Make Claude your AI podcast player.

Daniel Dines sizes US legal at $300B. He thinks about 30% of that labor is replaceable ($90B), and that vendors might charge about 10% of it as model usage, so roughly $10B in total. He asks whether Harvey's and Legora's investors are mistaken. Three days later on 20VC, Jason Lemkin said Legora had hit $200M ARR with its next round at $11B, and cited a report, which he wasn't sure of, of minus 50% gross margins at Harvey.

The Twenty Minute VC (20VC)

Why AI Cannot Replace Humans in Enterprise, with Daniel Dines of UiPath

Daniel Dines, CEO of UiPath

▶ Listen from 43:58

The Twenty Minute VC (20VC)

Meta's Muse Hits No. 1, Menlo Sounds the AI Bubble Alarm, and more

Harry Stebbings, Jason Lemkin and Rory O'Driscoll

▶ Listen from 1:00:24

Harvey raised $550M at a $15.5B valuation in September. Its co-founder said ARR passed $400M, up from about $190M in January; the company has not confirmed that figure in its own announcement.

Dines' $300B covers the US only, while Harvey sells in 60+ countries. His 30% and 10% are rough guesses. The ARR and margin figures are reported, not confirmed by the companies.

Dylan Patel says Anthropic's revenue per megawatt has gone from about $20M to as high as $100M, while customers like Jane Street capture an estimated $300M to $500M per megawatt. His view, which he calls non-consensus, is that labs will put less compute into selling and more into their own use, pointing to Anthropic's compute additions outpacing its revenue additions.

Dwarkesh Podcast

Dylan Patel: Anthropic & OpenAI will have most of the world's compute by 2028

Dylan Patel, founder of SemiAnalysis

▶ Listen from 0:00▶ Listen from 1:14:42

Anthropic's annualized revenue reportedly passed $100B in mid-September, up from $65B at the end of July and $9B at the end of 2025. Reports put an IPO in November at about a $2 trillion valuation. These are run-rate projections from press reports, not audited revenue.

The per-megawatt figures are Patel's estimates. The view fails if internal research runs into diminishing returns before outside demand does, and a public company faces pressure to show revenue.

Higgsfield's Alex Mashrabov says the company crossed $1B in annualized revenue in 18 months, and that margins are over 80% on its own post-trained open models versus 20% to 30% on closed models, because "PhD-level intelligence is not necessary to make a viral social media video." Ollama's Jeffrey Morgan predicts open models will carry 80% to 90% of a business's AI usage but only 10% to 20% of its AI budget.

The Twenty Minute VC (20VC)

$1BN ARR in 18 Months: The Untold Story of Higgsfield

Alex Mashrabov, founder of Higgsfield

▶ Listen from 27:13

Y Combinator Startup Podcast

Open Models Change The Economics of AI

Jeffrey Morgan, founder of Ollama

▶ Listen from 17:07

AT&T says about 40% of its AI usage now runs on open models, up from 20% in May, with a 60% to 70% target. It processes about 45B tokens a day, up from 8B a year ago, and reports coding costs down 56% for a 2% quality drop. It uses Gemma, Llama and Nemotron, not Chinese models. On OpenRouter, Chinese models carried 57% to 67% of tokens in mid-September, up from 6% to 13% in February.

Higgsfield's margins are self-reported and specific to video ads. OpenRouter is price-sensitive developer traffic, not a census of corporate spending. A 2% average quality drop can hide the tasks that fail.

YC says it has funded more than a dozen companies in two years that each make over $10M a year selling data or training environments to the labs, in many cases hundreds of millions, and that many of these companies stay quiet about it. On Dwarkesh, a researcher describes the order in which Anthropic built its training environments (coding first, then finance and spreadsheets, then PowerPoint and the long tail of office work) and says other labs have since followed it.

Y Combinator Startup Podcast

The State of Startups in 2026

YC partners

▶ Listen from 22:51

Dwarkesh Podcast

AI researchers debate how close we are to recursive self-improvement

A panel of AI researchers

▶ Listen from 33:26

Epoch AI's survey of vendors puts a simulated website at about $20,000 and a faithful copy of Slack near $300,000. Individual tasks run $200 to $2,000, and exclusive deals cost 4 to 5 times more. Anthropic reportedly discussed spending over $1B in a year. Vendors name enterprise workflows such as expense reports, CRM updates and spreadsheets as the 2026 growth area, and say the expensive part is the code that checks whether the AI actually did the task.

Labs build many of these in-house and are reportedly pushing standard formats that make vendors interchangeable. Revenue figures for these companies are mostly unpublished.

Greg Eisenberg reports that Thrive Holdings has bought close to 50 accounting firms in 24 months. At one, Larson Gross, AI cut one 180-hour annual job to 15 hours. He reports that General Catalyst's Long Lake bought 18 businesses and reached $100M EBITDA in under two years with margins doubling. He cautions that these numbers are self-reported.

The Startup Ideas Podcast

$5T opportunity: AI Roll Ups

Greg Eisenberg

▶ Listen from 4:40

Long Lake has now made 40 acquisitions and closed its $6.3B take-private of American Express Global Business Travel. PitchBook puts that business at $2.7B of 2025 revenue and $182M of operating profit, about a 7% margin. Long Lake was valued at $5.8B last November.

The small-firm results are self-reported and untested in a recession. Long Lake says it has never done a layoff, so any margin gain has to come from growth rather than cost cuts.