20VC: Thrive & OpenAI, Databricks Raising at $134BN & the TAM Trap

Summary

• Databricks at $134B / 32x sales may be “reasonably priced,” not cheap — the worked example is Snowflake: same ~$4B revenue, 28% growth, $80B, 20x sales, versus Databricks at 55% growth and still accelerating. Rory O’Driscoll’s killer fact: re-acceleration at scale breaks every valuation model — “by definition, if it continues to re-accelerate… it’s infinitely valuable” — and it’s exactly what forced the step-function repricing of Anthropic this year. Only ~1 in 3 companies re-accelerate for one year, 1 in 10 for two.
• Jason Lemkin’s blunt corollary: “seed’s for suckers.” Risk- and time-adjusted, Databricks at $134B or Anthropic at $180B (where Kleiner Perkins went in) can beat a pre-C SAFE at $60M post with 7-10 years of duration risk. There is literally one public company growing >30% — Palantir, at 50% growth, wildly profitable, 80x sales — so there’s no public dataset to price hypergrowth against.
• The majority of public SaaS companies are in a “TAM trap”: average public SaaS growth is ~16%, the slowest ever. Rory’s resolution of the puzzle — the CEOs aren’t idiots; venture funded so many companies that markets saturated (“Hunger Games in SaaS,” his 2019 post). Zoom is the archetype: everyone who needs an account has one, “and they’re done.” Investing rule that follows: “overpayment only works when the TAM is huge. In finite TAMs, you gotta bid more tightly.”
• Seat pricing faces an existential squeeze — everyone’s ARR per employee is climbing (HubSpot 2.8x more efficient than 2021, Salesforce 2x, Microsoft “permanently past peak employee”), so leaders simply run out of seats. Jason’s metaphor: “SaaS has become like Japan” — a great economy where everyone has .9 kids and there are only so many seats to go around. Rory notes rising efficiency may explain why SaaS multiples haven’t fallen as much as growth rates have.
• Security is becoming the incumbents’ weapon: Gainsight locked out of Salesforce for two weeks with no resolution date, Drift “completely dead” after 700 folks’ data was downloaded and ransom demands covered 700 orgs including Cloudflare, OpenAI permanently booted Mixpanel. Jason thinks this could be “the revenge of the enterprise” — and Harry voices the cynical read: use security as the excuse to cut off third parties, then sell your own agent product.
• Google cloned Lovable/Replit in under 10 months (albeit with no database, no OAuth at launch), and Datadog built a PagerDuty competitor in 24 months versus PagerDuty’s 2008 founding — “now you don’t even get a year.” But OpenAI’s code red is read as peak “models will do everything”: “win the ChatGPT wars and you are worth $2 trillion. Let’s not fuss around with… little vertical markets.”
• The next AI wave both would take the meeting on: AI that runs the business, not sells it software — Rory’s Range deal automating wealth management, taxes, trusts down the wealth continuum. Jason’s lived proof: three trusts took 11 months and his celebrated lawyer shrugged “most of my clients never even finish them.” If AI removes that friction for the ~$1.8M average American retirement pot, “you could build a 20, 40, $50 billion company” — with Wealthfront’s fate as the TAM-trap warning.
• Quick-fire split worth keeping: Rory takes Lovable over Superbase (if vibe coding is a category, the front end captures the money — “in for a penny versus in for a pound”), Jason takes Superbase because “hard problems are reassuring” and databases are hard to churn — “this was the year of growth but nothing… I’d love a little defensibility” going into next year.

Deep dive

1. Thrive × OpenAI: power law on steroids — but already stale news

• Rory’s zoom-out: the partnership “doesn’t matter” because 24 hours later it isn’t the OpenAI story anymore. The story is the code red — focus on the core: pushing back ads, pushing back healthcare agents, “no distractions.” “Google did a code red three years ago on them, and now they’re doing a code red back.”
• Jason’s read on the deal itself (OpenAI investing in Thrive Holdings): “there’s, like, only a couple deals that matter to VCs… it’s power law on steroids. It’s power law with what you do with your week.” Turn the fund into a holding company, put a billion or two in, go even deeper with your number-one company ever.
• Rory on who wins: great for Thrive — they put a bunch of money into OpenAI in the ~$70B round, stood by Sam through “the great fiasco of two years ago,” and now get a huge halo effect. “The whole trick in venture is we try and pretend we matter, but in our hearts we know our best companies matter.” The advantage to OpenAI is “much less clear” — he knows who was ecstatic at the announcement and who was “yeah, whatever.”

2. Databricks at $134B: the growth-premium question with a live worked example

• The setup is unusually clean: Snowflake public at ~$4B revenue, 28% growth, $80B (20x); Databricks rumored at $5B raise, $134B, 32x on $4.1B of 2025 sales growing 55%. Rory: “not cheap, but possibly reasonably priced” — the whole question is how much extra multiple you pay for 25-30 points of extra growth, and “if that extra growth lasts for any length of time, extra growth’s worth a shit ton, to use a technical term.”
• The public comps fail you: “there is literally only one public company growing more than 30%, and that’s Palantir” — 50% growth, wildly profitable, 80x sales. Jason: Databricks “would be the second-best public company if it were public today. It seems about right.”
• What breaks the model is modest re-acceleration at this scale — something “we just haven’t seen before.” Rory: any rational model assumes gradual deceleration; “by definition, if it continues to re-accelerate… it’s infinitely valuable, ‘cause that’s just what the math says.” Same dynamic made the foundation models hard to value — when Anthropic re-accelerated at scale this year, “everyone realized the model was wrong” and valuations step-functioned. Base rates: one in three companies re-accelerates for a single year, one in ten for two — and rarely from 50%.
• Jason’s conclusion, seconded by Harry: “seed’s for suckers.” Risk- and time-adjusted, Databricks — or Kleiner Perkins into Anthropic at $180B — can beat “a pre-C deal at 60 post on a safe” with 7-10 years of duration and far less certainty.

3. Snowflake vs Databricks: oligopoly, not coexistence — and agents change the data game

• On Databricks CRO Ron Gabrisko’s claim that their technology is five years ahead: Rory says Ron is right on AI-centric data work, but “no one’s gonna coexist peacefully. They probably hate each- in fact, we know they hate each other ‘cause they time their sales events to overlap.” Expect them to “slug it out” for ten years like SAP and Oracle did for twenty — margins get dinged, neither folds, because “the core value of relational database isn’t going away.”
• Jason’s genuinely open question: all the vibe platforms — Cursor, Lovable, Replit — can now directly access Snowflake data, something impossible a couple of weeks ago. “What happens when I can use easy-to-use agents… to access all of my data any way I want to build any report, any analytics, any workflow? I think we’re 1% on this journey.” He won’t bet against anyone managing huge amounts of structured and unstructured data.

4. Will CRMs become dumb databases under agents? Intentions vs ability

• Jason points to Benioff putting 2,000 people on Agentforce as “the future right there” — and Harry’s pushback lands: “it tells you his intentions. It doesn’t tell you his ability.” Jason concedes Salesforce has maybe two years to unlock it, “and two years is not a lot of time at Salesforce… traditionally that’s like a major release.”
• Rory’s architecture call: single-app agents get bundled by Salesforce, Microsoft-style; but enterprises wanting agents across five or six data sources will “stuff it all in Snowflake and run an agent directly against that” — throw $5M and a bunch of Snowflake or Databricks at it and build bespoke. “Agentforce can get a fair slug of the market,” but the high end will “make systems integrators rich for the next decade.”

5. Security becomes the incumbents’ moat — “the revenge of the enterprise”

• Jason’s exhibit list: Gainsight locked out of Salesforce for two weeks with “no known resolution time”; Drift — breached, 700 folks’ data downloaded, pirates demanding millions per instance — kicked off five months ago and now “completely dead”; OpenAI permanently removed Mixpanel this week. His escalation logic as a platform owner: one breach you blame the PE-owned vendor, “two times, I might start locking down my platform. The third time I might say, ‘I’m just gonna own all the agents.’”
• His worry generalizes: “with agents running everywhere with our data… I generally am worried we’re underestimating security and data residency.” How many board meetings have led with AI-age security? “For me, it’s been close to zero.” And the SecOps team at the average startup versus the size of the checks being written doesn’t compute.
• Harry sharpens it to the cynical version — “I’m using security as an excuse to cut you all off, but lo and behold, I have my own agent product right here, which you can now safely buy, Mr. Customer” — and Jason essentially agrees: “I think it’s the best excuse there is.” Very few incumbents grew materially this year; “this could be a piece of the revenge of the enterprise.” The irony Rory flags: both breaches came from mature, PE-managed SaaS companies, but new AI companies will eat the restrictive policy.

6. Eventbrite at $500M and the PagerDuty problem: smart money thinks 2x revenue is stupidly cheap

• Eventbrite went for ~$500M — 1.5x revenue, a 50% premium. PagerDuty sits at ~$1B on $500M ARR growing 4%. Rory on the mechanics of being public and cheap: someone offers a 50% premium, “the lawyer gives you the speech about fiduciary duties,” and unless you can argue you’ll beat that premium, “you’re very forced to take it.”
• The bullish read: lump in Semrush (bought by Adobe two weeks ago) and “smart, savvy money” is saying “two times revenue is stupidly cheap. I’ll have that” — an aggressive PE firm could buy PagerDuty, bolt on a hot AI startup, get it back to 20% growth and “look like a hero.” Jason’s check: “we haven’t seen pagerduty.com and pagerduty.ai magically mashed together into a winner yet, have we?”
• Rory’s confession is the sharpest specimen: they looked at PagerDuty ten years ago (“we should have let him pay a little more because he was right”), and the post-IPO autopsy found their old model was accurate within 3% on revenue — “all that happened was the market was just willing to pay more for the asset, and now it’s not.” The frustration stands: “literally every ops team on the planet uses PagerDuty. For God’s sakes, it’s pretty obvious what to add here, people. Get it done.”

7. The TAM trap: not stupidity, saturation

• Jason’s confession-as-thesis: “the majority of the public SaaS companies I think are in a TAM trap.” Average public SaaS growth is ~16% — “no one’s ever grown this slowly.” “How did the Aaron Levies and the Drew Houstons… not figure out the TAM trap? What hope is there for the rest of us?”
• Rory’s answer — worth keeping whole: maybe there’s no answer. It’s not that the CEOs were idiots; venture made so many companies that markets saturated, and by the time you needed to expand, another venture-backed SaaS company already held the adjacent market. He wrote this up in 2019 as “Hunger Games in SaaS.” Zoom is the archetype: “everyone who has a Zoom account has a Team account, the poor bastards, and they’re done… you gotta build a new thing” — and the obvious new thing, contact center, they couldn’t get done.
• The portfolio rules that follow: “overpayment only works when the TAM is huge. In finite TAMs, you gotta bid more tightly” — and start the second product far earlier than feels natural; the portfolio company that compounded best “continually added a new product that for the first year or two is a couple of million dollars” and is now many hundreds of millions.
• The open question for AI: can it tap labor budgets and support order-of-magnitude pricing — Gamma at $100/month versus $8 for Canva, Cursor at $500 versus $3 for Jira? Rory’s caution: labor-value pricing erodes the moment “there’s three providers of the same AI software and they’re all willing to do it for 100 bucks.” It hasn’t happened yet — “is a super question.”

8. Seats, Japan, and the efficiency ratchet

• On Workday calling seat reductions existential: Jason keeps returning to Jeff Lawson’s warning about the move away from seats, and his own data crunch — HubSpot 2.8x more revenue-efficient than 2021, Salesforce 2x, Microsoft “permanently past peak employee.” “If you’re a leader, you’re just gonna run out of seats.” Hence the line of the episode: “SaaS has become like Japan… a great economy, but if everyone only has .9 kids, there’s only so many seats to go around.”
• His operating bar has inverted since early 2023: then “I need 200% headcount to grow 100%”; now “I wanna see you grow 100% next year with 50% headcount growth.” A CMO who needs 50 people or a product lead who needs 80 more — “I think it’s time to part ways.” That 2021 DNA is “still ricocheting around probably in the majority of executives.”
• Rory’s synthesis: software prices on value delivered; seats were just the measurable proxy, usage (starting with AWS) tracked value better, and if AI does the work, per-seat becomes irrelevant — Workday probably lands on X dollars per employee served plus Y per HR user, with fewer HR users. The catch: “you can count butts in seats pretty easily… every login is a butt. When you’re trying to measure value delivered, that’s tricky.” He also notes rising efficiency may explain why SaaS multiples haven’t compressed as much as growth rates: slower-growth companies are at least “wildly more efficient” ones.

9. Growth vs efficiency — and why none of the winners need people

• Jason, on whether founders must now deliver both: “in the fastest-growing companies that I’ve invested in, no one gives a rat’s ass about the bottom line.” Per the likely ICONIQ data, the fastest AI companies carry the lowest burn multiples even with heavy inference costs, because revenue outruns compute. And the brute-force path is closed: “you can’t go from 1 to 100 in 10 months without massive inbound demand and a lot of AI” — maybe Larry Ellison or Marc Benioff could hire fast enough; nobody else.
• Rory’s three-category map: public companies grinding on FCF; model labs spending with Nvidia (“no one’s telling OpenAI, ‘Be efficient,’ or if they are, he’s clearly not listening”); and apps-layer companies like Gamma where “traction is ahead of their ability to hire — there’s literally no way to spend the money.” His uncomfortable realization: all three categories don’t need people. “That’s not great if you’re people.” He remains an optimist — “I think all this unemployment thing is bullshit” — but concedes the near-term labor-versus-capital balance favors capital.

10. Google clones Lovable in 10 months — but model providers may have peaked in ambition

• Jason actually tried Google’s new entrant: a Replit/Lovable clone with no database, no OAuth at launch — unimpressive in itself, and big companies “only have so many priorities.” But the clock is the story: Google shipped in under 10 months; Datadog launched its PagerDuty competitor within 24 months, versus PagerDuty founded 2008. “Now you don’t even get a year. If you go from zero to 200 million in a year, you should attract some competition. It’s not a free lunch.”
• Rory’s counter-frame: OpenAI’s code red is “tantamount to an admission” that the core mission needs a year of focus — “that sound you might hear is the consumer hardware product slipping out a little.” Board logic: “win the ChatGPT wars and you are worth $2 trillion. Let’s not fuss around with little vertical markets that can be worth a couple of hundred million bucks.” His call: “we may have seen peak the-models-are-gonna-do-everything” — coding yes, every vertical no.

11. The next act: AI that runs wealth management, not software sold to it

• Rory’s Range thesis: don’t sell software to wealth managers — automate the business itself and push down the wealth continuum: taxes (“in the UK now north of 50% and getting higher”), estates, filings — knowable-but-complex work “done expensively with humans that can be done really cheaply with AI.” The universal pattern: “whenever you see a product that only really rich people have, if you can find a way to get that in the hands of the rest of us, we all want it too.” Target: the doctor, the dentist, the entrepreneur — “more complex than nothing, but not where you can spend 20 grand on a lawyer.”
• Jason’s lived example carries the argument: three trusts took 11 months with a celebrated Silicon Valley trust lawyer, who consoled him with “well, good news, most of my clients never even finish them.” Morgan Stanley’s only real product is loans against stock (“they’ll tell you they’ll help you with your trusts. They don’t.”). If AI removes the friction around the average American’s ~$1.8M retirement pot, “I think you could build a 20, 40, $50 billion company. I would at least wanna take the meeting.”
• But he applies his own TAM-trap test: Range will charge $8-10K against the $30-50K “numbnuts” alternative — not 10x pricing — and Wealthfront, in theory a $10 trillion company because everyone could use it, has real TAM limits after 17 years. His prescription for the category: “maybe be capital efficient until you prove it… if you half prove it, you’ll be screwed” — a line Rory calls spot-on for all founders: capital discipline should be proportional to how hot your market is perceived to be.
• Rory’s defense of slow compounders: Wealthfront at 10bps needs 10x the assets of a 1% manager, so it takes a decade by construction — but “over the next 10 or 15 years, that generation’s gonna get rich… it’s gonna be a compounding machine,” like Schwab from the ’70s. His challenge to Harry: Schwab went public around ‘82-‘83 and is worth $60-80B today — “name me five tech companies that went public in 1983.”

12. Harry’s savage pushback: venture is a relevance game

• Harry’s honest framing of new-age venture: LPs are seduced by “incredible follow-on investors, quick up rounds, and numbers” — “I’d rather be playing that game than the ‘it’s coming.’” And the momentum data backs him: Rory corrects his own stat upward — 40% of Q1’s first-time unicorns have already had a follow-on round. His most cutting version: does Rory have to pick unglamorous compounders because at Series B he “can’t beat” Andreessen, Founders Fund, and Sequoia head-on?
• Rory’s answer: you do both, but the filter is singular — certainty of a big outcome, “because there’s a rule in engineering that you’re only as accurate as your least accurate variable.” Everything else — valuation, time — you can adjust for. He credits Peter Thiel’s version: “all that matters is can you build a big company here… once I filter for that, I can’t have any bullshit rules on stage, on sector. I just want big.”
• The scar-tissue line he’s kept for 20 years, from an LP: “there’s no such thing as blue collar venture” — you’re not choosing on value, you’re choosing on certainty of a big outcome.

13. Quick-fire: Lovable vs Superbase splits the table on risk philosophy

• Rory takes Lovable ($6B) over Superbase ($5B): either vibe coding is a category or it isn’t — “if it’s not a category, both of them are screwed,” and if it is, the front end captures the money (Lovable gets the $20, pays $2 to Superbase). “You may as well be in for a penny versus in for a pound.”
• Jason takes Superbase, “because of stability… hard problems are reassuring.” Yes, it’s “a fork of Postgres… it can be done again — Neon did it, and Databricks bought ’em for a billion” — but databases are brutally hard to churn off, and after “the year of growth but nothing,” his posture going into next year is: “I’d love a little defensibility. I would just love a little hard frigging problems.”