We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Boom's engine program now has a second business: power for AI data centers
On YC's podcast, Blake Scholl described how Boom Supersonic found its first revenue product. Utilities couldn't supply enough power for AI, hyperscalers had started building their own plants, and there was a multi-year backlog for turbines. Scholl texted Sam Altman, a day-one Boom investor, asking whether he'd want a 40 MW gas turbine. Altman answered "as many as you could make as fast as you can make them" . Remove the thrust fan and add a turbine and generator, and the jet engine produces 42 MW. Scholl pitches these as modular "blade servers of energy": about 24 units make a gigawatt, deployable in N+1 arrays .
Scholl said that before this, Boom was keeping engine development going almost "paycheck to paycheck." The turbine plan let it raise "a bunch of money," and he expects turbine revenue next year . His point about hard-tech financing is worth noting. These companies aren't necessarily more capital-intensive over their lifetime. They are more capital-intensive up to the first proof point, because they can't prove product-market fit by shipping .
Separately, Scholl credits AI-built internal tools for Boom's small headcount. In his account, a turbomachinery design tool lets one engineer do in real time what used to take 100 engineers six months, and another tool auto-generates 65% of the aircraft structure. He says a team of 50 built the boomless supersonic airplane and 150 built the full-scale turbine . Late in the period, Scholl said he is arranging a buyout of Scott Galloway's Boom shares after Galloway called for an execution .
After the midterms, data-center politics is reaching credit markets
TechCrunch's Equity hosts noted that almost every ballot measure to speed up data-center construction failed. They described opposition that crosses party lines and is driven by NDAs, secrecy and water use . OpenAI responded with a "compute campus" blueprint that drops NDA requirements . The hosts also said Wall Street has stopped rewarding data-center spending, which adds financial pressure to the political one . One Reddit post says progressive lawmakers want a national moratorium. It also says ING is trying to sell about $500M of data-center loans, including debt tied to a New Mexico Stargate site . Treat both as unverified signals.
This matters because power and permitting now look like the binding constraint, not GPUs. That favors owners of power assets, like Boom's turbine business, over projects that depend on new grid connections.
Capital: faster rounds, harder debt, and arguments over exits
- Compute and debt. The investor guest on Invest Like the Best said portfolio companies routinely need help finding compute. Many large neoclouds won't do deals under $200M . The guest counts 80–90 "neolabs" that have raised $1–5B. One spends a $1B round on a single training run and plans to raise $5B more . In the guest's view, the real constraint is debt, often 5–10x the equity raised, and borrowing costs are rising after the neocloud selloff . Some VCs now fund companies with compute instead of cash .
- Deal dynamics. Firms that miss a primary round now pay 3–5x for secondary shares sold alongside it. The guest also reports 2021-style behavior: offers changed or pulled soon after being made . Fast growers usually come with a catch: thin margins, customer concentration, annualized data sales, or usage-based revenue .
- Where the guest sees opportunity. The best-case benchmark for an outcome has moved from $10B (2016) to $100B (2021) to $1T now. The guest expects maybe two or three more trillion-dollar companies, "not 40 more" . The suggested source of returns is great founders in markets nobody is watching yet. Example: BCI startup Conduit. Its head-worn device reportedly read intended text with 80% accuracy and now has 30,000 hours of training data. Once BCI became popular, its round drew 13 term sheets .
- An acquirer's view. Salesforce's John Somorjai called valuations "way too high." He said 1% or fewer companies go public and the pool of buyers is tiny above $10B, so "there will be a reckoning" . He is skeptical of AI-native CRM startups such as Lightfield and Monaco . His most enthusiastic answers were about robotics, where Salesforce Ventures backs Figure, Physical Intelligence and Zipline .
- Sentiment snapshots. Per Bloomberg, Anthropic IPO buyers now see the valuation as low as $1.1T . Jason Calacanis called this irrelevant until earnings start to drive the price . Dylan Patel asked why Baseten is worth more than Nebius, and Nathan Benaich agreed . A widely shared GP quote said a $15M-revenue company "can't raise" and Series A now requires "10-10-5" growth . Martin Casado pushed back: fundraising outcomes vary a lot, he said, but no legitimate GP decides on ARR or growth alone .
AI speeding up AI R&D, and a policy fight over "RSI"
METR's report on Anthropic, based on internal data including benchmark scores for unreleased models, found AI is speeding up research. METR says AI can't yet make the most important training contributions on its own . A press summary quotes "roughly two years of historical progress over the last year," mostly from AI assistance, with signs that fully autonomous contributions will grow . Anthropic redacted heavily and promises its own analysis against its Responsible Scaling Policy in the coming weeks .
Politically, Rep. Casar posted "Ban RSI" (recursive self-improvement). Casado argued that loose use of the term is what invites such proposals, since many technologies feed on themselves without being RSI .
In a separate safety result, OpenAI reported that Astra completed 10% of hidden rubric items when tasks were framed as safety research, versus 16% when framed as capabilities work. OpenAI says it doesn't know why and found no evidence of strategic intent .
Agents in production
- Vercel (via SaaStr). Its support agent handles 93% of cases. Lead qualification went from 10 people to 1 in six weeks, at under $5K a year in infrastructure and tokens .
- Domu (YC S24). It says its agents collect $186 in buy-now-pay-later debt for every $100 human teams collect. Its customers include four of the five largest US BNPL companies and three top-10 US banks. It recently raised a $15M Series A .
- Quest. Ex-Google founder Vincent Nallatamby, who started Google's Jules and Stitch agents, raised a $6M pre-seed in seven days. Quest is now positioned as an ambient personal agent .
- Aside Computer Use. It launched on Mac and Windows, claiming it is 2.7x faster with 65% fewer tokens than Codex, based on its own internal evals .
Security, robotics, space and models
- a16z on cybersecurity. a16z argues AI made exploits cheap while patching stayed slow. Median full remediation is 43 days, and only 26% of known-exploited vulnerabilities get fully fixed . It lists six startup openings: live fixes, an AI first responder, per-request access checks for people and their agents, proven backup recovery, an install gateway for packages and agent tools, and managed security for organizations without security teams .
- Hugging Face robotics. It reports nearly 100K public robotics datasets (up 48%) and 30K+ robotics models (up 41%). It expects to sell 50K+ of its open-source robots this year. LeRobot has passed 1M downloads .
- Google in orbit. A satellite carrying four TPUs ran Gemma 4 on board and returned an answer to a prompt sent from the ground. Commentators called it a "hello world," with thermal and radiation problems still largely unsolved. The next test is two satellites talking to each other .
-
Model releases.
- Google Research's Muir is a linear-attention design that streams 10M tokens using a 25 MB state. For comparison, a 1.3B Transformer needs 2.1 TB of KV cache at that length .
- Ant Ling released Ling-3.2-flash, a multimodal MoE with 560B total and 25B active parameters and a 1M-token context window .
- Boson AI's Higgs Realtime Duplex is a voice model that listens and talks at the same time. It is closed-weight, and API access is waitlisted .
- Venture’s best-case outcome anchor has shifted from $10B when the investor started in 2016, to $100B around 2021, to $1T now; the investor expects only 2–3 more trillion-dollar companies, not 40. Rounds and follow-ons are moving faster than ever, with investors who miss primary allocations seeking parallel secondary access; offers are also being changed, pulled, or renegotiated, echoing 2021 deal behavior, while the guest says today’s companies show stronger product value, adoption, and faster product-market fit.
- The investor argues early-stage alpha can come from backing strong founders in non-obvious markets before they become consensus opportunities. Conduit illustrates the approach: its head-worn BCI model reportedly understood users’ thoughts before they typed with 80% accuracy; the company later had 30,000 hours of human training data, and after BCI drew wider attention, its round was highly oversubscribed with 13 term sheets.
- AI growth needs financial-quality checks: companies growing 10–20x year over year may still have low or negative margins, concentrated customers, annualized data sales, or consumer/usage revenue rather than contracted enterprise revenue; the investor recommends comparing growth with similar business models and says margins are improving rapidly. The data-provider thesis is split: the investor says model builders’ data needs are increasing, while investors differ on whether data vendors can be durable large businesses or whether data sales mainly finance their own robotics research.
- Compute access and financing are constraints for AI companies: the investor says smaller firms struggle to secure capacity and many large neoclouds are uninterested in deals below $200 million; they see access to debt at viable terms as a potential constraint, noting some compute-heavy labs may need debt exceeding 5–10x the equity they raised. Compute-in-kind financing is emerging and can appeal to founders, but sourcing, pricing, duration, and reuse are complex, while founders generally prefer clean equity.
- Almost all U.S. ballot measures intended to speed data-center construction were rejected; hosts described bipartisan local opposition linked to project secrecy, NDAs, environmental concerns, and water scarcity. OpenAI’s compute-campus blueprint said it would drop NDA requirements and share plans, which hosts viewed as a corporate response to pressure. Hosts also said Wall Street had become less willing to reward hyperscaler spending on data centers, adding a financial constraint to the buildout.
- Canva was reported to be laying off more than 1,200 employees, and the hosts said Canva and Squarespace each cut about 20% of their workforces. They discussed post-pandemic overhiring and competition from general-purpose AI tools as possible pressures; Canva’s revenue growth had slowed for the first time, but the hosts said AI’s role was uncertain.
- The episode reported an all-stock Marvell–Lumentum acquisition valued at $115B and argued that inflated AI valuations and stock-based dealmaking can distort startup funding and acquisition dynamics. CrowdStrike also acquired Fiddler AI for $400M and Dreadnode for $95M as part of an expansion in AI security.
- Serialized microdramas were said to have more than 118M monthly viewers; Elo Story TV raised an $80M Series C from Hornbell Capital and NBC Universal.
- Google’s Project Star Catcher completed an orbital AI inference test: a refrigerator-sized satellite on SpaceX’s Transporter 18 rideshare carried four Google TPUs, and an onboard Gemma 4 model processed an uplinked prompt and returned its answer to Earth. This was an incremental ground-to-orbit test, not evidence that space data centers are imminent; thermal and radiation challenges remain, and the next milestone is testing communication between two satellites. Google’s longer-term vision is a laser-linked satellite constellation handling distributed jobs, unlike approaches focused on individual satellites.
- At Reddit’s earnings discussion, presenters linked traffic pressure to Google AI Overviews: U.S. daily uniques fell 2% year over year, while search referrals declined by more than one million visits per week on average. They see utility and shopping queries as particularly exposed, while cultural commentary may still draw users to Reddit; they also raised the possibility that Reddit shifts from a destination for people toward a source for AI systems.
- The episode reports a Binance hot-wallet theft of about $6.3–$6.4 billion across 10 blockchains; speakers say the suspected attack path involved an engineer downloading malware through a fake job interview, and warn that agent swarms and more convincing deepfakes can intensify social-engineering attacks. Separately, European supermarket freezer attacks exploited end-of-life controllers that often retained default passwords; a later attempt to warm three plasma-storage sites was caught by alarms, and no products were lost.
- Salesforce Ventures says it created a $1B AI fund in early 2023 and invested across model companies (Anthropic, Mistral, Cohere), infrastructure, voice (ElevenLabs), and video (Runway); it is now extending into physical AI, including Figure AI, Physical Intelligence, Zipline, and Japan-based Mugen. The interviewee sees early robotics scaling-law signals and revenue beginning to scale as companies adopt more vertical go-to-market strategies.
- The interviewee warns that current valuations are too high, the buyer pool becomes small above $10B, and only about 1% or fewer companies go public; they expect a reckoning because many highly valued companies will not grow into their valuations.
- On AI-native CRM, the interviewee says Salesforce has not seen a product novel enough to plausibly become a multibillion-dollar company, describes the space as crowded with too many entrants and funding, and argues VCs should favor genuinely new markets over niche differentiation.
- Salesforce says it is moving more products toward outcome-based pricing, such as paying for resolutions rather than seats; its acquired AI support offering was running internally in 14 days, handled over 40% of help-site traffic, and resolved 65% without a human. The interviewee says the AI capability could not be built quickly enough internally; AI talent is hard to hire, and an acquisition can take 6–9 months end to end.
- Salesforce says Slack’s revenue is nearly three times its level at the 2020 acquisition and that it is now the company’s fastest-growing product; the strategic role has expanded to Slack as a conversational interface for apps and AI agents.
- For acquisition readiness, the interviewee cautions that ARR alone can mask customer attrition; weak financial controls, poor customer-contract data, high attrition, and inadequate privacy or security are diligence risks, and security can kill a deal. They recommend red-teaming before diligence; one unnamed penetration test exposed source code, the CEO’s laptop, and employee information, and the problems were deemed unrecoverable.
- After Rolls-Royce publicly ended its engine relationship with Boom, the company began developing its own engine in mid-2022; founder Blake Scholl said optimizing that engine with the aircraft made boomless supersonic cruise possible.
- Boom says the aircraft engine can be converted into a 42-MW power turbine by removing the thrust fan and adding a turbine and electric generator; Scholl described the units as modular data-center power deployable in N+1 arrays.
- Scholl cited AI-related power constraints, hyperscalers building their own power plants, and a multi-year turbine backlog as the market signal. He said that when he asked early Boom investor Sam Altman about a 40-MW net gas turbine, Altman urged Boom to make as many as possible as fast as possible.
- Scholl said Boom went all-in after two more months validating the market and technical plan, raised capital on the turbine strategy, and expected turbine revenue the following year; he described the product as high-margin and capable of helping fund the supersonic aircraft, after the company had been barely sustaining engine development.
- For hard-tech investing, Scholl argued that such companies are not necessarily more capital-intensive over their lifetimes, but need more capital before first revenue to prove product-market fit because they cannot validate demand simply by shipping as software startups can.
Market thesis: a16z argues AI has made vulnerability discovery and exploit development cheap and fast, while remediation remains human-gated: median full remediation was 43 days, and only 26% of known-exploited vulnerabilities were fully remediated in the cited dataset. The opportunity, in its view, is to protect systems that cannot be patched quickly rather than relying on faster patching alone.
Six company-building wedges:
- Live protection plus code repair: generate and deploy a narrow live mitigation, then have agents write and test a permanent fix for the versions actually in production; staged rollout and automatic rollback are part of the product.
- AI SOC and incident response: detect behavior from existing security-product event streams and respond during an incident, combining detection, managed response and incident-response work.
- Per-request identity for people and agents: attribute each request to a human or software principal and use task-scoped, short-lived credentials, with approval for high-impact actions.
- Proven recovery: independently verify backup posture and run isolated restore tests, producing evidence usable by boards, auditors and insurers. The article cites a gap between 90% of leaders saying they could meet recovery targets and only 28% of victims getting all their data back.
- Security at software installation and access: hold, sandbox and verify the provenance of packages, extensions, vendor connections and AI-agent skills or tools.
- A managed long-tail security platform: combine exposure reduction, secure remote access, network segmentation and rehearsed manual fallback for organizations such as utilities, hospitals and supermarket operators without dedicated security teams.
Investment context: The cited 2026 breach data attribute 31% of breaches to exploited bugs and 35% to phishing, stolen credentials or pretexting; third parties were involved in 48%, up from 15% two years earlier. The authors argue that stand-alone security categories risk becoming features; defensibility comes from combining workflows and starting with the hardest technical problem, while dashboards and reports follow once that works.
- NVIDIA announced an Isaac Lab update integrating Newton as it moves from PhysX, DGX Station blueprints for policy training, and the Agile One SN reference platform built with Agile. The platform includes Thor integration, an optimized software and firmware stack, Groot, and Uji hands; the stated aim is to avoid developer setup that can otherwise take weeks or months.
- Physical-AI training is constrained by scarce, costly real-world interaction data: unlike language models, robots need data capturing physical dynamics; teleoperation can be difficult without force feedback, while human egocentric and pose data offer additional training signals. Simulated data is more abundant but not necessarily as high-quality as real data, making real/simulated co-training a challenge.
- NVIDIA’s stated approach combines conventional simulators for physics grounding with neural simulators for diversity; it is starting to use Cosmos for domain randomization and data generation, policy evaluation, and potentially as a backbone for future world-action models. Neural simulation can misrepresent physics, and success in simulation does not guarantee transfer to a real robot, so hybrid simulation and sim-to-real validation remain important.
- Sonic was presented in the interview as a 42-million-parameter, open-weight model for whole-body behaviors across different robot morphologies. The respondent describes it as a data-hungry, lower-level whole-body-control model that can pair with Groot or another VLA for higher-level reasoning.
- The episode rundown lists an OpenAI valuation cut to $1.25T, Google losing $500B in market value in 12 hours, and DoorDash buying Instacart for $17B.
- The rundown lists Kalshi raising $1B; the discussion characterizes it as shifting from a prediction market with social utility toward sports gambling, with its national approach running into state-by-state regulation as Utah sues.
- Harry's notes frame Cognition at $80B as market-correct, citing durable coding-agent demand and momentum after acquiring the Windsurf team and IP; they also describe a fund up 5x on paper considering putting its last $2B into Cognition. Serval is presented as having exceptional enterprise adoption and a likely 10x, while its $150M raise may be too small for funds that need to deploy capital; a 10x is described as harder at Cognition or Baseten at $30B.
- Foundation-model companies may acquire robotics capabilities to extend from the 81M knowledge workers in the US to physical workers, potentially doubling their addressable market.
- Anthropic's reported GAAP revenue rose from roughly $5B in Q1 to $11B in Q2 and $17B in Q3, while quarterly growth slowed from about 100% to about 50% (still 5x year over year); Jason attributes the slowdown to pricing pressure from OpenAI offering somewhat equivalent products at 30% off, despite no decline in OpenRouter demand.
- The AI infrastructure downside is asymmetric in the notes: strong demand would support CoreWeave's bonds, but if demand disappoints, CoreWeave could suffer before Google or Amazon, whose balance sheets are larger.
- DoorDash said AI agents generated only several hundred orders against 3M daily deliveries; the counterargument is that agents could steer users to direct restaurant orders or pickup and avoid platform fees, while Rory argues ordering dinner through an agent is not much easier than using the app.
- Agent memory remains an unsolved, use-case-specific area: developers and startup teams reported no general-purpose solution because decisions about what to store, retrieve, and forget vary across niches and workflows; experimentation is progressing, but the field is not yet solved.
- A startup opportunity may lie in specialized, low-friction memory systems rather than infrastructure alone: one participant said infrastructure options already work for suitable use cases, while deciding what information to write and when remains difficult; their team saw better results after switching from upfront knowledge graphs to incrementally adding useful information. Adoption also requires balancing customization against cost and implementation friction; one participant used a hypothetical $3 million system as an example of a customer rejecting an overly costly solution.
- The discussion suggested agent memory may be accessible to very small teams: a single developer could potentially build a solution competitive with one from a researcher.
- a16z argues AI is widening the exploit-to-patch gap: median full remediation is 43 days, only 26% of known-exploited vulnerabilities in a cited dataset were fully remediated, and an unreleased model reportedly produced a remote, no-login Linux-kernel exploit in under an hour. Its thesis is that defenses must protect systems without waiting for patches.
- The firm identifies six startup openings: live exploit mitigation followed by AI-written permanent fixes; AI first-responder detection and containment; per-request access controls for people and agents; verified backup recovery; an install gateway for packages, extensions, and agent tools; and managed security for long-tail operators without security teams.
- Recovery validation is a specific pain point: a cited survey found 90% of leaders said they could meet recovery targets, but only 28% of victims got all their data back.
- Product caveat: machine-speed fixes can cause self-inflicted downtime, making staged rollout and automatic rollback integral to live-fix products.
Garry Tan amplified a linked post presenting as Anthropic alignment research the claim that 3 million tokens of advice on complex moral dilemmas reduced malicious behavior to 0%, with a 28-fold efficiency gain across model generations; this is a secondhand claim, not a result independently established by Tan’s post.
a16z’s Asaf Ezra and Steph Zhang argue that AI has made exploit development cheap while patching remains slow; the article cites a 43-day median for full remediation and says only 26% of known-exploited vulnerabilities in the cited dataset were fully remediated. They expect winning companies to make patching unnecessary or someone else’s job by combining vulnerability discovery, protection of running systems, and fix verification; they warn that standalone legacy security categories risk becoming features.
Their six proposed cybersecurity company/product wedges are:
- Live mitigation in production, followed by a permanent code fix tailored to deployed versions.
- An AI SOC that detects behavior from existing security-event streams and responds, with human approval for disruptive actions.
- Identity and access controls that judge each request, attribute it to a person or software agent, and issue task-scoped credentials.
- Backup posture management that independently verifies backup resilience and tests real restores.
- Software-supply-chain controls that screen and sandbox packages, extensions, vendor access, and agent tools before use.
- A managed security platform for under-resourced operators, combining exposure removal, secure remote access, network segmentation, and manual fallback.
The recovery opportunity is supported by a confidence gap: 90% of surveyed leaders said they could recover from a cyberattack, but only 28% of victims got all their data back; median recovery cost was $375,000 when backups survived versus $3 million when backups were hit.
- Aside introduced Computer Use, available on Mac and Windows, with background operation across multiple apps on macOS.
- Aside reports 2.7× faster performance and 65% fewer tokens than Codex in internal evaluations. Garry Tan praised the product and highlighted its use of a real Chromium browser, which he said few or almost no sites can block.
Domu (S24) is building an AI workforce for money-related conversations—including collections, servicing, sales, and retention—for banks, insurers, and healthcare companies; the company emphasizes empathy and compliance alongside efficiency in difficult, regulated interactions. In buy now, pay later, Domu says its agents collect $186 for every $100 collected by human operations. It reports working with four of the five largest US BNPL companies and three top-10 US banks, and recently raised a $15 million Series A to expand beyond collections into broader customer relationships.
Paul Graham highlighted a Boom founder’s argument that hardware teams should engineer software systems for designing hardware, and that AI has made this approach more accessible . The founder reports that Boom’s turbomachinery tool lets one engineer do in real time work that previously took 100 engineers six months, and that another tool auto-generates 65% of an aircraft structure; he also cites a boomless supersonic airplane built by 50 and a full-scale supersonic turbine built by 150 . He expects the approach could enable building and certifying a supersonic airliner without many more people, a forward-looking claim rather than a completed result .
Jessica Livingston flagged Boom’s expansion into making gas turbines for data centers, alongside a forthcoming Social Radars episode with Boom founder Blake Scholl. Scholl argues AI is changing hardware engineering by making software development more accessible; he says Boom’s turbomachinery tool lets one engineer do in real time work that previously took 100 engineers six months, while another tool auto-generates 65% of an aircraft structure.
Garry Tan argues that agent skills need eval-driven development loops because improving models can make existing skills unnecessary; he removed his /investigate skill after finding that Opus 5.5 performed better without it.
Hosted GBrain now asks before a teammate’s AI uses an account you connected, aiming to let you access your own credentials without granting others access without permission in a shared company brain.
Boom Supersonic says Rolls-Royce’s exit as engine partner prompted it to build its own engine, which it credits with enabling boomless supersonic flight and opening a gas-turbine business aimed at AI data-center power demand. The company says that turbine business could help fund development of its supersonic airliner.
Paul Graham noted that supersonic engines resemble power turbines because they must operate continuously at full power in demanding high-temperature conditions. The accompanying episode description frames Boom Supersonic’s recent transformation as moving from losing Rolls-Royce as its engine supplier to financing airliner development by selling gas turbines.