We did the 20VC x SaaStr episode this week with Harry Stebbings and Rory O’Driscoll, and by the end something funny happened. Every single segment, no matter where we started, collapsed into the same question:
What is the right split between humans and tokens, and what happens to everyone on the wrong side of it?
Nvidia, the OpenAI S-1, the SpaceX S-1, the layoffs, the infrastructure rounds, vibe coding. All of it was one conversation wearing different outfits.
Here’s what mattered.
Nvidia did $81.6B in revenue and $58B in profit in a single quarter, and the stock barely moved
Start with the profit number, not the revenue. $58B in a quarter makes Nvidia the most profitable company on the planet. Google clocks in around $100-120B in profit for a full year. Nvidia did half that in three months. Run that rate out and you’re looking at $200B+ in annual profit.
So why did the stock go flat on a quarter like that?
Because markets price the delta, not the news. Nvidia already 3-4x’d a year and a half ago when corporate America internalized the scale of AI capex. The business grows 80%. The stock grows 20%, because the market is saying “this is amazing and we don’t think it 5x’s from here.” At a mid-20s P/E, that’s a market that’s pretty happy.
The number that actually matters for everyone reading this: Jensen says $3-4T in AI capex by 2030. Extrapolating the last three years of growth gets you there mathematically. The question isn’t technical. It’s economic. Is the ROI there on the next $2T?
The ROI question is the entire ballgame for 2026
This is the heresy you’re not supposed to say out loud in San Francisco. Rory put it well: raising it feels like wandering around the Vatican asking if the Pope is the right guy. But it’s the only question that determines whether you get to $3T in capex or not.
The signals are already splitting:
- Uber’s COO said they burned a full year of Anthropic credits in 4 months and the productivity gains weren’t measurable. He thinks they’re there. He can’t prove it.
- Microsoft is reportedly moving off Opus, calling it too expensive.
- Door Dash, founder-led and aggressive, is spending everything and says the engineering savings clearly justify the token spend.
That’s the whole story. Companies already running at $1-2M in revenue per employee will token max until there’s no tomorrow. If you’re already hyper-efficient, you find more ways to use AI. Larger, more traditional orgs protecting 40% gross margins get skeptical fast, especially as Anthropic and others raise prices and the free-experimentation era closes.
Here’s the part nobody wants to confront. Once AI spend starts eating real chunks of payroll at big companies, the green-eyeshade people show up. Someone has to stand up and say “we’ve never done a layoff in 5 years, we’re about to cut 10% of our people, are we sure we’re getting the ROI?” When you’re spending $3M on tokens you can be lazy about it. When you’re spending $300M, somebody needs to know.
Anthropic margins went from 38% to 70%. Profitable. $44B ARR. And it lapped OpenAI on revenue growth.
Anthropic did as much in Q1 as it did in all of last year. Gross margins went from 38% to 70%, with a projected $559M operating profit in Q2. The trajectory was set: negative gross margin in 2024, positive 34% last year, 70% now. High growth plus improving margins makes profit close to inevitable. When you add $5B of revenue at any decent margin, it’s just hard to lose money.
OpenAI’s Q1 was bigger in absolute terms but only about 30% of its prior-year revenue. One company growing 10x. The other growing 2-3x. Play that out a couple quarters and Anthropic is visibly, obviously ahead: bigger, growing faster, and profitable. Pareto dominant on all three vectors.
That’s why OpenAI confidentially filed its S-1 at an $852B-$1T valuation for a Q4 listing. When you’re number two, you do not wait for number one to go public and then show up three months later as the smaller, unprofitable version of them. You go first, you tell the story, you remind everyone you built the category. Outside the valley you say “Claude” and people stare at you. You say “ChatGPT” and everyone knows. They earned the right to tell that story. They should go now.
Does Anthropic change anything in response? No. That’s the point of being number one. You don’t agonize over what the other guy does.
The “it’s just COVID overhiring” take on layoffs is wrong, and the math proves it
This was the dumbest take on the timeline this week. The idea that the ClickUp, Cloudflare, Intuit, and Coinbase cuts are all just unwinding 2020 hiring.
Run the numbers. Natural attrition runs 15-25% a year. Compound that over 5 years and you’ve already churned through far more people than these layoffs represent. So blaming it on overhiring isn’t just convenient, it’s not mathematically true. You had 5 years to manage out your low performers. That’s what performance reviews are for.
What’s actually happening is cleaner and more honest. ClickUp cut 22% so they could pay $1M to high performers. If a 10x engineer is now a 100x engineer, or a $1M-bookings rep is now doing $2.5M, you have to pay them. You funnel the comp to the people delivering the value.
$2M in revenue per employee used to be the someday number reserved for Apple. It’s becoming the new normal as you scale. And when it does, your high performers should make 3x what they used to, not 40% more than your mid-pack. The gap should be 5x.
The lingering worry: instead of agentic skill becoming Excel-level table stakes in 18 months, the agentic experts get another 10x better and everyone else falls further behind. Science advances funeral by funeral. Retraining past a certain age is hard, and it must be pretty rough to have to do it or get left behind.
SpaceX filed the largest IPO in history at ~100x trailing sales
Rory’s line: this could be the Geocities deal of the AI era.
Run sum-of-the-parts on the three businesses. A great, stable, boring launch business growing 10-20%. A genuinely strong Starlink business growing 30-40% with a much bigger TAM. And xAI, which at filing was a $15B hole in the ground with no obvious revenue. Add it all up and you’re so far below the proposed valuation that the only thing closing the gap is the Elon premium, which doesn’t show up in a prospectus.
But here’s the financial engineering you have to respect. SpaceX built Colossus fast, then rented it to Anthropic for $1.25B a month. That’s $15B a year, on roughly $19B of total xAI capex. He covered his nut on a product he built 18 months ago. Take a great private launch-and-comms business, bolt on a Geocities-era AI valuation, fold in the Twitter acquisition that’s down 50% since he bought it, and roll the whole thing public. It makes no sense to anyone except the people getting bailed out. And it might still work, because the one human on earth who’s earned the right to say “give me a trillion dollars, I’m going to make the bet” is Elon.
The thread that ties the conglomerate together is data centers in space, enabled by the launch business. I don’t ascribe a high probability to it. But the narrative is coherent if you believe the next thing happens. That’s always the way.
Pick-and-shovels keep printing: Exa raised at $2.2B building search for agents
The cliché is true. The pick-and-shovels of the agentic buildout are a good place to be. Exa raised at $2.2B. Open Router raised around $150M at $1.3B. There’s a whole layer of infrastructure one level below the foundation models that’s genuinely interesting.
The insight that made Exa click for me: agents need a completely different toolset than humans. Agents don’t hop on Google. They don’t create files in Dropbox. They don’t join Zoom calls. They don’t use a traditional CRM. But they absolutely need current information, and they need it in a structured way the foundation model doesn’t ship in the box. Exa nailed that. It’s a product that has essentially zero use without agents, which makes it a clean look into the future. The idea that we’ll all manage more agents than humans isn’t hype. It’s real.
The risk, and I’ve made a version of this mistake already: volume. These tools only get used for some apps and some workflows, not all of them. If you pick the primitive only 10% of developers need, you’ll command a premium and have a tiny, sad market. The winners are the 5-6 things every agent needs. Search. Model switching. Observability. A database. Over the next few years, revealed preference will tell us which ones made the cut.
And the volume is coming. At Databricks’ Neon, over 90% of databases are now built by agents, not humans. The number of apps, workflows, and agents being built is exploding at a rate we’ve never seen. It may not save Dropbox. But software creation is going vertical.
The closing rage bait: “I replaced my $600K Salesforce contract with a vibe-coded CRM in 3 weeks”
A CEO said three things: he replaced Salesforce with a CRM he vibe-coded in 3 weeks, he’ll kill 80% of his internal B2B tools, and he wouldn’t change his usage if Anthropic doubled prices.
The first one is dated 2025 rage bait, and it’s unhelpful. We moved to headless Salesforce. We don’t log into it anymore. We run our own autonomous agents on top of it. Could we rip it out and build our own database? We’re small enough that we could. It’s not worth it. Not worth maintaining the database, not worth rebuilding the connectors to our 10 external agents that are all natively built into Salesforce, not worth giving up Agentforce.
If you’re a vertical company, your team is small, the CRM doesn’t need to be collaborative, you get zero value from the thousands of apps that integrate with the incumbents, and you have time to burn because you’re not actually saving money? Then sure, vibe it away. But of all the real threats to Salesforce, HubSpot, Monday, and Atlassian, this one isn’t in the top 10. It doesn’t even explain the B2B software repricing. Other things do.
The third statement is the interesting one. If you’d genuinely pay 2x for Anthropic, you’ve just admitted the ROI is so good you’re underspending. So go tell your team to use twice as much, because there’s a return on the next project on your list. Do it.
But here’s where it actually lands for those of us already running lean. The constraint stopped being money a while ago. Our AI VP of Marketing costs $257 a month. I don’t care if it’s $500. The bottleneck is idle agents and human throughput. Every morning at 7:13am our AI VP of Marketing pushes three ideas we should be running, plus a note that we didn’t get to yesterday’s. You cannot implement 21 good ideas a week. The AI is now generating more good ideas than a human can process.
The answer to that isn’t fewer tokens. It’s hiring another human. But not a $125K one. I need a $1M person who can take those ideas and turn them into $5M. I’ll hire that person today. The catch is you have to actually be that person, not just want to be.
Why this was really one episode, not eight segments
Every topic came back to the same place. Nvidia is the picks-and-shovels winner of the humans-vs-tokens question. The layoffs are the human-side consequence of the same efficiency math. The infrastructure rounds are a bet on how many agents get built and how much they spend. Even the vibe-coding fight is a question about what work moves from humans to tokens and what stays.
In one frame we focused on the winner, Anthropic. In another we focused on the loser, the displaced employee. Same discussion the whole way through.
Quotable Moments:
Rory O’Driscoll:
“It could be the Geocities deal of the AI era. I love the S-1. I think it’s all madness. I wouldn’t buy a share. I just love the optimism.”
“The stock moves not on the total news but on the delta news. They beat slightly, it’s gone up slightly. A year ago it was a credible question whether the capex was sustainable. Now the market’s pretty happy.”
“Science advances funeral by funeral. People don’t learn and change their views. The people with the wrong views just die.”
Harry Stebbings:
“If you have to say you’re a people person, you’re not a people person.”
“It’s always about AI one way or the other, isn’t it? Directly, or our growth slowed because we don’t have an agentic product, or we’re not being renewed because they need budget for Anthropic tokens.”
“There’s a quantum reduction in risk when you go from no revenue to revenue, and then a linear reduction thereafter. The sweet spot is early product-market fit. The amazing thing about these AI companies is that moment used to take two years. Now it’s weeks.”
Jason Lemkin:
“The issue isn’t more money. The issue is idle. Our agents are idle. We don’t have enough brain cycles. The AI gives us more good ideas every day than we can process.”
“A $125K laid-off ClickUp-er is worth nothing to me. I need a million-dollar person to process these ideas and turn that into $5 million. I’ll hire them today. But you’ve got to be the million-dollar person, not just want to be.”
“You want to invest the minute a hot AI startup blows up. The minute it blows up, you want to get the DM and just wire the money. That’s the play.”
Have a question for Dear SaaStr? Submit it at saastr.ai/ai-mentor
