We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Nvidia: board investigation after a counterintelligence report on Huang
The New York Times reported that National Security Adviser Marco Rubio oversaw a highly classified memo, drafted this spring. It said the intelligence community assessed that a woman who had become close to Jensen Huang was working for China's Ministry of State Security . A statement attributed to Huang, posted on Reddit, apologizes for being unfaithful to his wife. It says Nvidia's board "has initiated an independent investigation" into the national-security concerns and that he will cooperate fully . That text comes from a Reddit post, not an Nvidia filing.
Investor reaction tied the story to the export fight. Jason Calacanis called it "defcon 1" and said Huang and Nvidia's technology "are a critical target" that needs protecting . He also reposted a May Culper Research claim that documents suggest Megaspeed was financed by Alibaba through a Singapore shell company. Culper describes Megaspeed as Nvidia's largest chip buyer in Southeast Asia and says it is under US and Singapore investigation. Nvidia had vouched that Megaspeed had no Chinese ownership . These are still allegations. For anyone exposed to Nvidia, the near-term issue is governance and US policy toward chip sales to China, not demand.
Salesforce's Cresta deal is already slowing startup sales
Salesforce agreed to buy Cresta, whose AI voice agents handle customer-service calls from start to finish. The Information put the price at about $2.7B. Earlier this year Salesforce bought Fin (text support) and Braze (marketing messages), so it now covers text, marketing and voice in one bundle . Vendors who sell into Salesforce accounts report buyers waiting to see how Salesforce prices the bundle. Some ask for shorter pilots or month-to-month terms instead of 12-month contracts . Another vendor says it is already moving to monthly contracts because buyers "want a kill switch" . For investors, expect softer contracted ARR and shorter commitments at AI-support startups that depend on Salesforce-heavy customers.
Diligence takeaways from SaaStr this week
Jason Lemkin's notes on the latest 20VC x SaaStr:
- Inference costs are finally falling in practice. Moving SaaStr Connect to Haiku 5.5 and a cheaper model mix cut total costs 72% for the same workflow, and it got faster and better. He says nothing they ran agentically had gotten cheaper per use since early 2025 .
- Anthropic's growth is slowing because of price, not demand. In his view, OpenAI sells roughly equivalent products at 30% off. Companies whose value is mainly model capability should plan for prices to fall 30–40% within a year .
- Always-on agents are expensive to serve. They can cost more than $20 per user per month, versus about $4–5 for Replit and Lovable .
- Vertical AI needs 2x the incumbent's ACV. He won't invest in an agentic B2B company unless its deal sizes are at least 2x what pre-AI competitors get .
- Acquirers want $100M in revenue. Most large acquirers won't put their go-to-market behind a target below that. Listen Labs, at about $20M ARR for roughly $2B, was the exception .
- A growth rate without a base looks weak. Cohere announced Canada putting in up to $1.14B and ARR growth of 165%, but didn't give the base .
A companion SaaStr piece uses Stripe's Privy as an example of how strategic buyers value companies. In the year or so since the acquisition, wallet assets are up about 25x, revenue about 9x, and monthly volume is nearly $25B . Stripe's stated reasons for the deal didn't mention ARR . The piece argues revenue sets the floor in a negotiation, and the premium depends on whether the company fills a gap in the buyer's roadmap .
AI for biology: technical progress and pharma pushback
The WSJ reports that no AI-designed drug has won FDA clearance . Vertex executive chairman Jeff Leiden said Vertex had about 50 AI companies work through its dataset of 10,000 compounds, and "not one managed to generate a working drug." He still finds AI useful as a scientist's co-pilot .
YC's Paper Club showed both what works and where the limits are. Agents reprocessed NIH transcriptomics data and grew a dataset from 36M to more than 500M cells. But immune cells, and T cells in particular, were heavily overrepresented . The speaker's proposed fix is "AI guided experimentation" rather than simply collecting more data . On the genomics side, the DNA model EVO 2 was trained on 9 trillion base pairs, and its developers report it is competitive with or better than human-specific supervised models on a cancer-risk gene benchmark. They said the data mix made the difference between state-of-the-art and "bottom of the pack" .
DeepMind's Pushmeet Kohli says the lab is seeing "a genuine phase transition in the pace of scientific discovery," with releases expected in the coming weeks . Demis Hassabis describes Isomorphic Labs' aim as cutting drug development from about 10 years to months . The difference between these labs' framing and pharma's experience is the main thing to check when diligencing AI-bio companies.
Is AI progress getting more expensive?
One analysis finds AI progress (measured by AECI) per $1B of R&D fell 63% from 2025 to 2026, even as progress per year doubled . Martin Casado asked how METR's report is useful if it doesn't normalize by spend . Taken with Lemkin's pricing view, frontier capability looks more expensive to produce and harder to charge a premium for.
Policy and regulation
- Mistral vs. France's Darcos bill. Arthur Mensch says Mistral Large 4 was trained entirely on Mistral's own infrastructure in Essonne, and that the company is building 1 GW of capacity, mostly in Europe and largely in France . He argues the bill's presumption that copyrighted work was used in training reverses the burden of proof. Developers would have to prove a negative without exposing trade secrets, and he says that risk will weigh on where data centers get built . His alternative is a revenue-based levy on all AI providers selling in Europe, including foreign ones .
- Suno lawsuits. Elvis Presley's estate and Eminem's publisher sued Suno, alleging copyright and likeness violations. Suno denies the claims .
- Local resistance to data centers. A Reddit post says voters backed restrictions on data centers in 40 of 41 local ballot measures across 10 states . This adds to the permitting pressure covered earlier this week.
Macro: AI growth is coming from the companies building it
A summary of the Bank of England's November report says AI accounts for just under a fifth of UK GDP growth since January. Almost all of that comes from companies that build, install and host AI. Users report about 0.5% total productivity gain over three years. Hiring is slowing through fewer replacement hires, but firing has not risen . Firms expect about 0.9 points a year of productivity growth, half of it from slower headcount growth. UK productivity figures due November 17 are the first test .
Smaller items
- Can an LLM fix another model's weights? MIT CSAIL researchers report GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights, with no training run, and lifted MATH-500 from 7.9% to 48.1%. Critics say Astra had Python tools and that the fix applies steering, an existing technique. The work is a blog post with logs, not a peer-reviewed paper .
- Smaller teams. Garry Tan says Apple should rebuild apps like TV and Calendar with "a few people and a hundred codegen agents" . He also says personal just-in-time software looks inevitable .
- Private vs. public valuations. WashU CIO Scott Wilson thinks Anthropic could have raised privately above $2T, a price he doubts public markets would pay. He blames weak DPI (cash returned to LPs) on too much capital chasing private winners .
- EVO 2 was trained on 9 trillion base pairs; its developers reported performance competitive with or better than supervised human-specific models on a cancer-risk-gene mutation benchmark across coding and non-coding variants. They flagged distal-enhancer prediction as a weakness, which another group improved through a different data composition.
- EVO 2’s few-shot prompting generated novel landmodulin candidates from 14 known examples and CRISPR-Cas9 systems from four examples. In a separate design task, an enhancer produced with prompting after reinforcement-learning post-training was reported active and specific in the mouse brain.
- Virtual-cell work faces a data-diversity bottleneck: agents expanded an NIH transcriptomics dataset from 36 million to 500 million-plus cells, but immune cells—especially T cells—were overrepresented, prompting a shift toward AI-guided experimental data generation. The Locus system also improved over off-the-shelf coding agents on molecular-binding and generative tasks as GPU compute scaled, but those results were in silico and still require lab validation.
- AMA co-founder Alfredo described building a personalized research lab for each cancer patient using AI agents to analyze patient data; he cited a rare brain-tumor case where cross-tumor analysis identified potential therapies despite limited relevant literature. Clinical translation remains a constraint: clinicians retain treatment authority, setting up a new NF1 trial is a separate challenge, and hospital tissue handling can prevent organoid work.
- Hassabis described Isomorphic Labs as a new spinout building on AlphaFold technology to reimagine drug discovery with AI; he framed reducing drug development from an average of 10 years and billions of dollars to months or possibly weeks as an ambition.
- AlphaFold predictions for 200 million known proteins were made freely available without restrictions; Hassabis said more than two million researchers use the resource and that it has become a standard biology research tool.
- Hassabis’s screen for applying game-derived AI to real-world problems is whether a task has a large combinatorial search space, a clear objective to optimize, and sufficient data or experience—or an efficient, accurate simulator.
- DeepMind’s stated next step is to combine agent-based search and planning with Gemini and world models for real-world tasks; Hassabis forecast major robotics advances in the next two to three years. Genie2’s generated game worlds, however, currently remain consistent for only a few seconds, with longer consistency an active development goal.
Garry Tan says his gstack-evals project tried to reproduce an article about using Jev with Kimi K3, and reports that it replicated; he says Particle-1 did “even better.” He gives no evaluation metric or numerical results.
- Google says it aims to compete at the frontier while using its full stack—from hardware and models to products—to serve users across a range of cost and quality levels and use cases.
- Google product leader Tulsi described a “60/20/20” approach to AI products: make roughly 60% work now, build another portion around capabilities that seem near, and reserve room for more ambitious features as models improve. She cautioned that rigid wrappers can constrain newer models, making repeated retesting important.
- For AI product evaluation, she recommended starting with small, hands-on tests—around 5–20 prompts—rather than beginning with a large benchmark, and evaluating the end-to-end experience because judge quality, serving, filters, and mitigations can all affect results.
- Google’s stated safety approach includes monitoring models for possible misalignment and intervening; the interview also cited work with Wiz to identify and patch vulnerabilities, including one found in a hospital system.
- Garry Tan argues Apple should retool its software team around “software factories,” saying teams of a few people with 100 code-generation agents could remake buggy apps such as TV and Calendar.
- He cites TV app defects: downloading resets the list, and trying to move a movie to the top of the download queue crashed the app and erased the queue; he says the reset is easy to identify in telemetry and fix.
- Tan says frontier models make software easier to change, making bugs less tolerable, and predicts personal just-in-time software will become inevitable.
Garry Tan urges founders to build their own “software factory” , amplifying Demis Hassabis’s forecast that one person who understands AI could outperform an entire startup team and that AI-enabled builders could run whole companies with one-tenth the headcount in five years .
Martin Casado questions METR’s report for not normalizing AI progress by spending, asking whether the analysis is useful without that adjustment . The comparison in the quoted post says AECI progress per $1 billion of R&D fell from 4.05 points in 2025 to 1.49 in 2026—a 63% decline—even as progress per year doubled .
Garry Tan says he is conducting open-source evaluations of harnesses with frontier models and links a GitHub project called GHarnessEvals, describing the work as being done in public.
Garry Tan said he was asking his “software factory” to test a prompt for Claude Code or Codex that asks agents to identify implicit state machines in an app and improve user, developer, and agent experience without breaking anything; he said he would report back, so no test result was included.
Y Combinator’s Paper Club framed AI biology as shifting from reading biological data toward modeling and designing living systems: sequencing costs have collapsed, data is outpacing researchers’ ability to interpret it, and models are beginning to catch up. The session covered cancer genomics, virtual cells and autonomous research, DNA language models, and few-shot prompting for biological design.
Paul Graham says YC fosters unusually close founder relationships: 20 years ago, many YC founders lived in a San Francisco building nicknamed the Y Scraper, and a reunion prompted his claim that YC founders become closer friends than founders backed by any other investor.
Martin Casado said he was unsure whether Seb Krier’s graph was correct, but that what it showed was what he would expect from prior technology trends; he called anticipating a different outcome this time an aberration warranting scrutiny. The post does not specify the graph’s subject or findings in text.
A podcast summary says Nikesh Arora grew Palo Alto Networks’ market value from $18 billion to more than $250 billion; after eight years and 35 acquisitions, the company had more than 70 products. The described acquisition playbook is to buy winning teams, put the acquired founder in charge, agree on a product roadmap before signing, and target only number-one or number-two players. The promotional post discloses paid partnerships.
David Sacks attributed Republicans’ midterm losses to the war and its economic effects, while saying job and GDP growth remained strong and inflation would drop quickly if the war ended . He cited Brent crude’s 10% election-day decline as a market reaction to the possibility of an earlier end to the war, and argued an end would rapidly reduce energy and transportation costs .
Sarah Guo says AI still cannot help with the iterative parts of writing an essay: ordering ideas, cutting excess length, maintaining coherence, and preserving substance.
Jason framed U.S.–China AI competition as a national-security race, calling AI the battlefield and saying Jensen Huang and NVIDIA’s technology were critical targets that needed protection, while urging the U.S. to accelerate toward superintelligence. The attached MaggieNYT post reported that a highly classified memo overseen by Marco Rubio said the intelligence community assessed that a woman who had formed a close relationship with Huang was working for China’s Ministry of State Security.
For evaluating efficiency-driven markets, Martin Casado suggests looking beyond price: efficiency can lower prices and create a rebound, so compare aggregate margin dollars before and after the efficiency gain.
A shared post says an open-world game was made using Opus 5.5, gpt-Astra, Three.js, Anatomy, and Suno, with an art direction inspired by Moebius and Roger Dean; Martin Casado praised it as beautiful.
Scott Kupor says Sequoia has handled generational and management transitions unusually well over its 50-plus-year history, a rare capability that has helped it remain a tough, respected competitor among venture firms.
- WashU’s Scott Wilson says the endowment’s AUM grew from $8B to $19B in nine years. It held SpaceX at more than 20% of its portfolio and used continuation vehicles to stay invested through the IPO; Wilson says the approach depended on having the right governance structure.
- On portfolio construction, Wilson says diversification benefits taper beyond 20–30 look-through exposures; WashU targets half its portfolio in its top 50 positions. He sees megafunds as a poor fit unless a round is too large for WashU to invest directly, and doubts they can repeat the multiples of smaller funds.
- Wilson believes Anthropic could have raised privately at more than $2T, a valuation he doubts public markets would support, and attributes weak DPI to capital chasing private-company champions. WashU also wants valuation discipline and avoids putting capital into deals at 100× revenue.
- WashU identifies direct opportunities by asking managers which five portfolio companies excite them most, then meeting promising companies and offering to help lead the next round; it has also led employee tender offers and invested directly rather than waiting for SPVs. Wilson names not adding to winners—partly because they were already 1–2% of the portfolio—as his biggest mistake.
