We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Capital is going to fewer companies
The 20VC x SaaStr discussion described an unusually narrow market. Harry Stebbings met a profitable company with $20M ARR, growing 600–700% and with a strong team, and no VC responded . Rory O'Driscoll says space data centers and quantum are still getting money. The squeeze is on vertical software, where investors quietly fear AI erases the category . Jason Lemkin's tactical advice: founders with top-decile numbers outside hot categories should plan to meet 100 VCs .
The panel also went through the math of funding at each stage:
- Large funds: $5B+ funds need billion-dollar positions, which makes a $2B outcome hard to use without majority ownership .
- Seed: entry prices of $50–100M and multi-stage firms crowding allocations leave seed investors with stakes too small to matter .
- The counterview: Rory cites history: 10–20 decacorns per decade, against a market behaving as if 90 will appear this year .
Inference vs. neoclouds. Modal raised a $750M Series D led by Accel at $15.75B post-money and says it crossed $1B in annualized run rate. Its May Series C valued it at $4.65B . The CEO says sandboxes, which run AI-generated code, now make up about a third of revenue. Demand came from vibe-coding companies, background agents and RL work at new labs . Rory's argument: if quality is similar, cost-sensitive buyers like Replit will move tokens from a 70%-gross-margin lab to a 30%-margin inference host . Replit did switch its default model to OpenAI, which Lemkin attributes to prices 30–40% lower for similar quality .
CoreWeave shows the other side. It fell 15.5% on October 14, some of its bonds yielded above 15%, and credit-default-swap spreads passed 1,000bp .
Anthropic. Anthropic reportedly told prospective investors Q3 revenue was $16.8B, up from $11.6B in Q2. IPO talk reportedly centered on about $1.5T, below the ~$2T investors expected .
Late-stage multiples. An Anduril employee tender implied about $90B on $2.2B of trailing revenue . Rory calls private multiples "almost certainly disconnected" .
Security: OpenAI says bug-by-bug fixing won't scale
Greg Brockman published OpenAI's account of PHP2Shell. The bug sat in PHP for 14 years. Calif researchers found a version using GPT-5.6 Sol in August, and OpenAI found a fuller version independently in September . He also mentions a Node.js RCE disclosed the same day. He calls the fixing process "slow and manual" and says it won't scale . He floats authorized defenders, such as government, scanning for vulnerable hosts and in extreme cases patching them remotely . Brockman says OpenAI moved 25% of its production engineers to model-driven hunting for holes in its own systems .
The fallout widened this period:
- Quest Diagnostics was reported hacked, with patient data exposed. Quest serves roughly half of US hospitals and physicians .
- Replit says it patched its infrastructure and blocked exploits after vulnerabilities in widely used third-party software .
- Jason Clinton (Anthropic) told security leaders to ask their boards for funding for the basics, not AI, and to "hold on for ~18 months" .
- Nick Dobos claims, without sourcing, that 3% of all recorded "critical" bugs were found in the past month .
CrowdStrike's view. George Kurtz says attacker breakout time now averages 29 minutes and can be as fast as 27 seconds . He says much of the attacking runs on open-weight models, because frontier models refuse cyber requests . He claims a general remediation task costs about $0.03 on CrowdStrike's post-trained model versus about $10 on a top frontier cyber model . One bank CIO told him open-weight models a few months behind the frontier are "plenty" for most work .
On the startup side, Cogent Security (CEO Vineet Edupuganti, formerly at Abnormal) has raised $53M, with Greylock and Bain Capital among its backers .
Agents in bank accounts. Axios reports major US banks are finding third-party AI agents inside customer accounts and can't reliably tell humans from agents . David Sacks says most of this is customers handing their credentials to agents. He argues banks should support user-directed agents, with read-only access as an interim step . The open question is how to identify and permission agents.
Frontier bets and new funds
- a16z Frontier: a16z renamed a16z Crypto to a16z Frontier and widened its mandate to companies building markets that may take years to develop . Named interests include AI-native uses, autonomous machines, bio, full-stack startups, and networks of agents that transact with each other . Chris Dixon leads it with Marc Andreessen, Ben Horowitz, Eddy Lazzarin and Guy Wuollet . Andreessen predicts startup formation will "skyrocket" because individuals are adopting AI faster than institutions .
- Warp Sciences: raised a $215M Series A led by Lightspeed to apply AI and robotics to drug development. Co-founders are Kevin Weil, DJ Patil, FAIR co-founder Manohar Paluri and Sol Reisberg . The pitch is a US robotics-first CRO, a contract lab that runs experiments for biotechs. Its founders say US CROs cost 4–5x more than offshore ones, but offshore labs carry IP and geopolitical risk .
- AMI Labs: a moderator restated details of Yann LeCun's company, a world-models bet against LLM scaling. It raised $1.03B in March at a $3.5B pre-money, described as Europe's largest pre-launch AI round .
- Tractian (YC W21): raised a $130M Series D at $1.4B. It sells sensors plus AI for predicting industrial machine failures, has passed $100M ARR and is growing 90% a year .
- Robotics caution: Jason Shuman says robotics venture funding is up over 400% in two years while installations grew less than 25% year over year .
How AI is changing incumbents
Podium says it crossed $150M in AI ARR in 2.5 years. Its first $150M business took eight years, and it had well over $100M in revenue when it chose to start over . It reports 20,000 agents working for customers, and says it is profitable and free-cash-flow positive . Paul Graham called it "founder mode": a hired CEO couldn't have abandoned the existing revenue .
Anthropic says its merged code changes went from about 2,500 a day in May to over 30,000 in September. It now sees source control, builds and CI as the bottleneck . It built a closed-source source-control system (Scion) and a code-review tool (Antique) to replace Git and GitHub . Together with the panel's view that making AI-written code production-ready is software's most important category , this points to developer infrastructure as an area to invest in.
Getty Images filed for Chapter 11, and public shareholders get nothing . Its active annual subscribers fell 25.2% to 240K, while editorial revenue grew 9.2% . That is one data point on generative AI eroding stock content while news coverage holds up.
Briefly
- Carbon-X: Hugging Face's 14.8B open DNA-and-text model averages 66.7% on five DNA benchmarks, versus 67.5% for GPT-6.1 Sol. It fits on one 80GB GPU and is not clinically validated .
- OCRia-1: LlamaIndex's document-parsing model claims top scores on ParseBench and three third-party benchmarks .
- Whop: reportedly raised $200M at $5.1B, led by Sequoia, as it moves into payments .