Why You Need a $1B Fund To Do Series A | SpaceX at $2TRN & Data Centers in Space | Groq's $20BN Deal

Summary

• Anthropic is taking the marginal enterprise dollar. Ramp data shows Anthropic capturing 73% of new spending among companies buying AI tools — 50/50 ten weeks ago, 60/40 in OpenAI’s favor in early December — and the panel treats the marginal buyer as “the most leading indicator,” consistent with Anthropic’s roughly $22bn run rate. OpenAI’s “lemonade stand” snark at Ramp was, per Rory, “just a bad look” that misunderstands statistics.
• Coding lock-in deserves a “code red” at OpenAI. Jason Lemkin’s mechanism: switching models is cheap, but re-QA’ing dialed-in agents is expensive — SaaStr’s AI VPs of marketing and customer success run on Sonnet 4.7 and “there’s no way we’re going to switch them to Codex.” Rory’s window: if OpenAI lets Claude stay perceived-best for another 6–12 months, “you’ve probably sacrificed value that you’ll never get back.”
• Rory’s base case on OpenAI is still constructive: fix two jobs — monetize the consumer base nobody has taken away, and win enterprise coding — and “you still have a comfortable chance… to exit two, three years from now as the largest market cap stand-alone foundation model player. You blow it for another year and you won’t.”
• SpaceX at $2 trillion rests on a probability you pick yourself. Elon’s Terafab announcement (~70% of TSMC’s volume, ~$25bn capex, 80% for SpaceX data centers in space) moved Polymarket’s $2trn IPO odds to 50–60%, but Tesla stock didn’t move — Rory’s tell that real money isn’t buying it. An optimistic analyst can still “do it on a spreadsheet”: 80% it happens, 30% on time, Starlink’s 53% margins as the template.
• Bezos’s $100bn manufacturing fund is the “Indian Creek Island investment” — the Walmart play, not the Amazon play: buy incumbents and inject AI rather than build full-stack, “inherently less disruptive and more financial engineering.” Every billionaire wants one of these; expect a bunch more.
• Nvidia’s ~$20bn Grok deal shows M&A’s new perverse math: Jason thinks Nvidia paid roughly 3x the last round ($6.9bn) for sub-$100m ARR because the value to a $5trn acquirer is enormous — then burn ~$4–5bn in double taxation (roughly a 60% effective rate on founder Jonathan’s ~$950m) purely to dodge antitrust review. “Wire me the money either way” — the government wins under both paths.
• Figma’s 22% drawdown on Google’s Stitch is a rational panic wrongly triggered — “massive market overreaction to a proof of concept,” yet the market is right that revenue durability is breaking: Figma Make is “one of the worst products I’ve used in the last 6 months,” and Jason’s test is blunt — “you’re not an AI company if you can’t charge for it.” Notion passes (ARPU doubled); most public software is “in terminal decline” until proven otherwise.
• Venture’s structural squeeze: leading Series A now takes ~$1bn of fund ($30–40m rounds, $25–30m checks, reserves), while the acquirer-to-unicorn ratio sits at “the lowest ratio of our careers” — PE is gone, hyperscalers won’t buy 100 companies, and incumbents can’t afford app-layer companies marked above them. “It’s so much easier to get a 9 billion dollar valuation than a billion dollar exit… It’s basically win or die.”

Deep dive

1. Ramp’s data says Anthropic owns the marginal enterprise buyer

• The claim, stated precisely by Rory (likely Rory O’Driscoll): Ramp — which Harry said he thinks processes roughly 0.5–1% of US GDP transactions — shows Anthropic capturing 73% of new spending on AI tools, versus 50/50 ten weeks ago and 60/40 OpenAI in early December. OpenAI still leads on total spend, but “the marginal buyer in the last six, eight, ten weeks has massively shifted — which is obviously the most leading indicator.”
• OpenAI’s response — a snarky line about “extrapolating from a lemonade stand” — was a self-inflicted wound: Ramp likely has a diversified base and good data scientists, and “sometimes the first thing you got to do in dealing with a problem is to face the hard facts in the face.” Anthropic at a possible $22bn run rate “certainly isn’t inconsistent” with the conclusion.
• Jason Lemkin’s caveat: both OpenAI and Ramp could be right, like the Cursor debate — inside tech portfolios “cursor’s dead, they’ve all moved to Claude Code,” but “for the normal world, they live in ChatGPT.” If Ramp skews tech, it may be measuring that bubble. Still, Jason describes “Opus 45 and after” as an under-discussed step function: “Everyone’s PRs exploded, everything got better.”

2. OpenAI’s whiplash vs Anthropic’s consistency — and a closing window

• Harry’s indictment is inconsistency, not capability: flat headcount for cost control, then doubling to 8,000 by year-end; deep into agentic commerce, then “basically canceling — and Walmart says it doesn’t work”; Sora folded into ChatGPT; hardware deprioritized. Anthropic, by contrast, “is very consistent about its ICP and goals. We know what it stands for.” The mood cost is real: “It’s just a downer to be around and I don’t want to hang out with Debbie Downers — I actually don’t want to try their new products.”
• The deeper read: OpenAI used to be “the exception that made the rule” — enough momentum to survive founder drama, board dysfunction, management turnover. “Now the downside is rearing its mind… this inconsistency is damaging the company.”
• Rory’s counter — “things are never as good or as bad as they seem” (he called the Jony Ive hardware deal dead on arrival, and noted the press “only has two stories: we love you, we hate you”). The fix is focus: job one, monetize the consumer business no one has taken away; job two, enterprise coding. Do that and “you still have a comfortable chance… to exit two, three years from now as the largest market cap stand-alone foundation model player. You blow it for another year and you won’t.”
• The stakes, in Harry’s first-mover frame: consumer mindshare is ChatGPT’s; coding — “the mother lode app within the enterprise spend” — was up for grabs 6–12 months ago and is now “half up for grabs.” “There was a tide in the affairs of men, as Shakespeare says… you don’t get to show up after a whole bunch of people have made enterprise decisions and say, ‘Now we finally got our shit together, please pick me.’”

3. Lock-in mechanics: soft costs beat token costs

• Two opposite things are true in the data. OpenRouter usage has exploded since the start of the year — cost-optimizers rotating between Kimi (“Kimmy” as spoken), Haiku, and Mini. Meanwhile quality-sensitive builders are locking in: since Sonnet and Opus “4.5 and 4.6,” Jason wants to “build all my scaffolding… about something that isn’t just good, but now is epically good.” Switching models is cheap; QA’ing and qualifying the outputs is not.
• His own specimen: SaaStr’s AI VP of marketing and AI VP of customer success, built since December on Sonnet 4.7 plus a little Opus — one defines every marketing activity and runs weekly team meetings, the other services ~200 sponsors 24/7 (“all the humans would quit cuz it was too much work”). “It took us weeks to dial it in… there’s no way we’re going to switch them to Codex. I would have a code red on this.”
• The metric both want tracked: AI token spend as a percentage of revenue. Plenty of apps build huge value at 5–8% of revenue versus coding apps at 40–50% — and at the low end, optimization is irrelevant: “You want to reduce my token cost from $2,000 a month to 1,500? Leave me alone… I got 99 problems. This isn’t one of them.”

4. SpaceX’s Terafab: $2 trillion is a probability you assign, not a fact

• The announcement: a fab near the Gigafactory at roughly 70% of TSMC’s total volume and ~$25bn capex, split ~80% SpaceX (data centers in space) / 20% Tesla. Polymarket’s odds of a $2trn SpaceX IPO jumped to 50–60% — but Rory’s pushback: “Tesla stock didn’t move… where actually significant money is changing hands, nobody blinked.”
• His framework for all things Elon: value the done things on multiples, then assign a probability to the announced things. “If the probability is 100%, announcing a fab means you own the fab. If it’s 1%, you own 1% of the fab.” TSMC — 30 years of fab-building — is just over $1trn; a $400bn pop on an announcement implies ~50% odds. Elon is “the most accomplished entrepreneur of at least the last 30 years” on hard engineering, but his timing record is spotty — Rory literally asked ChatGPT for the chronological list of FSD and Starship predictions.
• Jason’s counter-vision, as told: on Jay Leno’s Tesla Semi test, the lead designer said “the future is fusion… we believe the fusion’s from the sun” — chips to space to power fusion. “Now you see it all coming together for SpaceX for real… starting to sound cheap at two trillion. Who else can harness the sun?” And if Starlink really runs 53% profit margins, extending that vision arguably raises a believer’s DCF.
• The reconciliation: these are step-function companies — a technical leap every five-to-seven years, harvested while building the next one. Rory: an optimistic analyst can justify $2trn — 80% it happens, 30% it happens on time, Starlink-like margins within five years — “you can do it on a spreadsheet. And that bet will be available to you, and have at it.”

5. Bezos’s $100bn fund is the Walmart play, not the Amazon play

• The WSJ report: Bezos raising a $100bn manufacturing transformation fund to acquire companies across semiconductors, space, and defense and inject AI into their operations, touring Singapore and Middle East sovereigns. One complication: “SoftBank seems tapped out — they’re flashing above their covenants.”
• Jason’s instantly-adopted framing: “a great classic Indian Creek Island investment… You’ve already built Amazon. You get to think big at Carbone on the yacht and you don’t want to go small anymore.”
• Harry’s historical triptych, worth keeping whole: when the internet hit retail you could build Shopify (sell software to retailers — worth a couple hundred billion), buy Walmart (a 2x on half a trillion), or build Amazon full-stack from zero (the $2trn play). At 25 with no money you do Amazon; at Indian Creek with $100bn, you buy Walmart and inject AI — “inherently less disruptive and more financial engineering.” Jason: every billionaire not running a public SaaS company wants one of these plays; “we’re going to see a bunch.”
• The tangent that’s actually a thesis: billionaires are migrating to where they’re not vilified — Sergey Brin judging a grassroots Miami hackathon, Ryan Smith “beloved here in Provo” — while “in the Bay Area, billionaires are vilified… Who the hell wants to live where you’re vilified?” Jason’s warning: QSBS tweaks won’t fix it; comfort, not just taxes, moves the golden geese.

6. Grok’s $20bn: revenue multiples get invoked, then abandoned

• Jason’s answer to “when does someone pay $20bn for $100m of revenue?”: when the value to the acquirer is huge and they have the market cap — Nvidia at $5trn qualifies, and Jensen said Grok’s tech goes into production within a year, “that’s worth billions.” WhatsApp is the scale precedent: $16bn “and it didn’t have a dime of revenue.” Jason adds the pricing tell: last round $6.9bn, and he thinks the deal was roughly 3x — “my first startup was acquired for exactly 3x our round… they downloaded our certificate of incorporation and showed up unsolicited.” M&A anchors to round math, then abandons it the day after close.
• The structure is the story: an asset sale to dodge antitrust review — gain taxed at the company level, then again on distribution — wasting “plus or minus four or five billion bucks on a $20bn transaction,” a 60% effective tax rate for founder Jonathan ($950m), no Nvidia stock to roll into, and “you don’t even get the IP… you don’t get the company.”
• Rory’s institutional alarm: the government now wins either way — “you either lobby extensively at the highest levels and get a waiver from the top down… or you pay double taxation. Wire me the money either way. It’s a really perverse incentive.”
• The human note, kept as told: Jonathan “was in the desert in the dark for many years… a real cockroach who’s gone through the hard times. It’s nice to see good people win.” Chamath’s matching ~$950m gets less charity: “It’s okay, we believe you’re rich… Therapy will help.”

7. Figma’s drawdown: wrong trigger, right panic

• Google’s Stitch launch knocked Figma down 22% to $21.66 — and Harry bought on the way down after seeing Sequoia buy $35m of stock. Jason, who actually used Stitch: “massive market overreaction to a proof of concept. Give me an effing break.” Google launches and abandons constantly (see its audio competitor — “Sono” as spoken); the odds it commits to a Figma competitor for a decade “approach zero.”
• But the market’s underlying message is rational: “We no longer believe this revenue is particularly durable.” Figma Make is “one of the worst products I’ve used in the last 6 months” — the only vibe-coding tool that can’t pull context from 20vc.com to build a site, something “anybody can do today.” The credit market agrees on the sector: Qualtrics couldn’t finish its debt offering this week; Salesforce barely got its done.
• The diagnosis is decaying product-market fit, not go-to-market. Figma’s CRO admitting no AI in the sales team worries Jason less than the product — AI-hot companies are full of “recycled mediocre” CROs and it doesn’t matter, because “sales does not fix product market fit.” Rory’s escalation: AI in go-to-market or engineering is “jacks to open” — the only question is how AI changes the end product — and “if you’re a software product and you don’t think AI is going to disrupt not just how you build, but what you build, you actually probably want to actively short it.”
• Why a good CEO like Dylan lets this happen: the installed base is a trap — “the greatest opportunity” and simultaneously “50 years of debt, 50 years of features… if you’re not careful it will consume 98% of your resources.” Mike Cannon-Brookes alluded to it before Atlassian’s layoffs: “otherwise Jira and Confluence are going to suck it all up and I have no people.” Public companies can barely choose to let a billion-dollar core decline to fund the agentic product — “it’s what you got to do.”

8. The test: charge for your AI or you’re not an AI company

• Jason’s line in the sand: “If it’s not good enough to charge for, it doesn’t count. You’re not an AI company if you can’t charge for it.” Why price at almost nothing when Anthropic is at $22bn “in the blink of an eye” and buyers line up for agentic legal and healthcare products? “Very few public companies can effectively monetize AI, and that’s why they’re all in terminal decline” — he won’t buy their stocks until that turns. Figma, notably, isn’t really charging.
• The SMB benchmark: is ARPU 50%+ higher than pre-AI? Notion appears to pass — roughly doubled ARPU, $20/month with AI versus $10 basic, “not because they splashed an AI label on it, but because it’s so much better.” Microsoft Copilot is the counterexample the market already rejected; Jason suspects Slack’s AI version quietly passes next.
• Rory’s aggregate version: acceleration is the test — pricing, ACV, growth, something in the KPIs “has got to be working… otherwise you are falling behind, cuz the AI-first players are delivering utility and customers are giving them money.”
• On buying the dips: Rory wouldn’t touch Atlassian (“at 21 today” as quoted) — “I would rather miss the bounce right off the hard deck than invest in something that may fall below it.” Rory’s discipline: the 5–10–20% bounces are probably there, but “you only should buy things if you think 5 years from now they’ll be a winner” — and neither has seen the product evidence yet.

9. Series A now takes a billion dollars — the old fund math is broken

• Harry’s provocation off Mamoon Hamid’s new KP funds ($1bn early, $2.5bn growth): you can’t lead Series A with less than ~$1bn — A’s are now $30–40m rounds needing $25–30m lead checks, times ~20 positions plus reserves. Rory doesn’t entirely disagree: ~$20m average check with 50%-of-initial-capital reserves means ~$30m per deal, and with longer exits you’re “nearer to 30 than 20” positions — “there’s a certain scale required to play meaningfully in the Series A business.”
• The old heuristics are dead: “$60 million per partner for seed, maybe 100 for A or B… that math just is broken today.” Hummingbird’s initial capital was ~$200–300m, while its new fund is $800m — “We were really good at doing Revolut at four. Now billion dollar rounds are a good entry point for us.”
• Harry, “being a mild little jerk”: “the math is easy — you just don’t like the answer.” Rory’s response: we are leveraging up our risk unless outcomes get massively bigger. The Wiz datapoint cuts both ways — if Cyberstarts owned ~4% at exit, 3–4% of a $30-something-billion outcome still works for a seed fund, but only if exits clear $10bn.
• Jason’s cycle perspective carries weight: he lived ‘95–‘99, the ‘02 unwind, ‘07, and 2021 — “this feels faster than that.” Hence the episode’s mood: “every VC is stressed right now. No matter how successful you are… everyone wrestling with doing deals now is grumpy, stressed and feeling the pressure.”

10. The exit drought — and the momentum-investor confession

• Jason’s under-discussed risk: “there’s some ratio of potential acquirers divided by unicorns, and I think we’re at the lowest ratio of our careers.” PE is gone as a buyer, “Nvidia isn’t buying 100 companies, Microsoft is not buying 100 companies” — so who buys Replut—likely Replit—, Legora, Harvey, and Level if they don’t IPO? “We have outstripped any current ability for these companies to have any exit… it’s basically win or die.”
• The structural trap: the whole app-layer thesis is a bigger TAM (“eat the work”), which marks the new company above the incumbent it replaces — so the incumbent definitionally can’t afford to buy you. Harvey at $10bn versus legacy legal software at $2bn: “it doesn’t matter what you want — you can’t afford to buy them and they can’t afford to sell to you.”
• The math that should scare late-stage buyers: “It’s so much easier to get a 9 billion dollar valuation than a billion dollar exit — which should be terrifying if you’re the people giving the 9 billion dollar valuation.” Companies fundamentally worth $5bn but priced at $10bn will see last rounds “convert and take 50 cents on the dollar unless it has a meaningful block.” Maybe the market gets “chill with down M&As” the way it accepted down IPOs — Jason isn’t so sure; for early-stage holders, selling into secondaries is a possible answer.
• Harry’s confession closes it: his biggest regret is not breaking the Series A mandate to momentum-invest into “your 11 Labs, your Legoras, your Lovables” earlier. Rory’s honest reframe: taking five of twenty A-slots to do D/E rounds at $1–2bn pre would have delivered Series A returns “at much lower risk — and Anthropic will be the definitive version of that forever; the rounds at 14 and 60 billion were risk-adjusted freaking awesome.” His caveat is the season: “the definition of a momentum strategy is it only works in a rising market… for the last 3 years it’s hard to distinguish momentum players from very shrewd players.” Harry’s brutal coda: “Then maybe you and I aren’t shrewd. Look in the mirror.”