With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll
We’re back in the studio. Or at least the Riverside.
On the docket this week:
- Washington lifting the 19-day Fable 5 ban and what a structured pre-approval process means for shipping models going forward.
- OpenAI floating the idea that every frontier lab hand the U.S. government a 5% stake, and whether that buys alignment or just invites Bernie Sanders to ask for 50.
- The dilution insensitivity that has quietly become the default posture of the best AI founders.
- Meta and SpaceX turning failed proprietary bets into neocloud businesses and getting a 10% stock pop for the pivot.
- Nvidia financing its own demand with compute-now-pay-later. Anthropic and DeepSeek building their own chips.
- Kling raising at $18B while OpenAI kills Sora.
- China owning the top six models on OpenRouter.
- And how the the model companies are becoming product companies, and the trillion-dollar services layer between them and corporate America is up for grabs.
Top Takeaways:
1. The Fable Ban Was Minor. The Pre-Approval Precedent Is What Should Worry You.
Washington lifted the 19-day ban on Fable 5, but the resolution matters far less than the precedent it sets. As Rory framed it, the industry is now entrapped in some version of a structured pre-approval process that hasn’t been finalized yet. Six months ago you could ship software like a free man. Now you get permission from Washington first.
There are credible arguments for oversight on the cybersecurity side. But all else being equal, you would rather not clear your business with any administration before you pursue it. Rory’s point: a lot of what made the U.S. a dynamic economy was the absence of exactly this. Europe has it. Now we have it too. We sneered at GDPR, and here we are.
The specific impact of the Fable ban is small, because Fable is moving to variable per-token pricing in a week or two anyway. It will be a niche, expensive model until those gains percolate down into standard Opus and Sonnet over the coming months. Most builders won’t touch it at that price. The model barely moves the world. The precedent does.
2. Sam’s 5% Trial Balloon: Alignment or an Open Invitation to Get Taxed at 50?
Then OpenAI floated that frontier labs give the U.S. government a 5% stake. Ground it in fact first: the proposal wasn’t “OpenAI gives up 5%.” It was “companies should,” which implies Anthropic too. At Anthropic’s valuation that’s roughly $50B. Sam is, in effect, volunteering other people’s capital.
Jason’s read is that this is straight out of the portfolio playbook. You sell 5% to a giant strategic partner, the way founders give Shopify a slug so Shopify doesn’t crush them, and it buys an unexpectedly large amount of alignment. Five percent is immaterial to the government’s balance sheet and immaterial to the cap table, but Jason’s repeated experience is that it drags the big partner into the room in a way that consistently surprises him. If it plates the federal government and makes you the good guy, an investor takes that dilution.
Rory’s counter came in two parts. First, the business case: Microsoft owns roughly 30% of OpenAI and they are not besties, they are in a stale marriage looking for a divorce they can’t quite afford. If a 30% stake with a rational, profit-maximizing partner didn’t produce alignment, why would 5% with the U.S. government? Look at TARP: the government took small stakes and then told banks who they could and couldn’t pay. Second, the political case: you have simultaneously said give me 5% and published a nine-point plan arguing AI is so catastrophic to the labor market that Congress must restructure the entire tax system to shift the burden from labor to capital. If you really are destroying labor in a $30 trillion economy, the political monster does not settle for $50B. Bernie has already said he wants 50.
So why do smart people volunteer for this? Because they believe the technology is important enough that everything has to be on the table, and that belief is exactly the narrative that let them raise hundreds of billions in the first place. You needed the biggest story to get the biggest checks. Once you and your employees genuinely believe the thing is dangerous from a cyber and a jobs perspective, every downstream ask becomes logical. Jason’s read on the mechanics: this is Sam anchoring. The question isn’t whether 5% is a good idea. It’s that Sam is planting 5% as the number before someone else plants 50.
3. Dilution Insensitivity Is Now the Default Posture
The thread that ran under the 5% debate: nobody is money-motivated in the way the last generation was, because the cap tables have changed. Anthropic’s founder owns roughly 1.7%. Sam owns nominally zero. The most successful startups of our lifetime, and the founders own low single digits or nothing.
That reprices the whole conversation. Ramp has done a dozen announced rounds and probably more than 20 counting stubs, each at 5% or 6%, and every raise is a TechCrunch victory lap. Those add up. But founders no longer flinch. Jason’s honest math as a seed investor: he used to assume his real entry price was double the headline because of dilution. Now he assumes four times. A seed at 60 is really a seed at 240.
Two structural things changed. One, per-round dilution is actually declining (Carta’s data backs this), so founders can raise more dollars for the same ownership. Two, and more important, nobody worries anymore about returning their last high-priced round. Investors have learned to accept 1x without drama, without blocking, without threats. The exception is non-standard money: Jason is watching a non-traditional VC block round after round in one portfolio company right now. But the standard players don’t block exits. That removed the fear that used to terrify Jason’s generation of founders, and it freed up the risk. As Rory put it, what you’re really weighing is optionality versus upside. If the prize is a trillion dollars, and it has been in at least three cases, it doesn’t matter what it takes to get there, it only matters that you get there. If the prize is one to five billion, raising too much kills your ability to take the life-changing exit.
You see it in the contrast between Databricks, which named its rounds honestly all the way to a Series M with no performative discipline, and Karri Saarinen at Linear, who raised just two rounds and refuses VC intros. Jason loves the discipline but wonders aloud whether, in 2026, capital efficiency is the right call when the prize is this large. If the exit is $5B, discipline wins. If the exit could be $100B, you might have wanted the fuel.
4. Alex Karp Was Right on Both Counts
Karp’s CNBC hit went viral, and stripped of the German-philosophy references and the personal baggage, the two substantive points landed. One, corporate America is asking whether it is getting anything for all this AI spend, which is the ROI question. Two, corporate America is asking whether it is handing over its data to be trained on and resold. Palantir stock went up 9% on the day.
The second point is less obvious, and Jason tied it to the same week’s HubSpot episode. HubSpot announced it would pool everyone’s verified contact data into a shared prospecting product, the thing vendors have tried to pull off for 15 years, and customers erupted so hard the company walked it back inside a week. Every vendor facing slowing growth or intense competition will be tempted to cut the same corner on training privacy. The labs already got caught on books and on YouTube. HubSpot got caught on contacts. Salesforce will be tempted next. If you sell to the government and regulated industries the way Palantir does, “you can’t really trust these guys with your data” is a genuinely great wedge.
5. Those That Can’t Build a Model Open a Cloud Business
Meta launched MetaP Compute to rent out AI infrastructure, hosted or raw GPU by the hour, and the stock jumped 10% in a single day, its best in five months. SpaceX ran the same play. Both bought a mountain of compute to build proprietary assets, fell short of those assets, and pivoted to selling the compute instead. Both got rewarded for the pivot. The neoclouds, Nebius and CoreWeave, dropped 10% to 15%, which is the correct market reaction: you would rather compete with two players than four.
What the market is actually pricing is less clear. One reading is a Goldilocks story, where the market believes Meta has short-term excess it’s smart to sell and a long-term AI use for the compute nobody can see yet. The other reading is bleaker: both of you failed at your original goal, but being a hyperscaler is a good business, so go team. That second reading invites more entrants to run the same journey, and you end up with a few buyers of compute (OpenAI and Anthropic can obviously absorb it) and a glut of sellers.
For now none of that matters, because the core business is working. Facebook, WhatsApp, and Instagram are the engine, and Zuck is treading water on the AI side while the engine funds the search. Did he overpay for Scale? Probably. Did Llama pan out? Unclear. But the core throws off enough cash that there is no fatal-error risk in continuing to play, which is exactly the advice Jason gives founders whose main business is healthy: stay in the game. As a board member, Jason’s line would be, “I don’t understand where $70B of investment gets you, but you’ve earned the right to play.”
The bigger aha: this spending does not stop because the supply side says stop. Meta, Google, and Microsoft will not blink. It comes down to the demand side. As long as enterprise revenue at OpenAI and Anthropic keeps doubling and tripling, even at a fraction of capex, the spend keeps coming. Demand is the only thing that shuts off the spigot.
6. Nvidia Is Now Financing Its Own Demand
Nvidia’s compute-now-pay-later structure, used in two early-July deals including one with a next-generation neocloud, is the mechanism keeping the money machine rolling. Nvidia sells the chips up front and recognizes the hardware revenue up front, then gives the buyer put-back rights if they can’t use the capacity. It separates the revenue from the guarantee, which makes it ASC 606 legit, but it is aggressive. Nvidia is leaning backward to make these new neoclouds work as it diversifies away from the hyperscalers, and it is taking on a lot of contingent liability to do it.
Everything about it works right up until compute demand softens at the margin. The whole thing is a derivative bet on demand continuing up and to the right. If it slows and excess capacity appears, Nvidia stops growing and starts de-booking prior revenue as customers go bust. The structure isn’t crazy, but that’s the bet underneath it. Rory’s flag: a year ago he’d have said don’t do buybacks, but if you’re going to be aggressive with cash this is actually a better use because it keeps the flywheel going. And the time to manage for the downside is exactly when no one is managing for the downside. Right now, nobody is. We remember that stage of the cycle from 1999 and 2000. History doesn’t repeat, but it rhymes.
7. The Custom-Chip Push Is About Margin, Not “Special Needs”
Anthropic opened talks with Samsung to build its own AI chip, and DeepSeek announced its own silicon effort. Rory softened his “this is mad” take from last week after Anjney Midha’s two arguments: you have to own the compute, all the way down, because if you don’t own the compute you’re exposed; and custom silicon optimized for your specific model can be meaningfully more efficient than adapting a general-purpose Nvidia platform.
Jason’s pushback is sharper on one point. The argument that “we have specialized needs Nvidia can’t meet” makes zero sense to him. If you’re driving that much volume, Nvidia will do your tape-out and build you the custom chip. “Customized for us” is soft language for the real motive: the margins are so high that to survive, these companies want to recapture them. Everyone is dancing around saying that out loud while keeping the Nvidia relationship intact.
8. Kling at $18B, Sora Killed, and the China Valuation Question
Kling raised $2.8B at an $18B valuation, is doing $500M in Q1 ARR, is the biggest AI video business in the world, and is headed for a Hong Kong listing. That lands in the same window OpenAI shut down Sora, which raises the obvious question: if Kling can build a real business here, why couldn’t Sora?
Part of the answer is opportunity cost. If you’re OpenAI, that compute has a higher, enterprise-centric use where you make real money, so on a standalone basis consumer video gets cut. A rough estimate puts a 30-second generation at $1.30 to $2 in GPU cost, so with any reasonable charging model there’s a business, and Kling proved people will pay. Sora also gave away too much for free. For OpenAI, $500M in video revenue is below the materiality line and a distraction. For Kling it’s the whole company.
The sharper question Jason raised is valuation arbitrage. Higgsfield, where Harry and Jason are both investors and Jason was one of the first 10 users, just hit $500M in revenue and $2M a day in credit-card billings outside the enterprise, and Kling is only one of the models it wraps. If Kling is worth $18B at $500M and a partial layer on top that’s cash-flow positive is raising at $5B, is that a 3x arbitrage, or is there a Chinese AI valuation bubble? The counterexample: DeepSeek raising at $50B is a gross discount to any Western equivalent. Different market, different dynamics for capital raising.
9. On Open Source, China Is Ahead, and We Built That
The top six models on OpenRouter are Chinese. On frontier closed-source models, the U.S. is clearly running away with it. On open source, China is running away with it, distilled in some part from OpenAI and Anthropic, reasonable people can differ how much.
Jason’s on-the-ground point after two weeks in China and Hong Kong: this is a problem the U.S. created. On the other side of the Great Firewall, you cannot access OpenAI or Claude at all. It isn’t only blocked, it’s unavailable, and they actively shut down the VPN workarounds. So of course the second-largest economy in the world, with excellent engineers who’ve been building internet and software for decades, is going to build things as competitive or better. Jensen was right. If you don’t like what China is building, you built it by not letting them use ours. And the newest twist borders on absurd: China is now floating restrictions on overseas access to its own open-source models. We’re nervous about using them because we think it’s dangerous, and they’re nervous about letting us use them because they think it’s dangerous. Both can’t be true. If Chinese open source were pulled as an alternative, that would be a real gift to U.S. frontier and open-source providers alike.
10. The Model-Selection Tax Is Real, and Frontier Wins When the Problem Is Hard
Jason’s build story is the clearest framing of where the industry is heading on cost. He spent about 10 hours in Replit on sonnet-plus-open-source trying to crack an algorithm he admits he isn’t smart enough to fully understand, got nowhere, then passed it to Fable and Opus and solved it in 20 minutes. The cheaper model cost him a full day and $500. The frontier model was faster and effectively free inside his existing plan. So the frontier model came out cheaper on hard cost and soft cost both.
Decagon founder Jesse Zhang made the same point in a post: when the problem is new, unbounded, or full of unknown unknowns, you use frontier models because they’re smart enough to figure it out. As the answer commoditizes and you already know the shape of it, you push it to open source. The nuance is that OpenRouter token share is misleading, because all the tokens can sit in one place while all the dollars sit in the other.
The cost pressure is already bending the AI CX market. The industry is standardizing around roughly 50 cents per resolution, down from a dollar, which means your model cost has to be 25 cents or less. Everyone is racing to open source to hit it, and Jason is seeing genuine plateauing in resolution quality as a result. That may push the best players back toward frontier models to get from 40% resolution of not-that-hard problems to 95% resolution of genuinely complex ones. What Jason loves about CX specifically: it’s on the good side of the chasm. The ROI is legible. Customers already say “I get it, I’ll pay 50 cents to resolve what used to cost me three dollars an email.” The high-class problem is that going from 65% to 75% resolution might cost two dollars a pop instead of one, and customers will still happily pay it. Finn just sold for $3.6B, so the market believes.
11. Every Tech Giant Is Racing to Become the Next IBM
Microsoft is putting $2.5B and 6,000 people into embedding engineers inside enterprise clients, a services play, days after Amazon made the same move, against the MIT finding that 95% of enterprise AI pilots deliver no measurable P&L impact.
Jason thinks it fails, and the reason is talent depth. SaaStr works with some of the best forward-deployed engineers at the leading vendors, and even at the hot companies the bench is one or two people deep. His proof point: a public-company vendor told SaaStr it would take three months to fix a bug where the AI still references a 2026 event that already happened, because the good FDE was out on paternity leave and nobody else could touch it. You cannot scale enterprise transformation on a two-person-deep bench. Throwing 10,000 people who were weak at customer success at massive enterprise problems is not the answer.
Rory agrees it will be slower but thinks it works in a limited, important sense, and he put it in one line: every technology company either goes bust or lives long enough to become the next generation’s IBM. Corporate America is oil-and-gas companies and banks. On the other side of the table sit Anthropic and OpenAI, product companies to their core. Somebody has to sit in the middle and do the change management, and that somebody is the mature enterprise-trusted vendor pivoting into services, exactly what IBM Global Services and HP did for 20 years. The pitch runs: “Those Silicon Valley people are scary and talk about the end of the world. We’ve sold to you for two decades. We’ll make this work.” It won’t be as good as the enterprise wants, but it beats the enterprise trying alone.
The nuance that makes these services businesses hard to build: when all you sold was a database, you needed a database expert. When you’re selling intelligent answers grounded in the customer’s own business, you need a technical expert and a domain expert together. Harvey deploys an FDE and a lawyer on every single engagement. That combination works at a high price point. Staffing that with B-tier and C-tier people at scale is a lot harder. And it feeds the single biggest question hanging over all the demand math: the rate of diffusion. AI has diffused faster than any technology in history so far. If the next 10x takes three times longer because corporate America can’t absorb it, the path from $4.5B to $40B to $80B in lab revenue gets a lot bumpier. That slows the labs down without stopping them, and how fast they can go is what everything else rests on.
12. Ashton Kutcher Leaves His Own Firm, and Employee Liquidity Becomes the Whole Game
Ashton Kutcher is leaving Sound Ventures to start a new deep-seed firm with Morgan Beller (ex-a16z, ex-NFX). Walking away from your own firm looks strange at first. Rory’s read is that it’s simpler than the gossip suggests, and there’s no dark story. Firms hold together because the brand and the reference are the asset, so even partners who can’t stand each other paper it over to keep the engine running. None of that applies to Kutcher, because he is one of maybe a handful of investors on earth whose personal name is the brand. Founders say “Ashton Kutcher is on my cap table,” not the firm name. Google Trends him against Sequoia and it isn’t close. When you’re that famous and that good (and the sneering stopped years ago, because he’s killed it), you can do a clean new thing without carrying the firm.
Jason, ever the solo GP, pushed back on principle: if the engine is working, the CFO is working, IR is working, and you can stand your partners, he’d rather just stay, because you can’t rebuild everything overnight. But that’s a personal call, and Kutcher doesn’t have to make it.
That thread led to the most actionable idea of the episode. ElevenLabs did a secondary at $22B. The valuation is unremarkable. What matters is that they ran a tender at all. With the IPO window now stretching to 12 years in some cases, tender offers are the new proxy for going public. So Jason’s question to any hyper-talented operator: why would you join anything you don’t have high confidence will be tender-worthy inside 24 months? Employees are sequential VCs, one shot at a time instead of 20 in parallel, so stock-picking is the job. The trick, exactly as it is for VCs, is to join something that isn’t doing tenders yet, get a healthy grant reflecting that, and land somewhere that starts doing tenders once you’ve vested 50% or 60%. Clay ran a tender at $5B, so it’s not only the Databricks and OpenAIs of the world. But there’s a real line below which regular liquidity disappears, and if there’s no path to it, life’s too short. Cliffs, at least, are already gone at the top labs.
Key Takeaways
- The Fable ban is resolved. The pre-approval precedent is here to stay. Shipping frontier models now runs through Washington, and that is a structural shift.
- The 5% stake is an anchoring move. If AI really is a $30 trillion labor threat, 5% only opens the bidding, and Bernie has already said he wants 50.
- Dilution stopped mattering to the best founders. When the prize is a trillion dollars and investors no longer block 1x exits, the old ownership math dies.
- Excess-compute pivots get rewarded, for now. Meta and SpaceX turned failed proprietary bets into neoclouds and got paid for it. Demand decides how long that holds.
- Nvidia is financing its own demand. Compute-now-pay-later is clean accounting and a leveraged bet on demand never softening.
- China owns open source because we locked them out. You can’t use Claude or OpenAI behind the Great Firewall, so of course they built their own.
- The frontier model is often the cheap option. A day lost to the n-minus-one model costs more than the token premium on the right one.
- The model companies are becoming product companies. The trillion-dollar services layer between the labs and corporate America is the prize, and talent depth is the bottleneck.
- Tender offers are the new IPO. For operators, joining a soon-to-be-tender-worthy company is the whole stock-picking game.
Quotable Moments
Jason Lemkin
“I spent 10 hours with the n-minus-one model and solved nothing. Fable plus Opus got me to the heart of the problem in 20 minutes. It wasn’t just more expensive to use the cheaper model. It was cheaper to use the frontier one.”
“I’m going to throw 10,000 people who were terrible at customer success into solving massive enterprise problems. Good luck with that one.”
“If I were a hyper-talented employee, I would not go somewhere without liquidity. It just doesn’t seem worth it today. Life’s too short.”
Rory O’Driscoll
“Six months ago you could ship software like a free man. Now you have to get permission from Washington before you do it.”
“It’s like rewriting Atlas Shrugged, where John Galt goes to Washington and says, ‘Why don’t you regulate me more? Why don’t you take more? Grab some of my stuff.’ What are these people thinking, volunteering for this?”
“Every technology company either goes bust or lives long enough to become the next generation’s IBM.”
Harry Stebbings
“Those that can do, those that can’t teach. And I always liked to remind my teachers of this, which is probably why I was so unpopular at school. And then you look at the ecosystem today: those that can do, those that can’t open a cloud business to sell excess compute.”
“It’s like: ‘Hi, I’m a wolf, but I’m a good wolf.'”
“Is this purely a marketing exercise? And who exactly are you marketing to?”
This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll.
