AI economics at scale: Anthropic's numbers and Modal's round

Anthropic's Q3 financials are out. The last brief noted Anthropic would show investors its Q3 results ahead of a possible IPO. Ed Ludlow, citing Bloomberg, reports the figures from that meeting: $16.8B revenue, an 85% adjusted gross margin, an adjusted operating margin above 11%, and positive free cash flow and adjusted operating profit @EdLudlow. The profit measures are all "adjusted" and come from secondhand reporting, not audited filings. If the numbers hold up, they cut against the view that frontier labs can't make money on inference, and they set the comparison point for any IPO.

Modal says it has passed $1B in annualized run rate. It raised a $750M Series D at a $15.75B valuation, led by Accel. The company says usage has tripled since its Series C and that it grew more than 15x year over year @modal. Taken with last period's Exa numbers, this suggests compute infrastructure is capturing much of the spending driven by agents.

App revenue keeps growing past $100M. Nathan Benaich lists these figures:

• Legora and Sierra each doubled to $200M in about six months.
• Harvey is at $400M and Lovable reports $600M.
• Cursor has been reported above $4B.

He notes the companies use different clocks and revenue definitions. In his view, the real test is how much of this revenue survives renewals, competition and changes in model capability @nathanbenaich. On exits, Deedy Das counts nearly $200B of AI startup M&A over 18 months, with a median company age of four years. He argues funds can post strong returns without owning the big AI labs @deedydas.

Cheap and specialized models pressure pricing

Bindu Reddy calls DeepSeek Flash the most cost-efficient model at $0.13 per task. She also says OpenRouter data shows it is the most-used model by token volume @bindureddy. David Sacks cites Brian Armstrong's claim that Coinbase fine-tuned the open-weight Qwen-3.5-9B on its own fraud data. Per Armstrong, the model beat frontier models including Opus 4.5, ran cheaper, and cut fraudulent transactions by about 30% in testing. Sacks calls this "a direct threat" to the business of selling ever-larger general models @DavidSacks.

Agents also break flat-rate SaaS pricing. One founder says a single customer's research-agent loop burned $340 in one day on a $79 seat. That took the account from a 72% gross margin to negative. The company now gives each seat a compute allowance and meters usage above it at cost plus 15% r/SaaS. When evaluating agent startups, ask how they meter usage.

Gemini 4 Argon's coding results depend on the benchmark. It leads Opus 5.5 on DeepSWE (77.9% vs 74.2%) but trails on Terminal-Bench 4.0 (57.4% vs 66.4%) and on FrontierSWE r/artificial. Reddy says she moved Argon into production for data and Excel work and found it better and cheaper than Opus. She gives no numbers to back that up @bindureddy.

Nathan Lambert is launching Trillium, a nonprofit open AI lab. His views:

• Distillation gives only a few months' edge.
• Chinese labs have the talent but lack compute.
• The market for data and RL environments is worth billions, and some of it is "sloppy enough to teach models to reward hack."

Trillium plans to give its environments away for free @a16z. For RL-environment vendors, that means a free competitor and a public warning about quality.

Fusion: capital moves toward manufacturing

• Inertia Enterprises is a spinout from Lawrence Livermore National Laboratory (LLNL), led by Twilio founder Jeff Lawson. It reports $450M in Series A commitments, a 50,000 sq ft facility near LLNL, about 75 staff (up from five at the start of the year), and 200 licensed LLNL patents TechCrunch. Its plan rests on two changes: replacing vacuum-tube lasers with diode lasers, and mass-producing fuel targets. Targets now cost about $500K each and only six are made a year, while a power plant would need roughly a million a day TechCrunch. Lawson compares fusion to FDA approval: the key moment is when a private facility first puts electricity on the grid, and he expects early investors to exit around then TechCrunch. He plans to break ground on a pilot plant in 2029, with first light around 2035 TechCrunch.
• Helion says it has raised over $1.5B at a $15B valuation TechCrunch. It aims to deliver grid power under its 50 MW Microsoft power purchase agreement before 2030 TechCrunch.
• Commonwealth Fusion Systems (CFS) says its Spark device will produce net power next year, two years later than its 2020 forecast. It targets the early 2030s for its ARC plant, which has a 200 MW Google agreement TechCrunch.
• Cambridge Heavy Industries raised a $6.2M seed round led by Azolla Ventures for a benchtop laser fusion device. It says the device can run 50 to 80 shots a day, with the data feeding AI simulations r/Futurology.

Another pro rata dispute

Following last period's discussion of pro rata conflicts, a Reddit post describes Valar Atomics suing its investor Day One Ventures over a roughly $1B Series B. Day One invested $1M at seed and $15M in the Series A. Its terms included a "spring-back" clause: if any investor who waived pro rata still got an allocation, everyone's rights would return. According to the post, the majority voted the clause out the day before closing, while a large insider was set to receive an allocation r/startups. Seed investors should check what voting threshold can amend their pro rata rights.

Agent security and finance agents

TechCrunch hosts cite a colleague's count of about two dozen companies selling agent security, likely an undercount. Examples include Air ($50M raised), Rico ($55M) and Sequoia-backed Symphony TechCrunch. The hosts expect three to five eventual winners rather than a single dominant company, and note that incumbents like CrowdStrike will also have to address agents TechCrunch. The founder of Hidden Layer says more than 50 of its customers have agents in production touching critical systems TechCrunch. OpenAI's Joseph Kim predicts agents will handle "investment idea generation and trading" in 2027 Latent.Space.

Talent and founders

• Pay: A Disrupt panelist with 20 years of recruiting experience says startups now open with 90th-percentile offers TechCrunch. One founder said London AI salaries jumped about 30% the week OpenAI opened an office there TechCrunch. Founding engineers got 0.8–1% equity four years ago; 1.5–2% is now common TechCrunch.
• Team size and validation: At one Stanford program, average founding-team size has fallen from "2 point something" to close to one Stanford eCorner. Steve Blank says all eight teams in his class arrived with finished products, and instructors had to push them back to customer discovery. The class now requires each team to find a design partner willing to use and pay for the product Stanford eCorner.
• Product-market fit: Superhuman's Rahul Vohra argues that because products are now easy to copy, companies should start their next product at $10–15M ARR instead of $30M. He says 30–40% of Superhuman's new users each week use only a "headless" version driven from AI assistants, which weakens time-in-app as a metric TechCrunch.
• Geography: Palo Alto lawyer Anil Advani now hears from about a dozen Indian founders a week about moving to the US, more than triple the rate of two years ago. Newcomer counts more than 100 investments by Indian VCs in US companies over the past two years Newcomer.

YC S26 company Molagri raised $5.8M for biopesticides that act only on the target pest species and can be reprogrammed when pests develop resistance @ycombinator.