• The interview describes Modal as having reached $1B in revenue, after a $750M raise at a $15.75B valuation; its founder says it grew 50× in the prior 12 months and was still growing 30–40% month over month. 20VC with Harry Stebbings Its 2023 sandbox product took off after coding models improved in 2025 and reached a $300M run rate, about one-third of revenue; the founder hopes its recently launched training product will reach a $100M run rate within 6–12 months. 20VC with Harry Stebbings
• The founder expects inference to become a very large but increasingly specialized market. He argues that selling tokens alone risks commoditization, while a broader platform spanning inference types, sandboxes, post-training, and reinforcement learning can provide differentiation and staying power; he sees potential for 40–60% margins, but not legacy software margins above 90%. 20VC with Harry Stebbings He expects Nvidia and specialized chipmakers to grow in parallel; in his view, CUDA itself is less of a moat than Nvidia’s software ecosystem and Nvidia-first support, while alternatives remain behind. 20VC with Harry Stebbings
• Modal says it is profitable on a P&L basis but cash-flow negative because of GPU capacity spending. The founder cites 3–4-month GPU lead times, 30–40% monthly customer-demand growth, and forecasting 2–3× capacity; he says supply is not simply unavailable, but prices have risen and paying more can secure GPUs. 20VC with Harry Stebbings He favors raising less than offered in early rounds to limit dilution, but says GPU buildout has changed Modal’s capital needs; he cautions against businesses dependent on continual equity, recommends raising each round as if it were the last with a path to commercialization, and expects a possible 20–30% market pullback despite remaining long-term bullish. 20VC with Harry Stebbings
• He sees coding-environment businesses selling to frontier labs as lucrative now, but questions their long-term durability; moving into implementation, deployment, evaluation, and monitoring can support margins and a longer-term business. 20VC with Harry Stebbings He sees post-training and reinforcement learning as especially suited to tasks with finite feedback, such as coding, cybersecurity, and math, while longer feedback loops remain uncertain; he expects software-engineering productivity gains to expand demand for software. 20VC with Harry Stebbings

• Hypersonica, a Munich- and London-based startup building Europe’s hypersonic deterrence capability, is reported to have gone from a blank sheet to a Mach 6 test flight in Norway’s Arctic in nine months with less than 25 million raised. Lightspeed Venture Partners The company attributes its pace to combining specialist hypersonics expertise with rapid hardware iteration and testing; CEO Philip Kaat says it is moving from prototyping toward industrialization and weapon systems. Lightspeed Venture Partners
• Lightspeed and NGC co-led Hypersonica’s €100 million Series B; the company says the funding will lay the foundations for industrialization across Europe over the next 18 months. Lightspeed Venture Partners
• Hypersonica’s stated use case is that hypersonic systems complement shorter- and mid-range capabilities, traveling significantly more than 1,000 km in under half an hour and leaving adversaries less time to respond. Lightspeed Venture Partners The interview frames Europe as lacking a sovereign ground-launched long-range strike system; Kaat cites Germany’s move to procure Tomahawk, a 37 billion allocation for deep-strike systems and procurement across NATO, and UK defense-plan funding as demand signals. Lightspeed Venture Partners

• Tutor Intelligence, founded in 2021, has about 160 employees, primarily in Boston, and combines frontier robot-AI training, manufacturing, and direct deployment. a16z It says its robots are deployed across the U.S. at businesses ranging from small operators to Fortune 50 and Fortune 100 companies. a16z An interviewer also relayed a Claude estimate that Tutor robots had helped manufacture and ship more than 250 million products to American consumers; this was presented as an estimate, not a company-verified KPI. a16z Tutor frames robotics as a full-stack systems business and says it builds robots in Watertown, Massachusetts, with plans to manufacture more components in the U.S. a16z
• Its go-to-market targets a largely greenfield market: Tutor says almost all its robot buyers are first-time purchasers, and it uses a $16-per-hour proposition to open sales conversations. a16z It reports 40–60 days from first call to signed contract and about 40 more days to install; robots can enter production within a day, sometimes within hours, without placing economic risk on the customer. a16z The founder says ongoing costs currently exceed upfront costs, a notable shift from the traditional capex-heavy robotics model. a16z
• Cassie and Sunny target manufacturing and logistics; the founder estimates they can handle 80–90% of work within that scope. a16z Tutor has 100 Sunny robots in a Boston data factory for internal data collection and research, alongside a growing real-world fleet. a16z The company’s thesis is that tutors teaching robots new tasks and paid customer deployments generate useful real-world training data. a16z Current products are not suited to outdoor work, and home robots are not an immediate fit. a16z

• At ADIPEC, the interviewer said Abu Dhabi had announced OpenAI agreements spanning energy, education, investment and government, and described ADNOC as the first energy company to adopt “OpenAI frontier.” Mint Altman said OpenAI, ADNOC and G42 would use frontier AI on ADNOC’s hardest, highest-value problems, adding that the models had “caught up.” Mint
• Altman said AI had begun solving math problems people could not solve; he also described a forthcoming next-GPT model using 10,000 decentralized agents for about four days to prove the first Millennium Prize problem. Mint
• Altman called energy OpenAI’s current challenge and said that, despite improving query efficiency, growing AI use and the physical energy supply could make power a dominant factor in AI’s cost and availability. Mint He argued AI could help energy production through new materials, turbine designs and improved operations, and said OpenAI wanted AI’s net effect to increase energy availability. Mint

• Altman described OpenAI’s partnership with ADNO as a way to put ChatGPT, GPT-6 and future models in people’s hands, and argued that AI could help energy companies increase production and efficiency through advances such as new materials, turbine designs or better operations, as demand for energy grows. Talk Shows Central by DRM
• Altman said 10,000 agents of a forthcoming GPT model worked continuously for about four days and claimed they proved the first Millennium Prize problem; he described the agents as decentralized, sharing progress and choosing what to work on without bosses. Talk Shows Central by DRM

• ADNOC was presented at the event as the first energy company to adopt OpenAI Frontier; Altman said OpenAI, ADNOC and G42 aim to apply frontier AI to ADNOC’s hardest, highest-value problems. DRM News
• Altman framed AI as a tool for energy-sector R&D, citing potential work on new materials, turbine designs and operations, and said current models can address such problems. DRM News He also said energy use per query is falling while usage is growing, and identified physical energy supply as a likely dominant constraint on AI’s cost and scale. DRM News
• As a deployment risk, Altman said the industry will increasingly need to demonstrate it can avoid serious safety and alignment incidents; he also insisted that AI must serve people and not take over human decision-making. DRM News

• Blacksmith reports 20× growth over the prior year and a roughly $66 million Series C at a $1 billion valuation. Y Combinator
• The company says agentic coding is increasing code production and shifting the bottleneck to validation and CI. Its approach is specialized infrastructure for CI workloads, including gaming CPUs and keeping cached artifacts close to compute. Y Combinator
• Blacksmith says it secured large five- to six-figure contracts without a sales team; LangChain’s reported wait to acquire CI compute fell from 15–20 minutes on GCP to 10–15 seconds on Blacksmith. Y Combinator
• Blacksmith reports CI usage spikes whenever a new LLM launches, creating capacity-planning pressure because new machines take time to order and install. A Phoenix data-center cooling failure affected customer job startup times and highlighted the need for geographic distribution. Y Combinator

• Harry describes inference demand as effectively “infinite,” predicts “3–4 specialised inference providers worth $300–$500bn in this market,” and identifies Modal as one; he says Modal has $1bn in ARR, including $300m in sandbox revenue alone. @HarryStebbings The interview notes say the inference market has grown 10x–100x beyond Erik Bernhardsson’s expectations, with room for low- and high-latency, fine-tuning, and post-training segments. @HarryStebbings
• Erik argues standalone inference is a commodity with no switching costs and a race to the bottom, so providers need a broader suite spanning post-training, sandboxes, agentic tooling, reinforcement learning, and low-latency and batch inference; Modal has spent five years building this infrastructure. @HarryStebbings He expects Modal’s long-term margins to be 40%–60%, not the 90%-plus software margins of the past, as multiple stack layers take 20%–30% gross margins. @HarryStebbings
• Harry’s post calls Modal a “$15.75 billion company” and says it has never made a pitch deck for any round; the D round came together quickly after Erik had known Accel’s Rich and Gonzalo for at least a year. @HarryStebbings
• Bernhardsson says he enjoys the work, hopes to keep building Modal for 20–50 years, and does not want to sell. @HarryStebbings Erik also expects software engineers to remain fine for the next 10–20 years, arguing that greater programmer efficiency increases demand for software. @HarryStebbings

Garry Tan called for repealing CEQA, describing it as a primary NIMBY tool and arguing that it blocks hospitals, high-speed rail, and housing through legal challenges. @garrytan He amplified a post alleging that Sutter gave up on a planned Emeryville hospital after failing to obtain CEQA relief, and warning that the East Bay could face trouble if Alta Bates closes in 2030. @marcjoffe

• AI-app defensibility: Falling inference costs can improve margins, but stronger models and easier development also make applications easier to replace. A16z argues that apps must outperform credible in-house alternatives across the workflow, provide context that is costly to rebuild, and prove value with pricing tied to measurable completed work rather than token usage. @TugceErten2022
• In an August 2026 survey of 100 enterprise LLM decision-makers, 94% expected total LLM spending to increase and 37% expected it to at least double in the following year. @TugceErten2022
• Repsol’s blind evaluation compared Harvey with Microsoft Copilot and its internal Lexia tool; Harvey delivered the highest-quality output and then outperformed the alternatives on accuracy and time savings in an eight-week pilot involving 50 lawyers across 12 jurisdictions. Repsol proceeded with a full rollout; Harvey reports savings of four to six hours per lawyer per week and 96% adoption across the legal department. @TugceErten2022 Notion chose Decagon after considering an internal build; Decagon’s published case study reports up to a 34% improvement in ticket-resolution time. @TugceErten2022

Garry Tan highlighted direct weight manipulation as a new use for smaller models. @garrytan The linked post reports that GPT-6 Astra edited Qwen3.5-0.8B by changing 0.016% of its weights, claiming this fixed reasoning loops and improved math by up to 6× and instruction following by 3× without gradients or training. @rikiyatakehi

• Harry Stebbings says inference demand is far larger than expected, argues neo-cloud providers will struggle to enter effectively, and predicts “3–4 specialised inference providers worth $300–$500bn.” He identifies Modal as one and reports $1bn in ARR and $300m in sandbox revenue alone. @HarryStebbings
• Modal cofounder Erik Bernhardsson argues inference alone is a commodity with no switching costs and a race to the bottom; Modal’s proposed differentiation is a broader suite—post-training, sandboxes, agentic tooling, reinforcement learning, and low-latency and batch inference—which he says the company has spent five years building and is hard to copy. @HarryStebbings
• Bernhardsson says almost no company today has the 90%+ software margins common a decade ago, citing 20–30% gross margins at each layer from applications through models, infrastructure, cloud, and NVIDIA; he expects Modal to reach 40–60% margins long term. @HarryStebbings
• On AI’s effect on labor, Bernhardsson expects software engineers to be “fine” for the next 10–20 years, arguing that programmer productivity expands software demand and that engineering work continues moving toward understanding human requirements. @HarryStebbings
• The post calls Modal a $15.75bn company that has never made a pitch deck for any round; before its D round, Bernhardsson had known Accel’s Rich and Gonzalo for at least a year, and the deal came together quickly. @HarryStebbings

Bill's current portfolio includes two startups: Skyfi, a geospatial-intelligence platform described as “Uber for satellites,” where users can order satellite imagery by phone, and SIMAX, an intelligence firm; he also continues to run and trade his hedge fund. My First Million

Paul Graham said applications for YC’s Winter 2027 batch were due that day and linked to YC’s announcement @paulg; YC’s post linked to its application page @ycombinator.

Garry Tan says Capy Captains increased his issue-fixing throughput about 5× overnight, crediting coordination across captains and subthreads; this is his reported experience, not an independent benchmark. @garrytan Capy announced Captains and Issues as generally available in version 0.7.0: Captains are persistent agents that orchestrate hundreds of threads across devices, Issues adds filters, full-text search, labels, and due dates, and Dock CLI supports headless runs on Linux, macOS, and Windows. @capydotai

• Tutor Intelligence is deploying AI robots in American factories and warehouses for hard-to-hire work; it builds them in Massachusetts, charges by the hour with no upfront cost, and has a human “tutor” take over when the AI gets stuck, generating training data. CEO Josh Gruenstein says the company can go from a first sales call to a working robot in about three months. @a16z
• Gruenstein says factories cannot hire for this work; the post reports year-over-year hours per worker rose across all 10 manufacturing categories and grew faster than headcount in 7 of 10. He cites three-month worker turnover and an average worker age of 55 nearing retirement, and floats a $16-per-hour robot rate. @a16z

Paul Graham observes that in a successful startup, one month’s revenue growth can make the previous month’s towering revenue look small and the early part of the chart look flat—an illustration of how exponential growth appears over time. @paulg

Blacksmith (YC W24) is building faster cloud CI to address testing and validation becoming a bottleneck as AI coding tools accelerate software production; YC says the company grew 20× in the past year and recently raised a $66 million Series C at a $1 billion valuation. @ycombinator Its three founders began as DevOps outsiders, won their first 200 customers, and built the business with a 30-person team; Blacksmith says it cuts CI waits from minutes to seconds. @ycombinator

• Harry Stebbings says inference demand is far larger than expected and predicts 3–4 specialized inference providers worth $300–$500B; Erik Bernhardsson says the market has grown 10×–100× beyond his expectations and expects segments spanning low- and high-latency inference, fine-tuning, post-training, and hardware or technologies needing far fewer resources. @HarryStebbings
• Bernhardsson argues standalone inference is a commodity with no switching costs and a race to the bottom; he sees differentiation in a broader suite including post-training, sandboxes, agentic tooling, reinforcement learning, and low-latency and batch inference. Modal says it has spent five years building this infrastructure and calls it hard to copy. @HarryStebbings He also says 90%+ software margins are now rare, with each stack layer taking 20–30% gross margins, and sees Modal reaching 40–60% margins long term. @HarryStebbings
• The post reports Modal at $1B ARR, with $300M in sandbox revenue alone, and describes it as a $15.75B company. It says Modal made no pitch decks for its rounds; Erik had spoken with Accel’s Rich and Gonzalo for at least a year before the Series D. @HarryStebbings
• Bernhardsson expects more, not fewer, software engineers: as programmers become more efficient, demand for software rises, and he expects engineers to be fine for the next 10–20 years. @HarryStebbings

Rahul, founder of Aerodome, says he is joining a16z as an Entrepreneur-in-Residence two years after selling Aerodome to Flock Safety; he notes that Flock drones provide air support for 911 calls around the country and says it is a good time to build for America. @rahul David Ulevitch says he is excited to work with Rahul on a16z’s American Dynamism team. @davidu