Peter Thiel and Softbank Sell NVIDIA - Why? & Why VC Will Hit $1TRN and The Opening of Retail
Summary
Cursor’s $29.3 billion valuation works only if AI coding becomes universal infrastructure, not merely a productivity tool. Max reframed today’s 30%-70% developer uplift as tomorrow’s default workflow: 100 million-plus developers paying perhaps $5,000 annually creates a theoretical $500 billion-plus market. Against revenue reportedly moving from roughly $100 million to $1 billion, even the headline valuation can resemble “a classic bull-market bet” at about 10x prospective revenue.
The investment case turns on durability and free cash flow because demand and growth are already conceded. Cursor’s supplier is also its competitor, with model tokens representing perhaps 50%-70% of the product, yet distillation, proprietary models and hardware efficiency could lift margins toward 60%. That may be enough for a self-serve product with little sales-and-marketing expense: “Do they need to” reach legacy SaaS margins of 70%-plus?
AI-coding market share may soon congeal, but the panel split sharply over whether technological change still prevents lock-in. Tomasz expects memory, personalized tooling and enterprise standardization to let Cursor retain perhaps 75% of today’s users five years out; Harry argued for a familiar three-to-four-year land grab followed by stable shares. Max countered with Replit V3’s multi-agent architecture and months-long context: software is improving “20 or 30x faster than 24 months ago,” so another 30-person team could still reset the market.
A genuine AI price war—not merely more tokens for the same dollar—is the most frightening downside scenario. Portable prompts and thinner integrations could let third- or fourth-place vendors cut a $100,000 agent to $20,000 or even $2,000, turning software from Salesforce-like infrastructure into DRAM-like commodity supply with 50%-80% price swings. “To say that would be ugly would be an understatement. It would be terrifying.”
Late-stage venture currently looks effortless because marks rise quickly, but its liquidity is one-way. Ramp reportedly financed four times in a year and moved from $13 billion to $32 billion, while 15% of Q1’s newly minted unicorns had already stepped up by Q3. Max’s warning was that private investors can trade from $60 billion to $180 billion on the way up, but “you can’t execute a trading strategy” when the downside arrives and buyers disappear.
The clearest AI-cycle warning is not insider stock sales but leverage, customer concentration and inference utilization. Oracle credit-default protection repriced to roughly three times hyperscaler levels, data-center capex was framed as rising from $500 billion toward $800 billion annually, and Nvidia’s top two customers represent more than 40% of revenue. Capacity is still sold out and hyperscalers remain cash-rich, but one data center filling only 80% could trigger a “fast and brutal” correction across leveraged suppliers.
US venture could reach $500 billion by 2030, but increasingly as one correlated bet financed by retail capital and recycled through secondaries. Roughly half of a cited $184 billion annual total went into four companies; if OpenAI, Anthropic, xAI and SpaceX produce the hoped-for returns, those gains can swamp dozens of failed unicorns. The governing valuation rule emerged in quick fire: “Entry price counts when TAM is unclear. Winning is the only thing that counts when TAM is huge.”
Deep dive
1. Cursor’s valuation assumes AI coding becomes the default interface
The opening bull case paired Cursor’s $2.3 billion financing at a $29.3 billion valuation with unusually strong product-market fit: coding may be AI’s best application after search, developer productivity is reportedly up 30%-70%, and Cursor’s new model was described as four to five times faster in tokens per second.
Max argued that “productivity boost” is already the wrong frame. At SaaStr in May, developers still debated the uplift from Cursor and Windsurf; now he knows virtually nobody who codes without an AI tool. His endpoint is “100% penetration per developer,” potentially at $5,000-$6,000 per year.
The discussion moved from Microsoft’s recent discussion of 100 million-150 million developers on GitHub toward a possible 200 million global developers. Multiplying even 100 million users by $5,000 produces $500 billion; the 200-million scenario implies a theoretical $1 trillion market.
Tomasz pushed back on the largest number and narrowed the calculation to full-time professional developers. Even three million-four million serious US developers paying $5,000 annually still supports an enormous company. The panel also noted that the GDP comparison depends on whether the market is global or US-based.
2. Revenue velocity matters more than the headline multiple
Tomasz illustrated the revenue case with a hypothetical progression from roughly $1 million to $100 million, then from $100 million to $1 billion. If momentum carries revenue to $3 billion-$4 billion next year, the $29.3 billion valuation suddenly looks closer to 10x next-twelve-month revenue than an untethered speculative mark.
Harry described an agentic software seller whose supposedly mid-market contracts were high six figures to low seven figures. That is not conventional SaaS pricing; it reflects labor replacement and therefore “massive TAM expansion.” Max added that without such expansion, “there’s just no point in even playing as VCs.”
A separate creator market sits beyond professional engineers. Harry said he had shipped 12 Replit apps since June, used 700,000 times, despite building products without coding. The immediate bear case is retention: comparable vibe-coding products showed gross account retention around 50%, making enterprise penetration and standardization decisive.
3. Margin expansion depends on escaping the supplier-competitor trap
Cursor’s unusual platform risk is that its direct competitors also supply the models constituting perhaps 50%-70% of its product. Rory described the situation as a supplier-competitor problem. With lean staffing, labor is not the central cost; token payments flowing toward Anthropic and other model providers drive both the profitability question and the durability question.
Harry contrasted that structure with Replit and Lovable, which can default users to cheaper or N-minus-one models that work adequately for their use case. He put their gross margins north of 50% and asked the unanswered question: even after blending in its own model, “How do they get to 60% gross margins?”
Tomasz sees considerable architectural fat available for removal. His firm distilled tool-calling from a roughly trillion-parameter teacher into a 20-billion-parameter model and achieved 97% equivalence—despite being a venture firm, not a research laboratory. Rory likewise noted that Microsoft reported producing 90% more tokens per GPU-hour than 12 months earlier.
4. Legacy SaaS margins are unnecessary if distribution stays cheap
The panel did not assume AI applications will recover the 70%-72% gross margins of publicly traded workflow software. Its counterpoint, made by Harry, was that those businesses also carried large sales organizations and integration costs; a self-serve coding tool can generate attractive free cash flow at materially lower gross margin.
Harry’s conclusion was conditional but bullish: if users prove durable, a company selling billions of dollars of revenue with roughly 60% gross margin and only around 100 employees can throw off substantial cash. If digital optimization is the remaining obstacle between Cursor and a $50 billion-$60 billion outcome, there should be room to solve it.
Tomasz’s own behavior illustrated willingness to pay. His $200-per-month Claude Code Max allowance ran out two days into the week, leaving him considering several seats and perhaps $1,000 monthly despite the inconvenience of rotating keys. “I will never go back to using a computer without Claude Code.”
5. Personal memory and enterprise standards could freeze market share
Tomasz’s switching model has two regimes. When coding performance jumps sharply—as he said Gemini 3 had edged past Claude 4 Sonnet—users experiment. As model improvements asymptote, memory, coding conventions and personalized tools matter more, and only a significant advantage justifies migration.
His own Claude Code environment contains about 100 tools written by Claude, plus accumulated preferences such as linting and indentation. He asked Gemini to migrate that environment, but at his own expense; once Fortune 500 companies standardize on one product through an enterprise agreement, organizational switching costs become larger still.
On that basis, Tomasz guessed Cursor might retain roughly 75% of its current audience five years from now. His likely end-state was Cursor at 40%-60%, Microsoft eventually second through VS Code and bundling, and Anthropic remaining important because coding appears central to its model strategy.
Rory largely agreed but excluded OpenAI’s Codex from his improvised top three, favoring Cursor, Microsoft/GitHub and Anthropic, with Cognition as a differentiated possibility. That challenged Harry’s initial hypothetical split of Codex 60%, Anthropic 20% and Cursor 20%: existing distribution may matter more than model ownership alone.
6. Replit V3 shows why the skillet may stay hot
Max’s counterexample was Replit V3, which he called beyond “night and day”—“Pluto and Mercury.” Its agents call an architect, ask separate agents to find bugs and review work, and retain what appears to be months of context. As coding autonomy improves, functional QA becomes the new constraint and another potential 10x unlock.
Max argued that software has never improved remotely this quickly in his lifetime. Harry supplied the sharper comparison: the current pace looks “20 or 30x faster than 24 months ago.” If 30 people built Cursor, another 30-person team may still produce a discontinuity within 18 months.
Harry defended the conventional pattern: markets remain fluid for three or four years, then shares hold for 10 or 20 years even as the category expands. Intel’s processors improved rapidly without constantly reshuffling CPU market share. Jason called the timing question the “bacon-and-the-skillet” debate: when the heat comes off, the fat congeals.
7. Portable prompts make application moats thinner than they appear
Max’s Salesforce experiment weakened the lock-in case. SaaStr moved a prompt developed over months in another AI agent into Agentforce, iterated for roughly one day, and achieved comparable performance. “Don’t overestimate your moats today”—much of the apparent advantage may be portable meta-learning.
Harry distinguished two forms of deflation. The benign version gives a customer two million tokens next year for today’s price; spending remains stable while value rises. The dangerous version is explicit erosion, perhaps Anthropic cutting Claude Code from $100 to $50 to gain share and forcing competitors to respond.
Tomasz expects pressure to originate with players ranked third through fifth, which need share and can underprice leaders. Jason pushed back that cheap CRMs never prevented Salesforce reaching massive scale, but conceded AI portability makes the threat stronger: a customer might keep the same intelligence while replacing a $300 core seat with a $5 system.
8. A price war would turn software economics into semiconductor economics
Rory placed the risk on a spectrum. Salesforce, embedded through many integrations, is nearly immune to cheaper substitutes; commodity DRAM buyers remain loyal for “30 seconds,” while oversupply can drive prices down 50%-80%. If AI or GPUs acquired DRAM-like economics, “terrifying” would understate the damage.
The hinge is abstraction. Harry invoked Iceberg’s separation of storage from Snowflake compute: enterprises reclaimed control of data and selectively granted access. A comparable prompt database could route institutional learning among interchangeable agents, preserving customer value while stripping vendors of lock-in.
SaaStr already interacts with Agentforce, Qualified, Artisan and other agents rather than logging into Salesforce directly. Max called Salesforce increasingly “a database”; unless incumbent applications win the agent layer, logos may remain while value “slowly leaks out every week,” leaving growth and market capitalization structurally lower.
Current go-to-market agents still cost about $50,000-$70,000 plus roughly $25,000 for forward-deployed support, or approximately $100,000 to start. A future move from $100,000-$200,000 to $20,000—or $2,000—would deflate today’s spectacular ARR. Historically, Harry said, the best retention predictor was integration count: easy-to-remove software gets removed.
9. Late-stage venture is behaving like an illiquid trading market
Max said Cursor had completed at least three rounds during the year, while Ramp reportedly completed four and moved from $13 billion to $32 billion. Of roughly 24 unicorns minted in Q1, 15% had already received a step-up by Q3, some twice—compressing a traditional 12-to-18-month financing cycle into months.
Harry questioned whether his insertion point was fundamentally wrong. With media-driven access, he could write $10 million-$25 million checks into obvious high-flyers and capture rapid marks, yet instead chose “the craftsmanship of seed” and company-building in the trenches. Max responded, “Why do I do that?”
Established early-stage firms making enormous later bets reinforced the question: Harry cited Bessemer backing Anthropic and co-leading Ramp around $32 billion, along with other major firms entering Anthropic at vast scale. Max’s maxim was stark: “The late-stage business is either the best business in the world or the worst business in the world.”
Max described a billionaire managing the category as a ruthless book—buying at $60 billion and selling at $180 billion within a year. Max’s correction was that private markets mimic public trading only on the way up. When prices fall, the liquidity needed to exit will not be there; “you can’t execute a trading strategy” on the downside.
10. Credit and utilization—not insider sales—are the cycle’s tells
Max characterized Peter Thiel’s reported $100 million Nvidia sale as below 1% of an estimated $10 billion-$20 billion fortune: modest evidence that he preferred selling to holding, not wholesale capitulation. SoftBank’s Nvidia exit was even less defensive because the proceeds were being recycled from a profitable public chipmaker into OpenAI.
Max instead watched Oracle credit-default swaps rise to roughly three times Amazon and Microsoft levels within several days. Absolute default probability remained small, but the move repriced risk around debt-funded data centers supporting Oracle’s OpenAI agreement.
Harry added that the equity value associated with Oracle’s deal had unwound: the market cap of the core company was below where it had been when the deal was announced. Other warnings included record subprime auto-loan delinquency among cited borrowers over the prior 60 days, frozen redemptions in a Blue Owl non-traded vehicle, and the First Brands private-credit default.
AI data-center capex was described as moving from $500 billion toward $800 billion annually amid increasingly circular financing arrangements. Nvidia’s concentration is the larger structural issue: Max said two customers represented more than 40% of revenue and four more than 50%, roughly 10 times Lucent’s concentration in the dot-com era. The mitigating fact is that customers such as Google and Meta generate ample cash and can stop spending whenever economics weaken.
11. One underfilled data center could produce a fast, brutal correction
The merry-go-round stops when inference demand disappoints. Max said that if a hyperscaler builds capacity and fills only 80%, investors will immediately question the other data centers under construction. For now, GPU capacity is reportedly sold out for two years and hyperscalers continue asking for more, so weakness remains at the leveraged margins.
Max’s image was a tachometer at redline: the economy is traveling “a thousand miles an hour on a car that’s designed to go 999.” When GPU depreciation assumptions can move the entire US equity market, even mild deceleration becomes painful; any wobble could make the correction “fast and brutal.”
Harry’s upside scenario was physical scarcity: limited power connections may prevent the industry from overbuilding. Companies can say they would have built 10 more centers but lacked electricity, allowing supply growth to slow gradually without the catastrophic admission that a newly opened facility attracted no demand.
12. Corrections are expected even if the secular thesis survives
Asked for the probability of smooth sailing over the next three to four years, Max answered “zero,” then softened that to perhaps 10%-20%. He recalled SaaS falling 30%-40% in two weeks during 2016 before recovering; rapid AI progress makes several such corrections more plausible, not less.
Harry distinguished eventual recovery from the experience of living through it. The Nasdaq’s 2001-02 decline was roughly 70%-80%, and regaining the old level took about 16 years. A long horizon helps, but anyone nauseated by a 4%-5% fall should reconsider asset allocation before a genuine drawdown arrives.
Harry suggested holding more cash if current volatility already felt unbearable; Max’s veteran response was, “If you’re scared, don’t look.” Their joking synthesis—more agents than humans behind a “white GPU fence”—carried the macro caveat that agents do not pay car loans.
13. Venture’s path to $500 billion is narrow and highly concentrated
YC showed no fear from public-market weakness. Harry described founders raising $5 million and immediately opening “the next note on the note,” often near a $50 million post-money valuation, while treating investor meetings as auditions. Harry’s response was that this is rational in a capital-rich, entrepreneur-friendly market, although interpersonal behavior still matters because “life is long.”
Asked about the US venture market reaching $500 billion by 2030, Max answered “100% chance.” Harry supplied the historical series from roughly $8 billion in 2008 to $300 billion in 2021 and about $275 billion today; Max supplied the missing cyclicality—approximately $100 billion in 1999 collapsing to $8 billion several years later.
An Axial estimate cited by Harry and Tomasz put annual investment at $184 billion versus $183 billion in 2021, yet roughly half went into four companies. The remainder was described as about half of 2021 and consistent with 2020. Thus AI megadeals and YC can be overheated while the broader venture market remains far below 2021.
The industry’s doubling therefore reduces to the outcomes of OpenAI, Anthropic, xAI, SpaceX and a handful of peers. A $40 billion gain in one can swamp 40 failed unicorns. Max emphasized that the downstream beneficiaries include many LPs and smaller investors. The bet is “singular and utterly correlated.”
14. Retail capital can arrive years before poor returns become visible
The next supply frontier is retirement and retail money routed through ETFs, funds of funds and venture managers. Harry cited Coatue with roughly $3 billion in retail-oriented funds and described General Catalyst as moving aggressively; that flow could arrive within 24-36 months even though venture performance takes five to seven years to diagnose.
Max called it a “tsunami of retail capital” and compared the mismatch with Blackstone’s roughly $21 billion real-estate vehicle and its redemption problems. Private marks can remain unchanged for 12-18 months or longer, delaying the feedback that would normally restrain capital.
Max supplied the moral endpoint: raising giant funds feels delightful until managers must tell investors their money is gone. He recalled closing his own failed company and facing backers directly. Locking retail investors into subpar decade-long returns while managers collect fees would make the eventual annual meetings miserable.
15. GC AI shows how cash efficiency can justify leaning on price
Rory did not expect to lead GC AI’s financing at a stated $550 million post-money valuation. References for another legal-AI company repeatedly produced unsolicited customer enthusiasm for GC AI: strong adoption, low barriers and a product designed around daily work for in-house general counsel rather than outside corporate-law firms.
The discussion emphasized simple fundamentals: customers liked the product, the team was strong, growth was fast, and the business had barely spent its prior round while remaining profitable. The broader discipline is to avoid combining a high entry price with high burn; demand-led, cash-flow-positive companies provide more protection in a downturn.
Harry worried that future financing partners were already committed to Harvey or Legora. Jason viewed the markets as distinct and expected no sequence of giant rounds. He also rejected deterministic “kingmaking”: corporate buyers do not purchase bad software because Sequoia funded it—“the customers decide.”
16. Private access has inverted the old illiquidity discount
Stripe’s tender at an all-time-high $41 per share prompted Tomasz to call private markets “a new public market.” Earlier Microsoft-era IPOs reportedly required about $50 million of trailing revenue and six profitable quarters; Harry said an IPO can consume $25 million-$30 million, or 6%-7% of a $200 million-$300 million raise, versus roughly $1 million for a late-stage round.
Investors were once taught that private companies deserved a 20%-30% illiquidity discount. Scarcity has created an “access premium” that may be 20%-30% in the other direction: elite companies obtain cheaper private capital, recurring tenders and freedom from quarterly reporting.
Tom Loverro limited that privilege to a small set of highly desired companies. Navan and ServiceTitan could not indefinitely raise private rounds or run hundreds of millions in employee liquidity, so public markets became their lowest-cost capital. Max added that they are perfectly good businesses but lacked the access premium. Your dentist wants Stripe or SpaceX exposure, not every merely good cloud company.
Harry argued that retail capital could keep more companies private even if they were not top-tier names. Max agreed conditionally: if returns remain high, more capital will arrive, but when returns decline, the accumulated inflow will make the correction worse. The question is how long “in the end” takes.
17. Secondaries may turn venture holdings into synthetic public stocks
Max called the year’s IPO finish a whimper, citing troubled performances from StubHub and Navan despite Cursor reaching $30 billion in 22 months. Tom Loverro said IPO count is now the wrong lens: venture secondaries have risen from roughly 2%-3% of the asset class toward 10%-12%, versus about 25% in private equity.
Harry cited Goldman’s acquisition of Industry Ventures at what he described as an exceptionally high asset-manager multiple as evidence of demand for the infrastructure. He expects a secondary market not merely for names one through 20 but perhaps 20 through 200, creating a clearing price for some of the roughly 900 unicorns unlikely to IPO.
Tomasz’s model is gradual liquidation: after finding a fund-returning company that may need 15-20 years to mature, sell a quarter several rounds later, then more in subsequent rounds—“dollar-cost-average my way out.” The asset behaves like a public stock, but through periodic private transactions and a narrower buyer base.
Max still expects the biggest outcomes to list and argued that private ownership carries aggregate 2-and-20 fee drag versus roughly 60 basis points publicly. Harry said late-stage retail fees could compress substantially. Tom Loverro noted that private-equity funds can operate around 65-75 basis points, while Tomasz cited PE’s 2022 take-privates and scarce IPOs as evidence that the public software universe may keep shrinking.
18. The final valuation rule separates uncertain markets from obvious ones
Asked to choose Cognition at $12 billion or Cursor at $29 billion, Tomasz, Rory and Jason all chose Cursor despite usually favoring the cheaper asset. Cursor’s revenue and category scale were “godstopping” enough that the winner mattered more than the lower entry price.
Asked to choose Harvey at $8 billion or Legora at $2 billion, Tomasz instead chose Lovable, Rory seconded him, and Jason also backed Lovable. Jason said he could not yet see the $30 billion legal-AI exit needed to justify Harvey’s mark. The resulting rule was: “Entry price counts when TAM is unclear. Winning is the only thing that counts when TAM is huge.”
OpenAI IPO guesses included Q3 2026 from Tomasz and Q3-Q4 2026 from Rory. Tomasz then said alternative financing could push it into mid-2027, while Jason discussed a possible government guarantee.
The governance wrinkle is incentive alignment: Tomasz said a leader with little or no equity would not be dilution-sensitive, so the button becomes growth or “world domination.” Jason’s counterexample was a turnaround CEO guaranteed 7% fully diluted through IPO—an arrangement that visibly changed behavior.