M barrels in September, the lowest since October 1982, after Trump authorised a 172M barrel release in March as part of a coordinated 400M barrel IEA drawdown following Hormuz disruptions. The reserve is now testing its operational minimum of 250-300M barrels, with analysts warning further releases risk signalling the US is running out of options even as swap agreements with Shell, Vitol & Trafigura should return 133M barrels at 1.25:1 starting early next year (here)
Defence
Notable deal:
Tekever (AI-powered surveillance drones) raised $580M at $6.4B led by UC Investments & Baillie Gifford, citing 50,000 operational flight hours in Ukraine since 2022 & a £400M UK Ministry of Defence contract awarded in July. The Portuguese-British company joins Helsing ($1.8B), Quantum Systems (€1B), Cambridge Aerospace ($300M), & Stark (€500M) in a wave of European defence tech raises driven by rearmament policy (here)