Summary

Tesla’s $1 trillion compensation headline is a board-level double-or-quits bet on Elon Musk, not a generic template for CEO pay. Maximum vesting requires an $8 trillion market cap, $400 billion of EBITDA—four times Google’s cited $100 billion—20 million cars, 10 million FSD, 1 million Optimus robots, and 1 million robotaxis. If Tesla is 25% car company and 75% “Elon’s special sauce,” refusing him could erase 75% of the equity value overnight, leaving the board in a “prisoner’s dilemma.”
The Scale AI and Windsurf transactions expose a shift from missionary founders to mercenary liquidity, even when the economics are defensible. Harry argues both offers exceeded the companies’ likely standalone value and employees could have been made whole; a panelist notes the returns ultimately fund universities, hospitals, and foundations. Jeff’s harsher reading is that the buyers “eviscerated the brains and the heart” while leaving a nominal company “dead as the dodo.”
Ramp at $1 billion of ARR and Brex at $700 million, growing 50%, show AI capital cascading through the economy—not every software company suddenly working. Their interchange and credit economics permit rapid revenue growth but carry financial-services margins; once growth converges with Amex, Rory expects Amex-like valuation. The sharper test for B2B vendors: with OpenAI, Anthropic, and their ecosystems spending so heavily, anyone capturing no benefit “gets an F.”
Sierra’s $10 billion valuation on $100 million of ARR is the cleanest expression of today’s great-company, terrifying-price trade. AI customer service can plausibly automate 70–80% of work, Sierra is among the high-end leaders, and Bret Taylor is treated as a generational operator; the only unchecked box is whether investors are paid for valuation risk. At 100 times ARR and a prospective check near 10% of Greenoaks’ cited $2.75 billion fund, “close your eyes” underwriting leaves little room for execution merely to be good.
Venture capital has migrated toward confirmed AI winners because their risk-adjusted returns can look easier than seed investing. Kleiner Perkins putting $100 million into Anthropic’s $13 billion round at a $183 billion valuation was called both a “logo deal” and rational. The panel’s blunt taxonomy is now roughly “20% old-school venture capital and 80%” late-stage growth investing, with valuation risk expanding after other risks disappear.
AI threatens the seat-based SaaS model more fundamentally than copilots suggest. Incumbents want to promise humans will become 10% more efficient; customers will increasingly ask, “I don’t need 75% of these people anymore—give me that product.” Pure plays can replace both software and some labor spend, while Salesforce, Atlassian, and other incumbents must cannibalize seats and revenue before attackers do it for them.
The best developer businesses sell authority or complexity that developers cannot—or economically should not—recreate. Jeff’s three durable categories are “business development as a service” such as Twilio and Stripe, “capex as a service” such as AWS, and exceptionally difficult “algorithm as a service” such as DynamoDB or frontier-model training. Inference alone may commoditize through open-source models; the defensible product is the model whose “secret recipe” cost billions to train.
The boom’s weak control layer is diligence, not capital availability. Jason says rounds now close on Saturdays and “diligence isn’t even being attempted,” arguing prison should deter fabricated revenue and unpaid pilots presented as paid; Rory distinguishes aggressive presentation from forged evidence but agrees categorical lies warrant severe consequences. Jeff’s counterweight is equally important: when investors are too eager to verify users or cash, “their greed…gets rewarded with some amount of fraud.”

Deep dive

1. Tesla’s board bought the Elon bet at an unprecedented price

Rory’s governing principle is that “compensation is how boards reveal their real priorities.” After reading roughly 150 of the 3,322 proxy pages, he saw three priorities: honor Musk’s disallowed 2018 award, give him another 12% because the board feared he might leave, and pay him to double down again.
The hurdles make the trillion-dollar headline less arbitrary: an $8 trillion market cap, $400 billion in EBITDA versus Google’s cited $100 billion, 20 million cumulative cars from roughly 10 million already, 10 million FSD, 1 million Optimus robots, and 1 million robotaxis. Tesla must double its existing business and create major new ones.
Rory called the package intellectually coherent even while finding the wager terrifying: “I will give you a shit ton of money if you take our trillion-dollar company and double or quintuple it.” He expects a high-water mark that lifts other founders’ aspirations, not a universal standard—unless their boards consist of “their brother-in-law and other relatives.”
Jeff’s pushback came from compensation design: as a founder he already had ample equity, and every extra lever creates another way for pay to feel unfair. Jason nevertheless sees post-unicorn founders receiving 7–8% or more, tied to $10 billion, $20 billion, or $100 billion outcomes rather than the old 2–3% refresh.

2. Tesla’s personality premium leaves its board in a trap

Jeff reframed the award as downside protection: perhaps Tesla is overvalued as a carmaker and the personality cult supplies the rest. Rory accepted the premise—“maybe this is a car company worth 25% of our current market cap and then Elon’s secret sauce…is the other 75.”
The alternative would be to abandon robots and autonomy, appoint an operator for a great car company, and accept the repricing. Rory expects the stock could fall 75% the next morning, followed by endless litigation; shareholders have already voted for Musk’s prior package twice, signaling that they want the risk-on future.
That makes bargaining resolve almost impossible. The panel’s theory is that a person with three other equally exciting alternatives may walk; if the board refuses and he follows through, the directors crystallize the downside they were trying to avoid. Rory’s verdict: “It really is a prisoner’s dilemma.”

3. Missionary founders are giving way to mercenary exits

Jeff expects a founder CEO to be “the most committed, the most long-term oriented, and the most visionary.” If money is the objective, his probability-adjusted advice is to join a hyperscaler; the defensible reason to found a company is believing “the world needs to have the thing you’re building.”
Jason sees that ethic fading amid employees seeking $10 million of liquidity after 12 months and founders leaving for Meta or OpenAI. The transition from zero to $10 billion in roughly 18 months is not the Twilio or StubHub grind: teams may still be strong, but “they’re mercenaries.”
Harry defended the Scale AI and Windsurf founders on price. The attempted acquisitions were, in his view, well above present value and probably any future value; antitrust drove awkward structures, but the residual employees—perhaps under 5% of ownership—could have been paid as though acquired without breaking the economics.
A panelist noted that Scale’s venture returns flowed to institutions including the University of Wisconsin or Michigan, the Cystic Fibrosis Foundation, and Children’s Hospital of Atlanta. Yet Jeff’s objection survived: the buyer “eviscerated the brains and the heart,” leaving a legal carcass everyone calls independent although it is “dead as the dodo.”

4. Ramp and Brex are monetizing the boom, not proving every tide is rising

Ramp reaching $1 billion of ARR and Brex reaching $700 million while growing 50% are strong company-specific results, not evidence that all private technology is booming. Brex regrouped after a rough patch; both can grow quickly by extending credit and collecting interchange rather than selling pure software.
Rory places them between categories: financial-services margin structure and risk, but software-like growth. The governing valuation principle remains risk-adjusted free cash flow; absent that, growth supports the premium. “Once the growth rate slows,” if they grow like Amex, “they will be valued the same as Amex.”
Jason sees AI spending spreading from OpenAI and Anthropic through Broadcom, Cisco, Twilio, and the broader B2B stack. His test is merciless: if a vendor has gained no AI customer, workload, or spending tailwind, “you get an F.” The money is “fish food at the top” now drifting into darker parts of the ocean.
Jeff’s infrastructure analogy came from the mobile boom: startups could fail after paying Twilio millions. That is the attraction and risk of supplying a boom—providers monetize experimentation regardless of the ultimate winner, but must replace the revenue when failed customers disappear.

5. Sierra clears every underwriting hurdle except price

Sierra’s reported $10 billion valuation on $100 million of ARR—roughly 100 times revenue—buys a scarce combination: Bret Taylor, a strong team, a large category, and a leading position at the enterprise end. Jason joked that if Scale was worth $28 billion, “Bret’s got to be worth $56 billion.”
Rory’s deal sequence is explicit: can the category support a major winner; will this company be among those winners; and are investors paid for the risk? Customer support passes the first test because AI can handle perhaps 70–80% of calls and emails, converting a large cost center into an obvious automation target.
Sierra appears to pass the second test through its high-end positioning and Taylor’s ability to move between technology and business. The failure is the last box: at 100 times ARR, investors rely on the market becoming so large that even a $20–30 billion outcome produces an acceptable, if low, return.
Rory noted the opportunity cost: he estimated Greenoaks’ next check near $275 million, around 10% of a cited $2.75 billion fund. Jeff supplied the comparison VCs rarely make—an investor calls 10% concentration bold, while the founder puts “100% of my capital allocation” across time, career, and personal finances into one company.

6. Late-stage AI is swallowing the venture-capital identity

Kleiner Perkins reportedly put $100 million into Anthropic’s $13 billion round at a $183 billion valuation, its first model-provider investment. A panelist called it a logo deal: a large fund cannot meet founders without Anthropic or OpenAI on its website. Rory countered that the arithmetic itself is not crazy.
Rory’s case is that putting in $100 million and potentially taking out $300 million may be an easier dollar than fighting beside a founder from seed. It is not the same business as early-stage venture, but it can still be a rational late-stage bet.
Jeff’s broader diagnosis: “Venture capital is 20% old-school venture capital and 80%, plus or minus, late-stage” investing that once resembled Fidelity growth. It may differ from what firms sold LPs four years ago, but it is where most dollars now go and where firms can remain visibly relevant.
The danger arrives after category and company risk seem settled. “Valuation risk expands to fill a vacuum”; extrapolation then becomes the whole underwriting model. Figma at $25 billion can perversely feel niche beside Canva, Databricks at $4 billion of revenue and 50% growth, Anthropic, and OpenAI.

7. OpenAI’s private liquidity is extraordinary mainly because it is visible

The $10 billion OpenAI secondary sounds unprecedented, but Rory normalized it using public-market arithmetic. At a hypothetical $500 billion valuation with management owning 20%, employees hold $100 billion; selling 10% of those holdings is rational diversification, not necessarily an abandonment of conviction.
Apple was worth roughly $800 billion in 2018, and equivalent insider sales at a public company would barely register. OpenAI also dates to 2016, so it is not conventionally young—although someone who joined two years ago and receives $10 million will experience it as exceptionally fast wealth.
The opportunity cost cuts both ways. Employees leaving Nvidia when it was worth roughly $500 billion may now regret converting $10 million that could have become $60 million. Anyone accepting Sam Altman’s long-duration vision must recognize that today’s extraordinary liquidity might still mean “leaving money on the table.”
Liquidity also warps recruiting: a B2B company posting “triple, triple, double, double” was S-tier 36 months ago but cannot match eight-figure AI payouts. Jeff’s counterexample was Domino’s strong technology operation in Ann Arbor and its better-than-Google cited ten-year stock return: recruit elsewhere instead of imitating Silicon Valley offices.

8. The public-market quickfire revealed how little conviction survives the numbers

With Figma at $52 and roughly a $25 billion market cap, Rory predicted the mid-$40s in 365 days. His logic was banker vindication: a $35 IPO price, a normal 10–15% pop, and a year of compounding would imply the mid-$40s once the opening volatility clears.
Jeff took the interest-rate bet and chose $75; Jason chose $60. Their range captures the unresolved question underneath late-stage underwriting—whether lower rates and renewed risk appetite overwhelm fundamentals, or whether the IPO bankers’ original price ultimately anchors the stock.
On a Canva IPO in Q4, the panel split between one “yes” and two versions of “no,” including “0%.” The bearish reasoning was procedural: by September 9 the calendar was tight, and Cliff’s recent willingness to discuss not pursuing a direct listing seemed unlikely if lawyers were preparing an imminent filing. First-half timing looked more plausible.

9. Anthropic’s author settlement draws a clearer line around training data

Rory read the $1.5 billion author payout as a coherent boundary: buying a $15 book and using it for model training was treated as permissible, while downloading a pirated corpus was not. The cited arithmetic was 500,000 books at $3,000 each—roughly 200 times the purchase price.
Jason resisted sanitizing the conduct: “This wasn’t cutting a little bit of a corner.” Anthropic downloaded from pirate sites to accelerate the model, making it a textbook case of asking forgiveness. He also cautioned that the matter might produce additional costs.
Rory suggested vendors may create legally compliant training corpora by purchasing and scanning books for each model company. Harder cases remain where a model does not merely learn from an artist or author but returns something effectively reproducing their work; then damages might extend well beyond the original $15 purchase.

10. ASML’s Mistral investment looks more sovereign than strategic

ASML becoming Mistral’s largest shareholder at a $14 billion valuation puzzled the panel. Rory located ASML several layers upstream: its enormous lithography machines enable TSMC’s fabrication and therefore much of the semiconductor industry, but the strategic need to own an LLM vendor is not obvious.
Jason offered a corporate-accounting explanation. Cash on a profitable company’s balance sheet can be difficult to deploy without depressing earnings; exchanging cash for an investment that avoids impairment may feel “basically free.” Corporate venture therefore needs business relevance and capital preservation, not necessarily venture-style return geometry.
Jeff’s test is whether the investment gives the core business a fundamental advantage, as Salesforce Ventures gained ecosystem position, acquisition intelligence, and product information. With ASML and Mistral, nobody could identify the equivalent mechanism; tying up roughly $1.5 billion is especially notable in highly cyclical semiconductor-equipment markets.
Rory’s best explanation was sovereign AI. Europe may treat models like defense equipment: national champions are economically suboptimal under free trade but become attractive when governments fear foreign suppliers could withhold capability. Anduril’s European subsidiaries, London visibility, and Australian acquisition illustrated how even an American vendor must localize to overcome that fear.

11. Atlassian’s acquisition spree reflects urgency more than certainty

Atlassian reportedly paid $21 million in cash for Cycle and $610 million for the Browser Company. Jeff did not find “we need a different browser for work” persuasive enough to change behavior, but the panel saw an understandable bet: Atlassian has an enormous knowledge-worker and developer footprint and needs a credible AI play.
The pressure itself can distort selection. The panel’s formulation was that when executives or investors feel “itchy,” they may choose the best of the three deals currently available rather than the ideal deal. Atlassian is a great company that has not yet enjoyed the AI boom, increasing the temptation to act.
Twilio’s own M&A urgency came from knowing SMS was already legacy technology. SendGrid, Segment, and Zipwhip were attempts to parlay a customer and revenue base into the next era before the bridge disappeared—closer to Intel moving from memory into CPUs than simply adding adjacent features.
Serial acquirers accept a loss ratio. The conventional regret is not buying a deal that later mattered, rather than buying one that failed; even a 30–40% chance that the Browser Company materially changes Atlassian could clear that bar. Jeff admitted one missed Twilio acquisition still bothers him, without naming it.

12. AI forces SaaS incumbents to sell the destruction of their own seats

Jeff’s categorical call was that AI will “decimate” SaaS seat bases because it performs jobs currently executed inside those products. A copilot promising to make each employee 10% more efficient preserves the incumbent model; the buyer’s preferred product says, “I don’t need 75% of these people anymore.”
Twilio had an unusual advantage because infrastructure usage was not priced per employee. Jeff saw AI as its opening to bypass the difficult transition into SaaS: incumbents faced an innovator’s dilemma, while Twilio could build automation without directly cannibalizing a seat base.
Salesforce illustrates the conflict. If Service Cloud represents roughly one-third of revenue, selling customer-service automation means attacking that franchise; a pure play’s job is to take the same third and replace it with a smaller number that is nevertheless entirely new revenue.
The panel disputed how large that replacement can become. Rory argued Sierra’s valuation ultimately requires capturing software spend plus a share of displaced labor; Jeff expected buyers to retain much of the savings. Jason countered that even a fraction of the displaced SaaS revenue could create enormous companies while devastating incumbents.

13. Public incumbents need growth before they can fund disruption

Harry’s operating threshold was practical: a public SaaS company growing above roughly 30% has capital and shareholder permission to “swing for the fences.” Jeff agreed that if the growth rate is right, the company should be aggressive; at 10%, management must first repair the immediate growth narrative, making it difficult to finance a long-term reinvention simultaneously.
Harry’s addition was that moving from 10% to 14–15% without an AI strategy may still be insufficient. The slow-growth company therefore needs transformative action most precisely when its valuation, credibility, and managerial bandwidth make that action hardest.
Jeff expects Atlassian’s Mike Cannon-Brookes to remain acquisitive because M&A is embedded in its DNA. He sees an interesting AI vision at Dropbox under Drew Houston, but the company must prove it has permission to expand beyond sync and sharing after several previous attempts.
Dropbox and Box earned Jeff’s admiring label of public-market “cockroaches.” Drew Houston and Aaron Levie have survived roughly a decade of brutal competition while maintaining R&D and strategic motion; their AI breakout is more likely to come through product intelligence, timing, and “a little bit of luck” than a giant acquisition.

14. Product surface area determines whether a platform can compound

Twilio’s central constraint was hidden in its messaging API: from, to, and body. Once a developer specifies sender, recipient, and exact text, any deviation is failure; the product leaves almost no discretionary surface where Twilio can add value without asking customers to rewrite code.
File storage has the same prison: success means the file remains available, while “oops, I lost your file” is failure. For a decade, Twilio sought products with enough expressiveness to do more than fulfill a narrow, binary promise.
Jeff envied Cloudflare’s position between the internet and a customer’s website. Once DNS routes traffic through Cloudflare, it can add dashboard switches for new features without another line of customer code: “flip a toggle to do this and do that,” a structurally better expansion surface.
Stripe may have more degrees of freedom because money supports more operations than a text message, but Jeff suspects it faces a related concentration problem. He had heard that much of Stripe’s portfolio contributes less than the core payments business—a common pattern where adjacent products struggle to break through.

15. Durable developer platforms sell authority, capex, or impossible algorithms

Harry’s Replit example showed the AI-era expansion of developers: he integrated SendGrid in 60 seconds after being unable to do so six months earlier. Jeff’s underlying framework came from Twilio’s 2017–2018 choice between horizontal developer services and going vertically deeper into communications.
“Business development as a service” lets a developer activate relationships they lack authority to negotiate. Twilio supplies carrier agreements, Stripe provides financial access, and AWS provides infrastructure; once the developer presents a working product, management must pay the embedded vendor to keep it working.
“Capex as a service” converts an unauthorized $10 million data-center build into an authorized credit-card transaction. AWS and its peers let developers consume capital infrastructure incrementally, moving an organizational purchasing barrier behind an API.
“Algorithm as a service” is rarer because developers treat paid software as a challenge to rebuild—especially when the bill reaches $5 million annually. DynamoDB’s scaling complexity cleared Jeff’s bar; inference may not once open-source models plateau, but training a frontier model with a “secret recipe” costing perhaps $5 billion still does.

16. Fraud flourishes when speed eliminates verification

IRL’s founder arrest prompted Jason’s hardest call: rounds now happen on Saturdays, and “forget about no diligence being done two years ago—now, diligence isn’t even being attempted.” He wants prison to deter founders who aggregate a year’s revenue into one month or present unpaid trials as paid pilots.
Rory agreed that categorical lies and forged documents deserve severe consequences, potentially prison, but separated them from optimistic presentation that investors could have interrogated. He cited federal conviction rates around 70–80%, with white-collar cases harder because intent and accounting facts create more ambiguity.
Jeff’s counterweight was investor culpability: if VCs are too eager to check the claims, “their greed in that scenario gets rewarded with some amount of fraud.” His recurring tell is that alleged frauds often involve supposedly enormous consumer companies he has never heard of—suggesting nobody verified whether the claimed community used the product.
Rory confirmed that pattern through someone who researched IRL’s claimed demographic and could not locate users. Jason’s firm also began independently checking bank cash roughly 15 years ago after another investor relied on a false balance sheet: the founder owns the crime, but sophisticated capital managers owe the system a duty of care.

(Preview) New iPhones, Absence of Awe, How Oracle Wins in the AI Era Inside OpenAI Enterprise: Forward Deployed Engineering, GPT-5, and More | BG2 Guest Interview