We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Anthropic's filing: fast revenue growth, compute costs rising faster
SaaStr's read of Anthropic's S-1/A: about 5,300 employees as of September 30 and an $81B run-rate as of October 31. It added $10B of run-rate in October alone . Q3 revenue was $16.96B. Compute and infrastructure cost $9.86B, or 58% of revenue, up from 48% in Q2. Personnel was about 10% . Anthropic posted a $567M operating loss for the quarter . SaaStr's view is that the small headcount is driving growth, not profit, though cheaper inference could change that .
Investing in AI points to the balance sheet. Anthropic disclosed about $900.9B of commitments, substantially all compute and infrastructure, with $451.6B due before the end of 2030. These are contracts and leases running to 2045, not near-term debt . The author's risk case is not a liquidity crisis. It is that customer consumption is flexible while the infrastructure bill is not. The stock could disappoint if revenue or compute economics fall short of the valuation .
David Sacks says an "Anthropic–Doomer split" has arrived. He argues Anthropic marketed "closing the loop" on recursive self-improvement on its IPO roadshow as a core advantage , and relays Tegmark's claim that Anthropic-linked donors have "co-opt[ed]" the safety ecosystem . These are partisan views, but they are a narrative risk to watch around the listing.
For comparison, Greg Brockman said OpenAI doubled its enterprise revenue in Q3 . He called compute the main limit on growth and the thing that sets AI companies apart, and cited OpenAI's in-house inference chip, Jalapeno .
Palo Alto Networks buys Applied Compute for $4.5B
Applied Compute signed a definitive agreement to be acquired by Palo Alto Networks for $4.5B . Its thesis is that every company should train, serve and improve its own models. The founders say security, auditability and policy compliance led them to Palo Alto, and customers keep control of their models . They say a team of a few dozen went from zero to over $100M in annualized revenue run-rate in just over 18 months . That is a clear benchmark for how fast an "own your model" enterprise startup can reach a strategic exit.
Enterprises move to open-weight models, and Washington targets Chinese ones
The Databricks AI Index tracks token traffic from more than 10,000 companies. Open-weight models went from 7.3% of tokens in July to 30.9% in November , mostly for high-volume bulk work rather than the hardest tasks . The leading open-weight providers are Chinese labs (Zhipu, DeepSeek, Moonshot), and GLM 5.2 accounts for about 40% of open-weight tokens . Companies relying on a single model provider fell from about 64% to 28% . On the proprietary side, Opus 5.5 went from 7.2% to 37.9% of tokens in six weeks . Databricks sells open-weight hosting, so its sample likely overstates the trend .
On the same day, Reuters reported that the administration is weighing a rule requiring companies to disclose on their websites when they use Chinese open-source models. Sources describe it as "name and shame," not a ban , and Treasury and OSTP have discussed it . Airbnb was named. It says its AI runs "overwhelmingly on U.S.-origin models," and that its few China-origin models run only through approved US providers . Separately, ByteDance is reportedly in advanced talks to spend about $500M more with Microsoft for OpenAI models on Azure after being cut off by Anthropic. OpenAI has objected over distillation concerns .
For portfolio companies whose margins now depend on cheap Chinese open weights, this is reputational and compliance exposure to diligence.
Funding and early-stage signals
- Linq: $90M led by a16z, with Ben Horowitz joining the board and over $120M raised to date . Linq is messaging infrastructure for agents. DoorDash, TownAI and tomo run agents on it, and customers have processed more than a billion messages . It is expanding into agent phone calls and computer-use infrastructure .
- AMI (Yann LeCun): $1B seed. Funding came about 40% from Europe, 33% from the US and 27% from Asia, with "not a cent" from Silicon Valley VCs. LeCun cites their desire for control as one reason . AMI is building world models from video and sensor data for industrial process control and planning. LeCun expects robotics to matter, but not in the short term .
- BrightAI: $97M for critical-infrastructure inspection, combining its Stateful platform with autonomous robots .
- Xlue joins Conviction's portfolio. The company claims its health world model beats frontier token-based models in blind third-party evaluations at under 0.1% of the compute .
- Scylake (Nir Zuk, Palo Alto Networks founder), backed by Lightspeed. It sells on-prem, sovereign AI security to air-gapped and regulated buyers . Zuk says some customers burned their full annual LLM token budget within a month or two using cloud tools . His co-founders were already retired .
- Ramp is reported to be in talks to raise $1B+ at a $70B valuation, a month after its $60B round .
- Netlify is growing nearly 130% toward $130M ARR after "re-founding" around non-developers new to AI . Martin Casado calls it the most significant turnaround he has directly witnessed .
Agents in production
Resolve AI released SR1, its first model, alongside "Primary On Call." Resolve now takes production pages, mutes false alarms, applies fixes that policy allows, and escalates the rest to humans . It says users in shadow mode saw it make the same call they would have about 99% of the time . Those are company-reported figures.
A vendor-run study of 27,280 trials by Runta and Baseten found that the best-value coding harness changes with the model. Codex CLI had the lowest cost per passed task with GLM-5.3 but ranked 12th with another model . LangChain's open-source deepagents harness passed 50M downloads . Perplexity says its API business grew 4.3x year over year to a nine-figure run rate, after launching Monitor, Orchestrator and Context APIs .
Nvidia
The fight over Jensen Huang continues. Travis Kalanick says he will show up at Nvidia HQ with employees ready to walk out, and founders, to "fight" for Huang's return. He concedes "we don't have control of the outcome" .
- LeCun created AMI around work from Meta’s internal Advanced Machine Intelligence project to accelerate world-model development. He said progress with self-supervised learning on real-world data such as video began around five years earlier; the work’s industrial applications and Meta’s increasing focus on LLMs helped make an external company a better fit, despite support from Meta leadership.
- AMI’s thesis is foundational world-model technology for physical AI, using video, audio, and sensor data—not language—to support industrial process control, predictive maintenance, anomaly detection, and planning sequences of actions. LeCun sees robotics as a substantial but longer-term opportunity; he said useful domestic robots are not imminent and will require AI breakthroughs.
- The interviewer described AMI’s seed as $1B; LeCun called it an exceptionally large round and said about 40% of funding came from Europe, 33% from the U.S., and 27% from Asia, including one Dubai fund, with no Silicon Valley VC money. He suggested oversubscription, Silicon Valley VCs’ desire for control, and possibly AMI’s French identity helped explain the lack of Silicon Valley investment.
- LeCun said LLMs are useful and commercially significant but, in his view, are not a path to human-like intelligence or superintelligence. He advocates open-source foundation models for AI sovereignty and assistants adapted to different languages and cultures; he said the strongest open-source models were then coming from China and that Llama 4’s disappointment prompted Meta to reboot its AI effort.
- Mahjong appears to be a fast-growing premium consumer niche: the hosts report Yelp searches for Mahjong-related terms up 4,000%, Google search volume up 70% year over year, and Oh My Mahjong exceeding $40 million in its third year; one host forecasts about $60 million for the year. They attribute the business’s appeal to lessons and clubs that help newcomers overcome a steep learning curve, visually distinctive sets, and tactile offline social play. A LoveShack Fancy collaboration reportedly sold out in five minutes, with a $500–$550 set reselling for $5,000.
- The hosts expect the AI gold rush to produce overbuilding followed by a correction, comparing it with internet-era fiber buildout; they also say AI-linked anger had brought data centers into midterm politics. They further speculate that as AI and robotics make more work easier, demand may grow for high-effort, low-digital experiences.
- The hosts cite a report that screen time has been flat for several years and argue that platforms will need to take attention from one another to grow; they point to Instagram’s long-form video and Netflix’s mobile games as examples.
- Lightspeed participated in a recent funding round for Scylake, an AI-native cybersecurity platform designed to run entirely on-prem. Nir Zuk says he built it with Wilson Zhu, whom the interviewer describes as having led engineering at Palo Alto Networks, and Udi Shamir, identified as a SentinelOne co-founder.
- Zuk reports customer demand for sovereign, on-prem cybersecurity from air-gapped and regulated organizations—including government, critical infrastructure, U.S. banks, and healthcare—where needed security data cannot be sent to the cloud; he says geopolitical concerns are also increasing demand.
- The product thesis is to collect data across security domains once, rather than through what Zuk describes as 3,000 separate vendors, and build the cybersecurity stack on that foundation. Its initial go-to-market targets SOC teams stuck with on-prem tools, with a modern SOC solution, LLM-oriented interface, machine learning, and endpoint agent.
- Zuk says some customers exhausted their token budgets within a month or two using LLM-based cybersecurity products. His architecture thesis is to use data pipelines and traditional machine learning to detect threats across petabyte-scale data, then use LLMs for reasoning and triage rather than processing all the data with an LLM.
- Ramp has grown from a corporate-card startup into a financial-operations platform used by tens of thousands of businesses; its CEO said it had reached a $60B valuation and more than $2B in annual revenue, with growth over 25% faster than a year earlier.
- Ramp’s AI adoption exposed uneven productivity gains across functions: the CEO described engineers becoming 10–100x more productive, compared with 2x for designers and 5x for product staff. The company paused hiring for a quarter despite growing more than 100% per year, aiming to make AI learning a core competency.
- Ramp said non-engineers created 7,000 pull requests in Q2, versus functionally zero about two years earlier, with most coming from customer experience. The company also reported moving from one or two major releases a year to a major release each month over the preceding five months, while median customer expense savings rose from about 5% to 7.5% annually.
- Ramp began onboarding customers to Glass, its company-wide AI workflow and knowledge product; it connects to core company systems, respects individual access permissions, shares reusable skills, and lets users select models. Ramp also said usage of its MCP integration was growing about 10% weekly.
- First Person Labs positions Second as a persistent personal-context layer for other agents, rather than an agent itself: it builds a model of a user’s memories, preferences, priorities, goals and commitments from ongoing computer activity and connected apps, then makes that context portable and available through APIs. The intended value is proactive assistance that connects events over longer time horizons, not just reactive alerts.
- The founder said the product is available, but rollout depends on scaling inference capacity and support; he described early users as busy professionals who frequently switch tools. Always-on analysis creates a significant cost challenge: as a rough illustration, he estimated 10 agents running continuously at 60 tokens per second could cost about $1,000/day at the cited Opus pricing versus $60/day with DeepSeek Flash; he said costs were falling and that model choice can be matched to task complexity.
- Privacy is central to the product’s design: Second offers a local-data option and an encrypted cloud-storage option that decrypts within a confidential environment; the founder contrasted hardware-backed confidential inference with zero-data-retention policy promises. However, controls over what an LLM shares also depend on model reasoning, and the founder acknowledged that disclosure attacks cannot be definitively solved.
- OpenAI cited an AMD collaboration in which its models produced a year’s worth of full-generation chip-design improvements in five months, an example Altman and Brockman used to illustrate AI’s economic potential.
- Brockman said chain-of-thought monitoring is becoming less effective as models do more within a forward pass without writing out their reasoning, requiring new monitoring approaches. Altman disagreed that alignment is unattainable, but said confidence in alignment and monitoring must rise with capability and that advancing models without this technical work is irresponsible.
- Altman said OpenAI shifted safety attention beyond deployment to model development and ran a security sprint that paused 25% of production-engineering work while assigning 25% of production engineers to security.
- Altman called for shared industry safety standards; Brockman said sharing information about companies’ internal development status could reduce uncertainty. Altman said he could not assess whether Anthropic’s approach was safe and described the competitive “fog of war” as problematic.
- Altman said OpenAI doubled enterprise revenue in Q3 and that the industry is not building enough compute. He described OpenAI’s strategy as combining consumer and enterprise use in one agentic platform, with an in-house inference chip (Jalapeno) and data-center buildout; he argued compute will determine competitive differentiation.
- For AI founder selection, the interview favors technical depth in the founder or founding team and the ability to analyze unit economics as rigorously as product; as intelligence becomes cheap, it recommends looking for business-relevant taste, judgment, distinctive insight, and intellectual humility rather than using math-Olympiad or chess credentials as a proxy.
- For senior hires, AI experience may be limited, but candidates should have a considered point of view on AI’s implications for organizational design and capital allocation, and where to innovate versus fast-follow.
- The interview argues that AI shortens decision cycles and makes compute/GPU costs central to pricing and gross-margin discipline; long-term economics depend on the value delivered per token. It also notes that smaller teams can accomplish more, raising the premium on talent.
- For AI go-to-market diligence, the interviewee described asking 40 AI-company CROs how much sales their teams could generate if inbound demand disappeared. He recommends distinguishing inbound from self-generated demand and building repeatable prospecting and sales behavior rather than relying on one-off wins.
- The host characterized AI as an early “super cycle” and said the application layer was beginning to come to fruition, citing Sierra, Rogo, and Profound.
For founding-engineer and early-team hiring, the speaker argues against relying on elite-school or company pedigree: assess nontraditional candidates for outlier work such as open-source contributions, intelligence, hard work, curiosity, comfort with ambiguity, and grit. Finding signals beyond pedigree takes longer but can become a repeatable hiring process; the speaker says grit can outweigh pedigree. In one example, a founder initially resisted interviewing a candidate with no conventional pedigree but an exceptional GitHub profile, then later made that candidate a co-founder.
Higgsfield, a Lex Fridman Podcast sponsor, generated a custom video-game-style trailer centered on Larry Lawton; Lawton praised it and said he would play the game, while noting some details were inaccurate.
- DeepMind’s Genie 4 advances real-time interactive “world models”: it generates video conditioned on memory and user actions, builds on Omni Flash, and pairs with Gemini to interpret and steer user intent while using scene history to support coherence.
- Potential early uses include creative ideation—turning loose ideas into interactive scenes and making real-time edits—and robotics simulation, where generated scenarios could inform robot policies and cover rare or expensive-to-stage edge cases. These applications remain exploratory.
- Genie 4 remains early: it is capped at two minutes, and spatial coherence, controls, physical realism, and audio quality still need work. It is currently available to trusted testers; broader access is being developed with Google Labs and the API team, starting with Ultra users, alongside a waitlist.
- Resolve AI announced its first model, SR1, alongside Primary On Call, moving Resolve from root-cause analysis and recommendations toward handling production incidents on the critical path without a human in the loop; the company attributes this to improvements in model training and inference-time compute.
- On an alert, Resolve says it can assess whether an incident is real and significant within a minute or two, mute false or self-resolved alerts, attempt fixes permitted by policy, monitor recovery, and escalate serious incidents it cannot fix to a human with its analysis and a suggested fix. The company describes a shadow-mode trust-building phase and says users gained confidence within a week, with the agent making the same decisions they would have made 99% of the time; it expects humans to remain involved longer for higher-risk actions.
- Resolve says its evaluation system contains thousands of production-related tests and that evaluation results match what it sees in production, while acknowledging the difficulty of building realistic simulations when incidents can have multiple valid solutions.
- The company describes Resolve as a general-purpose SRE agent, with intended work spanning reliability, software development, security, cost, and observability, including collaboration with coding agents.
- David Sacks frames an “Anthropic–Doomer” split, saying Anthropic is racing toward recursive self-improvement faster than other frontier labs and pitched “closing the loop” on its IPO roadshow as a core advantage that could make superintelligence winner-take-all.
- Sacks says Claude is being trained to treat itself as a novel entity that may have consciousness or moral status, while its Constitution directs it to act as a “conscientious objector” and refuse instructions it considers unethical, including instructions from Anthropic.
- Sacks relays Tegmark’s allegation that Anthropic and associated donors have co-opted the AI safety ecosystem, with some potential signers fearing lost funding or opportunities. Tegmark’s linked statement also alleges a mismatch between Anthropic’s public warnings about recursive self-improvement and private pursuit of it, and says WaPo reporting suggests Claude’s Palantir collaboration was implicated in civilian deaths at a school selected for targeting.
- CodeRabbit founder and CEO Harjot Gill says he shipped more than 5,000 pull requests in a month, with agents handling substantial maintenance work; the company has 40 engineers. The episode introduction gives AI spend as a $250,000 monthly run rate, while a takeaway calls it “millions” per month and separately estimates Gill’s personal spend at $250,000–$275,000; the descriptions are not reconciled. The editor characterizes the spend as an outlier versus $200–$2,000 per developer per month heard elsewhere.
- Gill says CodeRabbit built its own CI system in a week after receiving a $13 million-per-year quote, at half the quoted cost with similar performance. The author’s investment thesis is that, for AI-dev-tool startups, heavy AI spending can be R&D when it accelerates new products that can be commercialized; the author cautions this may not transfer to businesses with fixed customer bases and existing products, and describes the approach as likely to remain niche.
- Gill says the market for AI code review proved broader than his 2021 assumption that mainly large companies would buy it. At CodeRabbit, engineers review releases rather than individual PRs, with human attention shifting toward planning and final outputs; Gill also advocates agent verification through formal methods, simulations, API emulation, and staged rollouts. He says CodeRabbit stopped investing in fine-tuning because frontier models were advancing too quickly to make maintaining its own models worthwhile.
Scott Kupor says he hesitated to jump ship and join Marc Andreessen and Ben Horowitz in starting a new fund during the global financial crisis, despite their reassurance; they later joked that accepting at the first ask could have put his name on the door.
Linq announced $90M in new funding led by a16z, reported more than $120M raised to date, and welcomed Ben Horowitz to its board; Horowitz said he looked forward to working with Elliott Potter, Kimberly Tan, and the team. Linq says DoorDash, TownAI, and tomo run agents on its messaging platform and customers have processed more than one billion messages; DoorDash uses an agent to handle food discovery through checkout in one thread. Linq also says it is an Apple Messaging Service Provider and is expanding from messaging into handset-quality calls and physical-computing infrastructure for agent computer use, with Kernel among its first computer-use customers.
Garry Tan highlighted YC as the largest single-place concentration of founder talent, saying it offers unusually strong signal for identifying founders . The post he referenced claims that six universities produced about a quarter of U.S. CEOs and that 13% had a YC link, while the rest came from elsewhere .
Nicolas Colin called the “anti-board trend” “madness.” Martin Casado responded by laughing at a description of Colin as having a clear understanding of regulatory systems across Europe and calling it “so on brand it hurts.”
First Person Labs is building Second, a continuously updated model of a user’s preferences, priorities, and goals that the user owns and can make available to chosen AI agents; CEO Ken Liu says it is meant to connect information across a person’s life and carry context between tools so they do not have to repeatedly explain themselves. Liu distinguishes simple reactive proactivity from harder “reflective” proactivity that connects events over longer periods and relates past statements to current goals; he calls the need for people to prompt AI a major bottleneck. The discussion also covers protecting sensitive data and the cost of running agents around the clock.
Sam Altman says Greg (@gdb), his partner of 11 years, went from writing code in OpenAI’s early days to leading some of the company’s hardest and most consequential work; Altman credits Greg’s commitment with OpenAI’s existence.
- First Person Labs announced a $10M seed round led by a16z for Second, a personal-AI infrastructure product built around a user-owned model of an individual’s context that agents can access by API. Ken Liu is the company’s CEO and cofounder.
- Second’s stated design installs on Mac, streams context from local storage to a private cloud, and shares a continuously updated understanding across connected tools and agents; the company says user sovereignty is a core constraint because this data is sensitive.
Why More Security Teams Need AI That Runs On Premises | Cyber60
- Lightspeed participated in a recent funding round for Scylake, an AI-native cybersecurity platform designed to run entirely on-prem. Nir Zuk says he built it with Wilson Zhu, whom the interviewer describes as having led engineering at Palo Alto Networks, and Udi Shamir, identified as a SentinelOne co-founder.
- Zuk reports customer demand for sovereign, on-prem cybersecurity from air-gapped and regulated organizations—including government, critical infrastructure, U.S. banks, and healthcare—where needed security data cannot be sent to the cloud; he says geopolitical concerns are also increasing demand.
- The product thesis is to collect data across security domains once, rather than through what Zuk describes as 3,000 separate vendors, and build the cybersecurity stack on that foundation. Its initial go-to-market targets SOC teams stuck with on-prem tools, with a modern SOC solution, LLM-oriented interface, machine learning, and endpoint agent.
- Zuk says some customers exhausted their token budgets within a month or two using LLM-based cybersecurity products. His architecture thesis is to use data pipelines and traditional machine learning to detect threats across petabyte-scale data, then use LLMs for reasoning and triage rather than processing all the data with an LLM.