20VC: OpenAI cut to $1.25T; Google rout; DoorDash buys Instacart
Topics
AI Venture Funding and Bubble Risks
Battle for Frontier AI Moats
Commercial Pitfalls of AI-Native Applications
Autonomous Agents Reshaping Enterprise SaaS
Early-Stage VC Investing Dynamics
Embodied AI Funding and Exit Scrutiny
Summary
- OpenAI’s $1.25 trillion reset reflects relative-value pressure as investors become less forgiving across the AI trade. The proposed round falls from $30 billion to $20 billion, with about $11 billion committed; Rory compares that illiquid investment with a larger, faster-growing, profitable Anthropic potentially available publicly at $1.5–$2 trillion. Google’s roughly $500 billion selloff and CoreWeave debt yielding 15% show the shift toward “tempered euphoria,” but Rory’s decisive variable remains whether demand holds.
- Anthropic’s growth can look extraordinary and sharply decelerating using the same numbers. Rory cites roughly $5 billion, $11 billion and $17 billion of quarterly GAAP revenue: approximately 100% sequential growth becomes 50%, still a roughly 5X annualized pace. Jason’s concern is pricing pressure: OpenAI offers an approximately 30% discount, while his own application became “faster and 72% cheaper” after a model switch.
- Agent businesses are racing to improve their economics before subsidies exhaust their financing. Persistent agents might cost $20 or more per user monthly, against proposed subscriptions around $30–$50; Jason contrasts that with $4–$5 infrastructure costs for carefully managed app-building environments. Rory argues waiting for perfect margins forfeits distribution: “Can the economics get better before the money runs out?”
- DoorDash’s reported $17 billion Instacart acquisition buys network scale while agents threaten the fees those networks extract. Jason’s cited 60% and 20% delivery shares imply 80% combined, conditional on antitrust clearance. Several hundred agent-originated orders against three million daily deliveries look negligible today, but an agent finding a restaurant’s $2 direct delivery could bypass the platform; Harry and another participant point to food discovery and taxi booking as reasons people may keep using existing apps.
- Kalshi and Cohere represent opposite relationships between venture capital and government. Kalshi’s $1 billion raise at $40 billion follows a $22 billion valuation only months earlier, despite state-level legal pressure; Rory calls it a “money deal” increasingly driven by sports gambling. Cohere’s Canadian backing of up to roughly $1.14 billion supports sovereignty, but government financing cannot compel private-sector adoption, and undisclosed revenue makes the commercial case harder to assess.
- Serval offers Harry a clearer return multiple, while Jason’s initial IC pitch puts forward Cognition’s coding-agent momentum and room to deploy capital. The comparison is Serval at $3 billion, Baseten at $30 billion and Cognition at $70–$80 billion; Harry thinks a Serval 10X is “very likely,” whereas Cognition would need roughly a trillion-dollar exit after dilution. The load-bearing question is whether AI expands customer spending or merely divides an incumbent’s existing market among more competitors.
- Physical Intelligence’s proposed $20 billion valuation rests on technological potential and strategic demand ahead of deployment economics. Rory sees a genuine dexterity breakthrough, but approximately 300,000 US robots and a possible $3-per-hour model fee require substantial deployment growth. Foundation-model companies might buy robotics capabilities to expand from knowledge work into physical labor: “I’ve doubled my TAM.”
- Reliable M&A matters as much as IPO enthusiasm for venture liquidity, and buyers often need $100 million of revenue before committing their distribution machines. Harry contrasts Cresta’s reported $2.7 billion price and more than $200 million ARR with Listen Labs’ $2 billion price on $20 million revenue. Jason’s integration warning explains the preference for scale: “They can’t even assimilate you. You’re too small.”
Deep dive
1. OpenAI’s markdown exposes relative-value and financing risk
- Harry opens with OpenAI’s valuation falling from $1.4 trillion to $1.25 trillion, its proposed raise shrinking from $30 billion to $20 billion, and only about $11 billion committed. Rory rejects the automatic AI-collapse interpretation: why pay more for a smaller, slower-growing, loss-making, illiquid company when Anthropic could offer liquidity at $1.5–$2 trillion? “There’s two securities that are gonna be sold in the next two months.”
- Jason still sees legitimate “wobbliness”: approximately $5 trillion erased from AI stocks in a week, CoreWeave debt paying 15%, and Broadcom guaranteeing Anthropic debt. Rory separates financing conditions from demand. If compute remains scarce, CoreWeave’s assets should attract a buyer; if demand disappoints, leveraged suppliers suffer before Google or Amazon with their “big-ass balance sheet.”
- Harry describes Google losing roughly $500 billion and falling about 12%, with search growth slowing to 12% and free cash flow turning negative. He cites cloud growth of 103% toward $130 billion; Jason describes approximately a $120 billion run rate growing 100%—“about as good as it gets” for cloud.
- Rory attributes the reaction to search cannibalization, general anxiety and CapEx anxiety, while noting Google says it can sell compute as quickly as it produces it. Jason points to Shopify falling 15% despite raising next-quarter guidance: investors are “nervous as hell,” and even a small miss gets punished.
2. Anthropic’s arithmetic supports both the bull and bear cases
- Rory insists on distinguishing GAAP revenue from run-rate claims: roughly $5 billion in Q1, $11 billion in Q2 and $17 billion in Q3. Similar absolute additions on a larger base mechanically reduce sequential growth from approximately 100% to 50%. Sustaining 50% quarterly growth still compounds to roughly 5X annually; the bearish question is whether 25% comes next.
- Jason says a startup with declining net-new additions would fail his investment test. Rory pushes back on the distinction: revenue itself did not fall—“They sold more shit in Q3 than Q2 by about $5 billion.” Their disagreement concerns the trajectory of incremental growth.
- Rory estimates approximately $56 billion of GAAP revenue this year, against $4.5 billion last year and a $9 billion exit run rate. Extending similar assumptions produces roughly $120 billion next year: growth falls from about 12X to about 2X. Both magnificent scale and substantial deceleration can therefore be true.
- Another participant asks why bookings momentum would slow when Anthropic represents only 0.1% of Fortune 500 spending. Jason’s potential explanation is pricing: demand indicators such as OpenRouter appear healthy, but OpenAI is discounting approximately equivalent products by 30%, apparently despite a higher cost base. Anthropic’s ability to command a premium may have reached a “local maximum”; he presents that as a thesis, not a demonstrated cause.
3. Falling inference costs destabilize familiar software forecasts
- Jason’s concrete example is SaaStr Connect: switching to “Haiku 5.5 plus Jev” reduced effective costs by approximately 70–72% for the same continuously running B2B workflow, with higher quality and speed. “It’s faster and 72% cheaper.” More agents and more usage push spending upward, but that degree of application-level deflation materially complicates revenue forecasts.
- Jason also cancelled Canva after 12 years because his agents do not need it. He treats reports about zero growth or layoffs cautiously—“I don’t know what’s true”—but is surprised that disruption appears to extend beyond people running dozens of agents. He expected mainstream customers to sustain these businesses longer.
- Rory’s warning is “quiet obsolescence.” Old mainframe vendors survived on 20-year contracts, but shrinking software businesses become terminal-value assets that are difficult for investors to monetize. Joining the AI transition might permit 5%, 10% or 20% growth; continually cutting costs and harvesting an installed base is a much tougher destination.
4. Agent launches precede workable unit economics
- Jason estimates persistent Instinct usage could cost $20 per user monthly; Harry thinks power-user costs could be higher. Jason had assumed something closer to Lovable or Replit, whose carefully managed environments cost roughly $4–$5 monthly. That changes the case for subscriptions around $30–$50. Harry’s defense is blended usage: occasional customers may be cheap, while he would be “the worst” economic customer.
- Rory recalls alarm over Facebook potentially spending $100 million on infrastructure. His argument is that digital costs fall 50–60% annually, letting economics improve over two or three years after users become attached. A competitor financed by venture capital can launch 12 months before margins work; waiting until everything makes sense can surrender the market.
- Jason distinguishes Cursor’s success from a hypothetical version perpetually reselling Anthropic tokens at a loss, which he says would have died. Rory would take the $60 billion outcome and “call it a win.” Jason nevertheless finds Elon’s decision to cap Grok bot usage revealing: even aggressive subsidizers face limits. Their disagreement is about which teams escape the cost structure before funding runs out.
5. DoorDash buys concentration while agents threaten its take rate
- Harry reports DoorDash buying Instacart for $17 billion, less than half its 2021 private valuation; D1 still owns 9.7% and Sequoia 8.1%. Using Jason’s cited delivery shares, combining 60% and 20% could create an 80% network. He questions whether that should pass antitrust, but says concentration explains the strategic attraction if clearance comes.
- Jason admits surprise that combining delivery networks appears to work: he expected acquisitions such as Uber Eats–Postmates to deliver little beyond a brand and customer list. Instead, well-structured deals appear accretive. Rory adds a possible timing motive: “Do it now while you can.”
- DoorDash’s several hundred agent orders against three million daily deliveries tell Rory what happens today, not what comes next. Jason’s sharper mechanism is direct restaurant delivery: an agent could discover that a nearby curry restaurant delivers for $2 and save the customer $14. “Agents are an existential risk to DoorDash because their number three competitor is, is nobody.”
- Harry notes that people love food; another participant adds that people enjoy pictures, browsing and discovery, and sees agents as more useful for scheduled repeat orders. Harry and Jason also agree that conversational taxi booking can be less convenient than opening an app. Harry remains positive on Uber—“a banger of a buy”—while recognizing that comparison across providers could weaken platform loyalty.
6. Kalshi and Cohere put government on opposite sides of the trade
- Kalshi raises $1 billion at $40 billion, implying 2.5% dilution, after a $22 billion valuation three or four months earlier. Rory calls it a “money deal”: strong growth has displaced the original social-utility narrative with sports gambling. State-by-state regulation collides with national distribution, including the Utah charges Harry references. “VCs will put a lot of money into shit that’s working.”
- Cohere’s appeal is different: Canadian sovereignty, survival among the early LLM companies, and government investment of up to approximately $1.14 billion. Rory sees Cohere and Mistral as possible “you’re-not-America LLM options.” Jason finds withholding revenue a poor signal; Harry mocks a percentage-only disclosure that ARR grew more than 165%.
- Harry questions whether repeated state injections can overcome weak enterprise adoption. Rory distinguishes keeping a company operating from creating private demand: governments can sustain spending, but “what they can’t do is make the private sector use the product.” A compelling business that eventually lists at 3X the financing valuation is “not at all obvious to me.”
7. Crypto losses and offensive agents expose different balance-sheet risks
- Jason describes a $6.6 billion Binance loss, apparently taken by North Korea, and questions how close a larger loss might have brought it to failure. Its balance sheet is not transparent; he speculates that $13 billion could have been harder to cover. Rory’s conditional counterpoint: replacing roughly $6 billion from company resources would demonstrate an “amazing freaking business,” assuming customer assets were not diverted.
- Jason stresses that assets held as custodian are not Binance’s own balance sheet. He then flags the open-source penetration-testing tool he calls “Ardex”: initially useful for testing one’s own applications, it can be redirected toward others using LLMs at scale. The change is the ease of launching repeated attacks, rather than the invention of penetration testing.
- Rory’s metaphor is blunt: “You built a gun, and you hoped it was in good people’s hands.” He thinks the next 12 months’ cyber dynamics are difficult to overstate, imagining a day-long Fedwire outage that prevents money moving. That is a hypothetical warning, not an outage reported in the episode.
8. The mock IC exposes three different ways to underwrite AI
- Rory frames the actual choice: Baseten at $30 billion, twice June’s valuation; Cognition at $70–$80 billion, up from $48 billion in September; or Serval at $3 billion, up from $1 billion in December. Inference infrastructure, coding agents and IT service management offer different markets despite sharing rapid financing cycles.
- Jason puts Cognition forward in his IC pitch, citing coding-agent momentum and its opportunity after acquiring Windsurf’s team and IP, even if the original Devin product may not have worked. He calls $80 billion “market correct,” while worrying about limited ownership of underlying IP and a roughly trillion-dollar exit requirement for 10X after dilution. His pitch ends with the joking assurance: “We can’t lose.”
- Harry chooses Serval: strong product, exceptional enterprise adoption and a much clearer multiple from $3 billion. He thinks a 10X is “very likely.” But a $150 million financing cannot absorb the $500 million or $1 billion a huge fund might want to deploy. The distinction between maximizing returns and moving AUM is explicit: “I need to do Cognition because I need to move money.”
- Rory says nothing currently matches inference for “revenue velocity,” but flags competition and commoditization. Harry prefers Modal’s suite and Fireworks’ inference product. He also challenges Cognition’s competitive position behind Codex, Claude Code and Cursor, while disclosing that he invests in Factory and supports independent coding companies.
9. TAM expansion requires customers to pay for work completed
- Jason’s Serval question is whether it genuinely expands spending or competes for a subset of ServiceNow’s existing market. Rory proposes a concrete test: would a company spending $200,000 on ServiceNow pay $300,000–$400,000 for Serval? Tracking tasks versus performing them should show up in comparable customer ACVs.
- Rory sees potentially 10–15 years of compounding through grinding down ServiceNow, whose value Harry puts at $110 billion. Cognition can instead collect large sums immediately for helping enterprises rewrite codebases and adopt agentic development. The trade-off is immediate money velocity against a long replacement cycle.
- Jason predicts 2027 will expose vertical AI companies that assumed pricing at 3–4X incumbent software but can charge only 1X, with five to eight times as many competitors and an incumbent still alive. The result is “slicing up the same old effin’ TAM 11 times.” Rory pushes back on applying this to Serval: he thinks it expands TAM and is cross-functional. Jason clarifies that his warning is adjacent, not a specific Serval verdict.
10. Crosby’s legal opportunity depends on which profits can be automated
- Harry describes Weil Gotshal’s abrupt collapse and AI law firm Crosby, which raised $85 million, bidding for its assets and brand. He resists the idea that a few departures explain a prestigious firm’s failure. Rory argues that four or five rainmakers can carry critical practices away; a PE practice might represent 20–30% of the business.
- Rory separates trusted, individual advice on a complex deal—potentially $5,000 an hour—from an NDA or Series B stock purchase agreement where clients need standard work completed correctly. Jason counters that incompetent junior lawyers have cost him seven figures. Rory argues that reducing precisely those routine mistakes is part of Crosby’s automation case.
- Jason, who has not spoken with Crosby, questions retaining lawyers on $170,000 and 0.01% equity when alternatives pay more and successful partners earn $10 million. Rory counters that many lawyers want 40 hours, predictable compensation and family time rather than rainmaking. “They’ll be able to get lawyers.”
- They converge on the harder question: can this support a $10 billion outcome? Rory’s illustrative downside leaves 70% of conventional billings attached to irreplaceable rainmakers, then halves the remaining 30% to attract customers. A viable service business can emerge while the pool of profits available to outside shareholders remains constrained.
11. Robotics combines a capability breakthrough with slow deployment
- Physical Intelligence’s prospective $20 billion valuation contrasts with Dropbox’s billions of revenue, 9% staff cut and departing CTO. Rory’s framing: “Revenue multiples don’t get you anywhere.” Future opportunity drives the comparison, and new robotics models demonstrate dexterity he regards as a genuine breakthrough after more than a decade investing in the field.
- Deployment remains much slower than conventional LLM adoption. Rory cites roughly 300,000 robots in the US and a possible model-access price around $3 an hour; supporting a $20 billion valuation requires many more deployed machines. BMW will not install 100 humanoids in a week simply because a demonstration works.
- Rory thinks major foundation-model companies will build or buy robotics capabilities. His simplified prospectus pitch contrasts 81 million knowledge workers with a physical-worker market he states as “87” (the transcript gives no unit), and argues robotics could double the TAM. He admits that tracing every deployment step can obscure the broader bet—extraordinary technology and teams might create value before the mature business model becomes obvious.
12. M&A liquidity depends on clearance, scale and acquisition discipline
- Jason flags a “tiny amount of anxiety” about DOJ scrutiny around a reported $12.9 billion Hugging Face transaction; he does not think a formal information request has occurred. If ordinary acquisitions face obstacles again, buyers may revert to convoluted team-and-IP structures. Venture portfolios need M&A “to work like clockwork,” rather than depending entirely on difficult IPOs.
- Salesforce’s buying provides the counterexample. Jason describes senior executives effectively receiving acquisition punch cards: each proposes the best purchase for their business, and management accepts that some will fail. These are individual operating bets with accountable sponsors, rather than every transaction representing Marc Benioff staking the whole company.
- Harry compares Cresta at $2.7 billion and well over $200 million ARR with Listen Labs at $2 billion and $20 million revenue. Rory sees a choice between scale and a newer architecture. Paying steep multiples becomes uncomfortable when the acquirer itself trades around 7X revenue; Harry considers Cresta’s roughly 10X price attractive.
- Jason says major buyers often wait for at least $100 million revenue before applying their sales teams and platforms. Smaller technology bets can underperform: buy a small team cheaply or acquire enough scale to matter inside a $50 billion business. His explanation is organizational: “The Borg otherwise just can’t assimilate you. You’re too small.”