
Jevons Weekly
Issue #005 · Nov 17, 2026
Week of Nov 9 to Nov 15, 2026 · 7-min read · Startups, AI and frontier tech through an investor's lens
The Jevons take
Nvidia lost its founder and its shares fell 13% in a week, the sixth straight weekly drop. In the same days, Anthropic's IPO book was reported six times covered. Demand for AI compute held up. The market is now repricing who captures it: the company that sells the chips or the lab that rents them.
Startups and their business

Image: Jevons Weekly. Data: Mirendil, Mercor, Torus, Q-CTRL, Clara
Mirendil raises $1B to let AI agents build AI models
Kleiner Perkins and Clearlake Capital co-led a $1B Series A, with Google among the investors. Mirendil, founded last year by former Google DeepMind and Anthropic researchers, has not disclosed a valuation. Its first product, Lor, automates model-training runs by splitting the work among hundreds of AI agents. The company says one researcher built a coding model, Mirendil-Base, in about 3.5 months for $10M, with agents writing more than 99% of the code. All of that is the company's own account.
What this means for investors
A $1B first round with no disclosed valuation shows capital paying for the tools that build models as well as the models themselves. The cost and speed claims are untested outside the company. Wait for customers.
Source: SiliconANGLE · Crunchbase News
Mercor raises $750M at $20B to sell AI training to every company
Mercor raised a $750M Series D at a $20B valuation, double the $10B of October 2025, when it raised $350M. Benchmark, Felicis, General Catalyst, Nvidia, Salesforce, Sapphire Ventures and Zoom joined, along with funds managed by affiliates of Apollo. The round includes an employee tender offer, and Mercor did not say how much of the $750M it is. The company pays a network of more than 5 million experts to train and test models. Until now its main buyers were the frontier labs. Co-CEO Brendan Foody wrote that Mercor already spends 3x as much on inference as on salaries.
What this means for investors
Mercor's buyers have been the frontier labs, so a $20B mark rides on lab budgets. The new money bets that other companies will pay for the same service.
Source: The Next Web · Mercor · Reuters via Investing.com
Clara raises $105M in LatAm after Brazil's weakest venture month in 13
Clara, the Latin American corporate-spend platform, raised $105M: a $65M Series C anchored by Citius, Kaszek and DEG, plus a renewed $40M facility from General Catalyst's Customer Value Fund. It passed $100M in annualized net revenue, has run positive EBITDA for more than a year and serves more than 45,000 customers. Its debt capacity is above $250M. In Brazil, October was the weakest month for venture in 13 months, per Sling Hub: 15 rounds totaling $118M, 33% below September and 89% below October 2025. Asaas and Aline made up more than 80% of the total.
What this means for investors
In a thin Brazilian market, the money that arrived went to a company with real revenue and positive EBITDA, and part of it is debt. Capital in the region is selective and prefers cash flow.
Source: Clara · Bloomberg Línea
AI
Jensen Huang resigns, and Nvidia reports today with an interim CEO
Nvidia's board accepted Huang's resignation as president, CEO and director on Nov 11, effective immediately, according to its 8-K. CFO Colette Kress, in the job since 2013, is interim CEO and stays CFO. Harvey Jones, on the board since 1993, becomes its first-ever chair. The announcement gives no reason. Axios ties it to reports of two relationships, one with a woman identified by U.S. intelligence as a Chinese intelligence asset. Nvidia fell 13% this week, its sixth straight weekly drop and the longest since 2012, about 30% and $1.7T below early October, per CNBC. Third-quarter results come Tuesday after the close. Analysts expect revenue of $108.9B, up 91%, per LSEG. Nvidia's latest 10-Q caps its guarantees tied to an OpenAI campus at $105B.
What this means for investors
The numbers are not in question. Who signs the commitments is. On the call, listen for guarantees and customer financing as well as demand.
Source: Nvidia · SEC 8-K · CNBC · Axios · 10-Q, via Stock Titan
Anthropic's IPO book is six times covered, at a lower valuation than hoped

Image: Jevons Weekly. Data: Reuters
Anthropic launched its roadshow on Nov 10: 825M shares at $70 to $78, with pricing expected Nov 18 and trading Nov 19 on Nasdaq as ANT. At the top of the range it would be worth up to $1.405T fully diluted, per CNBC, against the $2T once expected, per The Guardian. The book was about 2x covered after the first day and 6x by Thursday, Reuters reported, citing sources. SpaceX was 2x covered after its second day. Seven suppliers (AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC) agreed to buy $26B of stock in a private placement at the IPO price. Run-rate revenue was about $81B at the end of October, up $10B in a month. Investors' open questions, per Reuters: whether Anthropic keeps its lead over Google, and whether Claude holds up beyond coding.
What this means for investors
Demand is strong at a price below the hoped-for mark. With suppliers owning part of their biggest customer, the pricing on Nov 18 also marks their own positions.
Source: Anthropic · CNBC · Reuters via Investing.com · The Guardian
Google prices Gemini 4 Hydrogen at twice Flash's launch rate

Image: Jevons Weekly. Data: Artificial Analysis, via VentureBeat
On Nov 12, Google released Gemini 4 Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1M-token window. Gemini 3.8 Flash costs $0.75 and $3.75 through Dec 31, and the Flash price rises to Hydrogen's level on Jan 1. Google says Hydrogen finishes tasks 2.5x faster than Flash and scores 57% on Terminal-Bench 4.1, against 18.2%. On the Artificial Analysis index it scores 51 against 31 for Flash. The cheapest small models (Claude Haiku 5.5, GPT-6 Luna, Meta's Muse Ember) list at $0.10 and $0.50, so Hydrogen costs 15 times as much per token. Artificial Analysis estimates $1.87 per benchmark task for Hydrogen and $7.77 for Flash, both at high reasoning.
What this means for investors
Price per token no longer says what a task costs. A model that costs more per token can cost less per finished task. Sellers with the better model can hold list prices while buyers' bills depend on retries.
Source: Google · VentureBeat
How much of AI revenue is organic?

Image: Jevons Weekly. Data: Apollo
Apollo's Daily Spark published no new VC-share data between Nov 9 and Nov 15. We carry the Oct 22 reading: 80.4% of the top 1% of AI spenders are VC-backed, against 13.1% of median spenders. Change vs #004: none.
Ramp's November AI Index, published Nov 11, adds context. 1% of businesses drive 80% of revenue to OpenAI and Anthropic, and that top 1% is "largely made up of VC-backed companies in tech, finance, and technical services," in Ramp's words. AI spend rose 1.65% in October to a new peak. Engineering teams drove 72% of classified spend. Customer success and support rose to 13.3% from 0.8% in June. The average price per million tokens fell 21% in the month.
How we read it: the 80% from Ramp is a share of revenue, a different measure from the share of VC-backed buyers, so the mark stays at 80.4%. Both point the same way.
What this means for investors
The buyers that fund AI revenue are still few and venture-backed. Watch customer support, now 13.3% of classified spend, as the first use outside engineering with real size.
Source: Apollo, Oct 22 · Ramp, Nov 11 · Ramp AI Index
Frontier tech

Image: Q-CTRL
Salesforce Ventures led a $290M Series C, with Bullhound Capital and Affinity Partners. Q-CTRL says the round values it above $1B, without giving a figure. AMD Ventures, In-Q-Tel, Lockheed Martin Ventures and Australia's National Reconstruction Fund are among the investors. Its Ironstone Opal system uses quantum sensors to navigate when GPS is unavailable. It is being deployed on Airbus commercial aircraft and Lockheed Martin autonomous platforms. The company says it works 100 times better than the best alternative, which is its own claim. A Salesforce Ventures partner says it is the first quantum navigation system qualified for safety of flight.
What this means for investors
Quantum revenue shows up first in sensing and control software, with aviation and defense buyers. The money follows the product that has customers today.
Source: Q-CTRL · SiliconANGLE
Torus raises $500M to put compute where the power already is

Image: Torus
Magnetar led a $500M round, with ICONIQ, Origin Ventures, Ridgeline Partners and Trinity Capital. Torus launched Foundry, prefabricated modules of 1 to 20 MW that sit behind the meter at existing sites and use grid headroom, with 12 to 16 weeks from contract to running workloads, per the company. It has thousands of GPUs deployed and plans more than 47,000 by the end of 2027. First customers are Hydra Host and another AI cloud. Its Foundry production line at the Salt Lake City plant, GigaOne, will be able to produce 500 MW of IT load, with plans to reach 1 GW.
What this means for investors
After the Nov 3 votes, power and permits decide where compute goes. A pitch that avoids a new campus and new transmission may meet less local resistance. That is the thesis. The 47,000-GPU target is a plan, not a result.
Source: Torus · The Next Web
Event note
Today, November 17, in São Paulo, I talk about how to invest in the decade AI arrived: "A próxima década começa agora" (in Portuguese). Reply IA to this email and I'll send you the details.
The investor's read
Last week I said I would read Anthropic's IPO book as a credit document. First page: a range of $70 to $78, up to $1.4T fully diluted, against the $2T once expected. By Thursday the book was six times covered.
The line I keep rereading is another one. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC agreed to buy $26B of stock in a private placement at the IPO price. Seven suppliers become owners of a customer they all sell to.
In a good quarter that aligns interests. In a bad one, the landlord, the chipmaker and the memory seller hold the same position, and the price set on Wednesday marks all of them.
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.

