With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll
Anthropic just hit $30 billion in annualized revenue. Up from $9 billion at the start of the year. That’s 3.3x growth in four months. It took Salesforce 25 years to get to $30 billion. Anthropic got there in five. Maybe three, depending on how you count.
And here’s the part that should terrify OpenAI investors: Anthropic’s training costs are a quarter of OpenAI’s. A quarter. You can explain some of that away with focus. No video generation. No consumer image products. Less of the slop. But when you’re out-accelerating your competitor on revenue AND doing it at a fraction of the cost AND the competitor is cycling through its entire management team? That’s not a code red. That’s a double code red.
Meanwhile, SpaceX confidentially filed for a $2 trillion IPO that would be the largest in history. OpenAI bought a media company nobody asked for. YC kicked out Delve for breaking the founder code. And a two-person company used AI-powered marketing to hit $1.8 billion in GLP-1 revenue by doing everything your compliance team would flag in 30 seconds.
The numbers keep getting bigger. The teams keep getting smaller. And the AI build-out continues to rain money on everyone in its path while we all quietly wonder: when does the music stop, and who’s still standing?
Key Takeaways
1. Anthropic at $30 Billion Is the Most Staggering Growth Story in Software History
Adding roughly $10 billion of net new ARR in four months while still being capacity constrained. Claude still can’t finish chats sometimes. Every engineer at every tech company has been told to consume more tokens. And they still can’t sell as much as the market wants to buy.
The supply side tells the story. Anthropic isn’t just growing into demand. They’re allocating scarce compute based on revenue optimization. That’s why they pulled open Claude out of the base plan. Heavy agentic users were consuming enormous amounts of tokens on fixed-price plans. When you can sell more than you can produce, you stop subsidizing the heaviest consumers and start pricing closer to value.
“It’s exactly what anyone in economics would tell you to do,” Rory explained. “You deemphasize things that consume huge amounts of compute for small amounts of revenue. In OpenAI’s case, that’s video. In Anthropic’s case, it’s open Claude access on fixed plans. You’re going to see a continued trend toward pricing tokens closer to the value they deliver.”
The comparison to Salesforce is almost unfair. Salesforce is the largest cloud software company on the planet. Anthropic matched its revenue in a fifth of the time. And Anthropic is still constrained by how fast it can bring data centers online. The question isn’t whether they hit $50 billion. It’s whether the estimates from two months ago even belong in the same conversation.
2. Training Costs at a Quarter of OpenAI’s Makes This a “Double Code Red”
The Wall Street Journal leaked financial details on both Anthropic and OpenAI this week. The number that jumped off the page: Anthropic’s model training costs are roughly 25% of OpenAI’s.
Focus explains part of this. Anthropic doesn’t build video products. Doesn’t do consumer image generation. Doesn’t maintain the sprawling product surface area that OpenAI has expanded into. But when you combine lower training costs with faster revenue growth, the compounding effects are devastating for the competition.
“You usually don’t have both together against your competitor,” Jason said. “You’re out-accelerating your competitor AND your training costs are a fraction of theirs. That just compounds.”
Rory drew the Uber/Lyft parallel, but noted the critical difference: “In the Uber/Lyft struggle, Uber had the ‘we’re out-accelerating’ story, but they were spending every dollar they had to do it. In this case, Anthropic is out-accelerating the opposition while being more efficient on a bunch of interesting measures. That’s a scary fact pattern.”
If both companies were public right now, the hedge fund trade would be obvious: short OpenAI at $820 billion, go long Anthropic at $370 billion. Roughly the same revenue. Better trajectory. Stronger management continuity. Half the price. And by shorting one and going long the other, you diversify away the overall AI risk and isolate a pure relative performance bet.
3. OpenAI’s Round Was “Barely Real” and Anthropic’s Valuation Looks Far More Comfortable
Dig into the OpenAI round structure and the picture gets worse. The SoftBank money comes in tranches. They have to borrow to pay it. The Amazon money is tranched partly on IPO or AGI milestones. The Nvidia money is almost entirely compute offsets, not cash. Andreessen’s money was real cash up front. And they tacked on another $10 billion that appears to have been actual dollars. But the vast bulk of the round wasn’t cash.
“That’s not a sign of strength,” Jason argued. “When the majority of the round is not cash up front, classically that’s a sign of barely getting the round done. Why wouldn’t you want $140 billion up front?”
At current revenue trajectories, Anthropic at $370 billion feels substantially more comfortable than OpenAI at $820 billion. Not a guarantee. OpenAI hasn’t released updated numbers, and they did claim $2 billion monthly in their most recent disclosure. But based on what’s public, the last-round investors in Anthropic appear to have gotten a meaningfully better deal.
4. The OpenAI Management Reboot Is Risky but Necessary
The COO moved to “special projects.” The CMO stepped down. The CRO is out. The head of apps took a leave of absence. And Denise Dresser, formerly CEO of Slack, just got handed basically everything go-to-market after being at the company for a couple months.
The changes make sense in context. When your primary competitor has radically changed the competitive posture over the last six months, you don’t sit still. The “code red” declaration three months ago didn’t magically change the trajectory. Something had to give.
But bringing in the executive with the perfect LinkedIn and handing them a massive portfolio during a period of turmoil? Jason put the success rate at about 30%. “In my experience, bringing in Mr. or Ms. Perfect LinkedIn and giving them a massive portfolio when you’re in turmoil has about a 30% chance of success. When things are executing to perfection, it always seems to work. But in turmoil, there’s not a lot of time for the get-to-know-you tour.”
The deeper concern isn’t the individual departures. It’s the signal. As Rory put it, borrowing from Oscar Wilde: “To lose one parent might be an accident. To lose both smacks of carelessness.” OpenAI is getting into carelessness territory.
5. The TBPN Acquisition Was a January Deal That Wouldn’t Happen Today
OpenAI’s acquisition of TBPN generated a lot of noise. Rory’s take was blunt: buying a media company while declaring a focus mandate is “just insane.” Not because the dollars are material. Not because it will sink the company. But because it sends exactly the wrong signal.
“OpenAI is the most known company on the planet, maybe after Apple,” Rory said. “The CEO has met every world leader. They get constant attention. In terms of media minutes, there’s nothing left to get. If you were to pick the one company that doesn’t need media attention and does need to focus, it would be OpenAI.”
The bull case exists. Jason argued for a profitable public company with trapped balance sheet cash and no ability to increase marketing spend, buying a media asset at scale can convert balance sheet into marketing. The problem: OpenAI is neither profitable nor public.
But the real meta lesson is about deal timing. The outreach happened in January. A different world. Fidji was new and thought it would be a great brand play. It took months to close. By the time it did, the management team that championed it was already changing. “There’s no way that deal happens today,” Jason said. “It’d be dead not because it’s not a good idea, but because of management priority change.”
For founders: when you say no to an attractive deal, be sure you’re okay if it’s no forever. The odds that the VP who wants the deal is still there in 12 months with the same priorities approaches single digits.
6. SpaceX at $1.75 Trillion Would Be the Largest IPO in History, and the “Big Three” Dwarf Everything
SpaceX filed confidentially for an IPO targeting $1.75 trillion. That includes xAI (formerly Twitter), which was rolled in earlier. Revenue of $15-16 billion. $8 billion of EBITDA. At $2 trillion, that’s 125x revenue.
But the number that should mess with every venture investor’s head: the combined IPO value of SpaceX, OpenAI, and Anthropic will exceed every other IPO of the last 20-25 years. Combined. All of them. Every deal Rory’s firm has done, every YC exit, every Sequoia win. All of it, summed up, would be smaller than these three.
“Psychologically, the thing about a power law they don’t tell you is you can have the third-best outcome in venture history and be one-tenth as large as the largest outcome,” Rory said. “If you’re going to let that in your head, it’s just going to be very tough business for you psychologically. You can have a life-changing event that’s down in the noise of $10-20 billion outcomes.”
Will SpaceX actually trade at $2 trillion? The sum-of-the-parts analysis gets you to a much lower number. The gap between that and $2 trillion is pure Elon premium. Elon himself said on X that $2 trillion was too high. But whatever number he picks, he’ll will it into existence on IPO day. The underwriters won’t be able to argue for more than five minutes. Between 30% retail allocation and whipped-up demand, the target valuation will hold for at least one day. Whether it holds for 30 days is a different question.
The xAI inclusion is the wild card. Less than 12 months ago, SpaceX alone was valued at $400 billion in a meaningful transaction. xAI is number four at best in the LLM space, burning $12 billion a year. That asset probably won’t appear on pages 1-3 of the IPO slide deck. The entire addition may have been net negative to the investment thesis.
7. Doug Leone Returns to Sequoia: Gravitas in a Time of Change
Doug Leone came back to Sequoia in an investing capacity. Not leadership. Pat and Alfred still run the firm. But Doug is writing checks again.
Jason read it as an LP play. “LPs are uncomfortable with change. They say they’re looking at the new generation, but they are comfortable when old leadership is still actively involved.” Having made one transition to Roelof that required an abrupt second transition, the first one clearly wasn’t fully successful. Bringing Doug back provides continuity.
Harry pushed back: LP appetite for Sequoia has never been stronger. The real driver is competition. Founders Fund has Anduril and SpaceX as brand tailwinds. “Doug is the ultimate closer. When Doug Leone goes to that meeting, whether it’s Christian Hacker at Trade Republic or the team at Wiz, he closes the deal.”
The Sequoia ethos came through clearly: even if they’re winning on every round but one, they’d be asking how to win that round too. “Why not get an 11th great player?”
8. YC Kicked Out Delve Because They Broke the Code
Delve, a YC compliance startup, got ejected from the YC community this week. Not just for using AI to fabricate SOC 2 audits. That was bad enough. But the kill shot was stealing open source code from a fellow YC company, Sim Studio, not attributing it, and claiming it as their own product. To their own batch mate’s customer.
“I think it was the part two that did it,” Jason said. “You took the open source code from your batch mate and said it was your own software. And they were your customer. You can’t hand-wave that one away.”
Rory framed the math: YC does 800+ companies a year. Statistically, if you index to the general population’s incarceration rate, eight of those founders per year will commit some kind of crime over the course of their lives. At that volume, fraud is inevitable. The question is how you handle it.
The answer: you can’t police it a priori at scale. But when someone breaks the community code, you have to be ruthless. “It’s like the Old West,” Rory said. “You broke the code, you’re out.”
9. Open Router at $1.3 Billion: Epic Startup. But Low-ACV Investments Can Make You Nervous
Open Router hit $50 million ARR (up from $10 million in October) and raised at $1.3 billion. The product is brilliant. A single API that dynamically routes to 50-60 different LLMs based on your workload. Charges about 5% of what you pay the model provider. Elegant, cheap, and works.
But Jason flagged the tension that keeps him up at night with low-ACV AI investments: “I suspect the cheapness is why it’s not worth $10 billion — yet. When you have a relatively low take rate from such high GMV, you do naturally get a little nervous about the true TAM.”
The math is instructive. If Anthropic’s $30 billion enterprise business generates $300-400 billion in API calls by 2029, and 10-20% of that goes to open-source models routed through something like Open Router at 5%… you get to maybe $2 billion at 100% market share. That’s the ceiling, not the floor. For now. For the moment.
The Twilio comparison is apt but cuts both ways. Twilio was an interface between app builders and telco complexity with 20-40% gross margins. Open Router only books the net 5%. Revenue Cat, which Jason invested in and owns 50% of mobile subscription management, faces similar dynamics. Even with dominant share, the low take rate creates a structural revenue ceiling.
“Open Router could be one of the greatest $200 million ARR companies,” Jason said. “My learning from Revenue Cat is that sometimes when your product is very cheap in a massive market, the notional BIPS map doesn’t translate to massive real-world BIPS map. You have to expand far faster than others to make up for it.”
Rory’s counterpoint: Visa and Mastercard only get 15-20 basis points on every dollar humans spend. And they’re the two best financial businesses on the planet. The resolution? Open Router has to go truly multi-product. Not a feature. Not an enhancement. Five distinct products. Run the Rippling playbook at AI speed.
10. Supabase at $10 Billion Is Riding the Only Trend That Matters: Agents Create More Databases Than Humans
Supabase was founded in 2020 as another fork of Postgres. Nobody would have predicted a $10 billion outcome from “let’s make an open-source database easier to deploy.” Then AI happened.
Agents need to spool up databases without humans. Supabase built exactly that. They lost Replit to Neon (which Databricks bought), but everyone else standardized on Supabase. Lovable, Emergent, and a growing list of vibe-coding platforms white-labeled it.
“More databases are now being created by agents than humans,” Jason said. “We’ve already crossed that line. Database is a fundamental category of software. The number being created is an order of magnitude more than 12 months ago. Why wouldn’t you bet on the leader?”
The trajectory mirrors MongoDB in the SaaS era. Mongo was the right database for cloud. Supabase wants to be the right database for vibe-coded and agent-built apps. The playbook is the classic SaaS-and-cloud playbook, but running at 5x speed.
Rory’s caution: white-labeling to five or six vibe coders won’t be enough over the medium term. They’ll need to expand functionality, add complexity, and make themselves indispensable before the platforms decide to build their own. “In 10 years, someone will be saying, ‘Oh my god, those legacy Supabase products are almost as bad as MongoDB.’ But that’s just the movie. And this is Supabase’s time to crank.”
11. The Mercor Hack Is a Preview of What’s Coming for Every B2B Company
Mercor, the AI data labeling company, got hit by a ransomware group called Latipus. All private data exposed. Meta reportedly paused or dropped them as a customer.
Jason had recently been with a hyperscaler executive when a minor third-party security issue arose. The executive’s message was chilling: “There’s not much higher on our list with partners. We have no tolerance. It’s not worth it.”
In the old days, through 2023, vendors always got one pass. Even for the worst breaches. You’d get called into the CISO’s office, get yelled at, and survive. That era is ending.
“AI can now find every unsecured database, every misconfigured endpoint, every weakness in seconds,” Jason said. “My CTO told me years ago: ‘The only reason we haven’t been hacked is no one cares about us.’ With AI, they can hack anybody they want.”
The broader point connects to Sam Altman’s warning this week about massive AI-powered cyberattacks. Most B2B companies are relying on a hodgepodge of open-source security tools, barely monitored, with understaffed teams under profitability pressure. It used to be hacker farms of people in the Philippines or Russia. Now it will be hacker farms of AI agents running 24/7.
And this is why the security stock selloff after Anthropic’s product launch was backwards. Anyone cutting their security budget in 2026 is missing the point entirely. The attacks are getting more ferocious, more automated, and more relentless. And more of what matters in our lives is online every year.
12. The $1.8 Billion Two-Person GLP-1 Company: Fraud and the Future Wrapped Together
A two-person company (Medvy) scaled to $1.8 billion in revenue selling GLP-1 weight loss drugs using AI-powered marketing at scale. They used deepfakes. They made improper representations about doctors. They did all the crappy affiliate marketing tactics people have been doing for 20 years, except they did it with AI and reached everyone.
Jason’s take: “Obviously, people are so excited about the one-or-two-person billion-dollar company that facts were glossed over. But if you let’s not over-glamorize this company that may be at the edge of fraud in many ways. The fact that they could use AI to scale this with two people, we are seeing the future.”
The pattern repeats across every technology cycle. Affiliate marketing originated from porn and Viagra. SEO was a dark art that only lead-gen companies understood before Corporate America adopted it. The best marketers always operate at the frontier, and the frontier is always uncomfortable.
The GLP-1 economics explain why it worked: $300-400 per lead delivered to a customer who already wants to buy a fungible product in insatiable demand. That’s not a business. That’s a marketing arbitrage. But the tactics are real, and they’re coming for every category.
“I talk to so many CMOs who are still struggling to run the 2023 playbook and falling further and further behind,” Jason said. “In two years, only the creakiest companies will be doing marketing the way they do it today. It makes no sense.”
Rory connected the dots: “The marketing tactics that look very dark-arty today tend to be adopted by everyone over the next 10 years. Everyone in Corporate America is now an SEO expert, whereas 20 years ago it was a dark art. The correct play is to be the software provider helping the normies tool up for agentic marketing.”
13. VCs as Enablers: The Kumbaya Problem in the AI Era
Jason raised a concern that applies to every growth-stage board in B2B right now: too many VCs are running a pre-AI enabler playbook where they offer “great job, keep at it” encouragement to portfolio companies facing existential AI competition.
“I’m an investor in a company that’s crossed nine figures in revenue, but it’s hitting massive AI competition. They have a new investor on the board who doesn’t really know the space, isn’t close to the AI changes, and every email is ‘great job guys, keep at it.’ Don’t you understand the disruption?”
The problem is acute because in AI, if you stumble, you lose. Barriers to entry are low. If five companies start the race and one doesn’t stumble, that one wins. There’s no coming back from a year behind in a market that moves this fast.
Rory threaded the needle: “You’re not a useful board member unless you understand what the company does, how it compares to direct competitors (ideally with hands-on product experience), and you find a way without being a jerk to keep the company honest about where they are relative to the competition.”
The internal slogan at Scale Venture Partners: “Not founder-friendly. Founder-honest.”
Quotable Moments
Harry Stebbings
“Anthropic is accelerating faster than ever, and OpenAI is having more challenges than ever, all at once. Have we ever seen a bigger chasm between where they’re at?”
“With the greatest respect, the insatiable appetite from LPs for Sequoia has never been more prominent. I don’t think Doug coming back is about LPs. It’s about competition. Founders Fund has Anduril and SpaceX as tailwinds. Doug is the ultimate closer.”
“Someone in the comments said, ‘Harry, you shill Mercor so you can’t not talk about them when there’s trouble.’ I don’t like to do that. But you have to address it.”
Jason Lemkin
“Anthropic got to $30 billion in five years. Maybe three. You usually don’t have both together against your competitor: you’re out-accelerating them AND your training costs are a fraction of theirs. That just compounds.”
“When the majority of a round is not cash up front, that’s classically a sign of barely getting the round done. Why wouldn’t you want $140 billion up front?”
“My CTO told me years ago: ‘The only reason we haven’t been hacked is no one cares about us.’ With AI, you can hack anybody you want. It used to be hacker farms of people. Now it’s going to be hacker farms of AI agents running 24/7.”
Rory O’Driscoll
“The combined IPO value of SpaceX, OpenAI, and Anthropic will exceed every other IPO of the last 25 years combined. The psychologically weird thing about a power law is you can have the third-best outcome in venture history and be one-tenth as large as the largest.”
“Anthropic is out-accelerating the opposition while being more efficient on a bunch of interesting measures. If you’re running the game theory and you’re the other guy, that’s a scary fact pattern.”
“It’s not that we don’t care about defensibility. It’s that you don’t have the luxury. Most of the time right now, the barriers to entry are low. If you stumble, you lose. And the barrier to entry is speed.”
This post is part of the ongoing 20VC x SaaStr collaboration. For more, head to saastr.ai/ai-mentor.
