We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Ben Horowitz’s largest-ever check is backing Travis Kalanick’s industrial-AI vehicle. Kalanick defines the company as “systems of software, sensors, robotics and AI” for automating industrial sectors, mapping manufacturing, real estate, and transport/logistics onto physical-world equivalents of the CPU, storage, and network. Its food-infrastructure example targets robotic production and delivery at roughly $6–$8 per meal.
The disclosed financing signal is relative rather than a conventional price comp: Horowitz calls it his biggest check ever, while Kalanick says it is several orders of magnitude larger than Uber’s first $4 million pre-money round. The transaction also required merging separate entities with different investors. The team is recruiting from Uber’s operating bench, including former CFO Gautam, former engineering SVP Ganesh, and former Uber ATG lead Eric Myhofer. The diligence question is whether this is a scalable industrial platform or a collection of difficult vertical bets; the company’s own examples span food, mining, and specialized industrial transport.
2. Emerging Teams
Agent correctness is becoming an investable wedge separate from agent capability. An ex-PM for AI at a fintech describes a failure mode in which an agent changed loan records, returned “done,” and left a green trace even when writes had not reached the underlying system. He is building post-hoc verification against the real system rather than the agent’s own logs, but still lacks live production agents to test against and is seeking three to five design partners in billing, CRM, and internal operations. This is a particularly clear early-stage test: the product either catches silent state divergence on real traffic or it does not.
MARGINAL is taking a similar reliability-first approach for coding agents. The open-source governance layer watches for repeated actions, weak progress, redundant verification, and low-value continuation; it starts in Shadow Mode, earns permission to block only after proving reliability, and falls back when confidence degrades. Its proposed next layer is counterfactual evaluation: whether the agent would actually have done better without the intervention. The ensuing design discussion recommends paired governed-versus-ungoverned runs from the same state, repeated seeds, versioned policy and reason codes, and a regression corpus for false stops and missed loops. The investor takeaway is that governance overhead and task outcomes—not token savings alone—should determine whether these systems earn enforcement rights.
Valor Atomics is underwriting nuclear scale through iteration speed. The less-than-three-year-old company says its OR250 reactor reached first criticality in November and produced power “a few days” before the interview. Its internal metric is the time between atom-splitting events: two years and four months from incorporation to the first split, about seven months to the second, and an eventual goal of minutes. Founder Isaiah Taylor says the team deliberately recruits high-agency hardware builders, including people who entered nuclear without prior reactor experience.
The financing and deployment strategy is unusually execution-heavy for nuclear: Valor is using equity capital to put reactors on the ground before project finance becomes available, and proposes building gigawatt sites speculatively so customers can follow the power rather than determine the construction schedule. That makes the company worth tracking as a hard-tech formation signal, while keeping the commercial scale thesis separate from the demonstrated first reactor.
3. AI & Tech Breakthroughs
GLM-5.3 makes post-training, not just pretraining scale, the center of the open-model competition. Interconnects reports that Z.ai’s model is initially available through its coding plan, with API access and Hugging Face weights planned, and places it near the frontier of agentic coding benchmarks at roughly 750 billion parameters. The analysis says GLM-5.3 uses the same base model as GLM-5.2 with substantially extended post-training, and quotes Z.ai’s framing: “Scaling post-training is all we did.” The article also argues that Chinese labs’ faster public release cadence—days rather than the months typical of OpenAI or Anthropic—may compound if model-improvement loops depend on user data.
The release carries a concrete dual-use test. Z.ai describes GLM-5.3 as its most capable cybersecurity model and says it will stage access through security partners, then API availability, before releasing complete weights; the analysis notes that smaller, easier-to-modify open models may make any single lab’s safeguards insufficient.
Document AI is being evaluated against production degradation rather than clean PDFs. A current-period ExtractBench repost points to regulatory filings, hand-filled tax forms, photocopies, and sensor noise; its summary says Codex performs well on scans but poorly on rotated documents, while OCR systems show the inverse pattern. This is also a benchmark-governance signal: Vik Paruchuri said fixing scoring bugs moved Datalab from 65% to 93.6%, and LlamaIndex’s Jerry Liu said the team would incorporate the configuration changes and update the results. For diligence, reproducibility and failure-case coverage deserve the same scrutiny as the headline score.
AI-biology is moving from sequence prediction toward wet-lab design. A Lightspeed interview describes Stanford and ARC Institute researchers using a genomic model trained on roughly a couple million bacteriophage genomes to design new bacteriophages targeting E. coli; the speaker presents the resulting variants as a first-of-its-kind result and flags the biosecurity implications. The important qualification is that the model produced a design roadmap: human scientists synthesized the viruses in a wet lab, and manufacturing and regulatory controls remain meaningful bottlenecks.
Valor’s physical proof is more useful than its long-range projections. The company says it connected an Nvidia Blackwell system directly to the reactor and hosted a website from that chip, while planning a passive-safety demonstration that shuts off electrical supply and active cooling so natural circulation removes decay heat. Those are company-reported demonstrations, but they give investors more to underwrite than a reactor design deck alone.
4. Market Signals
Compute is becoming a longer-lived asset and a routing problem. CoreWeave says it remains sold out of prior-generation GPUs as well as current SKUs and has signed an A100 contract extending into 2029, despite the chip having launched in 2020. Clouded Judgement interprets the evidence as support for useful lives of nine years or more and points to a software layer that routes each inference request to the right chip as workloads spread from frontier-complex to trivial. The opportunity is therefore not only new capacity; it is asset utilization, workload routing, and economics across heterogeneous fleets.
AI coding is consolidating around compute, product data, and the outer loop. The Cursor/SpaceX account describes a $60 billion all-stock acquisition that gives Cursor “an order of magnitude more compute” for post-training and product-linked experimentation. This is a late-stage strategic transaction, not an early-stage financing comp, but it is directly relevant to AI competition: Cursor says agents have pushed productivity gains beyond tab completion, while the resulting code-review bottleneck has created demand for tools around testing, integration, and verification. Its revenue mix had shifted to almost 80% B2B, with large enterprise deployments expanding rapidly.
Agent products are being differentiated by domain feedback loops and APIs. Klaviyo reports that its Composer marketing agent reached 95,000 users in its first month, with about one-quarter returning weekly, and that the first prototype was built over a weekend by other agents. Its claimed moat is not the base model: Composer receives a proprietary real-time feed of consumer responses and uses a coaching agent to score proposed campaigns for engagement and revenue. The company is also treating every employee as an L3 agent operator and pushing toward headless, agent-friendly APIs rather than human-only interfaces.
The founder pipeline is becoming more singleton-heavy. Paul Graham reports that single-founder companies rose from 9% to 18% in recent YC batches, attributing the increase—tentatively—to AI enabling one person to get more done. He still argues that a cofounder’s value is stress-bearing, not just additional output. That is a useful counterweight to claims that agent leverage removes the need for complementary founding teams.
5. Worth Your Time
- Watch — Travis Kalanick on Building Atoms After Uber. Focus on the physical-world CPU/storage/network analogy and the definition of industrial AI; it is the clearest expression of the period’s industrial-execution thesis.
- Watch — AI Creates New Viruses, Building AI for Healthcare & Moats in Healthcare IT. The useful segment separates model-assisted biological design from wet-lab production and lays out the associated biosafety question.
Read — GLM-5.3: How Chinese labs keep stride with the frontier. Read it for the post-training thesis, release-cadence comparison, and dual-use caveat.
Read — ExtractBench. The current-period repost links the paper and site and gives a compact picture of why real document conditions defeat one-size-fits-all extraction claims.
Anthropic targets a record $2T IPO. The FT reports investors expect an October debut (six to eight weeks out), breaking SpaceX's $1.75T IPO record; Polymarket implies 80% odds of an IPO this year and 67% odds Anthropic still holds the top model at year-end . Chamath pegs year-end annualized revenue run rate at $100–120B — up ~10x year-over-year for three straight years, an unprecedented Silicon Valley ramp (16–20x sales) . Baker is skeptical of the $2T leak (bankers jockeying for lead-left) but expects it to trade to $3–4T . Sacks: the TAM supports continued growth — agents are just taking off, token demand grows exponentially — with compute/energy as the binding constraint ; Anthropic's premium pricing depends on staying ~6 months ahead of open models, and Dario's own "FAA for AI" regulatory push would erase that lead, though a top ~20% of buyers will keep paying ~10x for frontier intelligence (Apple vs Android) . Baker notes Anthropic is losing share on the margin to OpenAI, open source and Grok, yet still growing exceptionally . Anthropic is in talks to buy AI startup Descartes for $6B — software that lowers training/inference cost by making chips more efficient . Caution flag: Dario reportedly says Anthropic might become "the only private company in the world"; the panel calls it hubris — "SBF land" .
Grok 4.6 reached the frontier at lower cost. xAI released Grokbot and Grok 4.6 this week; Cursor Bench, Databricks (freshly funded at a $190B valuation) and Mercury evals show it at/near frontier quality at lower cost — ahead of the "Fable 5" gold standard at a slightly lower price . It is a 1.5T-parameter model; a "significantly more capable" Grok 4.7 follows in a few weeks . Grokbot is framed as an "OpenClaw moment" for personalized, democratized AI . The panel credits the six-month leap to Elon buying Cursor and installing SpaceX executives atop xAI management . Baker: investors sizing SpaceX ignore Grok, which is "Pareto dominant on a lot of measures" . Elon also holds a call option on frontier status and a put option on compute: the deal renting compute to Anthropic has 90-day cancellation on both sides, letting him repurpose capacity or charge near spot .
Nvidia is securitizing AI compute. Nvidia partnered with Goldman, BlackRock, KKR, Apollo and others to raise $500B for AI compute — treating GPUs as financeable income-producing assets: customers borrow to buy Nvidia systems, rent compute, repay from cash flows; Nvidia matchmakes and, per Baker, offers residual-value guarantees (bearing ~25% of risk) plus revenue share above a floor; Morgan Stanley sees Nvidia becoming a very large capital-light cloud on these royalties . Sachs: this removes the financing constraint on the AI TAM — e.g., Elon's ~6–8GW 2027 buildout would cost $300–400B vs the ~$100B he raised — and compares it to aircraft asset-backed lending; Baker calls Nvidia the "central bank of AI" . CoreWeave already rents 2020-vintage Ampere GPUs profitably through 2029 — a nine-year asset life . Sachs's biggest risk: a compute glut ("dark GPUs," like dark fiber) from overbuild on $30–50/watt spot assumptions, though political headwinds likely prevent oversupply .
Open source is the disinflation force. Zuckerberg's 6,500-word essay ("The Future is for Everyone") pledges open-source models, a free AI agent/tutor for everyone, safety via "balance of power" rather than a singular centralized intelligence, and a right to learn from anything observable (defending distillation) . Baker frames the split: Anthropic/EA believe AI is "too dangerous to distribute"; Zuckerberg, Elon and Jensen believe it is "too dangerous to centralize" . Chamath expects Anthropic's growth to slow sharply — open models like GLM are ~90% cheaper than Claude, are already used by the founders he backs, and corporate America will adopt them (Linux analogy) ; one hyperscaler told him it reversed from frontier training on ~$100B infrastructure to open source . Baker's equilibrium: frontier tokens capture 65–85% of economic value while open source is ~80% of volume — open source makes frontier models more valuable by giving them cheaper intelligences to orchestrate .
Competitive shifts. OpenAI pivoted to coding (GPT 5.6) and is reportedly growing >20% month-over-month — an extrapolated ~10x annual rate, matching Anthropic ; Anthropic's own prescient coding bet followed Cursor's usage signals . Silver Lake may be buying Workday (stock +17%); Baker reads it as private equity returning to software after a 12–18 month "death spiral," with open source a "godsend" for the American software industry versus a world of 2–3 dominant labs .
Anthropic's S1 is the industry's pace car. Baker and Sachs expect Anthropic's quarterly disclosures to be the most important signal for the entire AI capex food chain — a demand-driven stumble would cause a pileup (Anthropic makes ~$100B per gigawatt of compute, pays SpaceX ~$50B/GW, which pays Nvidia ~$30B for chips), while being overtaken by a competitor would not . Baker: macro bears wrongly assume tokens are subsidized — "the overwhelming majority of tokens are profitable for everyone in the chain," and Anthropic is cash-generative; the S1 will "break a lot of people's brains" .
Valor Atomics (Isaiah Taylor, <3 years old) is the first startup to make power with an advanced reactor: OR250 went cold-critical in November and produced power days before this recording, making it the first advanced reactor built outside a national lab, the fifth new US nuclear device to make power since 2000, and Valor the only private company founded since fission's discovery to have made nuclear power .
The company powered an Nvidia Blackwell AI chip directly from the reactor (100 kW, ~1e17 atoms split per second) and hosted a live website on it that counted uranium atoms split per page view — a proof point for AI-driven power demand .
Founder Isaiah Taylor has no credentialed nuclear background but cites a Manhattan Project great-grandfather and a decade of waiting for others to iterate; he built a "bias to action" team of hardware-builders, including two 21- and 23-year-old engineers who invented a no-rebar, gamma-blocking concrete after a "rock hunt." Valor verticalizes whatever blocks scale: it built its own reactor protection system in 6 weeks for ~$400k vs a $5M vendor quote .
The reactor is a Triso-fueled, graphite-moderated, helium-cooled design with passive safety: after a scram, all electrical and active cooling systems can be shut off, with natural circulation removing decay heat. The "modular Citadel" bioshield uses precast concrete blocks with tortuous-path seams, stacked in ~42 hours versus 3 months for conventional shielding. It operates under Executive Order 14301 via the DOE testing pathway, bypassing the NRC commercial-licensing chicken-and-egg .
Taylor argues most nuclear incumbents are "modeling and simulation" companies that won't scale by 2031 or even 2035; Valor targets a "tick rate" of minutes between reactor startups (currently 2y4m to first split, 7m to second). It is venture/equity-funded and plans to build gigawatt-scale "giga sites" speculatively, betting AI load will follow; he projects a $500B company in 10 years .
- SpaceX is acquiring Cursor (Anysphere), the AI coding company, in a $60B all-stock transaction — the largest acquisition of a venture-backed startup ever — after SpaceX (which absorbed xAI earlier in 2026) announced an April 2026 call option: the right to buy Cursor for $60B, or a breakup fee of $1.5B cash plus $8.5B compute. SpaceX's June 12 IPO was the largest in history .
- Cursor was founded by Michael Truell and MIT cofounders Sualeh Asif, Aman Sanger, and Arvid Lunnemark, who first tried an AI email client and an AI-powered CAD tool before pivoting to Cursor; the four dropouts built Anysphere into the fastest software company ever to reach $100M ARR .
- Cursor went from token reseller to model builder: Composer (Oct 2025), Composer 1.5 (Feb 2026), Composer 2 (five weeks later, built on Moonshot's Kimi 2.5 base with extra RL), and Composer 2.5 (eight more weeks later) — Composer 2 beat Anthropic/OpenAI on price-performance, and Composer 2.5 made Cursor look competitive in absolute terms. The strategy combined open-source base models with unique data from Cursor's large developer distribution while 'compute-starved' .
- SpaceXAI compute gives Cursor 'an order of magnitude more compute' for post-training; with this week's release of Grok 4.6, Cursor's ability to produce Pareto-frontier models plus SpaceXAI compute is already 'a real competitive threat to the frontier' .
- Cursor's enterprise traction: revenue flipped from >80% B2C at the start of 2025 to ~80% B2B by the deal, with early enterprise cohorts growing >10x in year one ; NVIDIA went from POC to 100% of developers (30k+ weekly active users) within months, and National Australia Bank finished a scheduled 9-month .NET migration in 9 weeks with 100–150 engineers, now at 5–6k engineers . Productivity data: tab-complete lifts developers 5–10% and agents ~35%, with a third wave of parallel long-running agents targeting enterprises, where average customers get ~65% of production code from AI . One Fortune 500 CIO's code output rose from 150k to 800k lines/week, creating a code-review bottleneck Cursor targets with 'outer loop' tooling like BugBot .
- a16z authors (Casado, Bornstein, Ding Wang) use the Cursor/SpaceXAI story to argue the winning bet in AI coding is the fastest-iterating team, not the smartest or most compute-rich one, calling AI coding potentially the biggest addressable market of our lives .
Brett Adcock — founder of humanoid-robot maker Figure, public eVTOL company Archer, and recruiting marketplace Vettery (sold for > $100M) — is now running a new AI lab, Hark. The podcast host puts Figure at a $30–50B valuation and says Hark has raised at a multi-billion-dollar valuation; Adcock describes Hark as the digital half of his AI thesis: a Jarvis-like AI-human symbiosis assistant in the digital world, alongside Figure's physical humanoid robots.
Hark shipped its first computer-use model/research preview a week before the episode and claims top results on leading browser/computer-use benchmarks. Its technical approach: spin up a virtual computer per agent and train on screen + cursor + keyboard rather than APIs/MCPs (only ~1 in 1,000 websites have APIs), with a post-training RL process Adcock believes is unique. A consumer app on browser/iPhone/Android is slated for about a month out, with new AI hardware in the lab.
Hark raised on team strength: the early deck was just Adcock on video plus a list of ~100 people who moved from China with strong backgrounds; the team is now ~80–90, including a hardware lead who designed recent iPhone/MacBook generations. Adcock's device bet is that the next AI-native device replaces the phone and computer (AI OS, own weights, near-perfect memory), not glasses or wearables — he calls Meta glasses one of the worst products he has bought and says the end state is BCI.
Figure has made 1,000 EVT Figure 3 robots, shipped to its third customer, and counts BMW as a commercial customer. A real customer logistics task (package sorting) ran 200 hours straight at 2.9 seconds per package — already human speed — with customers getting strong ROI; demand is driven by labor shortages and >100% annual turnover in some areas. Adcock cautions the robotics market is full of noise and bullshit, making it hard to separate signal.
The bottleneck is no longer manufacturing but onboard intelligence and generalization: Figure can fold laundry in known settings but fails out-of-distribution (new homes, lighting, table heights), and needed robot data is not on the internet. Adcock says the home is super solvable today and the goal is a robot that enters any environment via language — he estimates that can be done with a hundred robots and a 50-person team and would create a trillion-dollar company. He treats Figure/Hark as binary: either robots scale or they don't.
AI talent market: Meta is buying talent at extreme packages — Adcock lost a senior Hark candidate to a $36M/4-year Meta offer — and junior AI engineers now earn ~$750k–2M/year. Adcock estimates only ~20–30 people in California can build truly good AI models; he calls Meta's buy-talent strategy smart and working so far, but doubts mercenary hires will ship great new products.
Investment themes: Adcock argues hard problems are easier in relative terms (less competition, bigger TAM, binary payoff for investors); humanoids are only ~3–4x harder than robot dogs but a million-x bigger payoff. He dismisses AI slop startups and says coding agents are excellent but most other AI products are not good. He flags energy generation as a 30-year secular trend and says AI will be 100 times bigger than the Internet.
Network and outlook: Jeff Bezos invested in Figure and acts as a soundboard; Jensen Huang visited the week before. Bezos told Adcock the next 1–2 years are make-or-break for scaling. Adcock says his Jan 1 prediction of 2027 unsupervised multi-day home robot tasks is on track, expects a full human Turing test on speech by 2027, and says Cover (school-weapons scanning) may miss its 2026 K-12 beta by a quarter due to custom chip fabrication.
Founder signal: Cursor (Anysphere) was founded by Michael Truell and MIT cofounders Sualeh Asif, Aman Sanger, and Arvid Lunnemark — four dropouts in their early 20s — after earlier attempts at an AI email client and AI CAD tool. Their early 2023 HN launch got only 11 comments before the product became the fastest software company ever to $100M ARR and the only premium AI coding product that started as a product, not a lab . Truell's stated strategy was to build on the model side, become multi-product, and own the whole AI coding bundle (editor + review/collaboration) .
Technical trajectory: Cursor launched its own Composer model in Oct 2025, Composer 1.5 in Feb 2026, then Composer 2 five weeks later on Moonshot's Kimi 2.5 base with heavy additional training/RL, and Composer 2.5 eight weeks later — beating Anthropic/OpenAI on price-to-performance. The app was rearchitected around parallel agents (Cursor 2.0, Oct) and then as a unified agent workspace (Cursor 3, April) . Cursor's edge came from being 'compute-starved' and combining open base models with proprietary developer-distribution data .
Market signal: AI coding is 'the biggest addressable market of our lives' (only ~1% done) . Cursor's revenue flipped from ~80% B2C to ~80% B2B, enterprise cohorts grow >10x in year one, the average enterprise customer gets ~65% of production code from AI, and the scarce resource shifted from writing code to vouching for it (driving new tools: BugBot, Origin, SDKs, automations, cloud agents) . Concrete traction: a Fortune 500 CIO saw code output jump from ~150k to ~800k lines/week; NVIDIA went from POC to 100% of developers (30k+ WAU); NAB completed a 9-month migration in 9 weeks and scaled to 5-6k engineers .
Competitive dynamics: Anthropic's Claude Code and OpenAI's Codex became two of the fastest-growing products ever after Opus 4.5, and by Jan 2026 the discourse had written off Cursor in the 'IDE is obsolete' narrative — before Composer and the SpaceXAI deal .
Deal/compute: SpaceX (which absorbed xAI) took a call option on Cursor in April 2026 — $60B acquisition or a breakup fee of $1.5B cash + $8.5B compute — IPO'd as the largest ever on June 12, and four days later confirmed the $60B all-stock acquisition, the largest venture-backed startup acquisition ever, giving Cursor 'an order of magnitude more compute' . xAI had built Colossus data centers and more compute than OpenAI/Anthropic but couldn't close the coding gap; with Grok 4.6 out that week, Cursor's Pareto-frontier models plus SpaceXAI compute is described as a real frontier threat .
- Lightspeed health tech partner Dr. Brenton Farnoli breaks down a Science-published Stanford/ARC Institute study: researchers built a genomic language model (Evo 2, trained on 9.3 trillion nucleotides, fine-tuned on ~2 million bacteriophage genomes) and designed 16 novel bacteriophages targeting E. coli — the first AI-designed viruses not found in nature, aimed at drug-resistant bacteria. The AI produced design roadmaps only; human scientists in wet labs synthesized the viruses, and manufacturing/regulation remain bottlenecks.
- Healthcare AI adoption is leapfrogging in provider areas SaaS ignored — scribing and voice AI for scheduling are exploding, driven by a "10x" wow factor, buyer bottom-line pressure, and bottoms-up physician adoption. This is creating an opportunity for a "system of action" layer versus the EHR "system of record".
- Founder playbook: insert at strategic patient-interaction moments (calls, visits, record transfers) then expand toward platform; winning teams are either deep domain experts (often physicians) or AI-native first-principles builders, with domain expertise mattering more near clinical care.
- Incumbent watch: Epic is building "Curiosity", a model trained on its EHR data to predict the best next clinical action, and is expected to announce more at the upcoming HIMSS ("Health IT Super Bowl") after entering scribing last year.
- Lightspeed's active healthcare themes: AI-native care delivery, employer healthcare cost inflation, robotics in OR/manufacturing, and healthcare-specific data/models/infrastructure; it also sees strong AI-lab talent now founding healthcare startups.
- Meta released Glimmer, an open-source AI model that is a pared-down version of its closed-source frontier model Muse Spark (released in April). Glimmer is downloadable and runnable locally, while Muse Spark remains API-only; Meta positions it as enabling personal AI assistants, but it requires specific hardware and is not accessible to average users.
- Zuckerberg's "AI for everyone" manifesto (6,500 words) explicitly rejects the AI-safety slowdown narrative ("we cannot afford to give an inch to China") and promotes locally-run personal AI; it has drawn widespread negative reaction, with criticism centered on the messenger and Meta's lack of consumer traction.
- Anthropic added watermarking to AI content in compliance with EU law; the move sparked user pushback over changing rules around AI use.
- AI infrastructure energy demand is a growing market signal: data centers are expected to quadruple US electricity use by 2035; Amazon's proposed Texas data center includes an on-site gas plant that could be the largest carbon polluter in the US (Amazon's emissions rose 16% in 2023, moving away from its 2040 net-zero pledge); community protests are intensifying.
- The energy bottleneck is creating tailwinds for climate/energy startups: Reservoir raised $8M for water heaters that store heat to shift demand; Discovered Materials uses AI to design cooler chips for data centers; investors are increasingly funding energy storage and efficiency.
- Cautionary flag: The $250M acquisition of VideoVerse by Minute Media has devolved into fraud allegations (including fabricated deal terms) and lawsuits; the CEO was ousted in April and is unreachable. It highlights the limits of due diligence and fraud risk in private startup investing.
- a16z's piece frames AI coding as "the biggest addressable market of our lives" (~1% addressed) and argues the winning team is the one that iterates fastest, not the smartest or best-funded .
- SpaceX (which absorbed xAI in early 2026) will acquire Cursor in a $60B all-stock deal — the largest acquisition of a venture-backed startup ever — following an April 2026 option with a $1.5B cash + $8.5B compute breakup fee and SpaceX's record-setting June 12 IPO .
- Cursor's founders — Michael Truell and MIT cofounders Sualeh Asif, Aman Sanger, Arvid Lunnemark, four dropouts in their early 20s — first built an AI email client and CAD tool before pivoting to the AI code editor launched on HN in early 2023; Cursor became the fastest software company ever to $100M ARR .
- "Compute-starved" Cursor went from Composer (Oct 2025) to Composer 1.5 (Feb 2026), then Composer 2 on Moonshot's Kimi 2.5 base five weeks later — beating Anthropic/OpenAI on price-to-performance — and Composer 2.5 eight weeks later, approaching the frontier in absolute terms; the thesis: open-source base models + unique developer-distribution data .
- SpaceXAI access gives Cursor "an order of magnitude more compute"; its Colossus 1/2 data centers straddle the TN/MS border, and xAI had more compute than OpenAI or Anthropic .
- Enterprise traction: B2B went from ~20% to ~80% of Cursor revenue between early 2025 and mid-2026; early enterprise accounts typically grow >10x in year one; NVIDIA hit 100% developer adoption and 30k+ weekly active users; National Australia Bank compressed a 9-month migration into 9 weeks and scaled to 5–6k engineers .
- Productivity shift: agents give ~35% lift vs 5–10% for tab-complete, and average enterprise customers now get ~65% of production code from AI; code review is the new bottleneck — one F500 CIO saw output jump from 150k to 800k lines/week — and Cursor is responding with outer-loop products (BugBot, Origin, automations, cloud agents) .
- Competitive context: Anthropic's Claude Code and OpenAI's Codex became two of the fastest-growing products ever after Opus 4.5, threatening IDEs; Cursor responded with multi-agent workspaces (Cursor 2.0, Cursor 3), and with Grok 4.6 now out, Cursor+SpaceXAI is already proving a real competitive threat to the frontier .
- AI coding thesis: a16z's Bornstein, Casado, and Wang call AI coding "the biggest addressable market of our lives," roughly 1% solved, and argue the fastest-iterating team usually wins . They note only one gigawatt-scale data center is operational today .
- Founders: Cursor/Anysphere's MIT cofounders Michael Truell, Sualeh Asif, Aman Sanger, and Arvid Lunnemark first built an AI email client and AI CAD tool; they switched to Cursor after "catching the bug," launching on HN in early 2023 to 11 comments . The four dropouts in their early 20s made Cursor the fastest software company ever to $100M ARR . Truell earlier named his 2019 Neo accelerator team "Winning Team" .
- Product strategy/traction: Cursor was the only premium AI coding product that started product-first, not lab-first ; 2024 growth was "like the Beatles for software" . It acquired Supermaven in Nov 2024, made model-switching central so developers stayed "sovereign," and drew criticism as a token "reseller" .
- Competitive threat and response: Claude Code in spring 2025 and then Opus 4.5/Codex in late 2025 made agentic coding take off; by Jan 2026 Cursor had been written off . Skeptics called AI subscription businesses vulnerable to a "short squeeze" (a16z disagreed) . Cursor shipped Composer (Oct 2025), Composer 1.5 (Feb 2026), Composer 2 on Moonshot's Kimi 2.5 five weeks later, and Composer 2.5 eight more weeks later, beating Anthropic/OpenAI on price-to-performance . It was "compute-starved," relying on open-source base models plus unique distribution-scale data .
- SpaceX deal: xAI (founded March 2023) built Colossus 1/2 and had more compute than OpenAI/Anthropic but couldn't close the coding gap . In April 2026, SpaceX (which absorbed xAI) took a call option on Cursor — acquire at $60B or pay $1.5B cash + $8.5B compute breakup; after SpaceX's largest-ever IPO on June 12, 2026, Elon confirmed the $60B all-stock acquisition, the largest venture-backed startup acquisition ever . Cursor gains "an order of magnitude more compute" for post-training iteration . With Grok 4.6, Cursor + SpaceXAI compute is already a competitive threat to the frontier .
- Enterprise metrics and product evolution: Per COO Jordan Topoleski, tab-complete gave 5-10% productivity lift, agents ~35%, and the next wave of long-running parallel agents targets ~65% AI-generated production code in enterprise; one Fortune 500 CIO's code output jumped from 150k to 800k lines/week, creating a code-review bottleneck that Cursor addresses with its "outer loop" (BugBot, Origin, SDKs, automations, cloud agents) . Revenue flipped to ~80% B2B by the SpaceX deal; early enterprise cohorts often grow >10x in year one. Examples: a $1.2M legacy migration completed by 7-8 engineers in a 3-day hackathon; NVIDIA went POC to 100% of developers/30k+ weekly actives; NAB compressed a 9-month migration to 9 weeks and scaled to 5,000-6,000 engineers . Product evolved to parallel-agent workspace (Cursor 2.0) and "unified workspace for building software with agents" (Cursor 3); Grok Bot extends developer powers across the org .
Travis Kalanick unveiled Adams, his new company focused on "industrial AI" — "systems of software, sensors, robotics and AI used to automate operations of entire industrial sectors" . He frames it as continuing Uber's "bits to atoms" mission, treating physical-world industries as "atoms-based computers": manufacturing as CPU, real estate as storage, transport/logistics as network . Adams is building across food (robotic production and delivery targeting under $8/meal, including "autonomous burritos") , mining (automation to boost output ~20%/year and improve safety) , and transport (specialized robotics for industrial-scale tasks including food/parcel delivery, freight, and off-road autonomy) .
Funding: a16z's Ben Horowitz says the investment is the biggest check he has ever written; Kalanick says the round is "several orders of magnitude bigger" than Uber's first round at $4M pre, and involved merging separate entities with different investors .
Team: Adams has attracted Uber alumni including former CFO Gautam, ex-SVP engineering Ganesh, Eric Myhofer (former Uber ATG lead, now running Food Robotics), and Jessica Morton (ex-head of Uber Eats Japan) . Kalanick says he worked in stealth for 8 years with a "massive high fidelity playbook" to avoid attention .
Market thesis: He argues multiple multi-hundred-billion/trillion-dollar physical-world industries will be automated — "we're competing for the next Industrial Revolution" — and calls Elon Musk "the best of the best" / "final boss" in industrial AI . Kalanick also says he has evolved as a founder, now operating "a few inches off the line" and adopting a "Best Idea Wins" culture .
a16z partners Matt Bornstein, Martin Casado, and Sarah Wang published an in-depth analysis of Cursor's 2023–2026 arc, framing AI coding as potentially the biggest addressable market of the era and arguing the fastest-iterating team wins .
- Deal: SpaceX—which absorbed xAI earlier in 2026—struck a call-option deal with Cursor in April 2026 (right to acquire for $60B, or a $1.5B cash + $8.5B compute breakup fee); after SpaceX's June 12 IPO (the largest in history), Elon Musk confirmed the $60B all-stock acquisition days later — the largest acquisition of a venture-backed startup ever, giving Cursor "an order of magnitude more compute" .
- Founding team: Anysphere/Cursor founders Michael Truell, Sualeh Asif, Aman Sanger, and Arvid Lunnemark (MIT dropouts in their early 20s) first built an AI email client and an AI-powered CAD tool, then launched Cursor in early 2023; Anysphere became the fastest software company ever to $100M ARR and acquired Supermaven (Nov 2024) for its tab-complete model. They are the only premium AI coding team that started as a product company rather than a lab .
- Model strategy: "compute-starved," Cursor combined open-source base models with distribution-scale developer data to build its own Composer models — Composer launched Oct 2025; Composer 1.5 in Feb 2026; Composer 2 five weeks later (built on Moonshot's Kimi 2.5 base with heavy RL for coding) beat Anthropic/OpenAI on price-to-performance; Composer 2.5 eight weeks later was competitive in absolute terms .
- Competitive shift: Anthropic's Claude Code (spring 2025) and OpenAI's Codex became two of the fastest-growing products ever after the late-2025 frontier releases starting with Opus 4.5; by January 2026 discourse had written off Cursor and the IDE form factor, while the frontier labs raise tens of billions quarterly with locked-in compute .
- Enterprise signal: Cursor's revenue flipped from ~80% B2C to ~80% B2B, with early enterprise cohorts growing >10x in year one; NVIDIA went from POC to 100% of developers (30k+ weekly active users) within months; National Australia Bank's 9-month .NET monolith migration took 9 weeks and now spans 5,000–6,000 engineers; a Fortune 500 CIO saw code output rise from 150k to 800k lines/week; average enterprise customers get ~65% of production code from AI .
- Product & roadmap: Cursor 2.0 (Oct) rearchitected the app around multiple parallel agents; Cursor 3 (April) is a "unified workspace for building software with agents," no longer a classic IDE; new Grok Bot gives developer powers to non-developers, and this week's Grok 4.6 pairs Cursor's Pareto-frontier models with SpaceXAI compute .
- Backstory: xAI (founded March 2023, the same month as Cursor) built the Colossus 1 and 2 data centers and had more compute than OpenAI or Anthropic but could not close the coding gap, prompting Musk's outreach to Truell .
Garry Tan, President & CEO of Y Combinator, said opponents of San Francisco's Marina Safeway project are helping YIMBYs assemble a comprehensive list of California laws to repeal . His comment responds to a report that neighborhood opponents are asking San Francisco officials to halt the project's fast-tracked approval under California law over what they call an unresolved environmental testing gap .
An unreleased version of AI model Claude failed to prove the Riemann Hypothesis but improved a related bound beyond the human record, with mathematicians calling it "a massive leap forward" . The successful run was prompted by a non-mathematician who mostly sent encouragement like "keep going" and "believe in yourself" after 650 failed tries; similar encouragement helped Claude disprove the Jacob conjecture . A formalized version of the proof is available for automatic verification, and the AI-generated explanation of its findings is included in the technical paper . During the key breakthrough run, Claude had internet access but did not use it, recovered from wrong paths, and self-assessed the result as "too strong to be new," even underestimating the rate of AI progress .
Cursor was acquired for $60B, after Martin Casado / a16z led its Series A ~2 years ago at a $400M valuation — ~120x return after dilution, ~1000% IRR, and a 10-digit LP gain; the quoted poster calls it one of the highest-IRR VC investments of all time . Garry Tan called it "definitely GOAT level work" and said "Amazing work" to Casado .
Z.ai (Zhipu AI) released GLM-5.3, an ~750B-parameter open-weights model at the frontier of agentic coding benchmarks that beats Moonshot's Kimi K3 on many benchmarks and surpasses Claude Fable 5 or GPT-5.6-Sol on some; it's live in the coding plan, coming to API, with Hugging Face weights in two weeks . It reuses GLM-5.2's base with heavy post-training — 'Scaling post-training is all we did for GLM-5.3' — positioning post-training as Z.ai's edge over Kimi's pretraining strength .
Zhipu was founded in 2019; its GLM line originated with Tsinghua's THUDM research group (March 2021), and close Tsinghua ties / CS talent are cited as core to Z.ai's compute efficiency . The company reportedly reached ~$1B ARR on on-premises deployments; flagship GLM models are text-only, letting it focus on high-value use-cases vs broader omnimodal models .
The post identifies Chinese labs' faster release cadence (days vs months for OpenAI/Anthropic) as a major frontier-competitiveness factor, with faster cycles potentially compounding if self-improvement loops need user data . It also flags the RL data industry 'taking off' in China as American data companies sell RL environments to Chinese model labs, scale/impact still uncertain .
Cautionary: GLM-5.3 is framed as Z.ai's 'most capable model to date for cybersecurity tasks' with dual-use risks; release is staged (security partners → API → weights) and inference is monitored via request classifier and chain-of-thought monitoring . The post argues such safety 'barely matters' as capable model sizes shrink and become easier to modify/deploy, calling for government/industry guidance .
Travis Kalanick’s new company Atoms is an industrial AI company that frames manufacturing, real estate, and logistics as the CPU, storage, and network of the physical world . In a fireside chat with Ben Horowitz and Erik Torenberg, a16z highlighted that Horowitz wrote his biggest-ever check for Atoms, and Kalanick says he has been working continuously for eight years across multiple industries . Kalanick’s founder background: a pre-Uber peer-to-peer file-sharing venture drew a quarter-trillion-dollar lawsuit from 33 media companies , after which he built a 'revenge' business that turned his suers into customers ; he now says he creates from having 'fallen in love again,' not fear or revenge .
WatchMochi (retweeted by Y Combinator) launched Mochi, described as "the first platform for anime" microdramas, explicitly following ReelShort's model — ReelShort made $1B last year on microdramas . The launch positions Mochi to replicate that playbook in anime .
Martin Casado (@martin_casado) praised @Mascobot as the most enthusiastic and tireless early advocate of Cursor, well before it was an obvious phenomenon . The referenced flashback notes Cursor was <15 people when @amanrsanger and @Mascobot cohosted a recruiting event at ICLR 2024 in Vienna, just before bringing/leading the Series A with @a16z; Mascobot congratulated the "GOAT team" .
YC President/CEO Garry Tan asserts that one person plus 20 AI agents working simultaneously can outperform an entire engineering department at a Magnificent 7 big tech company . He links this to a Dealroom note titled "Garry Tan on the new rules for founders, the $2.4B Palantir miss, and why a markdown file is an employee" , signaling a shift toward AI-agent-led, ultra-lean founding teams.
In recent YC batches, single-founder companies rose from 9% last summer to 18% this summer, a trend Paul Graham calls conspicuous . He attributes the rise to more single founders applying, presumably because AI enables one person to get more done, but still advises against it: the point of a cofounder is to help bear the stress, not just to get more done .
