We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
WSJ reports Nvidia's board is preparing to oust Jensen Huang as Fractile raises $720M
•
6 min read
•
2199 docs
A WSJ report that Nvidia's board may remove Jensen Huang divided investors over founder succession. Separately, Fractile raised $720M for memory-stacked AI chips, Runway released open robotics weights, data showed older GPUs holding their rental value, and a16z laid out its case for agents as buyers.
Nvidia: a key-person risk at the center of AI
The WSJ reported that Nvidia's board is preparing to oust CEO Jensen Huang after "two extramarital affairs, including one with a suspected Chinese spy" . This follows the internal probe covered in earlier briefs. Venture investors quickly sided with the founder:
- Garry Tan called an ouster ill-advised and said it would be "hard to believe a transition happens without massive market cap implosion" .
- Naval cited Steve Jobs and Travis Kalanick: "one does not just simply replace the founder" .
- Suhail suggested boards wait two weeks after a PR crisis before moving on a CEO, unless criminal acts are involved .
- Jason Calacanis wants a summary of the investigation released to shareholders so they can vote. He predicts the board will act in "CYA mode" .
- Elad Gil used the episode to advise founders to keep boards small and high-trust, and to resist the "governance experts" who push outside directors at IPO .
Two related policy items. Axios reports that President Trump authorized a crackdown on companies suspected of diverting Nvidia's advanced chips to China, and that Commerce Secretary Lutnick "repeatedly stood in the way" . An unconfirmed rumor, labeled "wild" by the account that posted it, says DeepSeek is pretraining a ~2T-parameter V5 on Huawei Ascend 950 chips . Taken together, the headlines raise both the governance risk and the geopolitical risk around the AI supply chain's most important vendor. None of this is a confirmed leadership change yet.
Funding
- Fractile raised a $720M Series C led by Thrive, Founders Fund, Lightspeed and Redpoint . Sarah Guo describes the approach as stacking DRAM directly on top of the processor to meet frontier models' memory-bandwidth needs, "in close partnership with some of the top labs," which sets it apart from incremental chip bets .
- Imprint, a co-branded card fintech, raised $130M at a $1.88B valuation. It aims to double revenue to $400M while growing headcount only 30%, using AI internally and in agentic commerce .
- Nuance Labs raised a $50M Series A led by Lightspeed . The founders are ex-Apple researchers building a unified, full-duplex foundation model for real-time text, speech and video expression . They argue that cascaded avatar pipelines lose emotional cues and add latency . The team has reached 27 people; the original plan was about eight .
- Index Industries came out of stealth with a $50M seed to power USHP, pitched as the first investable index of US residential real estate. The CEO is former Polychain president Josh Rosenthal .
- Higgsfield reports annualized run rate rising from $1B in September to $1.35B in October, with a target of $2B by year-end . Two caveats. The figure is the last four weeks of revenue × 13, and about 60% of revenue is credits and usage, which can swing month to month . SaaStr, which reported it, is an investor .
- Standard Capital opened its next Series A cycle. Applications are due Dec. 1, with decisions by Dec. 11 .
Model capability and cost
Allie K. Miller highlights Fable 5.2 scoring 41.3% on the Remote Labor Index, up from 17.9% for Fable 5.1 and 2.5% a year ago. By her reckoning, that is a doubling time of about three months . Pushing the other way on price, Bindu Reddy says that in her blind tests DeepSeek matches Carbon/Opus on 95% of agentic loops at about 40x lower cost. It degrades faster on very long loops . These are one tester's results, but they support last brief's point that model routing is eroding frontier pricing power.
Runway released Praxis-1, a world action model that turns video pretraining into control for real robots. The weights are open under Apache 2.0 . Cristóbal Valenzuela says it ranks #1 on NVIDIA's RoboLab-120 leaderboard . Separately, experts OpenAI appointed to vet its AI-generated proof of the existence of quantum Yang-Mills theory say they have verified it, per Scientific American . That verification comes from OpenAI's own reviewers, not open peer review.
Suhail says his five-person startup "maxed out all our inference compute in 2w across customers." He is offering a $20K referral bonus for post-training or inference-optimization hires .
Infrastructure economics: GPUs holding value, data-center queues inflated
a16z data show GPU rental rates as of Oct. 19 sitting above a six-year straight-line depreciation path: A100s by 9%, H100s by 54% . Mike Volpi explains it as scarcity. Demand is high enough that even inefficient chips get used, and that benefits anyone financing the assets. He adds that once there are enough Vera Rubins, running H100s stops making sense . Neocloud lenders should note that the residual-value assumption depends on chip manufacturing staying constrained.
On power, Ascend Analytics' Brent Nelson says ERCOT's large-load queue is about 400 GW against a roughly 90 GW system peak. He thinks realistic buildout is "a tenth to a fifth" of raw queue figures . Hyperscalers can't find capacity for 2027, he says, but by around 2030 many data centers may be competing for tenants. He also expects many VC-backed AI tenants to disappear . Senator Josh Hawley is backing a nationwide data-center moratorium , which adds a political risk to that picture.
Agents as buyers
a16z argues that agents will increasingly choose products, subscriptions and spending on users' behalf, and that the business models this enables "mostly aren't visible yet" . Its evidence includes Shopify's AI-driven store traffic up 8x year over year in Q1 '26, with AI-search orders up nearly 13x . It also cites agents accounting for nearly 40% of Stripe documentation traffic and 70% of CLI requests for API resources . The firm's investment map covers agentic models, agent-native businesses, incumbents optimizing for agent-led purchasing, and agent infrastructure. It notes that payments, fraud, identity and communications are already strained at low agent volumes .
SaaStr offers a firsthand example. Its AI marketing agent went straight to Higgsfield and Gamma over MCP and never connected to Canva. SaaStr's conclusion: "the window to get picked is when an agent is first wired into a workflow" .
Agent security
- Lazarus (Microsoft Research): editing fewer than 10 dormant neurons can plant a trigger-activated backdoor in an agent's LLM that still passes standard safety evaluations . The attack requires write access to the weights, and Microsoft has seen no use in the wild . Microsoft's detector flagged 91.4% of edited checkpoints, with a 9.7% false-positive rate . This makes model signing and weight provenance an emerging product area.
- Perplexity SPACE: Perplexity red-teamed its open-source agent sandbox. It says non-cyber models found exploits better than cyber-tuned ones, the sandbox held against isolation escapes, and it found and fixed a DNS-spoofing network bypass .
- Fortinet: a commenter citing Fortinet's own post says its patches of nearly 100 vulnerabilities included fixes from a frontier-AI testing program with Anthropic, OpenAI, Google and SpaceX. The same post warns these tools can turn patches into exploits within hours or days .
- 1Password: the company says it has passed $500M ARR and that Q3 net-new ARR grew about 90%, driven by AI use cases . It reports "explosive adoption" of 1Password for Agents, a zero-knowledge way to give personal agents access to credentials .
Operator and investor lessons
Ro says it went from about $270M run rate growing ~9% in 2021, which it calls a "zombie company," to $3B+ in 2026 . Headcount fell from 596 at its 2022 peak to 406 . Elizabeth Yin listed Hustle Fund's misses, starting with passing on Shield AI at pre-seed; Shield AI is now valued near $30B . Her conclusion: "sometimes you should just invest in your super entrepreneurial friends even if you don't like the idea" .
WSJ reports Nvidia's board is preparing to oust Jensen Huang as Fractile raises $720M
TechCrunch
- Ascend Analytics power-market forecaster Brent Nelson says large-load queues are poor proxies for actual construction: ERCOT's queue is about 400 GW versus a roughly 90-GW system peak. He estimates plausible builds at only around one-tenth to one-fifth of raw queue totals, which are inflated by normal project duplication and speculative entrants; forecasts already discounted for plausibility should not be discounted again.
- Nelson expects political slowdowns to delay data-center development by roughly one to two years, affecting 2027 more than 2032, while access to power is a separate bottleneck. Long-lead gas projects also face political and permitting risk. Solar and storage are poorly matched to long overnight winter reliability needs, helping make gas the default despite constrained generation equipment, pipeline challenges, and fuel-delivery risks during winter storms. Nelson argues that a curtailable grid connection paired with behind-the-meter batteries could be a comparable, potentially cleaner, cheaper, and faster alternative to gas, while noting this is not a consensus view. Paying for grid upgrades is necessary but may not secure community support, given concerns about affordability, water, noise, emissions, and AI more broadly.
- AI-infrastructure demand depends on timing and customer durability: hyperscalers reportedly struggle to find capacity for 2027, but around 2030 many data centers may compete for tenants, while some AI startups may fail or be acquired, changing who needs compute. Banks generally require a creditworthy tenant lease to finance a data center, though the adequacy of lender due diligence is an open question; the interview cites a Meta-leased data center that struggled to close without substantial concessions amid a tightening debt market. A strong neocloud anchor tenant can make a project more reliable but creates concentration risk, so investors should assess who will pay the rent rather than relying on capacity plans alone.
Y Combinator
- Theranos’s pitch was that a single drop of blood could support central-lab testing in settings from homes to battlefields . Schultz says there were no tests validated when the Walgreens product launched, patient samples instead went to other labs or third-party equipment, and only seven tests were validated by the time he left . He also recounts that a syphilis test initially showed 65% sensitivity; after that data was deleted and the study repeated, sensitivity was 80%, while more than 20% of assumed-negative colleagues tested positive before a doctor signed off .
- Schultz describes an 11-year culture of fear in which employees who questioned the CEO were fired and sometimes sued; he says people who knew the device was inaccurate stayed silent . He says employees might have raised concerns with the board if they had trusted that they could do so safely and confidentially .
- Schultz recounts that Khosla’s investment process led the firm to pass, and says he heard similar accounts from Sequoia and Google Ventures; Tim Draper was an exception, making a $500,000 angel investment .
Lightspeed Venture Partners
- One Password’s CEO warned that AI agents’ need for secrets is colliding with growing exposure from long-lived credentials scattered across devices, while the cost of finding and exploiting them is rapidly falling. The company said it had built a zero-knowledge framework for agent access, announced One Password for Agents, and saw “explosive adoption” on platforms where it was released.
- The interview reports that One Password protects over 200,000 businesses and recently crossed $500 million in ARR. Its CEO said Q3 net-new ARR growth was about 90%, with AI-oriented use cases driving growth rather than token resale. He attributed the business’s ability to fund product and distribution investment while remaining profitable to strong retention and a profitable core consumer business.
- For early-stage diligence, the CEO recommends testing funnel progression and real customer pull before scaling, then examining gross/net retention and expansion across customer segments; he cautioned that sales through founder, portfolio-company, or board networks can give a false signal.
Lightspeed Venture Partners
- The interview presents Nuance Labs’ $50 million Series A as led by Lightspeed. Founders said the team had reached 27 after initially planning for about eight hires over six to eight months; they attributed the expanded scope to building custom inference and orchestration systems as well as the model and its personality and emotional-intelligence capabilities. They said the round would accelerate work on talent, data and compute. Fong Chong Ma described prior Apple work on Vision Pro digital personas, and Karen Yang cited audiovisual research experience at Adobe, Meta and Apple.
- Nuance Labs is building a unified foundation model for real-time understanding and generation across text, speech and video. The founders’ proposed differentiator is full-duplex operation—attending to a user while generating the system’s response—rather than a turn-taking pipeline; they argue that existing cascaded systems lose expressive cues when they are reduced to text and incur latency.
- The founders describe natural conversational dynamics as a data bottleneck: they say suitable data is not available at internet scale, and podcasts or interviews often differ from direct conversations because they are edited or audience-facing. They also say simulation is unsuitable for this problem, leaving data gathering as a core challenge.
- The founders said their immediate focus was releasing a research demo, with a consumer-facing interface as a path to enterprise use. At company launch, they reported inbound interest in customer support, user and market studies, and training and coaching; this was reported interest, not evidence of deployed customers.
Stanford eCorner
- Loyal founder Celine Halioua was introduced as having raised over $200 million; she described Loyal as a Series C company pursuing the first FDA-approved lifespan-extension drug, beginning with dogs. At the time of the talk, she said its first drug had cleared efficacy and safety approvals, with manufacturing approval remaining.
- Loyal’s dogs-first thesis uses shorter dog lifespans and cash-pay owners to avoid a human-development bottleneck: proving lifespan extension in people could take decades, nearly matching a 20-year patent term. The company’s “save the dog, save the world” thesis is that dog-longevity work could establish regulatory, social and biological precedent for humans.
- Loyal abandoned its original gene-therapy approach about six months in after its team cited safety, cost-of-goods and technical risks, switching to another way of targeting the biology. Halioua described a portfolio of about five drugs at different stages; one had recently been dropped over a theoretical safety concern, illustrating the company’s effort to avoid dependence on any single drug.
- Halioua sees opportunities for AI in clinical development and operations, and cautions that some AI-for-bio work may focus on areas that become commoditized without enough understanding of drug-development bottlenecks. Loyal’s veterinary trial team has built “vibe-coded” interface skins over difficult existing study software to reduce user errors, which she cited as an example of unmet workflow needs.
Invest Like The Best
- Ro reported growing from about $250M to more than $3B in run rate over four years while employing fewer people than in 2021; GLP-1s were its fastest-growing category. Its CEO credited a major GLP-1 tailwind but said credit-card data showed Ro signing up roughly 2–3 times as many patients as peers.
- The 2022 refounding followed a 9% growth year after a $500M raise at a $5B valuation; the CEO said the company had tried to do too much and hire too broadly. The reset included moving on from seven VPs and about 20% of staff, unifying brands under Ro, and clarifying the company culture; Ro later reported a 10x larger business with about 30% fewer people. The CEO also described prioritizing sequential execution and dropping long-term roadmaps in fast-changing markets.
- Ro said it routes every patient message through an LLM to analyze the issue; reported side effects become structured tasks for a healthcare provider with an SLA, and the company claimed 24/7, sub-30-minute response times.
- The CEO sees healthcare-AI opportunities in combining newly available symptom, lab, genetic, and wearable data with AI and new treatments; he pointed to variable GLP-1 responses as a potential personalization use case. He also cautioned that medicine is difficult for AI because much of its data is unstructured and outcomes such as heart disease take a long time to observe.
a16z
- a16z frames the B2A economy as an emerging distribution shift: as agents move from answering questions to completing tasks, they may choose products, subscriptions, and purchases for users. The business models enabled by this shift are mostly not yet visible, making this an early thesis rather than a mature market.
- Early signals include Shopify-reported 8x year-over-year growth in AI-driven store traffic in Q1 2026 and nearly 13x growth in orders from AI-powered searches; Human Security also saw agentic-browser traffic grow 8x from August to October, with Muse driving 60% of activity and OpenAI products 30%. The article says AI-referred consumer traffic has higher engagement and conversion and lower bounce rates than non-AI traffic, and estimates that 11% of U.S. shopper journeys visited AI before a transactional site. Coding agents offer a separate adoption precedent: as of September 2026, upwards of 40% of GitHub PRs started with an agent draft, versus 17% six months earlier.
- a16z identifies potential opportunity areas in agent models and harnesses; agent-enabled businesses such as self-running SMBs, AI-intermediated marketplaces or network products, and auto-optimizers; incumbents adapting for agent-led product evaluation and purchasing; and agent infrastructure. Payments, fraud, identity, and communications are already experiencing strain at relatively low agent volumes.
a16z
- The guest sees silicon—not power—as the main AI-capacity bottleneck, with fab capacity likely to remain a governing constraint for the next two to three years; he expects demand to become more price-sensitive as buyers route work to cheaper models.
- He expects an economic correction and flagged weak neocloud balance sheets, but believes the underlying compute assets will remain useful. GPU longevity depends partly on ongoing scarcity: abundant newer, more efficient GPUs could make older hardware uneconomic.
- He argues AI infrastructure should be designed as an integrated system, increasingly at rack or row scale; he sees patchwork systems as less efficient. Because AI data-center traffic is highly utilized and more predictable than conventional network traffic, he sees an opening for cheaper, purpose-built networking, although adoption requires coordination across the system and its buyers.
- For startup exit decisions, he emphasizes the founding team's appetite to continue and whether the business could plausibly grow 10×, rather than merely double. He advises founders to maintain regular conversations with large companies and competitors for market intelligence and relationship-building, not necessarily to signal a sale.
Lightspeed Venture Partners
- Index Industries emerged from stealth with a $50 million seed round to build a platform for index assets. Its first product, USHP, is a token tracking broad U.S. home-price indices and backed by real-estate-related and liquid assets; an income-bearing variant, USHPI/USHP Income, is designed to pay holders yield generated by reserve assets.
- CEO and cofounder Josh Rosenthal previously led at Polychain Capital and joined the company after investing; cofounder Brian Hardeker brings residential real-estate acquisition and operating expertise through a firm described as having deployed about $2 billion in U.S. residential assets. The execution caveat is substantial: Rosenthal says tokenized real estate had yet to produce a breakout product by 2026, citing both challenging tokenization UX and the difficulty of scaled real-estate asset management.
- Rosenthal's timing thesis is that tokenization technology is now usable, large financial institutions including BlackRock and Fidelity are adopting tokenized assets, and housing's rising barriers to ownership create demand for easier investment access. The company says the seed capital will fund legal/regulatory work and product structuring, a functional launch product, and hiring; it reported about 10 employees and hiring across core functions.
My First Million
- Robert Greene says public fear and anger over AI are widening the gap between the public and those controlling the technology; he sees demands for AI laws and data-center restrictions as a potential source of major change.
- He warns that AI could enable dangerous misuse, citing rising phishing and scams on phones and speculating about access to information for making biological or other weapons. He also fears that reliance on AI could weaken young people’s ability to write, read deeply, and think independently, while acknowledging possible benefits such as cancer treatment and climate solutions.
Not Boring by Packy McCormick
- Reducto’s sponsored launch of r-1 mini offers document parsing from $1 per 1,000 pages, preserving reading order, tables, and layout; Reducto says it processes more than a billion pages monthly for AI teams at Scale AI, Harvey, Toast, and Vanta.
- The essay argues that AI’s utility, revenue growth, and lab profitability do not rule out a bubble: those conditions can draw in capital and competitors, leading to overbuilding and price collapses even as consumers benefit. The scale of the compute bets is substantial: Anthropic reported $900.9 billion in commitments through 2045, about half expected before 2030 ends; OpenAI’s compute-spending target for 2030 was about $600 billion, down from a previously touted $1.4 trillion. Anthropic had operating profit in Q2 and, according to the essay, would have had it in Q3 excluding roughly $2 billion in noncash donation-matching charges.
- Packy’s investment lens is to favor durable competitive advantages, scale and cost advantages, and companies able to adopt better technologies rather than be displaced; he warns AI may accelerate both imitation and advantage-building, citing Claude’s weakening of NVIDIA’s CUDA moat. A cited Nordhaus study estimated that producers captured about 2.2% of innovation’s social surplus in 1948–2001, versus roughly 97–98% for consumers—an example of transformative technology creating value without equivalent producer returns.
Andrew Ng
Profile
- Andrew Ng says software engineering is among the roles most affected by AI, but software-engineering job postings are up as AI makes engineers more productive and companies want to hire more of them. He says his teams would not hire an engineer who cannot use a coding agent because their productivity is much lower; he also criticizes universities for graduating computer-science students without experience using coding agents or calling AI APIs.
- Ng distinguishes AI’s ability to get work done from its ability to support learning: he says students using AI for homework can earn better homework grades but learn less and perform worse on closed-book finals. He says Learn Vector, backed by Coursera, is working on AI that helps people learn and scalable learning experiences.
- On AGI, Ng argues that human-level AI capable of any intellectual task is not here yet and may be decades away; he notes that current AI still cannot perform some tasks humans can, while broader definitions of AGI vary.
@jason
- Jason endorsed the post as hard-fought founder advice.
- Ro says its 2022 refounding followed 2021 growth of ~9% YoY at a $270M run rate while burning after raising ~$800M; by 2026, it reported a $3B+ run rate and revenue growth above 100% CAGR over the prior three years. The company says it nearly 10x’d since the reset while headcount fell from 596 at its May 2022 peak to 406 in October 2026; ARR per employee rose more than 10x and People & Opex fell below 15% of net revenue.
- To narrow its focus, Ro shut down 10+ products, cut ~$50M+ in DTC revenue it described as vanity revenue with upside-down unit economics, and ended B2B and external API efforts. It also consolidated its health offerings under one Ro brand, aiming to provide a single trusted place for patients across health categories.
- The operating reset required business cases and functional-lead, CFO, and founder approval for headcount, ended automatic backfilling, and replaced detailed long-term roadmaps with annual priorities and precise 90-day planning. The account cautions that its lessons come from one company and one founder’s experience, with survivorship bias and no guarantee they generalize.
a16z
- a16z’s B2A thesis is that agents are moving from answering questions to completing tasks and choosing products, subscriptions, and purchases; the article says agents already account for most web traffic and production code, while noting that the business models enabled by this shift are still largely unseen.
- Early commerce signals include Shopify-reported AI traffic to stores growing 8× year over year in Q1 2026 and orders from AI-powered searches growing nearly 13×; the article also says AI-referred consumer traffic has higher engagement and conversion and a lower bounce rate than non-AI traffic.
- Coding agents offer an adoption precedent: as of September 2026, more than 40% of GitHub pull requests reportedly started with an agent draft, up from 17% six months earlier; reported infrastructure spillovers include AI tools launching over 60% of new Supabase databases and agents accounting for nearly 40% of Stripe documentation traffic and 70% of CLI requests for API resources.
- The article’s opportunity map includes agent models and harnesses; agent-enabled businesses such as self-running SMBs, AI-intermediated marketplaces, and auto-optimizers; incumbents optimizing beyond discovery for agent-led evaluation and transactions; and agent infrastructure. It flags payments, fraud, identity, and communications as strained even at relatively low agent volumes, while many enterprises still block agent traffic over fraud and chargeback concerns.
Y Combinator
YC resurfaced a Social Radars interview recorded December 4, 2024, in which former Theranos employee Tyler Shultz describes discovering that the lab’s reality differed from the company’s claims, identifying failed blood tests and other red flags, facing a culture of fear, trying to raise concerns internally and with regulators, and encountering legal pressure after speaking to The Wall Street Journal—a cautionary case on technology validation and startup governance.
andrew chen
Andrew Chen congratulated Azra Games and its leader Mark Otero on UnGodly, saying he looked forward to playing it on iPhone Duo . The game's account said it made UnGodly after tiring of waiting for God of War on mobile and urged users to download it .
Harry Stebbings
- Stebbings now sees AI sovereignty as a strategic imperative and describes Mistral as Europe’s remaining frontier-model hope. Mistral says capital—not energy—is its main constraint on adding compute: energy is about 10% of compute’s total cost of ownership, Europe lacks capacity, customers may need thousands of GPUs as use cases scale, and Mistral is investing equity ahead of demand.
- Mistral characterizes inference as a commodity-like market with price-sensitive on-demand sales and a premium tier for mature customers seeking SLAs, reserved physical capacity, and latency and generation-speed guarantees; more precisely specified demand helps Mistral price that capacity.
- Mistral says ambiguous enterprise problems and human-in-the-loop objectives resist self-improving loops, so it uses forward-deployed engineers (FDEs) to define problems better.
- Mistral warns that reliance on two or three centralized model providers could give them significant leverage and let them use customer-generated data to improve systems and compete for those customers’ business; Arthur Mensch reportedly considers that outcome already priced in.
- Mistral says models’ ability to carry out cyberattacks makes security of supply for both infrastructure and model IP important to customers.
- Mistral argues it can go public in Europe, citing worldwide investor appetite for European private and public equity, and says it is preparing for public-company operating constraints while pursuing commercial milestones.
Jessica Livingston
As renewed attention turns to Elizabeth Holmes, Jessica Livingston urges people to hear the story of Tyler Shultz, whom she identifies as the original whistleblower .
Garry Tan
Garry Tan argues that Nvidia’s board would be ill-advised to oust Jensen Huang, saying his leadership is central to the company’s operations and that a transition could cause a massive market-cap decline; this is a hypothetical warning, not a reported leadership change.
Dalton Caldwell
Standard Capital opened applications for its latest Series A funding cycle, with applications due December 1 and applicants hearing back by December 11 . Applications and additional information: https://www.standardcap.com/ .
