Jevons Weekly
Issue #001 · Oct 20, 2026
Week of Oct 12–19, 2026 · 5-min read · Startups, AI and frontier tech through an investor's lens
The Jevons take
Intelligence got cheaper again last week. Google launched its new frontier model at half the price of Meta's, and DeepSeek priced web search for agents at under a tenth of a cent per call. Meanwhile, capital flowed to the companies that sell the "picks and shovels" for that exploding demand, from GPU clouds (Modal) to finance software that runs on AI (Ramp). That's the Jevons Paradox at work: as the cost falls, usage grows, and the question is who captures it.
Startups and their business
Image: Jevons Weekly. Data: TechCrunch, Axios Pro, ICON, Startups.com.br
Ramp raises $1.85B at a $60B valuation
Dragoneer led the round, with Thrive Capital, Founders Fund, Sands Capital and Avenir Growth co-leading. The valuation is up from $44B in June 2026, and Ramp says annualized revenue now tops $2B. The company is pushing deeper into AI agents for finance teams.
What this means for investors
Late-stage money is concentrating in a few AI-native compounders with real revenue. A $60B mark on about $2B of run-rate revenue resets comparables for fintech and spend-management names, private and public.
Source: TechCrunch · Ramp (PR Newswire)
Modal raises $750M Series D at $15.75B
Accel led the round in the GPU cloud startup, which lets developers run AI workloads without managing infrastructure. General Catalyst, Lux Capital, Redpoint and Bain Capital Ventures were among the other investors named.
What this means for investors
This is a direct Jevons bet. As inference gets cheaper, the number of workloads explodes, and the layer that schedules and bills for that compute captures volume. Watch gross margins: GPU resellers live or die on utilization.
Source: Axios Pro · Crunchbase News
Alice raises US$40M Series D led by IFC
The São Paulo health plan, now positioning itself as "an AI company that runs a health plan", raised from the World Bank's IFC with Germany's DEG and Finland's Finnfund, plus existing investors Kaszek, ThornTree and Canary. Alice says it has 100,000+ members, recurring revenue above R$1.2B, and that AI cut the operating cost of its member service by 18%.
What this means for investors
Development finance institutions are now leading late-stage rounds in Brazil, where crossover funds have pulled back. For family offices, that means co-investment alongside DFIs is a realistic route into Brazilian growth deals.
Source: Startups.com.br
AI
The frontier-model price war: Gemini 4 Argon vs. Meta Muse Wonder
Google made Gemini 4 Argon generally available at an introductory $2 in / $10 out per million tokens, with cached input 95% off. Meta's new Muse Wonder, built for agents that run for hours, costs $4 / $20.
What this means for investors
Model pricing keeps falling, which hurts pure model sellers and helps everyone building on top. Value is moving to distribution (Google, Meta) and to apps with proprietary data and workflows.
Source: Google · Meta (Muse Wonder pricing)
Anthropic outgrew git and built its own version control
Merged code changes at Anthropic went from about 2,500 a day in May to 30,000+ in September, as AI agents write more of the code. A team of 7 built "Scion," a version-control system designed for agents, and says CI checkouts are about 100x faster.
What this means for investors
When writing code becomes cheap, the bottleneck moves to version control, testing, CI and review. Expect a new wave of developer-infrastructure startups and acquisitions.
Source: Anthropic Engineering
DeepSeek opens a web-search API for agents at $0.75 per 1,000 calls
DeepSeek also updated its Flash model with no price change. Failed calls are free.
What this means for investors
Search for AI agents is turning into a cheap commodity sold by model providers. That squeezes standalone search-API startups and lowers the cost of running research agents.
Source: PPC Land · Investing.com
Frontier tech
Image: ICON
Google shows "magic state cultivation" working on its Willow quantum chip
Image: Jevons Weekly. Data: Google Quantum AI
In a paper published Oct 15 in Science, Google Quantum AI showed the technique on real hardware, cutting errors about 40x. Producing these "magic states" has been one of the most resource-heavy steps toward useful quantum computers.
What this means for investors
This is a real engineering step, not a commercial product. It shrinks the size of the machine needed for useful work. Quantum remains a long-dated bet; position through diversified exposure rather than single names priced for near-term revenue.
Source: Google Quantum AI
ICON raises $200M to scale robotic 3D-printed construction
Baillie Gifford and funds managed by BlackRock co-led the Series D. CNBC reports total funding is now $700M, at a $1.4B valuation. The money scales production of Titan, its multi-story printing system, with first deliveries expected in early 2027.
What this means for investors
Public-market growth investors are writing checks into physical-world automation. Housing is a market where cheaper building can unlock large latent demand, but execution and unit economics are still unproven at scale.
AI is now training robots in simulation
Image: Jevons Weekly. Data: Google, via Interesting Engineering
Google says data generated through Gemini 4 Argon's simulations raised a robot's real-world task success rate from 56% to 75% in its tests.
What this means for investors
Cheaper simulated training data lowers the cost of robotics development, a tailwind for the humanoid and industrial-robot startups raising big rounds this year.
Source: Interesting Engineering
Sponsor
This slot is open. Reach tech-curious investors and Brazilian family offices. Reply to this email to book.
João's take
Every cycle has a moment when capital remembers that the world is made of atoms. Last week, Google's Willow chip cut errors about 40-fold, ICON raised $200M to scale robotic 3D-printed construction, and simulated data lifted a robot's real-world task success from 56% to 75%.
The Jevons logic carries over from software: when designing, testing and training get cheaper, more physical things get built, and value migrates to whoever solves the bottleneck of material and precision.
The calendar still belongs to the engineers. Quantum remains a lab milestone, and ICON's Titan only starts delivering in early 2027. For investors, that calls for patient capital with diversified exposure, reading the primary source before pricing any timeline. As always, to be confirmed.
João Piccioni, editor
Share Jevons Weekly
Know an investor who should read this? Forward this email, or share your personal link.
Every item links to its source. Nothing here is investment advice. Spot an error? Reply and we'll fix it.