We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
OpenAI's smaller round and Anthropic's slowdown point to pricing, not demand
OpenAI has secured about $11B of the $30B it was seeking. It now aims for up to $20B at a $1.25T valuation, after pitching $1.4T a week before the deadline, according to WSJ, FT and CNBC reports summarized by SaaStr . Public markets were falling at the same time. One widely shared post claimed $2.8T was wiped from US stocks in a single day, which would be the largest one-day loss on record .
On 20VC x SaaStr, Rory said the lower OpenAI price is ordinary relative pricing against a larger, profitable Anthropic that is about to list. His view is that demand, not financing conditions, will decide the outcome. If demand weakens, leveraged infrastructure companies like CoreWeave get hurt before Google or Amazon . Jason was more cautious. He cited about $5T lost from AI stocks in a week, CoreWeave debt paying 15%, and Broadcom guaranteeing Anthropic debt .
On Anthropic, the panel used GAAP revenue figures from the S-1 amendment: about $5B in Q1, $11B in Q2 and $17B in Q3. Quarter-over-quarter growth roughly halved, from about 100% to about 50% . Jason blamed pricing pressure, not demand. He said OpenRouter data shows no decline in demand, and that OpenAI is selling roughly equivalent products at 30% off. He also gave his own example: moving SaaStr's AI matching app to Haiku 5.5 and a cheaper model mix cut its costs by 72% .
Enterprises are reining in spend too. Apple is reported to have cut engineers' daily AI limits from $300 to $200, with $100 planned . One practitioner said switching a colleague to DeepSeek v4.1 Flash cut daily spend by 10–20x, with no complaints about quality so far .
For investors: if mature AI apps are seeing cost drops like these, gross margins at application companies should improve. Model labs' ability to charge premium prices is what's under pressure.
Late-stage rounds still closed, often on momentum
Several large private rounds were reported:
- Zipline: raised $1.5B at $21.5B post-money, nearly triple its valuation from about 10 months ago. Uber holds a strategic stake . The money is going toward a nationwide drone delivery network . Amazon plans to expand drone delivery from 11 locations to about 500 .
- Kalshi: raised $1B at $40B, led by Sequoia and Wellington. Utah filed criminal charges against the company and its co-founders one day earlier . Rory called it a momentum round for what has become a national sports-betting market .
- Serval: raised $150M at $3B, led by Sequoia, up from $1B in December .
- Cognition: received investor proposals at $70–80B, weeks after a round at $48B .
- Baseten: set to raise at up to $30B, up from $13B in June .
The panel's diligence test for Serval was whether a customer that spends $200K on ServiceNow would pay $300–400K for Serval . Jason's concern with Cognition: a 10x from here means nearly a $1T outcome for a company that owns only part of its IP stack .
Physical Intelligence is negotiating at least $1B at about $20B with little or no revenue. Rory pushed back on the size: there are only about 300,000 robots in the US, and he expects foundation-model companies to buy their way into robotics . Canada will invest up to $1.14B in Cohere. Cohere says ARR grew 165%, but it did not disclose the base .
On exits, the DOJ is reviewing Nvidia's $12.93B purchase of Hugging Face. Jason said most acquirers want at least $100M in revenue before they will buy .
Texas data centers: conditions, not a ban
In a CNN interview, Texas governor-elect Gina Hinojosa said the state needs to "stop what we are doing" on data centers until there are measures to protect Texans . Gavin Baker's reading of her position is narrower:
- Projects that already have permits can finish construction, which covers 2027 and likely well into 2028.
- Projects newly permitted for 2028/2029 delivery would have to pay taxes, avoid raising consumer power prices, protect water supply, use union labor and give Texans "a seat at the table" .
Scott Kupor called this a mainstream position. He said removing NDAs and guaranteeing water and utility costs addresses most local objections . For infrastructure investors, the near-term pipeline appears safe, while costs rise for capacity delivered after 2028.
Early- and growth-stage deals
- Alta Ares: won NATO's counter-UAS program of record five months after Air Street led its $60M Series A. It took 6 of 16 lots, the most of any winner, and was the only supplier picked for area-defence systems. NATO allies can buy directly from the catalogue without a new tender .
- Tractian (YC W21): passed $100M ARR, growing 90% year over year. It raised a $130M Series D at $1.4B for sensor-plus-AI industrial maintenance .
- Reflex (YC W23): raised a $14M Series A, led by Standard Capital. Its platform turns business problems into production apps, with AI writing the software and Reflex engineers deploying it. YC says the framework powers internal tools at about 30% of the Fortune 500 .
- Quest: a personal AI agent. Air Street led the seed .
- Sekai: Keith Rabois led its Series B after leading the A in February. The company reports 30M+ plays a day and daily users up 20x+ .
Agents doing real work
Perplexity launched Teammates, persistent work agents that take high-level goals instead of prompts. They are rolling out to Max users . Aravind Srinivas reported internal usage:
- About 10,000 teammates in the last 30 days, 20x Perplexity's headcount .
- The median employee gets 7x more agent-hours per week than in September .
These are internal figures, not customer adoption. LangChain says most of its internal agents now use async subagents . Replit rebuilt its sandboxes around microVMs, using credential brokers so agent workloads never get reusable account credentials .
AI has made exploits cheap; patching hasn't sped up
An a16z essay describes an August campaign in which one operator directed a few hundred agents running a cheap open-weight model. They went from an empty workspace to the first break-in in under four hours and compromised 395 organizations in 48 countries, half of them schools . Meanwhile, the median time to fully fix a known-exploited vulnerability rose to 43 days .
a16z's base case is outbreaks on the scale of the 2003 worm era or WannaCry, concentrated on systems that are hard to update. It does not expect the grid or the internet to collapse . The UK's AI Security Institute finds the length of cyber tasks frontier models can complete unaided doubling every 4.7 months .
Other signals
- Talent: OpenAI hired Austin Lau, Anthropic's "one-man growth team," and Cat Wu, Claude Code's head of product. It also hired former Uber president Jill Hazelbaker .
- AI building AI: METR believes most of the recent acceleration in AI capability likely now comes from AI being used in AI R&D. It says it isn't sure, and Anthropic staff did not agree on this .
- Post-training: Trajectory says it trained Harvey a model that is six times cheaper than frontier models, raising accuracy on Harvey's legal benchmark from about 4% to about 18% compared with an open-source base .
- The episode reported that OpenAI’s round was reduced from $30B to $20B, with about $11B committed, and its valuation from $1.4T to $1.25T; the hosts said this alone did not establish that AI demand was collapsing, but argued that demand would determine whether financing stress persisted or exposed leveraged infrastructure such as CoreWeave. Anthropic’s bookings growth was described as slowing quarter over quarter even as absolute bookings rose; speakers pointed to discounting and cheaper models as sources of pricing pressure, while saying available signals did not show that demand itself had declined.
- A host said switching models in his B2B AI recruiting app reduced costs by 72% for the same workflow at higher quality; the investment wager for early AI products is whether unit economics improve before their funding runs out.
- DoorDash was reported to receive only several hundred agent-initiated orders against 3 million daily deliveries, but the hosts argued agents could still bypass marketplaces by finding direct restaurant pickup or delivery options, putting platform intermediation at risk.
- An investment debate compared a $30B inference provider, a $70–80B coding-agent company and a $3B AI ITSM app. One investor saw a clearer potential 10x in the ITSM opportunity but questioned whether it expands beyond ServiceNow’s existing market; coding agents were associated with near-term enterprise spending, while inference raised competition and commoditization concerns.
- Cohere’s round was described as potentially including up to $1.14B from the Canadian government as a sovereign-AI investment; speakers noted that Cohere had not disclosed revenue and questioned whether government support could translate into private-sector adoption.
- Physical Intelligence was described as having relatively little revenue while reportedly discussing a $20B valuation. The hosts praised recent robotics-model capabilities but cautioned that deployment takes longer and the business model is less developed; they also saw potential strategic interest from foundation-model companies seeking to expand into physical work.
- The hosts warned that an open-source Chinese penetration-testing tool, initially described as defensive, could be repurposed for attacks at scale using LLM agents, raising concerns about risks to critical infrastructure.
- Trajectory released a one-click platform that uses production-agent traces and a customer’s definition of what “good” looks like to train, evaluate, and deploy models together, then sends deployed models back into production to gather new traces. It targets AI-native companies with an established agent harness; Harvey and Rogo are named users, as is larger company Nvidia.
- Trajectory’s founder says Harvey used its legal-domain expertise and legal-agent benchmark to train a model whose accuracy rose from around 4% for an open-source model to around 18%, at six times lower cost than frontier models; he noted substantial room to improve. For Rogo, he said the tailored model is several times more token-efficient, cheaper, and faster while improving on tasks Rogo values.
- Trajectory says it is advancing a group-size-one version of GRPO: learning from one behavior across 10,000 trajectories from different tasks, with the stated aim of removing the need to replay tasks in a constructed environment. For reward labeling, it proposes having customers rubric-score 100 representative traces from 10,000, then using a metagrader to generalize those judgments; the founder said this approach is not perfect.
- The founding team’s relevant prior work included learning from production edits and retries at Windsurf, robotics at DeepMind, and using artist feedback for multimodal models at Apple; the founder described real-world feedback as the shared challenge of improving models.
- Agentic trading: Robinhood says more than 150,000 customers opened agentic-trading accounts within a few months of launch, with about 30 million tool calls per day at its last reported count. Its in-account Robinhood Agents support recurring strategies and portfolio rebalancing, and an app store includes tools such as satellite-data analysis to count cars in retailers’ parking lots. Tenev says the product aims to bring hedge-fund and quant-trading capabilities to individuals; accounts are separately funded and sandboxed, and he said agentic trading remained a small share of activity and was not imminently nearing double digits.
- Mathematical AI: Harmonic aims to build mathematical superintelligence; Tenev said its Aristotle model was one of three to earn a gold medal at the International Math Olympiad, and that after its public release people used it on unsolved problems and mathematicians used it for new results.
- Retail access to private startups: Robinhood describes Ventures Fund One as a publicly traded, NYSE-listed vehicle that raised capital from Robinhood customers and invests in late-stage private companies, including OpenAI. Fund Two is being developed with YC to provide exposure to seed and Series A companies, with Rich Aberman running it. Tenev also said the SEC had directed FINRA to create a knowledge-based accreditation exam, which he viewed as positive but insufficient to make access low-friction.
- Tokenized markets: Robinhood says it relaunched stock tokens with DeFi-enabled functionality alongside its new Robinhood Chain; it reports high developer activity and early experiments combining stock tokens with NFTs and memecoins. The company presents low-cost access to U.S. equities through non-custodial wallets in countries with weaker financial systems as a broader use case.
- Figma finance teams are building workflow automations: a credit-and-collections employee rebuilt invoice tracking, cash collections, reconciliation, and credit checks; separately, an AI-assisted spend analysis joined data from multiple sources and suggested policy changes, with the CFO valuing a directionally useful result over perfect accuracy.
- Figma’s CFO describes AI adoption as being in a “messy middle”: teams can prototype quickly, but a centralized CIO function is needed to identify successful tools and turn them into shared, durable foundations.
- Figma uses the ability to “vibe code a product” as one diligence check, helping assess whether a product is real and how difficult it may be to build; the CFO says the product team still partners in the process.
- For early-stage startups, a16z’s Ivon advises establishing reliable data access, integrity, and tracking before building reporting and metrics; he notes that startups may not track data at all and can calculate revenue or ARR incorrectly.
- Women and non-binary investors rose from 9% to 18.8% of partner-plus roles at U.S. VC firms since Allra began tracking in 2018, but a senior title does not necessarily bring authority to lead deals, write checks, secure carry, or receive deal credit—all important to building a track record for future fundraises.
- Allra is tracking 303 women and non-binary investors who spun out to start funds one through three and 301 in principal-plus roles at firms with more than $1 billion in AUM. Among surveyed people who left firms, one in 20 said they felt pushed out; others described wanting to build a firm or gain ownership and better economics, despite the current difficulty of raising a fund.
- Paige Hendricks Buckner cited a PitchBook figure that women founders had received 1.1% of VC at the start of the year, and said most funding to teams with women founders was in AI, naming Anthropic and Scale AI as the top companies. She also cited data that firms with more than 30% female partners are 4.7 times more likely to invest in a female founder.
- Buckner said she has spoken with many people building VC firms who are using AI to scale operations before they have enough capital, making the cost of AI support a practical constraint.
- a16z says AI has compressed exploit development: its timeline puts the work at about three weeks for a professional in 2017, a week in 2023, and an afternoon with AI in 2025 . The article reports that in June 2026 a model turned a released Firefox patch into a working exploit in under an hour, while cautioning that lab results explain—but do not prove—field performance .
- Early field signals include Google’s assessment that a zero-day developed by AI was in criminal hands and intended for mass exploitation, though disrupted before use; an AI agent chaining two zero-days to breach the Dutch Institute for Vulnerability Disclosure; and a government warning about AI-generated scripts targeting exposed Siemens controllers .
- Patching remains slow: the 2026 Verizon DBIR reports a 43-day median to fully fix a known-exploited vulnerability . a16z expects attacks to become viable against smaller organizations, and bets on a prolonged “worm era” for organizations without platform protection—not a grid or internet-wide collapse . For systems where updates require physical work or are otherwise costly, it points to isolation, segmentation, network-level protection, and safe manual fallback .
- AlphaFold is used by about 4 million researchers; Hassabis describes protein structure prediction as critical but only one part of drug discovery, with further work aimed at biochemistry, toxicity and ADME prediction. He also identifies Isomorphic as a drug-discovery spinout.
- Hassabis calls closed-loop recursive self-improvement without human review the riskiest next phase, citing control challenges and the potential for increasingly hard-to-understand systems. He advocates stronger sandboxing, monitoring and formal safety bounds, and sees room for third-party AI evaluation and auditing companies, including red-teaming between labs before release.
- He expects six to eight frontier or near-frontier model offerings, including Chinese providers, with competitive leads of roughly three to six months; open-source models may trail by six months to a year. He sees enterprises as able to choose among model strengths, and argues that countries and companies can build value on sensitive data or downstream applications such as healthcare rather than each trying to build a frontier model.
- At its London Zero Click event, Profound announced Slides, in-product visual creation, a mobile app for launching agents, and access through email, Slack, and Teams; it described proactive, always-on cross-channel functionality as its “Aim” (“AI Marketer”) agent.
- Profound’s thesis is that personal agents will become customers, so brands must ensure agents can surface, compare, evaluate, and transact their offerings. The speaker cited Gartner’s prediction that agents will handle one in five online transactions by 2030, but said that could happen by 2028 or sooner.
- Profound aims to build a marketing workbench where marketers do their work, analogous to Salesforce for sales teams.
Sequoia is refocusing on high-conviction, selective VC, sunsetting its Arc accelerator and Scouts program, simplifying its fund and operating and investing teams, combining Early & Growth into one team, and closing its London office while saying it will support European founders from the Bay Area. The firm says it has flattened its partnership, raised about $10 billion in new funds, strengthened its team, and moved away from a Monday partner-meeting cycle; its stated objective is quality and LP returns rather than scale. YC CEO Garry Tan called the changes a winning “back to roots” strategy and praised Alfred Lin and Grady.
DeepFinance launched as an ultra-long-horizon benchmark for high-finance work, with 21 problems based on real deal data and designed by experts with backgrounds at firms including Evercore, JPM and UBS and institutions including HBS and Wharton; each represents 20–40 hours of human work. It aims to evaluate end-to-end finance workflows involving assumptions and conflicting evidence, beyond financial-document QA or simple-model numerical checks.
Garry Tan characterized solving high finance as a path to AGI.
- AI is compressing cyber offense: a16z says exploit development has shifted from days or months to minutes or hours, and attacks averaged seven days before vulnerability disclosure by 2025. It cites a campaign using cheap, open-weight-model agents in which 395 organizations across 48 countries were compromised, half of them schools. The article cautions that pre-disclosure attacks began before agentic AI; its claimed shift is widening who can exploit that window and speeding weaponization of known bugs.
- The main risk is a patch-speed mismatch, not an expected collapse of the grid or internet: the median time to fully fix a known-exploited vulnerability was 43 days, while the article reports an AI model producing a working exploit in under an hour; only 26% of such vulnerabilities were fully fixed within a year. a16z expects worm-like outbreaks to hit systems that cannot be updated quickly, while saying it does not expect the grid or internet to go down.
- Defense and thesis-monitoring signals: for systems where updates are costly or impractical, the article points to isolation, segmentation, network-level protection, and safe manual fallback. Its scorecard says the thesis would look more concerning if exploited-vulnerability counts exceed 2,000 in 2027, over half of exploited bugs are attacked on or before CVE publication by 2027, or median patch time exceeds 60 days in 2028.
An unnamed startup’s growth was not linear: after a year of smoothly exponential growth, it stalled, then growth returned five months later . Paul Graham says the same company had a fall 2024 crisis that drove several months of revenue declines, and the later chart nearly obscures it—an example of how growth curves can hide difficult operating periods founders experience .
Garry Tan sees a bipartisan opportunity to curb Big Tech’s market power and urges Congress to open markets for startups, arguing that a few trillion-dollar companies’ control over distribution can keep new products—including AI applications—from reaching users and businesses; he says broader startup competition could lower costs and spread AI’s gains more widely.
- A few hundred AI agents using a cheap open-weight model turned two recently disclosed bugs into a first PaperCut-server breach in under four hours; the campaign compromised 395 organizations in 48 countries, half of them schools, and the agents ignored a 28-country exclusion list. a16z argues that AI is lowering the cost of every offensive step while defenders lag: the 2026 Verizon DBIR put the median time to fully fix a known-exploited vulnerability at 43 days, with only 26% fully fixed within a year.
- a16z does not expect the grid or internet to collapse; it expects the heaviest impact on systems that cannot be patched quickly and smaller operators, with added risk from attacks spreading beyond their intended targets and more actors gaining advanced capabilities. In healthcare, a16z reports confirmed ransomware attacks on providers grew about sixfold from 2018 to 2024; average US provider downtime per attack rose from 4 days in 2018 to 17 in 2021 and 18 in 2024.
- Capability shift: a16z reports that the longest cyber task a frontier model could reliably complete, measured in expert time, grew from 5.5 minutes in late 2024 to 40 minutes in early 2026 and 4.8 hours by April 2026; its article says cyber and coding task lengths have been doubling every 4–5 months. It also cites field reports of AI-chained zero-days at a Dutch vulnerability-disclosure organization, a criminal AI-developed zero-day assessed by Google, and AI-generated scripts targeting exposed Siemens controllers.
- Security implications: a16z says exploit development has fallen from a skilled person’s month to a few hours, while the median time to fully fix a known-exploited vulnerability is 43 days and only 26% are fully fixed within a year. It expects cheaper attacks to reach smaller organizations and unsupported systems that were previously less attractive targets. Its base case is a long, worm-like period of harm rather than the grid or internet going down, while flagging attack spillover and aggregation of capability against smaller critical operators as risks.
YC W21 alum Tractian builds physical AI for asset-heavy industries, using proprietary sensors, software, and AI to monitor hundreds of thousands of assets in real time; it has crossed $100M in ARR, growing 90% year over year, and raised a $130M Series D at a $1.4B valuation. Tractian is moving from equipment monitoring toward factory autopilot, with AI changing the economics of industrial maintenance.
Garry Tan highlights Capy.ai’s “captains,” with communicating subthreads and subagents that coordinate and manage merge queues and CI, calling this a “huge gamechanger.”
Reflex (YC W23) raised a $14M Series A; it builds custom software for companies’ operational needs, with a platform that takes a business problem to a working app using AI-written software and Reflex engineers to put it into production. The open-source framework has surpassed 29,000 GitHub stars, powers internal tools at about 30% of Fortune 500 companies, and has been used to build more than 1 million apps.
Standard Capital led Reflex’s $14M Series A. Reflex has evolved from an open-source, pure-Python web-app framework into a platform that takes business problems to working apps, with AI writing the software and Reflex engineers putting it into production. YC’s announcement reports 29,000+ GitHub stars, more than 1 million apps built, and internal tools at approximately 30% of Fortune 500 companies. Caldwell pitches the opportunity as replacing SaaS with customized software; this is an investment thesis, not evidence that enterprises have already replaced their SaaS.
Martin Casado endorsed the view that formal methods should complement, not replace, guardrails, tests/evals, real-world use, and human judgment: verification can show that code matches a specification, but not that the specification captures user intent. The linked post describes model checking (e.g., TLA+/TLC) as useful for checking system properties and finding bugs, while noting possible spec/code drift; theorem proving and bounded model checking have different effort and coverage tradeoffs. It also points to a Cursor plugins pstack /statechart skill for making implicit state machines legible and checking whether implementation matches intent.
