We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
The capital stack after Jensen
Huang's resignation was covered in the last brief. Two things are new. First, Nvidia's governance: Harvey C. Jones is the board's first-ever chair. Independent directors will keep investigating even though Huang has left, and Q3 earnings are still scheduled for Nov. 17 . Second, who will fund AI now. Newcomer reports that Huang personally led Nvidia's strategy of investing in its own customers, and that interim CEO Colette Kress is likely to rein it in . One podcast panelist estimated Nvidia often took 10–20% of the rounds it joined, and said "that money is going to dry up" . Databricks' Ali Ghodsi called the crisis "bad for the ecosystem, bad for neoclouds" . General Catalyst's Hemant Taneja suggested a weaker Nvidia could mean more competition in AI chips . Garry Tan signed the "StandWithJensen" letter calling for due process .
The capital Nvidia may stop providing is about to come from Anthropic instead. Its prospectus sets a $70–78 range, with pricing Nov. 18 and trading Nov. 19. Seven compute suppliers, including Nvidia, Google, AMD, Microsoft and TSMC, will buy $26B of non-voting stock at the IPO price. Bloomberg reports orders at about 6x the shares available . One trillion-dollar allocator told Harry Stebbings it is worried less about this $80B raise than about whether anyone can fund the next $100B, because that layer of capital is "very thin" . Separately, Greg Brockman said OpenAI's annualized revenue grew 70% in Q3 and that enterprise now accounts for most of it .
The mood at Newcomer's Cerebral Valley summit:
- Enterprise buyers are now focused on cutting AI costs, using purpose-built chips, open-source models and optimization .
- Startups worry that the frontier labs and Big Tech will control who gets access to the best models .
- Speakers said coding is the one clear killer app and that applications beyond it are "fuzzy" .
- 80% of attendees said AI is in a bubble, though most thought it was a problem for next year or later .
- The median company on the new AI 100 list is valued at $9.3B, almost double the year-ago figure .
Frontier prices are falling
Anthropic priced Fable 5.2 and Mythos 5.2 at $8/$40 per million input/output tokens, a 20% cut from the $10/$50 frontier level. The models are also about 30% faster, which lowers cost per token by roughly 23%. Jamin Ball reads the cut as passing efficiency gains through to customers during a land grab, and expects more of the same .
Garry Tan's open-source harness evals show that the agent wrapper around a model now drives cost more than it drives success rates:
- With the model and tasks held fixed, seven model–harness pairs all solved 71–79% of 20 tasks, and their confidence intervals overlapped.
- Cost per solved task was $0.17 for Codex, $0.22 for Claude Code and $0.44 for OpenClaw .
- When a knowledge base was connected with a company-wide token, OpenClaw and Hermes returned 121–123 records the requester was not authorized to see .
- Tan discloses that YC incubated QM, YC invested in Hermes, and gstack and gbrain are his own projects .
In a separate run of 52 tasks, Fable 5.2 matched Opus 5.5 (82.7% vs 84.6%) but cost 2.8x as much per solved task .
Aravind Srinivas says most of Perplexity's revenue growth this year has come from usage-based workflow automations, and calls net dollar retention "very high." He adds that routing tasks to cheaper models cuts usage revenue in the short term, but usage then grows enough to make up for it . He also says Perplexity's coding agents built a database that replaced DynamoDB for its trillion-page index, saving "several tens of millions" a year .
Deals
- LangChain: $300M Series C at a $3B valuation. It says more than 7,000 companies use LangSmith , including over 35% of the Fortune 350, and that its team has grown more than 4x this year to over 450 .
- Zipline: $1.5B at $21.5B post-money, 2.8x its January valuation, after Walmart signed on for 300 more sites . Rory O'Driscoll says the price is a heavy bet on the Walmart contract .
- fal: $1.3B at $16B, about 20x its reported $800M annualized revenue .
- Path Robotics: $300M Series E led by Georgian at $2.45B. Revenue is up 17x in 18 months, customers include HII (a $600M deal) and Lockheed , and its welding robot dog has 50 preorders .
- Intelligence (YC S25): reached a $60M annualized run rate in about a year. Its Design Arena product has 7M+ users and turns their real-world use into evaluation data for model labs .
- argenx–Unnatural Products: $550M upfront plus up to $595M in milestones for the macrocycle company, which is working on an oral drug against FcRn .
- Dialpad–Ada: an all-share deal valuing Ada at about $430M, against a reported $1.2B in 2021 . Jason Lemkin argues it offers no real liquidity .
Palantir grew Q3 revenue 96% to $2.313B , with net dollar retention of 165% . Bookings tell a different story: total contract value rose only 6%, and U.S. commercial contract value fell from $2.13B to $1.46B in a quarter. Management attributes this to shorter contracts .
Security and fragility
Wiz CTO Ami Luttwak says AI can now carry out in 20 minutes attacks that used to take skilled attackers weeks. He tells startups to build automated testing and patching into development from day one . a16z adds that developing a working exploit used to take a median of 22 days, and recent AI-assisted examples took hours .
Two large outages were not caused by AI or attacks, as far as anyone has said. Microsoft attributed its 365 outage to an earlier network-capacity upgrade, "not related to cyber or AI / agents" . The FAA grounded flights nationwide days after rolling out its SMART AI air-traffic tool across the country. The Transportation Secretary blamed a communications issue, DHS says it is "looking at" SMART, and the tool's maker denies any link .
Open models and frontier tech
- llama.cpp reached v1, its first fully stable major release .
- Thinking Machines released Inkling 2 Small's weights under Apache 2.0, about a week after launch. It is a 594B-parameter MoE with 17B active parameters .
- Fortune reports that OpenAI disbanded its robotics team, betting that its general models will run robots. This is from the headline only .
- Google DeepMind's AlphaGenome 2 identified 30% of variants believed to be causal across 28 immune cell types, up from 16% for the previous version .
- a16z's AI-pharma thesis says AI could raise drug-approval odds and shorten timelines. It currently takes about $2B and 13 years to develop a drug , and the firm expects wet-lab demand in the $90B CRO market to "explode" . It warns that biology is short on data and that how much feedback loops can actually be shortened is still unknown .
- The episode reports that Nvidia CEO Jensen resigned after 33 years and CFO Colette Crest became interim CEO; panelists said allegations of smuggling were unproven and the reasons for the departure remained uncertain. One panelist estimated Nvidia had supplied 10–20% of AI funding rounds it joined and predicted its investments could contract under CFO leadership; the discussion framed reduced circular financing as a possible de-risking, but warned that less Nvidia capital could also weigh on AI startups and infrastructure.
- Demis Hassabis described current coding tools as making human engineers roughly 3–5x more productive while people still review, architect, and write code; he identified closed-loop recursive self-improvement without human review as a riskier next stage, citing difficulty maintaining control and potentially alien outcomes.
- Anthropic put Claude Whiteboard and Sheets into beta, made Claude Data generally available, and launched Claude Meetings inside Google Meet on paid plans. The episode also reports Claude connectors for apps including Notion, Asana, Canva, and Dropbox, supporting a strategy of doing work through an agent interface; panelists noted that incumbents retain advantages in distribution, enterprise permissions, compliance, and installed base.
- Planet Labs launched Planet AI, a natural-language search app for a 10-year archive of more than 3,000 layers of daily satellite imagery; it was free for eligible users with usage limits. Separately, a researcher said 48 agents searched 1,200 km² of public NASA lidar data and flagged a possible Mayan site; Mexico’s National Institute of Anthropology reportedly said it appeared to be an original finding.
- The episode presented Jensen Huang’s reported departure from Nvidia as a breaking story based on partial information: it said Nvidia’s CFO was serving as interim CEO while the board searched for a successor, and the stock had fallen 15%; the speakers cautioned that the story was incomplete.
- Panelists said Huang had been central to Nvidia’s capital backing of startups, data centers, and infrastructure, so his departure could slow AI investment. They cautioned that rivals could not immediately replace Nvidia’s supply: scaling chips and securing foundry and HBM capacity remain constraints, though Huawei could benefit if Nvidia’s pace faltered for years.
- The discussion said Anthropic’s IPO target was 46% above its last private-round valuation after its run rate doubled; it cited an analyst bull case of a $4.2T valuation on $281B in projected 2028 revenue and said demand was six times oversubscribed. Speakers also warned that Nvidia-related and broader AI-market risks could affect IPO pricing.
- Anthropic was described as expanding Claude into a broad software bundle—including Meetings, Whiteboard, Sheets, Personal Finance, and Legal—with speakers arguing that bundling could challenge SaaS incumbents and reduce interoperability friction. The discussion also characterized Anthropic as marketing recursive self-improvement and winner-take-all prospects, while raising safety concerns.
- A panelist working in biology said AI was already accelerating molecular discovery, in-silico testing, and product development, citing weekly breakthroughs and use of Anthropic for substantial biology workloads; this was presented as personal experience, not a quantified industry-wide result.
- The guest says AI models can now find and exploit complex vulnerabilities, open-source models are catching up, and attacks that took top attackers weeks or months can be carried out in about 20 minutes. The tracked list of attacked AI-development platforms grew from 9 to 15 in October; the guest identifies the platforms’ AI keys as a reason they are targeted.
- The PHP incident illustrates the patching gap: about half of customers patched within the first 2–3 days, but the guest says resolving the widespread vulnerability could take years and some older PHP versions receive no patch; vulnerability information can leak before a patch is available, and Linux fixes may take weeks or months to reach distributions.
- For startups, the guest recommends designing from the outset for automated patching across development, testing, CI/CD, and DevOps. For startups holding AI keys, the advice is to limit key usage to cap charges if keys are stolen and to prioritize secure design, including sandboxing and isolation.
- The guest argues that defenders should continuously test unauthenticated, internet-exposed attack surfaces from the outside, prioritizing this over broad code scanning. The guest says red-agent use is nearing full coverage among its customers, while adoption across the industry remains insufficient.
- For AI products at scale, the guest describes cost-aware model routing as a core engineering challenge: combine models with different capabilities, speed, and cost, using deterministic and cheaper steps before escalating to more expensive models rather than running the most advanced model on every event.
-
Garry Tan shared the open-source
g-harness-evalsrepository, which compares agent harnesses while holding models and tasks fixed. In 20 development tasks, seven model–harness configurations scored 71–79%, with overlapping 95% confidence intervals; reported cost per successful task was $0.17 for Codex, $0.22 for Claude Code, and $0.44 for OpenClaw. In the same evaluation, initial context sizes ranged from about 12,900 tokens for Codex to 20,613 for Hermes; in one Sonnet test, prompt-cache reads and writes accounted for 87% of OpenClaw’s bill. - The tests also surfaced agent execution and access-control risks: gstack’s instructions loaded into Opus 5.5 but its steps were never executed across 46 tests, while a hook was bypassed through Bash 40 times. With gbrain using a company-wide token, OpenClaw and Hermes returned 123 and 121 records, respectively, that requesters were not authorized to see; scoping retrieval to the requester’s permissions reduced unauthorized results to zero.
- The repository reportedly includes raw records that can be checked locally without an API key. Tan disclosed that QM is YC-incubated, YC invested in Hermes, and gstack and gbrain are his own projects—relevant context for weighing the comparisons.
- Perplexity is shifting focus from mass-market consumers toward high-value business users, entrepreneurs and enterprises. Srinivas said usage- or token-based revenue is growing through repeat workflow automations such as research, monitoring, marketing and corporate development; customers may use up included credits and pay for more. He described revenue as sticky with high net dollar retention, while noting that lowering task costs can temporarily reduce usage-based annualized revenue even as usage expansion lifts total revenue.
- Perplexity’s product strategy extends from accurate answers into deep research, browsing agents, cloud-based agents and persistent work teammates. Srinivas argued that a model-agnostic harness—backed by its own sandboxes, routing, connectors, search and browsing—is a key source of value as models improve, with safety and enterprise access controls also important.
- Srinivas said Perplexity used coding agents to build a database from scratch and replace its use of AWS DynamoDB for its web-page index, saving the company several tens of millions of dollars annually.
- Srinivas said NVIDIA is one of Perplexity’s largest investors and described multi-year collaboration on inference libraries, open-source work and Neotron training data, with further projects planned. He also said Perplexity had been post-training its own models and lowering serving costs, and put a possible IPO window at anytime from then through 2028, while declining to share numbers.
- Market backdrop: a16z puts the top 30 pharmaceutical companies at roughly $5T in market capitalization and Eli Lilly above $1T; analysts cited in the article expect Lilly’s GLP-1 portfolio to exceed $100B in annual sales by 2030.
- a16z’s AI thesis is that improving drug-candidate success rates and shortening development timelines could lower the risk-adjusted cost of medicines; it cites an average 13-year development process, nearly $2B present-value cost, and fewer than 5% of candidates reaching FDA approval. It expects faster candidate generation to increase demand for experimental data, pointing to data providers and lab automation in a $90B CRO market ($30B discovery/preclinical and $60B clinical trials).
- The startup opportunity spans selling data and technology—including lab automation, predictive biology, scientific-reasoning agents, molecular design, and model-deployment tools—or building higher-risk, more capital-intensive full-stack drug developers with greater potential upside. Examples of full-stack companies named include Recursion, Isomorphic, insitro, Xaira, Relay Tx, and Octant Bio.
- Pharma partnerships offer an early commercial signal: a16z cites Isomorphic Labs agreements with Novartis and Eli Lilly worth up to $1.2B and $1.7B, respectively, in potential milestone payments, but says the key test is whether model gains produce better drugs entering clinical trials. The thesis remains uncertain: biology is data-poor and complex, drug designs still require wet-lab and human validation over long timelines, and how much startups can shorten feedback loops is unknown.
- Drug development averages 13 years and nearly $2B in present-value cost, with fewer than 5% of drug candidates reaching FDA approval; AI could increase success odds and shorten timelines, potentially making otherwise uneconomic targets viable. Biology remains data-poor, and drug designs still require wet-lab and human validation over long feedback loops.
- Startup opportunities span biopharma tools and infrastructure—data and lab automation, predictive biology, scientific reasoning, molecular design, and enterprise deployment—and full-stack companies developing their own medicines; the latter require more capital and risk but offer greater upside. The article puts the CRO market at $90B and expects demand for experimental data and wet-lab services to grow as candidate generation speeds up.
- Pharma partnerships provide early commercial signals: Isomorphic Labs has agreements with potential milestone payments of $1.2B with Novartis and $1.7B with Eli Lilly, while Chai, Boltz, BigHat Bio, Nabla Bio, Genesis Molecular AI, and Profluent also have pharma partnerships. The article says the key test is whether model progress produces better drugs entering clinical trials; Anthropic has opened a wet lab and announced drug candidates, while OpenAI acquired Orchestra Bio to build biotech execution capabilities.
- a16z frames pharma as a hits-driven market: the top quarter of drugs generate about 85% of revenue, while successful drugs can deliver years of patent-protected cash flow. It argues AI could raise drug-development success rates and reduce time and cost; the article puts current development at nearly $2B in present-value cost and 13 years on average, with fewer than 5% of candidates reaching FDA approval.
- Startups can sell data and technology—such as lab automation, predictive biology models, scientific reasoning, molecular design, and enterprise deployment—or build vertically integrated AI therapeutics companies. The latter approach requires more capital and carries more risk but offers greater upside through ownership of drug assets.
- Pharma partnerships offer an early commercial signal: Isomorphic Labs has agreements with Novartis and Eli Lilly involving $1.2B and $1.7B in potential milestone payments, respectively; a16z says the key test is whether model advances produce better drugs entering clinical development. Translation remains uncertain because biology is data-poor and drug designs ultimately require wet-lab and human validation on long timelines.
- Google DeepMind’s AlphaGenome 2 predicts DNA effects for a specified tissue, cell type, cell state, or data from previously unseen cells; trained on more than 107 million cell profiles across 68 human and mouse tissues, it identified 30% of variants considered causal across 28 immune cell types, versus 16% for AlphaGenome 1. The source says it still misses most variants, but could help researchers computationally prioritize disease-linked variants that often require real-world experiments to interpret.
- Argenx announced it would acquire Unnatural Products for $550 million upfront plus up to $595 million in development and regulatory milestones. The companies had partnered since the prior year; their work includes an oral macrocycle targeting FcRn, and the Unnatural Products team is to become argenx’s peptide-discovery hub.
- Path Robotics raised a $300 million Series E led by Georgian at a $2.45 billion valuation; the article reports revenue grew 17x in 18 months. Its Obsidian system plans and adjusts welds for variable joints rather than repeating a fixed path; named customers include Lockheed Martin, Saronic, and HII, with the article also reporting a $600 million HII deal. Its welding robot dog had 50 preorders for shipyards and construction sites.
- SpaceX ordered a turnkey heterojunction solar-cell production line from SINGULUS for delivery to its Bastrop satellite factory in the first half of 2027; the first line is worth a little over €60 million, with further orders and more than 100 GW of production equipment described as the plan. The article cautions that the current commitment is €60 million and treats the much larger scale as an ambitious promise, not established deployment.
YC CEO Garry Tan publicly backed Jensen, saying he deserves due process. The linked open letter calls for a fair process and the truth, citing Jensen’s 33 years building NVIDIA.
In Garry Tan’s 52-task coding evaluation on Claude Code 2.1.322, Fable 5.2 solved nearly as many tasks as Opus 5.5 (82.7% vs. 84.6%) but cost 2.80× more per solved task; Tan recommends Opus 5.5 as the coding default. He cautions that Opus already solves most of these tasks, so harder tasks are needed to distinguish the models.
Arav Srinivas highlighted a Perplexity demo in which rendering was orchestrated across 48 sandboxes in parallel, completing the process in three hours . The linked report describes a “Video Production teammate” used to recreate the Pantheon, gather survey dimensions and recent photos, build a three.js renderer with procedural shaders, and produce 60-fps video with motion blur; each frame took 80 seconds to render .
Garry Tan says he made an “idea-minder” skill intended to let a personal agent recursively self-improve as it is used, and links a page titled “idea-miner-what-to-build.”
Intelligence (S25) is building benchmarks for hard-to-measure AI capabilities; its first product, Design Arena, lets people use frontier models to create websites, images, videos, and other creative work, turning those interactions into data that helps model labs assess where models succeed or fail. Design Arena has more than 7 million users across 192 countries, and @Intelligence_ai reached a $60 million annualized run rate in about a year.
- a16z says a working exploit that once took a professional a median of 22 days to develop has taken hours in recent AI-assisted examples.
- Wiz co-founder and CTO Ami Luttwak warns that after attackers gain access, an attack can take 20 minutes to steal secrets and customers’ AI keys; startups can be targets even if they are not large. He advises limiting AI-key usage to curb charges if keys are stolen and prioritizing security when targeted.
Jason says the leak of Rubio’s memo prompted internal finger-pointing in the administration and estimates a greater-than-50% chance that the New York Times’ source was China hawks inside the administration; this is his prediction, not a confirmed attribution. In the linked post, he says those hawks may favor a total ban on chips to China for national-security reasons, and separately speculates that the CCP could seek to weaken the U.S. AI lead; neither is evidence of an announced policy change or a confirmed leak motive.
Srinivas called a VC-diligence vignette “Interesting”: a Series B SaaS company claims an organization as a customer, but a colleague says it still has a two-year contract while having stopped using the tool months earlier because “Computer” did the job 3× better. For investors, the scenario cautions that contracted customer status may not reflect active use or retention; it is presented as a vignette, not a verified churn event.
- Wiz co-founder and CTO Ami Luttwak says AI can compress cyberattacks that once took skilled attackers weeks or months into 20 minutes; he describes attackers stealing startup secrets and customer AI keys within that window, and warns that any startup holding something valuable can be targeted, not just large companies.
- For startups, Luttwak recommends limiting AI-key usage to curb charges if keys are stolen, while a16z’s discussion emphasizes building automated security testing and patching into development from the outset.
Jason says many people were tuning in live on X and that he is focusing more on X Live now that it has a new Clipping feature; this is an anecdotal creator-engagement signal, not evidence of broader adoption.
Wiz co-founder and CTO Ami Luttwak says AI can carry out cyberattacks in 20 minutes that previously took skilled attackers weeks or months. He argues companies need to continuously test their own systems with the latest AI models, and startups should build automated testing and patching into development from the outset; the discussion also emphasizes security architecture over simply buying more tools.
