Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/NEWCOMER NEWSLETTER/13 Videos from CVAI: Igor Babusc…
NEWS
// NEWSLETTER ISSUE
NEWCOMER NEWSLETTER

13 Videos from CVAI: Igor Babuschkin, Ryan Daniels, Guillermo Rauch & More

DATE November 17, 2026SOURCE NEWCOMER NEWSLETTERPARTICIPANTS ERIC NEWCOMER
In this episode
// SUMMARY

1. Key Themes

Theme: Multi-model, cost-optimized AI stacks are replacing single-model reliance

Enterprises are mixing models to cut costs

Several operators independently described the same shift: customers are no longer committing to one frontier model.

  • Deeter: "most companies that are customers of the big labs aren't sticking to a single model type, but are mixing and matching different models for different tasks."
  • Rauch: "companies are moving past expensive 'token-maxxing' and mixing small, large, fine-tuned, and open-weight models to cut costs."
  • Ghodsi: "Databricks routes more than four quadrillion tokens across OpenAI, Anthropic, Google, and open-source models, and its new router sends simpler requests to cheaper models."

Open-weight adoption is accelerating

  • Grinberg: "open models have grown from almost nothing at the start of 2026 to well over 10% of the tokens running through Factory, with some customers using them more than half of the time."
  • Ghodsi: "companies are focused on cutting costs to make AI pay off."

Theme: Concentration risk, in labs and in Nvidia leadership

Decentralized, owned models as a safety and business thesis

  • Babuschkin: "concentrating AI in a few labs is the biggest risk in the industry today."
  • River AI "wants to give businesses, and eventually individuals, models they can own and improve themselves," and "its first product lets companies post-train and run models on their own hardware."
  • Safety angle: "a world with many moderately capable models handling work like insurance claims and accounting is also safer, because no single company or person could do outsized harm."

Jensen Huang's resignation as a systemic shock

The summit took place against Huang's "forced resignation," and speakers framed it as a financing and ecosystem risk, not just a personnel change.

  • Ghodsi: "Huang's exit removes a financial backstop for the AI industry, since Huang was quick to support companies whenever investors hesitated."
  • Taneja: the shakeup raises "huge concerns" about continuity in AI leadership, "but could also open the door for more chip startups."
  • Wachen: the resignation "hit hard at a company that employs more than 100 former Nvidia workers."
  • Arora: "expects Huang's resignation to cause short-term instability but doesn't think it will seriously hurt the company."

Theme: Agents are reshaping labor, services, and the data/identity layer

Agentic workers create new white-collar capacity and new business models

  • Wachen "predicted the largest growth in white-collar work in history, driven not by people but by agentic workers."
  • Daniels: Crosby is an "agentic law firm" that is "the first company he knows of that sells legal AI directly to people who aren't lawyers," and chose to work on "getting rid of the billable hour" instead of buying law firms.
  • Foody: "today's agents need top engineers, lawyers, doctors, and bankers to design evaluations and reinforcement learning environments."

Proprietary data and identity become competitive moats

  • Rauch: "To stay competitive, businesses will have to give AI agents access to proprietary data, with identity protocols on top so agents can act for users with secure permissions."

Theme: Security is shifting to AI-vs-AI defense

Defending against agentic attackers requires many tailored models

  • Arora: "defending against AI agent attackers will take many models, not one," with systems trained "on how data normally moves through a given industry, then tailor[ed] to each customer, one at a time."
  • Qiao described Fireworks' incident "as an attack aimed at stealing compute and tokens to resell," handled with Palo Alto Networks forensics and "detection tools from CrowdStrike and Wiz."

Theme: AI value is vast but concentrated, with IPO timing in question

Valuation concentration in foundation labs

  • "The companies on the list are worth a combined $3.6 trillion, 36 times the value of the first Cloud 100 in 2016," and "a lot of this value is concentrated in the largest foundation model labs."

IPO window debate

  • Ghodsi "still thinks 2026 is a bad year to go public, but said Databricks won't stay private forever."
  • Deeter discussed "why 2027 could bring a wave of AI IPOs."

2. Contrarian Perspectives

Perspective: The humanoid robotics boom lacks a business model Younis is "skeptical of humanoid robots, saying no one has figured out how to make money from them yet and that scale alone won't solve robotics." He sees stronger pull elsewhere: "more demand for his company's autonomous technology from drone manufacturers than from self-driving carmakers."

Perspective: Great chips won't matter as much, but the physical integration problem will Wachen "expects highly capable agents to make chip design less valuable, while physical problems, like integrating chips and interconnects into data center racks, stay hard to solve." This challenges the view that silicon design is the durable moat, pointing to systems integration instead.

Perspective: Small research labs can out-innovate frontier labs on efficiency Spector "isn't worried about frontier labs like OpenAI and Anthropic moving into the same space," because "the lab's edge isn't being smarter or working harder, but having different incentives and no innovator's dilemma."

(Also notable: Srinivas claims "anyone who bets against Perplexity now will lose money again," after the audience voted it the company they'd most like to short last year.)


3. Companies Identified

Palo Alto Networks, cybersecurity platform. Mentioned for its CEO's views on AI-era defense and for assisting Fireworks. Quote: "Fireworks worked with Palo Alto Networks on forensics."

Bessemer Venture Partners, VC firm. Co-produced the AI 100 with Newcomer. Quote: "The companies on the list are worth a combined $3.6 trillion."

General Catalyst, VC firm. CEO commented on Nvidia and China. Quote: "America's 'bottom-up entrepreneurial ecosystem' is still far ahead."

Vercel, developer/web platform. Mentioned for its view on multi-model cost reduction and agent access. Quote: "companies are moving past expensive 'token-maxxing.'"

River AI, startup building customer-owned models. Mentioned for its decentralization thesis. Quote: "Its first product lets companies post-train and run models on their own hardware."

Etched, AI chip and rack maker. Mentioned for its Nvidia-alumni workforce and infrastructure outlook. Quote: "employs more than 100 former Nvidia workers."

Crosby, legal AI startup. Mentioned as a novel "agentic law firm." Quote: "it has lawyers, liability insurance, and a law firm entity."

Fireworks, AI inference platform. Mentioned for its security incident and open-model ecosystem role. Quote: "an attack aimed at stealing compute and tokens to resell."

Factory, AI software-development agent company. Mentioned as an open-model adoption data point. Quote: "well over 10% of the tokens running through Factory."

Applied Intuition, autonomy software. Mentioned for drone demand and humanoid skepticism. Quote: "more demand... from drone manufacturers than from self-driving carmakers."

Mercor, expert-labor/AI-training marketplace. Mentioned for the shift in who trains models. Quote: "today's agents need top engineers, lawyers, doctors, and bankers."

Flapping Airplanes, AI efficiency research lab, backed by Nvidia. Quote: "Its framework makes chips both more efficient and more capable. Nvidia is an investor."

Databricks, data/AI platform. Mentioned for token routing scale and IPO stance. Quote: "routes more than four quadrillion tokens across OpenAI, Anthropic, Google, and open-source models."

Perplexity, AI search company. Mentioned for its pivot. Quote: "shifted to serving mostly enterprise customers who can afford the token costs... more of a harness and infrastructure play than a chatbot."

Anthropic / OpenAI, frontier labs. Quote: Deeter discussed "why Anthropic ranks ahead of OpenAI."

Nvidia, chipmaker. Central to the Huang resignation discussion. Quote: "Jensen Huang's forced resignation."

Stripe, payments. Quote: Deeter discussed "how Databricks and Stripe fit into the AI economy."

CrowdStrike and Wiz, security vendors. Quote: Fireworks "used detection tools from CrowdStrike and Wiz."


4. People Identified

Nikesh Arora, CEO, Palo Alto Networks. Quoted on AI-agent defense. Quote: "defending against AI agent attackers will take many models, not one."

Byron Deeter, Partner, Bessemer. Quoted on multi-model usage and IPO timing. Quote: "mixing and matching different models for different tasks."

Talia Goldberg, Partner, Bessemer. Presented the AI 100 alongside Deeter. Quote: Deeter and Goldberg "gave a rundown on the AI 100."

Hemant Taneja, CEO, General Catalyst. Quoted on Nvidia continuity, China, and public messaging. Quote: the industry "should talk less about job losses and more about the 'abundance' and 'agency' AI can bring."

Guillermo Rauch, CEO, Vercel. Quoted on model mixing and agent identity. Quote: "identity protocols on top so agents can act for users with secure permissions."

Igor Babuschkin, xAI co-founder, now River AI. Quoted on lab concentration risk. Quote: "concentrating AI in a few labs is the biggest risk in the industry today."

Robert Wachen, co-founder/president, Etched. Quoted on agentic white-collar growth. Quote: "the largest growth in white-collar work in history, driven not by people but by agentic workers."

Ryan Daniels, CEO, Crosby. Quoted on legal AI and the billable hour. Quote: "getting rid of the billable hour."

Lin Qiao, CEO, Fireworks. Quoted on the security incident. Quote: "an attack aimed at stealing compute and tokens to resell."

Matan Grinberg, CEO, Factory. Quoted on open-model usage growth. Quote: "some customers using them more than half of the time."

Qasar Younis, founder, Applied Intuition. Quoted on anti-AI sentiment and robotics. Quote: "America is the most anti-AI country he's seen so far."

Brendan Foody, CEO, Mercor. Quoted on who trains models. Quote: the market "is limited by how many tasks humans can still do better than models."

Asher Spector, co-founder, Flapping Airplanes. Quoted on research-lab incentives. Quote: "different incentives and no innovator's dilemma."

Ali Ghodsi, CEO, Databricks. Quoted on the Nvidia backstop and IPO timing. Quote: "Huang's exit removes a financial backstop for the AI industry."

Aravind Srinivas, CEO, Perplexity. Quoted on critics. Quote: "anyone who bets against Perplexity now will lose money again."

Jensen Huang, Nvidia (departing). Central subject of the week's turbulence. Quote: "Jensen Huang's forced resignation."

Eric Newcomer, host and author. Quote: Deeter "joined Eric Newcomer to discuss why Anthropic ranks ahead of OpenAI."


5. Operating Insights

Route workloads by difficulty to cut inference costs

Ghodsi's "new router sends simpler requests to cheaper models," and Rauch says companies are "mixing small, large, fine-tuned, and open-weight models to cut costs." Operators should build or buy routing layers rather than defaulting to a frontier model.

Make proprietary data and agent permissions a first-class product investment

Rauch: "businesses will have to give AI agents access to proprietary data, with identity protocols on top so agents can act for users with secure permissions." Plan the data-access and identity architecture now.

Rethink service economics around AI-native delivery

Daniels built the structure of a law firm ("lawyers, liability insurance, and a law firm entity") but aims to replace the billable hour by "hiring lawyers who want to hand their work off to AI agents." Consider whether regulated services can be rebuilt with agents doing the work and licensed professionals providing accountability.


6. Overlooked Insights

Security incidents are increasingly about stealing compute

Qiao described the Fireworks breach as aimed at "stealing compute and tokens to resell," suggesting inference capacity is now a directly monetizable attack target, which has implications for how AI infrastructure providers should monitor usage.

The labor market for AI training is capped by human advantage

Foody said the market for expert training work "is limited by how many tasks humans can still do better than models." This implies the data-labeling and RL-environment business has a built-in ceiling as models improve, worth considering when evaluating the sector.