Summary

AI-native growth can make horrific cash burn look capital-efficient, but the burn multiple is no longer a plug-and-play valuation tool. ICONIQ’s data put sub-$100 million ARR AI-native companies at -126% free-cash-flow margins versus -56% for non-AI software, yet their extreme growth can produce more ARR per venture dollar. The discussion warned that ARR quality, hidden churn, shifting gross margins and capex can corrupt the comparison. Harry’s framing applied: “All ratios are wrong, but some ratios are, at times, useful.”
In 2025, cash balance matters more than elegant efficiency ratios, and a merely good $15 million ARR business may have “zero value to a VC.” Jason said venture prices either “on hope” or on current multiples; subscale companies without a credible route to a large IPO lose the upside option that makes VCs care. His advice to a company at $15 million ARR, growing 100% with a decent offer at a $250 million valuation: “Take that deal now.”
Being the category leader matters more in AI because capital, buyer trust and brand recognition compound into a wall of money. Ketty’s long-run market-share heuristic was 67% for No. 1, 20–30% for No. 2 and 10% for No. 3 in business markets; her operating corollary was blunt: “If you’re not number one, don’t spend like you’re number one.” A leading investor’s first $20 million can attract another $60 million, turning a reputational advantage into an economically unequal fight.
Today’s AI prices only work if software captures a profound share of labor spending or the market accepts an enormous reset. Harry cited $25 million ARR companies valued at $5–10 billion, plus xAI at $30 billion and Mistral at $10 billion with little or no revenue. The discussion said either productivity and revenue explode, potentially carrying OpenAI toward $200–300 billion, or valuations prove wrong “by an order of magnitude.” Investors must separately get company selection and entry price right.
Public markets remain generous, but that generosity is the bedrock beneath private valuations. Even after falling 63% since IPO day to roughly $53, Figma still traded around 26 times revenue; top public B2B names averaged roughly 30% growth at about 20 times ARR, versus a historical median nearer six or seven times forward revenue. If that anchor reverts to seven or eight times, “everything drifts down with it.”
OpenAI may win AI without fulfilling every trillion-dollar infrastructure assumption currently embedded around it. The discussed plan implied 125 times more energy capacity in eight years, potentially exceeding India’s present supply, while one 10-gigawatt NVIDIA commitment alone would consume more power than New York City. Harry argued Sam Altman is “willing the trillion into existence,” but the system could stop at $400–700 billion; Ketty’s distinction was that a visionary roadmap is not the same thing as a purchase order against which everyone should borrow.
AI attacks the old SaaS buyout model twice: products can become obsolete quickly, and agents do not need human seats. Pipedrive once took four years to ship a mobile app, while today a company waiting that long for an AI copilot would be “dead in the water”; even product-market fit can now unlock as models change. Jason’s own operation runs 12 agents while needing only two Salesforce seats, making static products and predictable seat expansion much less dependable for both venture and private-equity underwriting.

Deep dive

1. Hypergrowth makes terrible cash margins look efficient

Jason’s starting point from ICONIQ’s 73-page software report: AI-native companies below $100 million ARR showed -126% free-cash-flow margins, materially worse than the -56% for non-AI software. Yet because AI companies are growing so quickly, their burn multiples can still be lower — the afterburners are expensive, but the company reaches “Mach 10” faster.
Jason’s simplified example: spend $2 to add $1 of ARR, receive a 10-times ARR valuation, and the expenditure has theoretically created $10 of market capitalization. It is “one of those weird multiples where lower is better”: one beats two, and a negative number can indicate profitability.
Jason’s Lovable thought experiment carried the argument: even burning $6 million a month after raising $200 million could look efficient if the company adds $300 million of ARR. For venture, the theoretical attraction is leverage — more valuation-bearing recurring revenue for each dollar “you’re lighting on fire.”
The metric remains useful for comparing companies of different sizes and growth rates, but the discussion rejected using it mechanically. It worked best when SaaS companies shared seat-based pricing, 80–90% gross margins, minimal capex and low-churn enterprise contracts; those simplifying conditions no longer hold.

2. ARR quality and cash can invalidate a beautiful ratio

Jason unpacked the first hidden assumption: the reported ARR must be real. Hypergrowth can conceal churn because customers acquired when the company had $2 million ARR are small relative to a current $10 million base; nominal net ARR can therefore look strong while underlying stickiness remains unproven.
Moving gross margins create another distortion, particularly when token costs are material. Burn multiple also ignores capex, which may not matter for every $100 million ARR application company but plainly matters when model developers are committing billions to compute. Comparing unlike businesses now introduces “a lot more noise in the system.”
Jason’s preferred honesty check is GAAP revenue: if ARR supposedly moves from $2 million to $10 million, verify whether recognized GAAP run-rate revenue shows a comparable January-to-December increase. The ratio survives, but only alongside accounting, cohort retention, margin and capex analysis. “We are not in Kansas.”
Absolute liquidity overrides theoretical fundability. A strong burn multiple says the company should be financeable because it creates venture value; cash in the bank determines whether it survives. Rory’s recurring reaction to founders presenting the ratio without their balance was: “Yippee, you could have a great burn multiple and you still could be out of cash on Friday.”

3. Venture now prices companies on hope or current fundamentals

Jason’s framework was categorical: “There are only 2 ways of pricing a deal. You price a deal on hope or you price a deal on the multiples.” Hope supports apparently irrational prices when growth might eventually justify them; once investors fall back to current fundamentals, subscale revenue offers little venture-style option value.
Jason’s example: a $400 million revenue company can be marked at four times revenue and called worth $1.6 billion. A perfectly good $15 million revenue company with reasonable growth may be “of zero value to a VC,” because today’s revenue cannot support an IPO and the investor no longer believes the path from here to a large outcome.
Ketty noted that this does not mean pre-AI software is worthless. Ten of 15 year-to-date IPOs had almost no AI story, and many recent IPO companies were approximately 10 years old. The harder proposition is funding a company at $10 million today on the assumption that seven or eight years of compounding will eventually produce an IPO.
Jason called lingering 2021 advice “borderline inadvertently toxic.” The roleplay used an AI-enhanced mug company with $15 million ARR, 100% growth and a good burn ratio; if a credible investor offers a round at $250 million, the board should not delay to optimize price. “Triple-triple-double-double or better” no longer guarantees demand.

4. Cash scarcity divides good companies from fundable companies

Rory agreed that a doubling, non-AI business can remain excellent while attracting far less capital. His prescription was to close a reasonably priced round, keep growing and operate capital-efficiently: founders may ultimately prove the market wrong, but they should behave for the next several years “as if cash is pretty damn tight and scarce.”
Ketty argued that the distinction between AI and non-AI is disappearing. The ICONIQ report said 94% of public software companies describe themselves as AI companies, most mentioning agents, while Adobe cited $5 billion of AI-influenced revenue. Investors increasingly assume that every credible software company includes an agent or AI layer.
That ubiquity does not eliminate selection; it intensifies it. The market still funds cybersecurity, fintech, B2B and consumer companies, but merely attaching AI is insufficient when capital concentrates around a small set of perceived breakouts. Earlier in the discussion, Jason cited 70% of venture money going into fewer than 20 deals.

5. Kingmaker backing creates both brand and financial momentum

Harry observed that companies running second or third against a “kingmaker” leader such as Harvey or Abridge can struggle to raise at all. Ketty accepted the effect but rejected a universal veto: top-firm backing matters most when early customers are themselves Valley-connected startups; oil-and-gas buyers may barely distinguish Sequoia from Kleiner Perkins.
AI magnifies category branding because confused buyers have new budgets and pressure to act quickly. Ketty’s example was a customer choosing Bolt over Lovable: both names entered consideration, but Lovable did not call back, while Bolt supplied humans and trust. Buyers will not repeat Adobe’s five-year wait to implement Salesforce.
Ketty’s mature-market heuristic was stark: in business software, No. 1 may take 67%, No. 2 gets 20–30%, No. 3 gets 10%, and the rest barely matters; consumer markets skew still harder. A narrower segment that a company can lead may therefore be worth more than fourth place in a larger category.
The kingmaker’s real advantage is reflexive capital. A company receiving $20 million from a prestigious firm may collect another $60 million from followers three months later. Ketty revised her initial purity accordingly: money alone cannot manufacture a winner, but a differentiated challenger now has to beat “the wall of money,” not merely the original investment.

6. AI valuations encode an all-or-nothing labor substitution thesis

Harry’s “peak madness” evidence included $25 million ARR businesses valued at $5–10 billion, plus xAI at $30 billion and Mistral at $10 billion with little or no revenue. His alternative interpretation was that investors may still be underestimating the transfer from human-labor budgets into software.
Harry framed one of two paths within five to seven years. Either AI creates profound productivity gains and companies such as OpenAI reach $200–300 billion of revenue unusually quickly, or AI remains important but valuations undergo a readjustment “that’s going to make your head hurt.” Without labor substitution, current prices are wrong by an order of magnitude.
Jason leaned toward the technological direction being right and still early: his team had replaced 11 people with AI agents, and he described the B2B opportunity as “just getting started.” His concern was portfolio construction — whether venture funds have correctly modeled how many richly valued unicorns can implode or fail to generate venture returns.
The discussion distinguished company-selection risk from entry-price risk. Rory warned that Alexandr Wang’s acquisition could become an anomalous precedent used to justify later prices; Jason invoked Irving Fisher’s ill-fated 1929 claim of a “permanently higher plateau.”

7. Public-market generosity is carrying the private valuation ladder

Figma illustrated both quality and price risk. It had fallen 63% since IPO day to roughly $53 a share, yet still traded near 26 times revenue. Jason’s point was not that Figma was weak — “it’s as close to as good as it gets” — but that a 250% opening pop naturally leaves room for painful retracement.
Top public B2B companies were trading around 20 times ARR while averaging only 30% growth. Rory contrasted that with the pre-2019 mental benchmark: roughly 30% growth at six or seven times next-12-month revenue. Investors then built a ladder, paying incrementally more for 60%, 100% or 200% growth.
If the 30%-growth anchor is correctly higher because companies generate more cash, the ladder may hold. If it reverts to seven or eight times, every faster-growing private comparison falls with it. Rory compared the anchor to the 10-year Treasury: when the bedrock price moves, linked assets cannot remain untouched.
Weak trading in Klarna and StubHub should make IPO buyers demand more discount. If established public peers cost 10 times forward revenue, a new and unseasoned issue might need to price at eight — or now seven — to compensate for risk. That may reduce seller willingness, but Jason expected a pricing adjustment rather than a closed IPO window.

8. EA’s $55 billion take-private rewards a shrinking asset

Harry framed EA’s proposed $55 billion transaction, associated with Jared Kushner and Silver Lake, as momentous. Jason called it the largest LBO in history, with $18 billion of leverage: ordinary for a stable industrial company at roughly six times EBITDA, but notable for a hit-driven gaming business.
Jason gave Silver Lake credit for unusually strong prior decisions, including the Dell take-private and EMC acquisition. He also questioned the operating thesis: EA was roughly flat or shrinking, yet commanded about five or six times revenue, while “more AI” did not itself explain how old franchises would restart growth.
The venture analogy delighted Jason. If he could sell portfolio companies growing around -1% for five or six times revenue, he would “just send me the e-signature” without debating the disclosure. The deal is healthy evidence that mature, iconic assets can still command liquidity even when growth has stopped.

9. OpenAI’s compute ambition exceeds familiar financial scale

The discussion framed OpenAI’s infrastructure plans as roughly 125 times current energy capacity in eight years, potentially requiring more supply than India has today. Ketty added that the discussed 10-gigawatt NVIDIA commitment would consume more electricity than New York City.
Ketty’s image was a country dotted with “Stargates” the size of cities but occupied by only hundreds of humans, each producing the equivalent output of billions of digital minds. “The cities of the future don’t even have humans in them” captured how difficult the physical scale is to intuit.
Harry traced the chain cleanly: energy demand follows compute demand, which follows ambition. As long as revenue and capital keep arriving, OpenAI and the other five or six actors able to make the wager will continue until “unequivocal feedback” says it is not working. Repeated wins have earned Sam Altman another roll.
The technological evidence keeps improving — Claude coding for hours without intervention was one example — but Ketty separated science from finance. Through a science lens, the plan looks possible; through an economics lens, power generation, data-center deployment and enterprise adoption may constrain it within two or three years. Her investable expression was deciding between NVIDIA puts and calls.

10. OpenAI can win even if the trillion-dollar forecast fails

Ketty moved onto the limb: the adoption rate implicit in four or five years of data-center assumptions is probably too optimistic, and forecasts will be revised down. Harry’s objection was financing scale — a trillion dollars for OpenAI infrastructure alone dwarfs even heavily oversubscribed $50 billion fundraising demand.
Harry argued Altman is “willing the trillion into existence” by stating the need early and clearly. But the result need not be binary: if only $400 billion, $600 billion or perhaps $700 billion materializes, the world can use GPUs longer or accept slower gains. GPT-5 and Claude 4.5 already make his two daily hours of vibe coding highly productive.
Ketty endorsed the entrepreneurial mechanism: OpenAI’s broad direction has been vindicated, and “whatever the prize is for being the best company in AI, OpenAI is going to get that prize.” Altman is executing his CEO role better than any other CEO of this decade by keeping the company ahead of the technological train.
The CFO caveat is equally important. A CEO’s plan to triple does not mean the company should buy property and hire as though the triple is guaranteed; it might plan for a double and expand as evidence arrives. OpenAI could still be extraordinary at $30 billion and 50–60% growth while needing far less than a trillion dollars of immediate capex.

11. Meta has the right to wager $20 billion, not proof it will win

Harry disclosed Meta as his largest public position while saying his confidence in its AI strategy had “dwindled and dwindled.” Alexandr Wang’s role, the treatment of Yann LeCun and the team structure looked less like coherent execution than an attempt to assemble expensive talent around a problem.
Ketty’s criticism focused first on communication: Mark Zuckerberg may understand the technology, and Facebook’s engine may be “unkillable,” but neither Ketty nor Harry could explain where the strategy was going. Harry was blunter, calling it a desperate effort to throw money and dream talent together without early evidence that the organization worked.
Zuckerberg did supply one Altman-like statement of intent: he would rather burn $20 billion of operating income and fail than let Meta become irrelevant. With unparalleled power over a business producing enormous revenue and free cash flow, the wager will happen regardless of outside skepticism.
Ketty distinguished earning another roll from being correct. Harry said he would take the bet that this $20 billion produces little meaningful revenue and resembles Meta’s VR effort more than Instagram or WhatsApp; a 50% hit rate can still generate extraordinary returns. The existential logic is attention: two hours on ChatGPT are two hours not spent on Facebook.

12. ChatGPT commerce is inevitable, but may remain a feature

Ketty saw buying directly in ChatGPT as part of an unavoidable monetization push. Free users can principally be monetized by selling them products or selling advertising to them, and “$1 trillion isn’t going to cover itself.” Commerce will be tested, with advertising likely to follow.
Harry treated the launch as an experiment rather than proof that ChatGPT becomes the new commerce interface. OpenAI could generate substantial publicity through partnerships, but a new initiative must produce roughly $2 billion just to matter to the following year’s numbers. Anything smaller may be an integration or feature, not a business line.
Harry noted counterarguments from how users purchase through Instagram and Pinterest and questioned whether ChatGPT commerce would become a major business. The head of apps therefore needs several multibillion-dollar streams within two or three years — “non-trivial” when a trillion contains 1,000 separate billions.

13. Fiverr and DBT show how venture clears its unicorn backlog

The discussed Fiverr-DBT combination looked strategically adjacent and potentially “better together.” Fiverr was discussed at roughly $400 million ARR, while Contra had previously said $100 million; given growth rates, the combined business was described as exceeding $500 million. Ketty’s initial verdict was simply: “Smart.”
Scale is necessary because approximately 6,700 unicorns face an exit bottleneck. At around 15 IPOs year to date and perhaps 20 for the full year, Ketty calculated roughly 30 years of inventory. Combining two midsized businesses may create the critical mass needed to go public.
Shared ownership by Andreessen Horowitz simplifies the arithmetic and negotiation. An investor holding 20% of each company can preserve meaningful exposure to the combined entity; an investor owning only one may see 20% diluted to 8%. The correct comparison is nevertheless “20% of something that’s not going public” versus 8% of something that can.
The fatal risk is not dilution but integration: turning a well-run 20% position into 8% of a combined disaster. Industrial logic, CEO conviction and partner selection matter more than an investor’s desire to protect ownership. The discussion contrasted this with mechanically mashing together slower-growing properties such as Clari and Drift, even where the product logic might appear adjacent.

14. AI destroys both SaaS product durability and seat predictability

Harry asked whether technology-focused private equity is now underpaid for displacement risk while other strategies move more capital at better multiples. Ketty said buying Pipedrive- or Copper-like SaaS companies is harder than ever because capital can move elsewhere at better multiples and startups increasingly attack the installed base. Harry cautioned against abandoning buyouts for late-stage minority investing, where established specialists have the advantage, but agreed that every underwriting must explicitly model AI downside.
Ketty identified the old model’s hidden gift: business-software products barely changed from roughly 2008 to 2023. Pipedrive once took four years to launch a mobile app; today, taking four years to ship an AI copilot means “you’re dead in the water.” Jason doubled down on the insight. High net retention was valuable, but technological stasis made the spreadsheet dependable.
Risk has also increased for new AI companies. Harry noted that traditional SaaS could lock into product-market fit for a decade; post-LLM startups may move into and out of fit as base models improve and last year’s approach becomes obsolete. Jason responded, “That’s why it’s good the growth is higher” — the underlying business is less stable.
Seat economics are weakening without disappearing. Jason runs 12 AI agents but needs only two Salesforce seats; Salesforce could ultimately charge more if Agentforce supplies those agents, but today the software agents do not each require a license. Harry’s Accenture example made the human consequence stark: reskilling may be “not a viable path for the skills we need.”

15. CEOs retain political speech, but companies should stay neutral

Harry’s starting instinct was that becoming CEO should not eliminate a person’s right to political opinions. At the corporate level, however, Jason argued that recent experience favors restraint: companies and universities that entered culture wars discovered the value of sticking to institutional purpose, resembling the University of Chicago’s principles.
Jason then revised his view slightly. Some trusted roles historically required political reticence: in 1952, both parties reportedly approached Dwight Eisenhower because his affiliation was not publicly known. Jason still resisted policing clearly personal speech, but acknowledged that leadership can blur the distinction between individual and institution.
Jason offered the tactical evidence from private social-media interventions. He has quietly warned roughly 10 executives when a post landed worse than intended; only one ultimately welcomed the feedback. The others understood that they might alienate 40% of customers or upset underrepresented employees and responded, in substance, “I feel so strongly I don’t care.”
That leaves boards with a hard boundary question: personal speech may be protected, but what if it loses half the customer base or drives out 10 great engineers? Harry’s closing observation softened the practical risk — attention moves rapidly, as controversies around Deel and Rippling or Elon Musk and Donald Trump vanish into the rearview mirror. “Just keep moving forward.”

(Preview) Sora 2 and an AI Video Boom, Meta’s Vibes One Week Later, Questions on TikTok, Solar Power, and the iPhone Air Taste is your Moat (Dylan Field of Figma)