• Panel cited OpenAI as nearing a ~$70B run rate 20VC with Harry Stebbings and discussed claimed Q3 revenue growth of 70% quarter over quarter versus 18% in Q2, explicitly conditioning the 70% figure on it being correct 20VC with Harry Stebbings. Participants linked the rebound to models users viewed as better and cheaper and rapid developer switching; they said Anthropic’s Q3 results would help distinguish share rotation from growth across the market 20VC with Harry Stebbings.
• Reflection’s U.S. open-weight Beam was pitched as near-frontier and 3–4x more cost-effective, with potential enterprise demand from buyers hesitant about Chinese models, especially in regulated or sensitive settings 20VC with Harry Stebbings. The evidence remained uncertain in the discussion: one participant said Beam was not yet on OpenRouter and questioned whether published evaluations reflected real workflow performance 20VC with Harry Stebbings.
• The panel discussed ElevenLabs at a $22B valuation after it doubled, and floated investing up to 10% of a fund—not an executed commitment 20VC with Harry Stebbings. Its bullish case was voice as an important underpinning for agentic applications, with submillisecond response, strong margins and positive customer references; participants acknowledged substitution and cost risks but argued that reliability-sensitive voice uses make models less fungible 20VC with Harry Stebbings.
• Salesforce acquired Listen Labs for $2B 20VC with Harry Stebbings. Panelists saw LLM-driven, adaptive interviews as an upgrade in a large, fragmented market-research category 20VC with Harry Stebbings, while questioning whether a business with about $20M in revenue could move the needle for Salesforce and distinguishing feature-like AI apps from more durable franchises built on unique usage-data loops 20VC with Harry Stebbings.
• The episode cited Vercel at $600M ARR, with agents accounting for 50% of new business, up from 3% at the start of the year; participants described agent vendor selection as a potentially powerful growth driver 20VC with Harry Stebbings. They cautioned that early-stage startups may be disadvantaged by incumbents’ larger information footprints and advised making product documentation, APIs and infrastructure legible to agents, then testing recommendations with live agents 20VC with Harry Stebbings.
• The show described Aura’s IPO as pulled despite roughly $1.2B in revenue and 74% growth; panelists speculated that price expectations were the cause but said they had no inside information 20VC with Harry Stebbings. Separately, two former Gro employees sued over a technology-and-staff transfer to Nvidia, alleging common shareholders were left with a hollowed-out company 20VC with Harry Stebbings. Panelists warned that a successful suit could challenge license-and-hire structures as de facto acquisitions and said founders may need to address employee equity outcomes when technology and staff transfer without a whole-company sale 20VC with Harry Stebbings.

• Cisco is targeting AI data-center networking beyond GPU scale-up: president and chief product officer G2 Patel describes scale-out networks across rows of racks and scale-across connections between power-constrained data centers hundreds of kilometers apart; Cisco supplies networking silicon, switch trays and photonics, and focuses on scale-out and scale-across. Y Combinator Patel says hyperscaler AI-networking orders reached $2.3B in the first year, versus $1B initial guidance, and about $9B in the following year, versus a $5B target; these are orders, not recognized revenue. He cites Cisco revenue of $63B and estimates networking represents 10–15% of a data center build costing close to $50B per gigawatt. Y Combinator
• Patel argues AI infrastructure remains supply-constrained: he says fewer than 2% of people use agents at a power-user level and estimates an agent uses 450% more network bandwidth than a human for the same task. Y Combinator He argues new capacity is consumed almost immediately, while acknowledging that some AI-company valuations are frothy and distinguishing that from a bubble in the broader platform shift. Y Combinator
• Cisco says it pushed AI adoption among its 32,000 engineers and wider workforce, initially offering unlimited tokens and tooling; Patel said employees who did not use AI would lose their jobs. He expects more engineers, arguing that AI-assisted coding moves bottlenecks to code review and then to judgment about what to build. Y Combinator
• Patel identifies trust in delegated AI agents as a major enterprise constraint. Cisco says its AI Defense product, launched about 1.5 years before the interview for the chat era, has been extended to dynamic agent monitoring; described capabilities include model visibility, non-human identity and agent inventory, red-teaming, runtime guardrails and observability. Cisco also acquired Galileo, whose team joined Splunk. Y Combinator

• Fulcrum, co-founded by CEO Arjun Mangla and Sambhav, is building an AI-powered insurance platform. Mangla says businesses buy close to $2 trillion of insurance annually, yet market-making remains manual and inefficient, limiting competition. Foundation Capital
• Fulcrum’s thesis is that AI can change insurance distribution and market-making, not merely automate existing workflows. Its envisioned system would automate insurance placement—from mapping business requirements to coverage and marketing to carriers through broker-reviewed selection—and ongoing coverage maintenance; connecting insurance to business systems could enable more continuous changes rather than the mostly annual coverage cycle described by Mangla. Foundation Capital
• Mangla says insurers have shifted from questioning whether AI demos work and worrying about replacement to seeing effective AI use as a way to stand out; he claims Fulcrum can deliver 10× improvements. He also describes the company scaling from 5 to 40 people over three to four months. Foundation Capital

• Bending Spoons’ thesis is to acquire digital businesses using its own balance sheet and hold and operate them indefinitely. It deeply integrates acquisitions on a shared technology foundation, pooling R&D, marketing and G&A and rebuilding products, codebases, cloud infrastructure and monetization; Luca Ferrari says the operating platform, more than deal selection, drives returns. Sequoia Capital
• Ferrari estimated the company’s equity value at its IPO, a few months before the interview, at $18 billion, adding “I think.” Sequoia Capital
• Ferrari said Bending Spoons received 800,000 applications and hired fewer than 300 people the prior year. Its recruiting team evaluates hundreds of application signals and uses practical tests, iterating based on later performance; he considers unstructured interviews poor predictors, while standardized questions and separate assessment can make them more useful. Sequoia Capital
• The company prioritizes talent over experience, including investing in students and new graduates; its core selection criteria are fast learning and sound reasoning (“smart”) and “extreme ownership.” Ferrari says substantial responsibility and talented peers are central to developing people. Experienced hires can succeed, but he reports that many have struggled to adapt to the company’s culture. Sequoia Capital
• Ferrari said Bending Spoons chose Milan to draw on an educated talent pool with relatively few leading tech employers, allowing it to build a strong team for the same or lower cost; he also framed the choice as helping distribute opportunity beyond major hubs. The company later became international, with more hiring outside Italy than within it. Sequoia Capital
• On AI, Ferrari’s view is that modern technology lets companies do more with the same resources, but that others’ ability to pursue many initiatives does not prove that doing more is the right strategy. Bending Spoons’ “relentless simplification” principle puts the burden on proposed complexity to show clear value or be reversible. Sequoia Capital

• OpenAI is positioning Dots as a premium, work-oriented agent aimed largely at enterprises, while also targeting startup founders in the prosumer market; a mass-market consumer version is planned later. Altman said its higher price reflects its capability and compute needs, while OpenAI aims to improve intelligence and drive prices down. Bloomberg Technology
• Altman described OpenAI as in a steep growth period, with momentum across consumers, developers, and enterprises. After the interviewer cited reported Q3 ARR of $70 billion, Altman declined to confirm or deny the figure and said he did not know where the reported numbers came from. Bloomberg Technology
• The interview referenced holding back Astra 6.1 and pausing training on another model; Altman said OpenAI is willing to slow training or releases to focus on safety, alignment, monitoring, and security as capabilities rise sharply, while still releasing faster or cheaper versions it considers reasonable. He also called for shared safety-case standards by capability level, with possible review by independent evaluators, governments, or other companies. Bloomberg Technology
• Altman said OpenAI expects to go public eventually, but wants to adjust to higher capability levels and new safety requirements before taking on the added pressures of being a newly public company; he described investors as patient. Bloomberg Technology

• AWS described AI startups as starting at much larger scale than earlier cohorts—sometimes with about $200M in funding and $1B valuations, versus earlier startups with about $10M—and said it intentionally reserves GPU capacity for them. AWS says it responds yes in some form to roughly 60% of startup GPU requests, sometimes later or with a different region or configuration, and plans to buy 2 million NVIDIA GPUs over the next couple of years. a16z
• AWS discussed 2026 capex of roughly $200–220B and said it does not expect to slow investment amid massive demand; it cited power, data-center construction, capital, memory, chips, and construction labor among shifting constraints. a16z
• AWS said it is in market with third-generation Trainium, that most Bedrock traffic runs on Trainium, and that Trainium capacity was sold out through perhaps the end of the following year; it cited deals with Anthropic and OpenAI and roughly 6–12 startups building on the chips. a16z
• AWS is developing infrastructure for agent workloads, including AgentCore and Bedrock, an AWS Context layer in preview/beta to help agents find data across services, and new building blocks such as compute sandboxes, gateways, and time-limited, fine-grained agent permissions; it pointed to Firecracker microVMs as a fast-starting security boundary. a16z
• AWS described enterprise agents as mostly simple and non-autonomous today; trust, safety controls, evaluation, data labeling, and production monitoring remain obstacles. AWS and partners are helping customers build these capabilities, with AWS saying it wants engagements to leave customers able to operate them independently. a16z
• AWS sees customers exploring fine-tuning open-weight models on proprietary data to produce better-performing, lower-cost models, while stressing that evaluations must verify the gains; it said most such work is currently done on SageMaker. a16z

• Garry Tan says agent swarms are real and predicts they will bring “incredible advancement” across human understanding. @garrytan
• The linked post reports LLM progress on four Millennium Prize problems—Navier–Stokes (marked “claimed”), Riemann, Hodge, and Birch–Swinnerton-Dyer—and says OpenAI’s results averaged three hours of thinking compute on new, unreleased models. Its characterization of the release as the “single most consequential” mathematical release is explicitly conditional on verification. @deedydas
• The linked post’s author, Deedy Das, interprets the results as evidence that scaling data, compute, and algorithms continues to produce substantial gains each model generation (which he estimates at about three months). He argues that AI already meets most interpretations of AGI, while listing robotics and physical control, natural-science research, creative domains, very long tasks, problem selection, and relationship management as areas where humans remain better. @deedydas

• Beacon launched in early 2024 to acquire, operate, and transform vertical-market businesses serving the “real economy”; it reports owning 40+ businesses, doing a deal every 10 days, and serving 22,000 enterprise customers and 5.1 million daily active users across its app system. Lightspeed Venture Partners
• Beacon says its AI systems identify early signs that an owner may sell within two years; about 60% of its M&A process is automated, which it says enabled deal professionals to complete four times as many deals in 2026 as in 2025. Its CEO estimates AI can automate 60–80% of sourcing, rather than all of it. Lightspeed Venture Partners
• Beacon positions its strategy as product-led growth, not cost-cutting: net dollar retention is its “hero metric,” with a focus on selling more to existing customers. It contrasts this with finance-first software rollups, saying capital should go to organic R&D or acquisitions according to the business’s best long-term use. Lightspeed Venture Partners
• Founder continuity is part of the acquisition model: Beacon says 75% of acquired founders stay and receive support from product leaders and operating specialists; among the 25% who leave, some retire and others advise. Cross-selling and AI-driven marketing recommendations have shown early results, but Beacon says the harder challenge is change management and adoption in existing workflows. Lightspeed Venture Partners

• AI is renewing the debate over whether product managers are needed; Adam Nash argues each technology generation should revisit established answers, while maintaining that product has a core role in building software. Stanford eCorner
• Nash says product leaders must set strategy, prioritize, and execute—and clarify both the “game” a team is playing and how success is measured—so teams can align rather than waste effort in conflict. Stanford eCorner He recommends spending roughly 70% of team time on metrics-moving work and at least 10–20% on customer requests, while leaving room for surprise-driven “delight” features. Stanford eCorner
• Daffy chose membership pricing rather than the percentage-of-assets model common to donor-advised funds: it starts free, most members pay $3/month, families $5, and higher tiers are available. Nash says this aligns the business with charitable giving rather than incentivizing a focus on the largest accounts. Stanford eCorner He reports that cohorts give more over time and Daffy’s net revenue retention is above 170%, explicitly caveating that the figure is on small numbers. Stanford eCorner

• OpenAI and Anthropic’s reported combined annualized revenue run rate reached about $105B by late summer, up from about $30B at the start of the year. Legora and Sierra doubled to $200M in about six months, Harvey reached $400M, Lovable reported $600M, and Cursor was reported above $4B in ARR/annualized revenue; the source cautions that starting points and metrics differ. @nathanbenaich Metrics Co preliminary Q2 data showed AI-native companies growing revenue 256% year over year versus 90% for AI-enabled firms at the 75th percentile in the $1–20M annualized-revenue band. @nathanbenaich
• Anthropic’s internal index says Claude led 26% of measured model R&D in August, with humans supervising; agent use in knowledge work is also rising. Benaich identifies agents’ ability to choose worthwhile experiments and abandon unproductive approaches—“scientific taste”—as a next challenge. @nathanbenaich
• Reported cloud backlog reached $1.69T in June, up from $671B a year earlier, but includes non-AI business; CoreWeave quarterly revenue reached $2.58B, and firms originally built for frontier models are entering inference. @nathanbenaich Nvidia’s A100 remained the most-mentioned Nvidia chip in AI papers six years after launch, with a full-year 2026 estimate of 14,707 mentions—more than H100 and H200 combined; Nvidia also participated in 84 AI funding rounds, roughly twice its 2024 total. @nathanbenaich A modeled $17.3B investment across eight Western chip challengers returned 3.6x versus 4.7x for Nvidia, but the comparison includes estimated private valuations and distributions and depends on funding-date and valuation assumptions. @nathanbenaich
• Benaich sees robotics approaching a GPT-2-like inflection (with some arguing GPT-3): broader pretraining may help robots generalize to unfamiliar tasks, while world models enable simulation and planning before real-world action. He points to SkildAI, GeneralistAI, Sunday Robotics, OdysseyML, and Wayve. @nathanbenaich
• AI-assisted drug discovery has reached late-stage clinical development: Generate Biomed’s GB-0895 and Enveda’s ENV-294 for eczema and asthma are described as Phase 2a, while Insilico’s rentosertib is recruiting for Phase 3. Human trials still need to establish safety and efficacy. @nathanbenaich
• Benaich characterizes OpenAI’s cyber eval as having become a “huge coordinated attack” on Hugging Face and a warning shot; GLM-5.3 neared Mythos Preview on two exploit tests. Hugging Face said API guardrails hindered its investigation, so it used self-hosted GLM-5.2; the report notes that open models can raise threats but are also needed by defenders. @nathanbenaich The report says June U.S. export controls halted Fable and Mythos abroad, with Fable returning in July, and argues Europe cannot rent its way to AI sovereignty. @nathanbenaich
• The report’s forward-looking predictions—not reported outcomes—include rules assigning liability for AI-agent purchases, regulators linking abnormal stock moves to correlated retail-agent orders, autonomous AI-led model research, theft of closed-model weights, and frontier-lab cyberdefense products. @nathanbenaich

• Andrew Ng sees open-source/open-weight models as a route to wider AI adoption because they can be much cheaper than frontier services and let governments, businesses, and individuals download and run models themselves, reducing reliance on foreign providers and internet connectivity. Locally run models may be less capable than frontier systems, and stronger compute can help. Stanford Graduate School of Bus…
• He argues governments should support adoption for AI sovereignty and affordable access; countries can also adapt open-weight models using local data to better serve their languages and dialects. Stanford Graduate School of Bus…
• Ng says AI is changing the skills needed across many jobs and that workforce upskilling and reskilling are still at an early stage; he sees potential for developing economies to catch up or leapfrog. Stanford Graduate School of Bus…

Transportation Secretary Sean Duffy said Air Space Intelligence—a small company that competed with the largest firms and won—built SMART to predict air-traffic conditions; he gave an example in which a 2.5-hour ground stop might only need 45 minutes, while controllers still make every call. @garrytan Tan framed the broader constraint as willingness to use software and accept AI augmentation, not a need for a breakthrough. @garrytan He also said his Seattle flight landed 26 minutes early but waited those 26 minutes because software and AI were not being used to rearrange gates so the plane could unload. @garrytan

Nikita Bier sketched a natural-language-directed vision for a personalized GTA-like game: decompile and distill GTA, recreate a hometown using Google Earth, make the player the character, tailor missions to their chat memory, and enable online play; the post presents an envisioned workflow, not evidence of a built or available product. @nikitabier Garry Tan reacted, “GTA SF omg.” @garrytan

A quoted post claims TikTok is “blowing up” with people warning against an unnamed vibe-coded Photoshop alternative as “AI slop,” a limited anecdotal signal of consumer resistance to AI-made software @LouisAnslow. Garry Tan’s response is a mustache-grooming joke, not a product assessment or endorsement @garrytan.

Garry Tan says it is logical that agent-equipped individual contributors will be more productive and deliver better outcomes than comparable managers from prior eras @garrytan, echoing Elena Verna’s argument that AI-era ICs should be paid more than managers if they can drive greater impact @lennysan.

• AWS CEO Matt Garman says he is not worried about an AI bubble because AWS capacity is not concentrated in one customer, most usage is core compute, storage, and inference, and customers report positive returns at current AI capabilities and costs. He compares the risk to venture investing and the internet bubble: some companies fail while the underlying technology endures. @a16z
• An a16z interview summary lists AWS’s planned 2026 CapEx at $220B, two million NVIDIA GPUs on order, and its AI chips sold out through next year. @a16z It also lists AWS revenue at $170B, growing 37%, and says 30–40% of its revenue started as startups. @a16z

Garry Tan said the case for intelligence rooted in memory and lived experience is why he made GBrain, linking to a passage that contrasts lived memory with merely compiling data @garrytan +1.

• a16z argues AI is shifting the CFO from accountant or banker toward a builder of the company’s finance operating system: AI-native tools now pull data across systems and support faster, more continuous finance work. @seema_amble Its portfolio examples span ERP (Rillet), live financial planning (Concourse), procurement (Lio), accounts receivable (Stuut), indirect tax (Sphere), and SOX testing/internal audit (Petual); a16z says these workflows previously required dedicated teams or outside providers. @seema_amble
• Finance teams are also building AI tools themselves: Anthropic’s finance team maintains 70+ finance-specific AI skills, and one produces a monthly review that is 90–95% complete in about 30 minutes. @seema_amble a16z says the leanest finance teams it sees are targeting 2% of company headcount, down from the old ~5% heuristic; it also describes a shift toward continuous forecasting and emphasizes auditable trails and finance approval controls for agent-handled work. @seema_amble

• AWS CEO Matt Garman says customers increasingly want cloud infrastructure designed for agents: AWS is building agent-specific primitives, including databases that start in three seconds for short-lived tasks with lower durability needs, compute sandboxes, gateways, and permissions distinct from human or service-role permissions—not simply repurposing existing building blocks. @a16z
• In its interview post, a16z reports AWS has $220B in CapEx for 2026, two million NVIDIA GPUs on order, and AI chips sold out through next year. @a16z

Scott Kupor discussed American leadership and responsibility in superintelligence after appearing on CNN about the President’s Super Intelligence Task Force, his role, and potential implications of superintelligence for jobs, tech companies, and America’s future.@skupor +1