Summary

Cursor’s $2.3 billion raise at a $29.3 billion valuation is defensible if coding spend becomes a vast labor-replacement market. Tom Tunguz called agentic coding the strongest product-market fit after search, while Jason framed the bull case around a company moving from 1 to 100 and then to $1 billion in revenue, making $3–4 billion next year conceivable. Jason’s governing rule: “If you’re not seeing massive TAM expansion, there’s just no point in even playing as VCs.”
Cursor could retain 40–60% of coding-agent share if performance gains flatten before customer habits do. Tom Blomfield put Cursor at 40–60% share, while Tom Tunguz argued developers switch when models improve dramatically but stay once memory, personalized tooling and Fortune 500 standardization create inertia; he guessed Cursor might retain 75% of today’s audience five years out. Jason’s counter is that Replit V3 feels “Pluto and Mercury” ahead of predecessors and that software is improving two orders of magnitude faster than in prior eras, leaving the market’s “fat” far from congealed.
Gross margin looks improvable, but portability and price wars are the existential risks. At Tomasz’s venture firm, a 20 billion-parameter model reached 97% equivalency to Claude Code’s tool-calling capability, while Microsoft reported 90% more tokens per GPU-hour than 12 months earlier; Rory argued 60% gross margin on $1 billion of revenue with 100 employees would already produce substantial cash. The frightening case is that transferable prompts turn $100,000 agents into commodities: “If GPUs became more like DRAM, it would not be pretty out there.”
Late-stage venture currently resembles a private stock market with liquidity only when prices rise. Ramp reportedly completed four financings while moving from $13 billion to $32 billion, and 15% of Q1’s newly minted unicorns had already marked up again by Q3. Rory’s formulation was categorical: “The late-stage business is either the best business in the world or the worst business in the world,” because investors can trade step-ups but cannot count on exits during a reversal.
Thiel and SoftBank selling NVIDIA were weak top signals; the credit complex supplied sharper evidence. Thiel’s reported $100 million sale was below 1% of an estimated $10–20 billion net worth, while SoftBank was rotating into riskier OpenAI exposure, but Oracle’s credit-default-swap pricing rose to roughly three times peers as its AI commitments were repriced. NVIDIA also has extraordinary concentration—two customers above 40% of revenue and, as stated in the transcript, 4% representing above 50%—so any inference-demand wobble could make the correction “fast and brutal.”
US venture could reach $500 billion by 2030, but the outcome is becoming one correlated bet on a few companies. Tomasz Tunguz put annual deployment near $270–275 billion versus $8 billion in 2008; Rory cited Excel’s GlobalScape at $184 billion this year, with roughly half directed into four companies. If OpenAI, Anthropic, xAI, SpaceX and the other concentrated winners trade upward, their returns can swamp dozens of failed unicorns; if they do not, retail capital may discover the loss only after a five-to-seven-year feedback lag.
Stripe and the most coveted private companies may avoid IPOs because private capital now carries an “access premium.” A private round might cost about $1 million versus $25–30 million for an IPO, avoids quarterly-market burdens and now supports recurring employee liquidity; venture secondaries have consequently risen from 2–3% to roughly 10–12% of venture dollars. Tomasz sees investors “dollar cost” their way out across successive rounds, although Rory warned that a genuinely down market will test a system whose public-market behavior lacks public-market liquidity.

Deep dive

1. Cursor’s valuation asks whether growth can overwhelm two real risks

Harry opened with Cursor’s $2.3 billion financing at a $29.3 billion valuation, naming Andreessen Horowitz, Thrive, Coatue, DST and Accel among the participants. Tom Tunguz’s bull case combined exceptional agentic-coding product-market fit, 30–70% developer productivity gains, rapid revenue growth and a new Cursor model running four or five times faster by tokens per second.
The financial profile looked unusually clean for AI. Tom cited total employee count at around 30, limited ESOP dilution and little of the capital-expenditure dilution borne by foundation-model companies. Rory later used 100 employees as a hypothetical for a $1 billion-revenue business, not as Cursor’s current headcount. Tom could “see a 3X,” although whether Cursor can raise prices remained an ultimate test.
Jason’s valuation bridge used an illustrative case: if something moved from 1 to 100 a year ago and from 100 to $1 billion this year, Newtonian momentum could take it to $3–4 billion next year, turning today’s headline valuation into roughly 10 times next-twelve-month revenue rather than an obviously absurd price.
The unresolved operating evidence was retention: Tom said vibe-coding companies can show gross account retention around 50%. In the quickfire, Tomasz and Jason still chose Cursor at $29 billion over Cognition at $12 billion; Jason said Cursor’s numbers were simply “jaw-dropping.”

2. Coding spend expands from a software seat into a labor budget

Jason rejected “30 to 70% productivity boost” as an increasingly backward framing because AI coding is becoming mandatory infrastructure. He expects nearly 100% developer penetration and ultimately $5,000–6,000 of annual spend per developer, regardless of which vendor captures it.
Tom said market-sizing models five years ago assumed 25–30 million developers, whereas a recent Microsoft earnings transcript discussed 100–150 million developers on GitHub alone. Jason’s aggressive arithmetic—100–200 million people eventually spending $400–500 monthly—put the category somewhere between hundreds of billions and “coming up on a trillion.”
Willingness to pay already exceeds published tiers. Tom exhausts a $200-per-month Claude Code Max allowance two days into the week and contemplated buying several seats to reach $1,000 monthly: “I will never go back to using a computer without Claude Code.” Jason said his Replit bill was higher still.
The market extends beyond professional programmers. Jason, who said he builds products but does not code, shipped 12 Replit applications since June that were used 700,000 times. Replit, Lovable and Base44 therefore address another population beyond the estimated professional-developer base.

3. Agentic pricing supports TAM expansion far beyond coding

Tom contrasted traditional $20,000–50,000 mid-market software deals with an agentic-software sales leader whose mid-market contracts were all high six figures to low seven figures. The mechanism was straightforward: these products replace some form of labor, so software pricing escapes the historical per-seat budget.
Jason said even 2% of GDP sounded low in a world where fewer people want “hands-on-keyboard” work. Rory rejected the trillion-dollar extreme but agreed that several million serious US developers spending $5,000 annually can still support an enormous company.
The episode’s underwriting distinction followed: “Entry price counts when TAM is unclear. Winning is the only thing that counts when TAM is huge.” That is why the panel accepted Cursor’s price while preferring Legora at $2 billion to Harvey at $8 billion, where Jason could not yet see the required $30 billion category exit.

4. Margin improvement need not recreate old SaaS economics

Cursor’s awkward structure is that its direct competitors also supply the tokens responsible for perhaps 50–70% of its product. With little labor expense, profitability and competitive durability collapse into one platform-risk question: can Cursor reduce dependence on the model vendors whose coding products challenge it?
Jason contrasted Cursor with Replit and Lovable, which can default users to inexpensive or “N-minus-one” models and already produce gross margins north of 50%. His unanswered question was how a frontier-model-heavy product reaches 60%, even after blending in its own model.
Tomasz offered an efficiency specimen from his venture firm: it used Claude Code to teach a 20 billion-parameter model how to call tools and achieved 97% equivalency to the much larger system. “There is so much efficiency to squeeze out of these model architectures,” even if nobody returns to the prior software norm of 70–72% gross margin.
Rory argued old SaaS margins also funded large sales forces and complex integrations. A self-serve product with $1 billion of revenue, 100 employees and 60% gross margin would still generate substantial cash; Microsoft’s reported 90% increase in tokens produced per GPU-hour over 12 months strengthens that path.

5. The coding market may congeal before technical progress stops

Tom’s switching model is performance-sensitive: developers move when Gemini 3 or another release materially outperforms “Claude 405 Sonnet,” but stay once gains asymptote. His Claude Code installation contains 100 self-written tools, coding memories and linting preferences, making migration worthwhile only for a significant improvement.
Enterprise purchasing adds another layer of inertia. Fortune 500 companies will choose one vendor, standardize and buy something resembling an enterprise license agreement; Tom therefore estimated Cursor could retain roughly 75% of its current audience five years from now.
Tom Blomfield’s five-year ranking put Cursor first at 40–60%, Microsoft second if it improves and bundles through VS Code, and Anthropic third because of its coding strength—roughly a 60/20/20 shape. Rory saw first-mover advantage for Cursor, enterprise distribution for Microsoft and model-level strength for Anthropic.
Jason resisted premature consolidation. Replit V3 was not merely “night and day” better but “Pluto and Mercury”: agents summon architects, bug hunters and reviewers, while an apparently unlimited context window remembers months of work. Fully agentic functional QA could create another 10X productivity jump.

6. The dispute is about when the “bacon in the skillet” hardens

Rory’s base case was the familiar three-to-four-year land grab followed by a decade or more of stable shares, even while the market multiplies. Intel’s performance doubled repeatedly without dislodging its market position; similarly, customers may let their chosen AI coder improve rather than continually migrate.
Jason countered that software historically achieved a major release or integration every five years, whereas today’s products are improving roughly two orders of magnitude faster than software did in prior eras. Cursor itself shows what 30 people can build, so another small team could still disrupt the apparent leaders.
Tom Loverro named the disagreement “the bacon in the skillet debate”: everything is hot, fluid and sizzling until the heat falls and the fat congeals. Jason believes the skillet stays “on 10” much longer; Rory thinks enterprise adoption can solidify shares even while the underlying technology continues advancing.

7. Portability is the wedge that could start an agent price war

Jason transferred a prompt trained for months in one AI agent into Salesforce Agentforce, iterated for about a day and obtained comparable performance. His lesson was not that moats are absent, but that transferable prompts and history make them lower than conventional SaaS switching costs suggest.
Rory separated benign Bureau of Labor Statistics-style deflation—twice the tokens for the same spending—from actual price erosion. The dangerous version begins when players three through five underprice to win share and leaders respond, creating a price war that SaaS largely avoided.
Current GTM-agent economics leave ample room to cut. Jason said implementations generally start around $100,000: perhaps $50,000–70,000 for software plus roughly $25,000 of forward-deployed-engineer support. In calmer budget conditions, moving a portable workload from a $100,000–200,000 agent to a $20,000 alternative could become compelling.
Tom Loverro did not predict broad collapse, but thought one or two categories might experience it first. Rory nominated core API pricing, coding agents and Lovable-like products; all expose relatively standardized digital inputs and therefore less friction than deeply integrated enterprise systems.

8. Integration depth separates sticky software from commodity DRAM

Rory placed Salesforce and DRAM at opposite extremes. Salesforce survives cheaper alternatives because ripping out its integrations is painful; commodity memory can swing 5X and then fall 50–80%, with buyers loyal to Samsung only until Hynix or another supplier becomes cheaper.
His historical SaaS evidence was that integration count best predicted retention. An interchangeable prompt sitting in a vendor-neutral database effectively says, “Cut me now when you have to save 80 grand,” whereas five operational integrations force enough IT work to preserve the incumbent.
Tom Loverro invoked Iceberg within the data ecosystem: enterprises reclaimed control of data previously bundled with Snowflake compute and storage. A similar abstraction could let companies own prompts and selectively feed them into competing agents, shifting value away from application vendors.
Jason already sees the migration. His organization runs roughly 12 AI agents and separately has five SDR/BDR agents running through different instances and vendors; employees increasingly talk to Agentforce, Qualified or Artisan rather than Salesforce itself. Incumbents may retain their logos while “the value’s just slowly leaking out every week.”

9. Late-stage venture trades beautifully only while marks rise

Rory found roughly 20–24 newly minted unicorns in Q1; by Q3, 15% had already raised at a higher valuation, with some completing two step-ups. Ramp reportedly financed four times during the year while moving from $13 billion to $32 billion.
That velocity made Harry question seed-stage craftsmanship when his platform could instead place $10–25 million into established high-flyers. Bessemer’s Anthropic and Ramp investments, alongside participation from Kleiner, Lightspeed and other historically early-stage firms, suggested these were being underwritten as risk-adjusted opportunities rather than crossover speculation.
Rory’s warning was structural: late-stage venture is “either the best business in the world or the worst business in the world.” A $100 million position can become $200 million without operational effort on the way up; in a reversal, private-market liquidity disappears precisely when a trader most needs it.
Harry described one investor as a ruthless book manager who bought at 60 and sold at 180 in the same year. Rory accepted the “new public market” analogy with one decisive qualification: private holders can trade upward step-ups, but cannot assume commensurate downside liquidity.

10. Credit markets carry sharper warnings than NVIDIA sellers

Peter Thiel’s reported $100 million NVIDIA sale was below 1% of an estimated $10–20 billion net worth. It registered as a minor negative—people rarely sell stocks they expect to rise—but not a wholesale exit; SoftBank’s sale was less bearish because proceeds were being rotated into riskier OpenAI exposure.
Tomasz instead watched Oracle credit-default swaps trading around three times Amazon, Microsoft and other peers. Absolute default probability remained low, but the move showed creditors repricing the debt financing behind Oracle’s data-center commitments to OpenAI.
Rory linked debt and equity signals: the market-cap gain created by Oracle’s announced deal had completely unwound, leaving the core company below its pre-announcement value. Equity investors were discounting a risky contract while creditors demanded more compensation to insure the associated borrowing.
Other marginal warnings included record auto-loan delinquencies among subprime borrowers over the cited 60-day period, Blue Owl freezing redemptions in one non-traded BDC vehicle while moving it into another, and the First Brands default. None was a “big screaming flag,” but together they showed risk perception spreading.

11. AI infrastructure is running above its comfort speed

Tomasz said data-center capital expenditure was moving from roughly $500 billion annually toward $800 billion or more, amid circularity questions around a cited $15 billion investment involving Microsoft, NVIDIA and Anthropic. Yet hyperscaler GPU capacity remained sold out for two years, and their debt was small relative to free cash flow.
NVIDIA’s concentration was the harder structural concern: Tomasz stated that two customers represented more than 40% of revenue and that 4% represented more than 50%. He calculated that concentration at roughly 10 times Lucent’s during the dot-com period, although NVIDIA’s largest customers—Google, Meta and peers—generate ample cash and can stop spending voluntarily.
The trigger would be an inference-demand shortfall: if a hyperscaler built capacity and could fill only 80%, investors would question every unfinished data center behind it. “If there’s some wobble, the magnitude of the correction will be fast and brutal”; the economy is traveling “1,000 miles an hour on a car that’s designed to go 999.”
Rory’s upside case was physical constraint. If unavailable power prevents another ten data centers from connecting, spending can slow without anyone admitting demand vanished; that is less destabilizing than opening a completed facility where “nobody came” and immediately impairing the other 20 in development.

12. A secular AI boom can still contain repeated 30–40% drawdowns

Rory began with “Zero” when asked about three or four years of uninterrupted smooth sailing, then floated “maybe 10%, 20%.” Jason recalled SaaS falling 30–40% in roughly two weeks during 2016 and expected several similar corrections en route to a world where “data centers are the new cities.”
The long-term chart can conceal intolerable holding periods. Harry noted the Nasdaq’s 2001–02 decline was roughly 70–80% and took 16 years to recover; he said anyone nauseated by a 4–5% move should reassess asset allocation and that he was adding some cash. Rory’s advice was not to look when scared.
Rory repeated the behavioral advice, “If you’re scared, don’t look.” Early-stage founders appeared to follow it: Harry described YC companies raising $5 million, immediately opening subsequent notes and treating $50 million post-money as standard despite public-market weakness.
Rory saw no contradiction in founders exploiting plentiful capital just as VCs become hard-nosed when money is scarce. The character test is interpersonal—“life is long”—but the market was indisputably pro-entrepreneur, making nostalgia for a different balance of power economically irrelevant.

13. Venture’s path to $500 billion depends on a few correlated winners

Tomasz put US venture deployment around $270–275 billion today, versus $8 billion in 2008 and roughly $300 billion in 2021, and asked whether it reaches $500 billion by 2030. Rory added the missing cycle: the industry had already reached $100 billion in 1999 before collapsing to $8 billion.
Rory cited Excel’s GlobalScape estimate of $184 billion invested this year versus $183 billion at the 2021 peak, but roughly half went into four companies. Outside those names, conditions remained closer to 2020, producing a bimodal market of major AI companies and select accelerators against approximately 900 unicorns with no obvious IPO or private-equity buyer.
Rory reduced the industry forecast to returns: capital will keep arriving until excess money kills them. Because four or five companies represent roughly 40% of the industry, the pooled outcome increasingly depends on OpenAI, Anthropic, xAI, SpaceX and their peers overwhelming failures elsewhere.
Harry emphasized the downstream multiplication through LP portfolios and SPVs, reaching “dentists” and thousands of indirect holders. Rory’s summary was stark: “The bet is on, and the bet is singular and utterly correlated.”

14. GC AI shows how to pay up without underwriting burn

Rory said his firm discovered GC AI while referencing another legal-technology company. Customers consistently knew and liked the product, adoption was strong and barriers were low because it addressed daily work for in-house legal teams rather than corporate law firms. GC AI raised from Scale at a $550 million post-money valuation.
Comfort came from efficiency as much as growth. GC AI was profitable, had not spent its previous round and used an elegant, demand-led distribution strategy; Rory’s rule was to avoid combining a high entry price with high burn.
He also rejected venture “kingmaking” as dispositive. A fashionable firm can improve recruiting and visibility, but corporate buyers do not purchase bad software because Sequoia invested: “The customers decide,” and sustained customer love can overcome a rival’s financing advantage.
The quickfire reinforced the boundary. Tomasz and Jason preferred Legora at $2 billion to Harvey at $8 billion because the category’s $30 billion outcome remained uncertain—precisely where Rory believes entry price must still govern underwriting.

15. An access premium gives Stripe little reason to list

Stripe’s tender at an all-time high of $41 illustrated Tomasz’s “new public market.” He compared roughly $1 million of legal expense for a late private round with $25–30 million to go public, including the traditional 6–7% fee on a $200–300 million offering.
The old rule assigned private companies a 20–30% illiquidity discount to public multiples. Harry and Tomasz argued it may have inverted into a 20–30% “access premium”: coveted private shares cost more, while issuers receive cheaper capital, fewer transaction costs and no quarterly-earnings burden.
Rory limited that privilege to a small group with persistent Silicon Valley cachet—Stripe, the leading AI models and similar names. Most merely good cloud companies still need public markets because private buyers will not supply repeated $200–300 million financings and employee tenders indefinitely.
Tomasz guessed OpenAI could list in Q3 2026; Rory said Q3 or Q4 2026, while Jason expected alternative financing to push it into mid-2027. Tomasz argued that if he had no shares, he would not be dilution-sensitive; Jason found it bizarre for a leader to be motivated by “world domination” rather than conventional economic incentives.

16. Retail and secondaries are constructing a private public market

Tomasz’s retail pathway runs from a 401(k) into an ETF, then a fund of funds and ultimately venture. Harry cited Coatue’s $3 billion in retail funds and GC’s expansion efforts, arguing the money could arrive within 24–36 months while poor-return recognition takes five to seven years.
The cautionary analogue was Blackstone’s roughly $21 billion retail real-estate vehicle and its redemption problems. Venture marks may remain untouched for 12–18 months or longer, so liquid retail liabilities can sit against assets whose weakening values remain invisible.
Secondaries are already expanding: Tomasz put them at roughly 25% of private-equity dollars, historically 2–3% of venture and now 10–12%. He preferred measuring total liquidity across IPOs, M&A and secondaries rather than treating a weak IPO year as proof that exits disappeared.
For a fund-returner with a 15-year path to liquidity, Tomasz described incremental selling: a venture firm sells a quarter several rounds later, then additional pieces at subsequent marks—“dollar cost my way out.” Rory accepted that private investors are learning to behave like public shareholders through a less efficient venue.

17. The IPO debate turns on fees, retail demand and the next crash

Tomasz saw Goldman’s purchase of Industry Ventures, reportedly at an exceptionally high asset-manager multiple, as evidence that incoming retail money needs secondary exposure. A market-clearing price should emerge for companies ranked roughly two through 200, while the 900 stranded unicorns create a restructuring or buyout business somewhere between zero and their old marks.
Rory still believes the largest venture exits ultimately need IPOs because venture returns depend on a few genuinely extraordinary companies, not private-equity-style packaging of consistently adequate assets. He also noted private capital’s aggregate two-and-20 burden versus roughly 60 basis points in public markets.
Tomasz’s answer was fee compression: late-stage retail products could approach the 65–75-basis-point load of publicly traded private-equity managers. He also cited private equity taking 12% of publicly traded software companies private in 2022; if there are only eight IPOs, publicly traded software risks becoming “a dying breed.”
The unresolved variable is a real down market. Until one removes private liquidity, every trend favors recurring tenders and fewer listings; afterward, holders may rediscover why public markets exist. Rory’s warning to future retail managers: eventually they may spend ten annual meetings explaining why “you’ve made a ton of money and they’ve lost.”

Escaping AI Slop: How Atlassian Gives AI Teammates Taste, Knowledge, & Workflows, w- Sherif Mansour No Priors Ep. 141 | With Sunday Robotics Co-Founders Tony Zhao and Cheng Chi