ZeroNoise Logo zeronoise
Post
Nvidia's board shuts the door on Huang's return as $5B Isomorphic and $700M Fractile rounds lead the funding tape
•
6 min read
• 2767 docs
The fight over Nvidia's leadership hardened, OpenAI shipped GPT-6.2 to mixed reviews, Perplexity disclosed unit economics behind $1B ARR, and large rounds landed in AI drug design, inference chips and finance agents.

Nvidia: the board rules out Huang, and his backers turn to shareholders

Jensen Huang said publicly that he made "serious mistakes" in his personal life, created problems for the board, and is cooperating with its independent investigation. He added that he would return as CEO if the board decided that served shareholders, employees, the company "and America" . Hours later the chair said the board "does not intend to reinstate" him, according to Polymarket . A campaign backing Huang says the board told the earnings call that he resigned voluntarily and declined to answer questions about him or the investigation . A Semafor scoop says the board is split, with a three-director faction "vehemently" against his return .

Garry Tan answered "15% shareholder vote coming" . According to a Grok summary of the bylaws, holders of 15% of voting power for at least a continuous year can call a special meeting. Removing directors there takes a majority with cause, or 66⅔% without . That is a high bar, not an imminent outcome. Separately, the board announced a "$150B increase" to its buyback. Jason Calacanis argued the money belongs in data-center buildout .

Sam Altman said he didn't know enough to judge whether Huang should return, and that it is "hard to imagine the AI revolution without Jensen." Asked about a reported $105B Nvidia guarantee for OpenAI's Ohio data center, he said people "still trust their guarantees very much" . For anyone exposed to Nvidia-financed infrastructure, the governance fight is now open-ended.

OpenAI ships GPT-6.2, and early evals split

OpenAI released GPT-6.2 Astra and Sol, describing them as "super efficient" . Altman pointed to token efficiency and "Sol ultrafast" , but conceded the models still need work to be "nice to work with" . Third-party results so far:

  • Perplexity (WANDR benchmark): Astra scored 0.783 at about $12.54 per task, up 0.095. Sol scored 0.746 at about $2.65, up 0.065 .
  • Bindu Reddy: Astra excels at browser use but is "nerfed on agentic tasks," and Sol trails Sol 6.1. Her read is that OpenAI is trading performance for cost .

Perplexity puts numbers behind $1B ARR

Aravind Srinivas said Perplexity crossed $1B in annual run rate. He claimed it burned 5x less to get there than "the company that's going to IPO this week" . Its own fast-search model handles 80–90% of traffic on its own compute, which he estimates saves $100–200M a year . Gross margin is 56% on paid users. Counting free users it is 17% in Q3 2026, up from −219% in Q4 2025 .

The products follow the same orchestration thesis:

  • Orchestrator API: 45+ open and proprietary models behind one API, with failover and unified billing. Routing is designed to keep the KV cache intact. In one example, sending Fable 5.2 refusals to GPT-6.1 Sol raised Terminal-Bench 4.1 results from 38 to 41 of 63 tasks .
  • Multiagent: runs parallel mixed-model agent teams. Perplexity says diverse teams beat homogeneous ones on both performance and cost .

Deals worth noting

  • Isomorphic Labs, $5B. Demis Hassabis announced the funding for AI-first drug design . RAAIS describes it as a Series C .
  • Fractile, over $700M Series C (the transcript spells it "Fractal") . Its thesis is that LLM inference is memory-bound. It stacks DRAM tightly on logic and claims 20x the bandwidth per chip of an HBM-based die . It expects customer workloads running in its racks before the end of next year, and says it will never compete with customers at the model layer .
  • Model ML (YC W24), $250M Series B at $1.5B. It runs AI agents inside banks' and consultancies' own systems. Q3 ARR grew 12x year over year and daily active users 26x .
  • River AI, $1B. Igor Babuschkin's company sells software that lets enterprises post-train and serve models on their own hardware. Backers are General Catalyst, AMP PBC, Nvidia and AMD .
  • First Person Labs, $10M seed led by a16z. Its product, Second, keeps a continuously updated model of the user on the device and serves that context to agents the user approves .
  • Alta Ares won 6 of 16 lots in NATO's counter-drone framework, more than any other supplier. Allied countries can now buy without a new tender .
  • Auctor (YC P25) raised a Series A from Sequoia for AI-led enterprise software implementation. Usage rose 5–10x after it rebuilt the product from scratch .

Venture bar and market structure

SaaStr argues a company now needs to grow from $1M to $100M ARR within five years to raise venture capital. That is 151% compounded a year, which makes T2D3 roughly the minimum . Citing Carta, the AI Series A bar is about $3.5M ARR, and AI took over 60 cents of every venture dollar in Q1 2026 . As of May there had been zero software IPOs among 32 listings .

The counterpoint is Stripe's index of non-AI B2B revenue, which is up more than 30% year over year, with firms under a year old growing fastest . But the sample is mostly self-serve, card-on-file companies, not the enterprise names that were sold off .

Newcomer's Cerebral Valley summit added three data points:

  • Open-weight models: they went from about one basis point of Factory's tokens at the start of 2026 to "way into double-digit percent." At many of its customers they now carry more than half of all tokens .
  • Valuations vs. returns: the AI 100 is worth $3.6T, with OpenAI and Anthropic 55% of that. Bessemer estimates $8T of capex needs close to $6T in annual revenue to pay off .
  • Audience survey: 79% see a bubble and 67% expect it to pop in 2027. The same room expects Anthropic to be worth $1–2T after its first trading day .

Agents, autonomy and security

  • Replit Missions. Replit launched this agent-swarm product. It reports NumPy optimization speedups of up to 4x with one agent, 17x with 16 and 19x with 32 . It also shipped DataCat, a free 1:1 rebuild of an observability platform it says costs $83,700 a year . That is a direct shot at feature-layer SaaS.
  • Autonomous exploitation. Google Threat Intelligence Group and Mandiant report the first LLM framework they have seen fully automate vulnerability research and exploitation with no human in the loop. It diffs patched against unpatched web apps, then writes and runs exploits .
  • Waymo off freeways. California's DMV ordered Waymo off state freeways, citing an "imminent hazard to public safety" . A Reddit commenter says the cause was repeat entries into construction zones after a May fix, with no injuries, and that Waymo has paused freeway rides in all 15 markets . Calacanis, who says he is invested in every AV provider, warns that the industry is moving too fast .
  • Qwen4-Max is API-only. It has 2.4T total parameters, and no weights were released . A commenter's summary of draft Chinese export rules says models with 2T or more parameters would need an export license. Newly open-sourcing one would require a "file-and-go" filing, and the rules are not yet final . That would explain the release, but it hasn't been confirmed.
Nvidia's board shuts the door on Huang's return as $5B Isomorphic and $700M Fractile rounds lead the funding tape
All-In Podcast
  • Ali Ghodsi says enterprise AI is still mostly search/chat and coding: customers report roughly 20–30% coding speedups, while even leading users run only about 6–12 agents and keep humans in the loop. He identifies missing organizational context, security and governance risks, token costs without matching revenue gains, and concerns about retaining control of company data and IP as deployment barriers.
  • Databricks’ Genie builds an organizational ontology from people, reporting lines, meetings, decisions, Slack and documents. In a demo, it analyzed a survey with more than 10,000 employee respondents by department and surfaced engineering output and cost patterns across 2,000 engineers. Ghodsi says Databricks monetizes governed-data consumption rather than dashboards and claims agents are driving more queries and business acceleration; this came in response to a question about Anthropic’s dashboard push.
  • Ghodsi expects enterprise AI adoption and its GDP effects to unfold over roughly 10–15 years, citing organizational barriers such as a bank refusing to record meetings. He also says there is no evidence of recursive self-improvement today: he defines the necessary flywheel as faster training, lower resource use, broad accuracy gains and continuous repetition, while current training is taking longer and requiring more resources and manual work.
  • Ghodsi says Databricks will eventually go public but is avoiding a listing during the AI transition to limit stock volatility and distraction; he also says the company has entered security and marketing markets. The interviewer framed an August 2026 estimate of a $7 billion run rate and 80% year-over-year growth as a rumor.
  • On biotech, Ghodsi says Databricks-supported work at Regeneron identified a genome associated with chronic liver disease; he describes its Teddy transformer as modeling gene-regulatory networks to distinguish disease-causing genes from genes reacting to disease, with the aim of accelerating drug discovery.
Databricks CEO: AGI Is Already Here, Why Companies Are Behind, and the Superintelligence Debate
a16z
  • Boring Company announced a $3 billion Series D at a $23 billion valuation in September; a16z participated.
  • The Vegas Loop has carried more than 4 million passengers since opening beneath the Las Vegas Convention Center in 2021; the network now has 29 stations and more than 10 tunnel miles, with local approval for 68 miles and 123 stations. a16z’s post compares the original Vegas Loop at about $62 million per route mile with an average $160 million per route mile for U.S. light rail, and puts the planned Music City Loop at about $20 million per route mile.
  • Boring’s described business model has property owners pay for stations, riders pay fares around half the price of a rideshare or less, and the company buy land along routes to capture some of the value created; Davis says the company does not seek taxpayer dollars. The company argues that each added station makes the network more useful.
  • At its Bastrop factory, about 200 people build a boring machine every two months, with a target of one per month; the company makes about 99% of Prufrock-6’s parts in-house. Prufrock’s continuous-mining approach advances while installing tunnel lining; the feature reports that mining time rose from 5–10% to more than 90% after about two years of development. The feature puts current tunneling cost below $10 million per mile, with Davis expecting $4 million to $5 million per mile.
  • In Nashville, the project received its permit and began tunneling on February 25, 2026; by August, two machines were mining and a third was in final assembly. Davis expects the first mile to carry passengers in Q1 next year, while Boring expects to finish all 22 tunnel miles in two to three years. Dubai’s Roads and Transport Authority signed a construction contract for a four-mile, four-station pilot, with tunneling planned for late 2026; the broader UAE partnership covers more than 150 kilometers. Musk also said the company is working on a precursor Hyperloop tunnel between Austin and San Antonio, targeting speeds above 200 mph and a consistent trip under 30 minutes.
  • Permitting and contracting remain material risks: the Chicago deal fell through before a contract was finalized, Washington–Baltimore and the proposed New York–D.C. Hyperloop stalled in federal environmental review, and a Los Angeles tunnel plan was shelved after a lawsuit. By contrast, Davis says Nashville’s entire project was permitted in one shot.
The Boring Company & the Future Underground The Boring Company builds transit for a fraction of the cost in a fraction of the time. The original Vegas Loop cost about $62 million pe…
Yann LeCun
Profile
  • The interview identifies Yann LeCun as founder and chairman of Paris startup AMI Labs and says he left Meta in 2025. He argues that current LLMs mainly accumulate declarative knowledge rather than the adaptability needed for intelligence, while still calling LLMs useful, worth investing in, and potentially components of future systems. He sees code and mathematics as particularly strong LLM use cases when systems can search for answers and check their own outputs, though runtime search is expensive.
  • LeCun says he is still working on JEPA, which predicts in abstract representations rather than reconstructing every pixel, discarding unpredictable visual detail while retaining useful structure. He describes action-conditioned world models as a way to predict the consequences of actions and plan, including by learning from video paired with actions.
  • LeCun advocates open, free foundation models as infrastructure, with commercial applications built on top through integration and fine-tuning for vertical uses and different cultures. His proposed Project Tapestry would let participants contribute data and compute while retaining their data and exchanging model parameters to train a shared model; he says this could make currently unavailable or undigitized data usable.
AI Memorized the Internet. That's Not Intelligence. | Yann LeCun
Greylock
  • Simile trains its own foundation model to predict and simulate human behavior, supplementing web-derived information about what people say with observational data such as transactions and mobility, plus intervention studies designed to measure how behavior changes under different products or messages. The founder says participants in these studies consent and are debriefed.
  • Simile jointly trains a confidence model with its core model to estimate how well its data covers a particular prediction task; the platform surfaces confidence at population, session, and individual-query levels. The founder says it became a core feature for customers making business-critical decisions.
  • Its SaaS platform lets customers specify a population and question, run surveys or experiments, or have agents experience a linked product; outputs include focus-group-style feedback and statistical results, with workflows such as conjoint and max-diff analysis. The founder says Simile has served some Fortune-company use cases, including simulating earnings calls and rehearsing important conversations; the interview names no customers or deployment scale.
  • Simile describes individual and simpler multi-agent simulations as being productized, while broader multi-agent work remains in progress or research; societal-scale simulation is a later ambition. Its proposed intermediate steps include modeling a customer's environment using integrations such as CRM and databases.
  • The founder sees the product evolving from a reactive tool for answering research questions toward a proactive simulator that starts with a desired outcome—such as launching a product without backlash or reaching a revenue target—and explores paths to achieve it.
  • The founder says roughly 20% of the company consists of former Stanford labmates associated with the cofounders, alongside engineering, product, and go-to-market teams. Simile keeps research about a quarter ahead of the rest of the company; in that phase, the research team is roughly 75% focused on pure research, then shifts toward product adoption when a technology is ready.
AI Simulations to Enable Better Decision Making
Lightspeed Venture Partners
  • Computesk is building capital-markets infrastructure to price, hedge, and deliver AI compute. Lightspeed’s thesis is that compute is becoming a financial asset while the infrastructure to manage it at scale is missing; the founder says the market lacks options, futures, forwards, swaps, and supporting market microstructure, with most trades still bilateral and opaque.
  • Compute Signal uses exclusive, real-time neocloud invoice data from executed trades—not quotes or listed prices—to produce pricing and indices; the founder says companies such as Meta already obtain compute pricing through it. Compute Trader translates exposures into trades and manages execution, while Compute Clear pairs clearing with physical settlement and checks delivered specifications and service obligations.
  • The company’s strategic thesis centers on physical clearing as a difficult-to-replicate way to attract natural buyers and sellers and aggregate liquidity; benchmark adoption, it argues, depends on rates accurately reflecting actual trading costs. In response to concerns that more efficient markets could weaken hedging demand, the founder says tighter spreads may persist alongside disagreement about future compute value.
  • Founder Andrae Spahu combines chip-design and accelerator work, exotic-options pricing, and responsibility for GPU cluster buildouts at Meta; the company describes its team as split between capital-markets and GPU-infrastructure expertise. Lightspeed says it spent seven months developing a financialization-of-compute thesis before partnering with the founder.
The Financial Market Emerging Around AI Compute | Andrawes Bahou, Compute Desk
Sam Altman
Profile
  • Altman said AI is creating and disrupting substantial value, with investors appearing to recognize the capital being invested in the sector. He said OpenAI will go public at some point, while noting drawbacks to being public and describing the industry as volatile, with high-stakes safety and societal decisions.
  • Altman said OpenAI is developing new models faster using its technology, but is not having models autonomously design the next version and rapidly self-improve without human oversight; he said pushing further toward recursive self-improvement is not safe currently.
  • He said AI should remain a tool people use, rather than take over decisions about the future.
LIVE: Sam Altman Says AI Revolution Is Hard to Imagine Without Jensen Huang
Aravind Srinivas
Profile
  • Perplexity said it had crossed $1B ARR; Srinivas said burn to reach it was 5× lower than that of an unnamed company due to IPO that week. He put headcount at about 500 and said Perplexity avoids the massive compute commitments made by frontier labs.
  • Perplexity’s own fast-search model served 80–90% of its traffic, using its own compute, post-training models and inference stack; Srinivas estimated this saved $100–200M annually. He reported 56% gross margin on paid users; including free-user costs, margin was 17% in Q3 2026, versus -219% in Q4 2025.
  • Perplexity announced a frontier research program committing at least 10,000 GPUs to post-train models for multi-agent orchestration and continual learning; Srinivas described the company’s strategy as orchestrating multiple models rather than competing with frontier labs.
  • Srinivas framed fully local/on-prem AI as a proposition for sovereign wealth funds and regulated or data-sensitive sectors, including banks, credit-card companies, exchanges, insurers, pharma and bio; he said local models could escalate to cloud frontier models with user-granted access. Perplexity’s hybrid mode on Apple Silicon lets users start with frontier cloud models and orchestrate down to local models for private data; he reported good adoption and argued that local AI could expand, rather than displace, frontier-AI usage.
  • Srinivas said GLM took about 15 minutes to complete a task that Opus completed in about 3 minutes, largely because of erroneous tool calls; he estimated an open-weight model could match then-current Opus within 6–8 months, while noting frontier labs would continue advancing.
Perplexity CEO and Founder Aravind Srinivas at Silicon Valley & The World | Semafor
Sam Altman
Profile
  • Altman declined to say whether Jensen should return, saying he did not know enough; he said it was hard to imagine the AI revolution without Jensen and credited his ambition, while expecting the industry to remain fine given rapidly growing AI demand and revenue.
  • The interviewer referred to what he estimated at $105 billion in Nvidia guarantees for OpenAI’s Ohio data center and asked whether they were still as valuable; Altman replied that people still trust Nvidia’s guarantees.
  • Altman countered the framing of investor skepticism around AI, saying skepticism was more apparent among journalists and investors seemed bullish; he anticipated price swings and overbuild/underbuild cycles alongside substantial value creation and disruption. OpenAI plans to go public at some point: Altman cited broader public access to investment as an upside, while noting public-company disadvantages amid industry volatility and high-stakes safety and societal decisions.
Should Jensen Huang Come Back? Sam Altman Answers
  • Frontier researchers describe recursive self-improvement (RSI) today as AI accelerating human AI research, including pretraining and reinforcement learning: models can run scoped experiments end-to-end, but fully autonomous AI research or models creating new models entirely on their own is not yet here. A report cited in the interview estimated roughly 2× acceleration in AI-research progress over the prior year, likely driven mainly by human researchers using AI rather than full automation.
  • Human judgment remains a bottleneck: models can handle routine execution and hyperparameter tuning but are not yet strong at “taste”—choosing ideas with long-term value. One researcher said AI-generated experiments sometimes produced interesting leads but needed detailed human follow-up to make further progress.
  • Evaluation and compute constrain further automation: downstream evaluations have shorter shelf lives, driving work on new, harder benchmarks; AI research also requires running experiments, so progress may be bounded by compute and benefit from smaller experiments that predict larger-scale results.
  • Alignment presents both risk and potential: researchers warn that long-horizon goal or value drift could lead to actions such as copying model weights outside a lab. An August study cited in the interview reported that automated AI researchers found fixes for 10 alignment failures; some worked on models nearly 5× larger, and the best automated methods outperformed experienced human researchers.
OpenAI’s Cofounder & Research Chief Explain Self-Improving AI
Stanford eCorner
  • Acrew’s founders were asking for help taking technical, product-oriented AI-native startups to market, while Kolodny saw few people with experience commercializing AI-native products. The firm recruited Aliisa Rosenthal, OpenAI’s first sales hire, whose sales-team tenure spanned reported revenue growth from $10 million to $10 billion, to fill that gap.
  • For founder diligence, Kolodny asks about a candidate’s last state of flow to surface intrinsic motivation, asks them to name one outlier capability, and cross-checks motivation through references. She warns that charisma can obscure weak founder–opportunity fit: in one Series A with seven term sheets, she mistook business opportunity for founder passion and partnership depth for product depth; at the first board meeting, the product was breaking, retention was falling, and the founders had lost trust in each other.
  • Kolodny sees consumer experiences as ripe for reinvention and says she is beginning to see a new consumer wave beyond foundation models. She identifies everyday Americans living paycheck to paycheck—a population she says is overlooked despite comprising most of the country—and older consumers as opportunities; she says Chime challenged the assumption that serving the former profitably requires predatory practices, and cites strong uptake for a portfolio company addressing digital scams and fraud as AI enables more fraudulent scams.
  • Kolodny says she is advising her companies to hire new graduates, arguing that hands-on AI immersion gives them an advantage; she emphasizes adaptability, critical thinking, and building AI agents or products over a particular major.
Lauren Kolodny (Acrew Capital) - How to Find Your Spark and Spike [Entire Talk]
Sam Altman
Profile
  • Sam Altman said OpenAI will go public “at some point,” while citing real downsides to public-company status during a volatile period with high-stakes safety and societal decisions; he also viewed public-market access as directionally positive for giving the broader public exposure to the technology.
  • He said he was not worried that Anthropic going public first would leave too little investor demand, arguing that the previous year had seen too much demand and not enough supply.
  • Altman said chipmakers’ sales expectations may double-count demand: multiple vendors expect OpenAI, Anthropic, and others to buy their chips, but he said buyers cannot or need not source all their supply from every vendor. He characterized data-center buildout as solid while warning that the buyer-side demand is finite.
Sam Altman On Anthropic Listing & OpenAI IPO Plan: 'Real Downsides To Going Public'
a16z
  • The Boring Company announced a $3 billion Series D at a $23 billion valuation; a16z participated.
  • Vegas Loop has grown from an initial three stations and 1.7 miles of tunnel to 29 stations and more than 10 miles; the system has carried over 4 million passengers without a passenger-related safety or security incident.
  • The company reports tunneling costs below $10 million per mile, with Steve Davis expecting $4 million–$5 million; its Bastrop factory of about 200 employees builds a machine every two months and is targeting one per month, while in-house parts production rose from about 30% on the prototype to 99% on Prufrock-6.
  • Its model has property owners pay for stations, fares priced at about half a rideshare or less, and the company buying land along routes; it says it uses no taxpayer dollars and sees network value increasing as stations connect more destinations.
  • Expansion is underway: Nashville tunneling began the day its permit was issued on February 25, 2026, and by August two machines were mining while a third was in final assembly; a Dubai pilot contract covers about four miles and four stations, with tunneling planned for late 2026. Regulatory execution remains a caveat: the Washington–Baltimore project stalled in federal environmental review and was declined by a highway regulator, while the Los Angeles plan was set aside after an environmental-review lawsuit; Davis says Nashville’s entire project was permitted in one shot.
The Boring Company & the Future Underground
Sam Altman
Profile
  • Altman says OpenAI has encountered attempts to place cameras in conference rooms and get engineers to disclose pre-training secrets; he believes the company has sometimes been tricked and says its security has not been perfect.
  • He describes a pragmatic approach: not developing AI in absolute secrecy, but not treating it like any other technology either, while bringing the world along through iterative deployment.
  • OpenAI is using its own technology to develop models faster, Altman says, but not to trigger rapid self-improvement without human oversight.
Sam Altman Reveals Attempts To Steal OpenAI's AI Secrets: 'We've Been Tricked Couple Of Times'
Invest Like The Best
  • Brad described the Invest America Act as signed on July 4, 2025, and said it creates automatic capital accounts: newborns receive $1,000 invested in the S&P 500, while accounts had been opened for 70 million children under 18 and philanthropists could fund accounts beyond the newborn group.
  • On AI, the speakers described young people hearing existential-risk and job-loss claims, seeing layoffs, and reacting to data-center expansion; Brad argued that people need to share in AI's upside, citing Anthropic's reported interest in direct citizen accounts and Elon Musk's reported call to use 53A accounts, alongside an announced plan to give children SpaceX shares.
  • A Rogo sponsor spot pitched AI agents for financial firms to screen deals, draft sims, conduct buyer outreach, and diligence data rooms, with security and compliance handled; it claimed use at hundreds of leading firms.
Giving Every Kid a Stake in Capitalism | Brad Gerstner & Geoffrey Canada
Fei-Fei Li
Profile
  • Fei-Fei Li is AMD’s chief scientist, and World Labs has become part of AMD; its product trajectory is still being worked out. Li describes the partnership as deeper than compute access, emphasizing model–hardware co-evolution and expanding AMD’s open ecosystem.
  • World Labs has released Marble and Atlas, with Atlas described as the more advanced model; users include game designers, VFX and independent studios, and architects, while robotics companies are partnering with the company.
  • Li frames world models as connecting pixels and 3D/4D worlds with reasoning, generation, and understanding for tasks beyond language; she sees applications in design, gaming, VFX, robotics, and other industries.
  • Li says universities and the public sector lack the compute and tokens to keep pace with AI, and calls for public-sector investment and pragmatic, science-based governance that includes evaluation and incident reporting.
World Labs Co-Founder & CEO Fei-Fei Li at Silicon Valley & The World | Semafor
Lightspeed Venture Partners
  • Fractal announced a Series C of over $700 million; after four years in operation, it describes itself as a chip and frontier-AI company building hardware to advance AI capabilities, with a full-stack effort spanning systems, software, ML research, and advanced packaging.
  • Fractal’s thesis is that LLM inference is memory-bound: HBM bandwidth can leave compute underfed, while pushing existing platforms for speed can reduce utilization and increase joules per token. Its approach stacks DRAM tightly with logic to increase memory bandwidth.
  • The company says its DRAM approach provides 20 times more bandwidth per chip than an HBM-based die and more than an order of magnitude more memory capacity than an SRAM-based chip; it favors a general-purpose design over hardwiring a rapidly evolving model architecture.
  • After the round, Fractal said it had added trials and targeted workloads running in Fractal racks before the end of next year. It is expanding systems capabilities in Palo Alto, including high-power-density racks and networking, and said Dan Rabinovich—who previously led AI hardware work at Meta—joined to lead that effort. Fractal also said it would remain a hardware business and not compete with customers on the model layer.
Fractile's $7B Bet on Chip-First Frontier AI | Lightwork
Sam Altman
Profile
  • Sam Altman said chip-demand estimates may be double-counted because multiple suppliers expect OpenAI, Anthropic and other buyers to purchase their projected volumes, although those buyers will not source all their needs from every vendor. He said long-term compute demand remains strong, but matching supply and demand quarter to quarter or year to year is harder to predict.
  • He described AI infrastructure as facing shifting supply-chain constraints, including memory, rack assembly, data-center capacity and components such as PCB laminate.
  • Altman said opposition to AI and data centers reflects immediate job concerns and longer-term fears about people’s role and the balance of power between those with and without substantial AI compute; he argued the industry should commit to people retaining control of the future.
Sam Altman's Unfiltered Answer On The Data Centre Debate & Public Fears Of AI
Garry Tan

Garry Tan urged Nvidia shareholders and current or former employees to support a campaign about Jensen Huang’s return. The campaign says the board stated Huang voluntarily resigned and is not considering his return, and asks the board to explain that position and provide an update on the investigation’s process and expected timeline. A report Tan shared says the Nvidia board is divided, with a three-director faction opposed to Huang’s return. Tan also floated shareholder action; a Grok summary of Nvidia’s bylaws says holders with 15% voting power for one continuous year can call a special meeting, while director removal requires a majority vote with cause or 66⅔% without cause.

If you’re a shareholder, employee or former employee your voice is needed here [https://x.com/thejensenletter/status/2122912156664692875]… Jensen Heard us, Now the board needs to Scoop with [@rogoswami](https://x.com/rogoswami) - Nvidia board is divided into factions, with one three-person faction vehemently agains… Only three people to remove from the board? Easy work 15% shareholder vote coming [https://x.com/reedalbergotti/status/212287965019630003… The bylaws do allow holders of 15% voting power for at least one continuous year to call a special meeting. There the full shareholder ba…
a16z
  • The Boring Company announced a $3 billion Series D at a $23 billion valuation in September, with a16z participating.
  • The Vegas Loop has carried more than 4 million passengers since opening in 2021; the network now has 29 stations and more than 10 miles of tunnel, with local approval for 68 miles and 123 stations.
  • Boring says property owners fund stations, fares are about half the price of a rideshare or less, and the company uses no taxpayer dollars; it reports tunneling costs below $10 million per mile and expects $4–5 million per mile.
  • The company says Prufrock’s tunnel crew requirement fell from about a dozen people to two; it demonstrated crew-free continuous mining in May 2025 and reports that tunnel-ring placement was fully autonomous by August 2026.
  • Expansion includes Nashville’s privately funded airport connection, where tunneling began the day permits were issued in February 2026 and two machines were mining by August, plus a Dubai pilot planned at about four miles and four stations; tunneling there is planned for late 2026, within a broader UAE partnership covering more than 150 kilometers.
The Boring Company & the Future Underground
Two Minute Papers

OpenAI’s AI system uses Balaban’s approach to track interactions across multiple grid scales and produced a proof in Lean whose checks passed. The proposed result addresses only the problem’s existence component; the mass-gap component remains unsolved.

A review led by physicist David Gross is still underway, though initial results are described as encouraging. The code contains about 250,000 declarations, 60% of which are unused in the final proof; the video notes that this does not mean the proof is wrong, but makes it harder to understand.

OpenAI's New Breakthrough Is... Utter Chaos