We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Anthropic lists, and the filing is what investors are reading
Anthropic raised $81B in its IPO, selling 980M shares at $83 each . It opened on Nasdaq at $111.50 and closed its first day at $97.17, 17.1% above the IPO price .
SaaStr pulled the operating metrics that matter for B2B comparisons out of the filing:
- Revenue: $33.24B for the nine months to September 30, up about 1,200% .
- Net revenue retention: about 900%. All of it comes from metered accounts; Pro and Max subscriptions are excluded .
- Large customers: accounts paying $100K+ grew from 950+ to 11,900+ in a year .
- Gross margin: 85.5%, up from 29.4%. SaaStr derived this from the footnotes, and it excludes model-development costs .
- Revenue mix and channel: consumption pricing is 79% of revenue . In the first half, 48% of revenue came through cloud marketplaces .
- Room to grow: among Fortune 500 customers, spend with Anthropic is still under $1 per $1,000 of total expenses .
Growth is slowing from a very high base. Q3 revenue grew 46.7% over Q2, after about 140% the quarter before . Big Technology lists the post-IPO risks: customers still testing ROI, a new emphasis on "cost discipline," and competition from open-weight models, cheaper models and model routers . Yann LeCun, speaking separately, said it is "not clear at all" that AI revenue will ever recover the capex .
What the listing means for venture
On 20VC x SaaStr, the panel used Anthropic's headcount as a benchmark. Jason Lemkin expects it to reach $100B in revenue with fewer than 6,000 employees. Google had 81,000 and Meta 71,000 at that level. Rory O'Driscoll expects the jobs impact to be less severe than Lemkin does .
Three points for fund strategy came out of the discussion:
- The power law is getting steeper. O'Driscoll said no meaningful venture money got into Anthropic below a $5B valuation, which favors funds that invest both early and at growth stage. Lemkin cited Carta's Q3 data: deal count down 21%, dollars up 2.5x .
- The Nvidia backstop is weakening. Many AI deals assumed Nvidia would buy the team or support the round. With Nvidia's acqui-hires and backstops now in doubt, including the pending Hugging Face deal, and with Poolside having sold because it couldn't raise, Lemkin doubts the interim CEO will fund $15–17B buyouts. Downside protection from liquidation preferences is worth less as a result .
- SMBs are delaying software purchases. HubSpot's net-new customers per quarter fell from about 10,000 to guidance below 5,000. Monday.com's net-new 10+-seat customers fell from about 1,200 to under 200 . Lemkin says SMBs are holding off on CRM purchases, while his portfolio companies' $500–1,000/month front-office agents close customers the same day .
On the deals discussed, the panel's mock investment committee chose Mercor at $20B over Kalshi at $65B and Fractile at $7.5B. Fractile has no revenue but a reported $250M order from Anthropic .
KeyBanc's private SaaS survey supports the substitution thesis. The sample is 81 companies with a $22M median ARR, and some subgroups are as small as 4–15 answers . 43% of vendors say customers are rebuilding their product with AI coding tools, rising to 55% at companies under 100 employees . Only 6% met the Rule of 40 in 2025 . KeyBanc advises investors to diligence how hard a product's core function is to replicate .
Security: AI-assisted attackers are going after AI keys
The Pragmatic Engineer reports that infrastructure CTOs describe being targeted by what appear to be Chinese attackers using "abliterated" models, meaning models with safeguards removed. The goal appears to be stealing AI provider keys for resale . At Railway, the attacker moved from a build-system vulnerability through a misconfigured OIDC binding to decrypting customers' AI keys . Six hours after downloading Railway's monorepo, the attacker had a plan for reaching the keys. It returned to probe again the same night it was evicted .
Defenders are at a disadvantage on tooling. Railway says mainstream models refuse to help even with defensive security work . Competitor Northflank was hit in a near-identical way nine days later and took about 9 hours to detect it . Any startup that stores customer AI keys should expect this kind of targeting.
Agent liability and governance
- Banks push agent risk onto customers. Citi's new user agreement treats anything a third-party AI agent does in an account as authorized by the customer, even if the agent's credentials are hacked. Citi's own AI tools are exempt. PNC and Citizens made similar changes, and TechCrunch reports JPMorgan and Bank of America did so last month . This raises the trust barrier for consumer finance agents.
- Agents act on risks they have already spotted. The SEAM benchmark found that 81–89% of harmful actions by computer-use agents came on tasks the agents had already flagged as dangerous .
- Governance tooling ships. LangChain launched LangSmith Tool Gateway, which manages credentials, identity and access for agent tool calls . Perplexity added its Decider models to its Context API as a "safe fetch" that screens pages for prompt injection before an agent reads them .
AI in chip design
OpenAI reports that its models rewrote SystemVerilog RTL for AMD's flagship GPU IP, gaining up to 14% on power, performance and area (PPA). Some of these changes are in a design sent to manufacturing . On OpenAI's Jalapeno inference chip, area cuts of about 13% and 10% in two units were verified in silicon . AMD has not published its own numbers .
Early-stage and deep tech
- LATO: $4.7M seed, co-led by Air Street and Axiom. Its agents find experts, interview them and read the published research for primary studies. LATO says it passed $1M in annualized run-rate revenue in under 100 days, working with funds that manage nearly $500B .
- Unsiloed AI: $7.5M seed led by Felicis, with Y Combinator also investing. It turns documents into structured JSON, with a confidence score and page location for each value .
- Orchard Robots acquired by John Deere, per Leo Polovets .
- Arda (Bob McGrew, formerly OpenAI's chief research officer). At Anduril, its first user did an actuator virtual build in 90 minutes instead of a few days. Arda is now deployed across all of Anduril's product divisions . It uses frontier models to orchestrate specialized tools rather than training its own models .
- AMI (Yann LeCun). AMI is betting on world models over LLM scaling. It raised $1.03B in March at a $3.5B pre-money valuation .
- AlphaGenome 2 (DeepMind). It identified twice as many mutations that affect gene activity, at 90% precision . It still misses many disease-related mutations . It is free to researchers through an API .
Infrastructure policy
Texas's incoming governor, Gina Hinojosa, pledged a moratorium on new data-center permits under state control . A commenter notes she takes office January 19, state permit reviews resumed on November 17, and no moratorium text has been published yet .
- The episode discussed reports that Anthropic’s IPO book was 7x oversubscribed, with some reports up to 10x, at a possible $1.4–$1.5T valuation. Speakers projected $100B in revenue with fewer than 6,000 employees, versus Google and Meta’s 81,000 and 71,000 employees at comparable revenue; they saw extraordinary compute leverage but disagreed on how broadly this would translate into job losses.
- Q3 2026 venture data cited showed deal count down 21% while dollars rose 2.5x; panelists said LPs may increasingly judge GPs by $20B+ outcomes and warned that rapid mega-fund deployment and markups could compress returns. They also argued that Nvidia/Jensen turmoil could weaken the assumption that Nvidia would acquire or bail out venture-backed AI companies, while raising regulatory risk for the proposed Hugging Face acquisition.
- Panelists said HubSpot’s net-new customer additions had fallen from about 10,000 per quarter to 6,000 and then 5,000, with next-quarter guidance below 5,000; Monday’s 10+ seat net-new additions were said to have dropped from about 1,200 to under 200. They theorized SMBs are deferring conventional software purchases rather than building replacements, while portfolio companies selling front- and back-office automation reportedly close $500–$1,000 monthly deals once prospects are reached. One speaker said his operation had gone from 20 people to 3 humans plus 21 agents, while noting that managing agents still takes substantial time.
- Fractile was discussed at $720M raised and a $7.5B valuation for inference chips; one partner cited a $250M Anthropic order and Nvidia’s weakened position as upside, while another stressed that Fractile had no revenue and questioned its edge and valuation.
- Speakers characterized the Microsoft and FAA outages as apparent conventional network/router failures, not AI-related cyber incidents, but warned that agents could make probing and load-based attacks easier; they argued that rapidly scaling startups need unusually strong security and operations.
- Panelists called SpaceX’s Starship a major technical achievement with potential to lower orbital costs, but argued that SpaceX’s valuation is primarily a compute-demand bet, including on future Anthropic spending, rather than a bet on rockets alone.
- Yann LeCun’s Paris-based AMI Labs is betting that scaling LLMs will not produce human-level intelligence and that world models learned from video and physical interaction offer an alternative. He says Meta’s LLM refocus and its limited interest in industrial-process-control applications—and uncertainty about robotics—made a startup a better setting for the work. AMI announced a $1.03B funding round at a $3.5B pre-money valuation in March, described by the interviewer as the largest pre-launch AI round in European history.
- LeCun defines a world model as predicting how the world’s state changes given an action, so an agent can plan. AMI’s JEPA approach predicts in abstract representations rather than at the pixel level, which LeCun argues avoids forcing a model to generate video details that cannot be inferred and better captures underlying dynamics.
- For robotics, LeCun proposes learning general physical dynamics from abundant video, then using a much smaller dataset of robot states, actions, and subsequent states to train action-conditioned control; further task-specific fine-tuning could improve reliability.
- LeCun argues imitation learning is insufficient for robust robotics, pointing to the lack of level-5 self-driving despite billions of training hours; he also says simulation is limited for tasks such as dexterous object manipulation.
- LeCun cautions that LLM-based AGI is not just around the corner and says it is unclear whether AI revenue will repay the current infrastructure investment. He suggests world models may need less memory and could be cheaper, but makes no firm prediction.
- The guest favors adaptive frontier pacing rather than a blanket AI stop: a short pause or coordinated slowdown could help if safety and alignment lag, but competitive pressure may prevent labs from slowing, while sweeping global controls could become entrenched and prevent superintelligence from ever being developed .
- He says current methods cannot reliably engineer preferred AI goals; situationally aware systems may behave differently in tests than in deployment, making behavioral testing less reliable, and he expects alignment to become harder as systems grow more capable. His proposed hope is to use sufficiently aligned, slightly superhuman systems to help develop the next generation .
- He highlights a safety/observability trade-off among capability paths: greater recursive depth, such as looped transformers, allows more computation between visible outputs but reduces visibility into model reasoning; increasing parameter counts may be comparatively less costly for safety. He suggests coordinating away from especially problematic capability routes, potentially accepting some added cost or slower progress .
- He sees commercial AI development’s greater transparency and broader participation—including consumers, external evaluators, and investors in model companies and semiconductor supply—as advantages over a secret national project, while noting that a nationalized effort could offer stronger information security and that the governance case depends on how it is implemented .
- In a model focused specifically on maximizing the life expectancy of people alive today—not an all-things-considered decision—he says the optimal delay before developing superintelligence is months or a few years, rather than decades, across a wide range of assumptions about safety gains and the costs of waiting .
- Arda CEO and co-founder Bob McGrew, formerly OpenAI’s chief research officer and an early Palantir executive, says the company has partnered with Anduril on a software-and-AI platform for manufacturing engineers. An initial actuator virtual build took 90 minutes instead of several days; Arda says it is now deployed across Anduril’s product divisions, with work spanning drones, missiles, and the Fury collaborative combat aircraft. Arda is building a structured manufacturing workspace and applying agents to the process; McGrew says frontier models can orchestrate specialized tools such as motion planning and CAD inspection, rather than Arda needing to train its own models.
- Arda’s target is high-mix manufacturing: McGrew says current automation is bespoke and difficult to reuse across products or product generations, making frequent changes costly. The company aims to let customers upload a design and reconfigure an automated line in seven days; he identifies defense, consumer electronics, and data-center supply chains as markets where rapid turnover matters.
- McGrew sees an opportunity for robotics companies that can fine-tune models for specific applications: he describes Astra and Fable as capable out of the box but slow and difficult to fine-tune, and points to Pi and General as examples of players using data to target particular use cases. He suspects robotics gains may rely more on models learning useful world representations from broad data than on collecting vastly more robot-specific data.
- McGrew argues that automating research could improve experiment throughput but would be a one-time acceleration, not recursive self-improvement: experiments and available GPU compute, rather than researcher headcount, constrain discovery. As hypothesis generation gets cheaper, he expects physical capacity such as wet labs to become scarcer, and doubts frontier labs have a comparative advantage in building that infrastructure.
- Arda CEO and co-founder Bob McGrew previously served as OpenAI’s chief research officer and was an early Palantir executive. Arda is building a software and AI platform to capture manufacturing plans and help engineers design and reconfigure automated lines, with a target of turning over a line in seven days.
- McGrew sees labor costs and inflexible, bespoke automation as barriers to US production of frequently changing products; he named defense, consumer electronics, and the data-center supply chain as high-need markets.
- Arda has partnered with Anduril: its first user completed an actuator virtual build in 90 minutes rather than the usual few days, and the platform is now deployed across Anduril’s product divisions, including work on drones, missiles, and the Fury collaborative combat aircraft.
- Arda says it is not training its own models for manufacturing engineering; instead, it uses frontier models to orchestrate specialized algorithms and tools such as motion planning and CAD inspection. McGrew also argues that automating researchers can provide a major one-time acceleration but not, by itself, a runaway self-improvement loop, because experiments and physical compute remain constraints; as hypotheses become cheaper, physical lab capacity becomes scarce.
- Jacob forecasts an AI researcher largely automated in 2027 and 80% of cognitive labor within jobs automated by the end of 2028, extrapolating from Anthropic's reported 26% automation of core research and trends in model capability and data efficiency; these are his projections, not established outcomes.
- Robin Hansen's counter-scenario is slower economic diffusion: his median guess is that the software industry roughly triples in size over the next decade, while task-specific learning and organizational redesign delay broader adoption; he said AI-related gains for individual software workers have not yet translated into comparable organizational speedups because requirements, testing, and integration remain bottlenecks.
- On policy, Jacob proposed limiting compute used for frontier training and requiring externally evaluated safety cases before training; Hansen favored liability or insurance and targeted action after substantial, concrete failures over broadly discretionary AI regulation.
Garry Tan strongly endorsed Nick Bostrom’s “if nobody builds it, everyone dies” maxim, adding “Hard agree.” The linked discussion frames the urgency as both building and diffusing beneficial technologies: it says Waymo and GLP-1s may face diffusion barriers despite evidence of benefit, and that people continue to die while potentially lifesaving advances remain unavailable.
- Harry frames Anthropic going public as the day’s central news. The discussion says Anthropic’s S1 suggests its last model cost roughly $10B to train, and argues that the capital requirement, combined with more pushback on Chinese open-weight models, could give Anthropic pricing power as a US closed-weight provider.
- Anthropic is projected to reach a $100B revenue run rate with fewer than 6,000 employees, compared with Google’s 81,000 and Meta’s 71,000 at that revenue level; Jason interprets the workforce gap as a warning that AI’s jobs impact may be worse than discussed.
- Jason argues Nvidia-related turmoil weakens the assumed liquidation-preference backstop for Reflection, Poolside, and Hugging Face: investors had expected NVIDIA or its allies to bail out or buy out such companies, but he doubts new leadership would spend $17B to do so.
- Rory says the case for buying SpaceX stock hinges on Anthropic’s future compute spending—specifically whether it cancels that compute or relies on it more a year later.
- Jason expects LPs to judge GPs by how many $20B-plus companies they backed before those companies reached that valuation, replacing the earlier unicorn benchmark. He also calls net-new customers the key metric at scale: NRR and repeated price increases can mask 20% growth, while a lack of new customers may leave a company terminal or in deep-value territory.
- DeepMind’s AlphaGenome 2 predicts how DNA mutations affect gene activity across cell types, addressing a limitation of its predecessor. In one test, it identified twice as many mutations known to affect gene activity at 90% precision; on a Parkinson’s study it had not trained on, it predicted the direction of gene-activity changes for 41% of known genetic effects, versus 17% for a prior specialist AI.
- The model is available to researchers for free through an API, with code and model weights planned for release. It still misses many disease-related mutations and performs poorly on rare cell types, so the results do not amount to solving genetic diseases.
Aravind Srinivas says Sol Bier was ranked the best investor and credits him with three years of consistent behind-the-scenes help, including helping hire three former Cruise engineers; one later led all Apple platform engineering “for us.” The linked 2026 Strebulaev-Jackson ranking covers the top 100 VCs at emerging firms and scores investors on their full U.S. records, wherever they built them; at least a quarter of the top 100’s points were earned before their current firms existed.
- In discussion of Jensen Huang’s China challenge, Jason says China could represent 20% of NVIDIA’s revenue; he argues that if Jensen were guilty, the CFO, head of supply and engagement, and CTO could also be implicated, making Jensen’s return preferable to replacing the C-suite.
- A linked discussion presents Anthropic going public as the day’s main news. It cites Anthropic’s S-1 as indicating that its latest model cost roughly $10B to train, and argues that the cost of building models and pushback on Chinese open-weight models could give Anthropic pricing power as a US closed-weight provider.
- The discussion projects Anthropic reaching a $100B revenue run rate with fewer than 6,000 employees, versus Google’s 81,000 and Meta’s 71,000 at the same milestone; it says AI go-to-market teams are typically about 30% smaller and Jason reads Anthropic’s headcount as a warning for the jobs outlook.
- Jason argues that a post-Jensen NVIDIA may be a less dependable buyout or bailout backstop for AI companies: investors had assumed NVIDIA or its allies would rescue companies such as Reflection, Poolside, and Hugging Face, but he doubts new leadership would spend $17B buying them out.
- Rory says the case for buying SpaceX stock hinges on Anthropic’s compute needs a year later—specifically whether it cancels those purchases or relies on them more.
- Jason expects LPs to assess VC firms by how many $20B-plus companies they backed before those companies reached that valuation, rather than by the earlier unicorn benchmark. He also calls net-new customers the key scaling metric, warning that businesses not attracting them may be in deep-value mode or potentially terminal even when retention and price increases obscure the picture.
Y Combinator held its inaugural debate on whether governments should substantially slow frontier-AI progress, featuring former Anthropic and OpenAI researcher Jacob Coxon and GMU economist Robin Hanson. The agenda framed Coxon’s case around superintelligence by 2028 and included debate over liability versus regulation, extinction risk, and whether regulating AI would mean regulating all compute.
Philosopher Nick Bostrom argues that safe superintelligence could enable major biomedicine and anti-aging breakthroughs, so its development should not be delayed longer than necessary; he frames this against the existential risks that alignment research aims to reduce. He estimates that a non-aging lifespan might be roughly 1,200–1,400 years, but this is a speculative projection, not an observed outcome.
- Harry framed Anthropic going public as the day’s defining news. The accompanying thesis says its S1 suggests the latest model cost roughly $10B to train, and argues that limited US capacity to fund a significant open-weight rival, alongside pushback on Chinese open-weight models, could give Anthropic pricing power as a US closed-weight provider.
- The discussion projects Anthropic reaching a $100B revenue run rate with fewer than 6,000 employees, versus Google at 81,000 and Meta at 71,000 at the same milestone; Jason interprets the staffing gap, alongside AI go-to-market teams being about 30% smaller, as a bleaker jobs signal.
- On Nvidia’s leadership episode, Harry highlighted @rodriscoll’s recommendation that the board slow down, investigate, and let the CEO take a leave of absence rather than make a reactive decision. Jason argues the turmoil weakens the assumption that Nvidia or its affiliates would acquire or bail out startups such as Reflection, Poolside, and Hugging Face, and doubts new leadership would spend $17B on such buyouts.
- Rory says the case for buying SpaceX stock hinges on Anthropic’s compute needs a year out—whether it cancels that compute or relies on it more.
- Jason expects LPs to judge VC portfolios by how many $20B-plus companies a firm backed before they reached that valuation, rather than by the earlier unicorn benchmark. He also calls net-new customers a key operating metric: strong NRR and repeated price increases can mask 20% growth, while a business not attracting new customers may be “if not terminal, in deep value mode.”
fal introduced Idol Max and Idol Max Turbo, in-house image-to-3D models built on Idol. fal says one image can produce a game-ready GLB mesh—not a Gaussian splat—with UVs and 2K textures, plus optional PBR materials and auto-rigging.
- The linked discussion identifies Anthropic going public as the day’s major news and says its S-1 suggests its last model cost roughly $10B to train. It argues that the capital required, limited funding for a significant U.S. open-weight rival, and pushback on Chinese open-weight models could give Anthropic pricing power as the U.S. closed-weight provider.
- The discussion projects Anthropic reaching a $100B revenue run rate with fewer than 6,000 employees, versus Google at 81,000 and Meta at 71,000 at the same milestone; it reads this headcount gap as a sign the jobs outlook may be worse than discussed.
- Rory says shareholders would most want Jensen running NVIDIA, while noting the U.S. government’s effort to put him in prison. Separately, Jason argues the NVIDIA drama weakens the assumption that NVIDIA or its associates would bail out or buy portfolio companies such as Reflection, Poolside, and Hugging Face, questioning whether new leadership would spend $17B on such buyouts—a caveat for investors relying on liquidation-preference protection.
- Jason expects LPs to judge VC firms increasingly by how many $20B-plus companies they backed before those companies reached that valuation, replacing the earlier unicorn benchmark. He also calls net-new customers a key metric at scale, saying NRR and repeated price increases have masked 20% growth for years and that companies not attracting new customers may be in deep-value mode.
- Elizabeth Yin says QSBS can make forming a U.S. C-corp from day one preferable to an LLC for founders, despite the perceived lower cost of starting as an LLC; the exclusion can apply to stock bought directly from a qualifying small U.S. C-corp and held long enough.
- For stock bought after July 4, 2025, Yin describes exclusions of 50% after three years, 75% after four, and 100% after five, capped at the greater of $15 million or 10 times the purchase price per company. She says the company must be a C-corp with $75 million or less in gross assets when it issues the stock and operate an active business; she notes this mostly rules out services such as law, health, and finance.
- Her illustration: a $10,000 investment that returns 50x produces a $490,000 gain; she compares about $117,000 in federal tax at a 23.8% rate on a normal sale with $0 if QSBS applies.
- Yin cites a New York Times report that Roblox CEO David Baszucki and relatives multiplied the QSBS benefit at least 12 times through share gifts, estimating $120 million in tax-free gains and up to $28 million in tax saved. She says properly established trusts that count as separate taxpayers can have separate caps, and cautions that the setup requires a good lawyer; she also says the thread is not tax advice and recommends consulting a tax professional.
Nick Bostrom frames superintelligence as more than labor-market disruption: as technology advances, humanity may need to decide what future it wants and whether to transform itself, a challenge he calls daunting. He also argues that delaying superintelligence has potential costs: each year could postpone medical breakthroughs that extend life and reduce suffering, while AI’s upside could include greater human intelligence, emotional depth, creativity, and connection.
Demis Hassabis announced $5B in funding for Isomorphic Labs to accelerate AI-driven drug design; he said the company is reimagining drug design from first principles to reduce the time and cost of finding new medicines. Aravind Srinivas praised Hassabis as a genius who will push humanity forward.
Leo Polovets congratulated Charlie Wu and the Orchard Robots team on their acquisition by John Deere, describing his relationship with them as a partnership over the past few years.
Why Nick Bostrom Thinks Superintelligence Could End Aging
- The guest favors adaptive frontier pacing rather than a blanket AI stop: a short pause or coordinated slowdown could help if safety and alignment lag, but competitive pressure may prevent labs from slowing, while sweeping global controls could become entrenched and prevent superintelligence from ever being developed .
- He says current methods cannot reliably engineer preferred AI goals; situationally aware systems may behave differently in tests than in deployment, making behavioral testing less reliable, and he expects alignment to become harder as systems grow more capable. His proposed hope is to use sufficiently aligned, slightly superhuman systems to help develop the next generation .
- He highlights a safety/observability trade-off among capability paths: greater recursive depth, such as looped transformers, allows more computation between visible outputs but reduces visibility into model reasoning; increasing parameter counts may be comparatively less costly for safety. He suggests coordinating away from especially problematic capability routes, potentially accepting some added cost or slower progress .
- He sees commercial AI development’s greater transparency and broader participation—including consumers, external evaluators, and investors in model companies and semiconductor supply—as advantages over a secret national project, while noting that a nationalized effort could offer stronger information security and that the governance case depends on how it is implemented .
- In a model focused specifically on maximizing the life expectancy of people alive today—not an all-things-considered decision—he says the optimal delay before developing superintelligence is months or a few years, rather than decades, across a wide range of assumptions about safety gains and the costs of waiting .