With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll
Every week, Harry, Rory, and I sit down to argue about what’s actually happening in AI and B2B. This week I was back from two weeks in China, where it’s a parallel universe of DeepSeek and Gemini, and where I finally understood the sovereignty argument in a way I never did from San Francisco.
The throughline of the whole episode was the same force showing up in five different costumes. Capital and talent are flooding into AI so hard that everything else bends around it. DeepSeek raised $7.4B at a $50B price where only the Chinese government gets voting rights. Google lost two of the most talented researchers alive to Anthropic inside 48 hours. DRAM prices are up 90 to 95% in a single quarter, which means your next iPhone costs more because of a data center in Tennessee. And Wall Street is finally asking the only question that matters: who is actually going to pay for all of this?
Here’s what we got into.
Top Takeaways
1. Google Lost Two Generational Scientists in 48 Hours. The Whole Story Is Momentum.
Inside two days, DeepMind lost Noam Shazeer (co-author of the original attention paper, character.ai, brought back to Google in a clever multi-billion-dollar acquihire) and John Jumper (Nobel Prize winner, co-creator of AlphaFold). Both went to Anthropic.
Harry’s read was the simple one: the best researchers alive only want to work on exactly what they want to work on, and the new labs can promise them that. I think that’s true but one-dimensional. Rory caught the second dimension, which is the frustration of not being able to ship. Google had a ChatGPT alternative and the bureaucracy smothered it while OpenAI jammed theirs out the door and took the lead.
There was a rumor that Anthropic had a secret breakthrough that pulled Jumper over. Rory was skeptical, and the reason is the lag time. In core LLM work, the gap between invention and revenue is one to three years. In medical work, it’s ten to fifteen. Protein folding collected its Nobel and still has no drug in production. So this isn’t “join in the next 90 days or miss it.” It’s a multi-year science bet and a question of where you want to spend the next five years.
My honest aha was momentum. When you start winning, everything goes your way. The talent, the breaks, the projects. Winners win and they compound until something breaks the chain. Anthropic and OpenAI can let people work on whatever they were promised because other people spent $300B in capex on their behalf. They have a luxury the incumbents don’t.
And to be fair to Google: over the last 18 months, only two of the Mag 7 outperformed the S&P, Nvidia and Google. They are very much in the frame. But you don’t wake up and try the new Google harness or the new Google coding tool. You try Claude Code. You try Cowork. You try OpenAI. Being number three in innovation beats being Microsoft or Meta and saying “we spent $70B, we might have something next year.” It’s still number three.
2. Why #3 in AI Is the Most Dangerous Place to Be
Tech markets don’t end up with millions of players. They end up as tight oligopolies: a leader, a number two, and maybe a three and a four, with the vast majority of revenue going to one and two. Cloud went AWS, then Azure, then Google Cloud.
What makes number three uniquely exposed right now is routing. Three months ago it wasn’t clear how big a deal multi-model routing was. Now everything except the smallest startups is routing workloads to different models. And here’s the problem for a closed-source number three: number two is usually simpler, number three is usually cheaper, and the cheaper slot is exactly where open source attacks. Google Cloud a generation ago won by being cheapest and simplest. Now your number three closed-source model is competing for that same slot against six Chinese open-source models being subsidized into the ground.
Rory’s frame: when an industry gets ground down by a 5x-cheaper alternative, the number one guy makes a little less, the number two guy makes a lot less, and the number three guy goes bust. Google isn’t going bust because it has the balance sheet. But the downward pressure on profitability is real, and the implicit conclusion is brutal: it’s now almost impossible for a closed-source number four to emerge and catch up. The market is set.
This morning I got an email from Anthropic telling me my prompt cache hit rate was low. That’s not a Gemini shot. That’s Anthropic fighting open source directly, getting you to cache prompts so heavily that they can be cheaper than open source. That’s the actual battlefield.
3. DeepSeek’s $7.4B Round at $50B: Only the People Who Don’t Need Votes Get Them
The round itself is wild. $7.4B at roughly $50B. The founder is committing about 20B yuan himself, close to $3B, around 40% of the round. Fewer than ten investors including JD.com. None of them get rights. The only entity with governance control is the Chinese state.
Rory nailed the punchline: the only people getting voting rights are the only people who don’t need them. The Chinese government doesn’t need votes. They have sovereignty and an army, and as we saw with Meta, they can make you hand back your $2B after you’ve taken it. So the structure isn’t surprising at all.
Two weeks in China made the sovereignty argument real for me. Anthropic and OpenAI intentionally block access there, even in Hong Kong. So you live on DeepSeek and Gemini. DeepSeek is intentionally crippled inside China, can’t search the web, trained on different data. This is China deciding it will not be reliant on Anthropic and OpenAI to run the next-generation economy. Whatever it costs to subsidize that, $10B, $20B, $50B, it’s a rounding error next to an aircraft carrier.
And before anyone gets too smug: we’re doing a version of it too. Anthropic couldn’t ship its most recent model until it satisfied the U.S. government. Leopold’s situational awareness called this. Both governments now treat this as existential. China solves it their way, we solve it ours, Europe pays a tax to be first in a market where it would otherwise be fourth.
4. The $725B Question: Who’s Actually Going to Pay for AI?
Goldman projects $7.6 trillion in cumulative AI capex from 2026 to 2031. David Cahn’s Sequoia piece started at $600B. Now it’s the $725B question.
Here’s the plain-English version. The hyperscalers are spending roughly $700B a year in capex. You only do that if you expect revenue to exceed expenses. So at some point someone has to spend $700B a year in revenue for the industry to make a buck. Right now total AI revenue is well under $100B. So AI is pulling in maybe $100B and spending $700B a year. That’s not a great business yet.
And it got more aggressive, not less. A year ago capex was about 60% of Mag 7 free cash flow. Now it’s 120%, and they’re borrowing to fund it. That’s the classic bull market trap: you can be intellectually right that this is stretched, and the narrative keeps going for years.
Run the math forward. Round to $1 trillion of revenue needed, because you have to pay for electricity on top of capex. For companies to hand over $1 trillion, they need more than $1 trillion of value back. Total U.S. labor spend is under $20 trillion. So you’re talking about 7 to 8% of the labor force getting replaced by tokens for the math to work. That’s a dauntingly high bar, and the only way that last capex dollar earns a return is through enormous productivity gains and the labor displacement that comes with them.
There’s also a parity tax nobody wants to say out loud. If everyone deploys the same AI, the cost savings get competed away and show up as lower prices, not higher profits. Every bank adding ATMs at the same time didn’t make banking more profitable. But the company that doesn’t adopt is dead. That’s why you lean in even when the industry-wide return is questionable. Your team has to be 15% leaner next year, not because the ROI is proven, but because your competitor’s will be.
5. Gross Margin Is the New Growth (And Why That’s Only Half True)
A YC partner posted that the most common reason good companies fail to raise their A or B right now isn’t growth, it’s margin. Founders doing real revenue, growing fast, but once you remove delivery costs there’s nothing left. Investors fund the margin, not the revenue.
I liked it. Rory pushed back hard, and he’s right historically. Companies with brutal gross margins and hypergrowth have been getting funded and have been pulling it off. That describes the foundation models, the inference providers, and the coding agents. Cursor got to $60B with negative gross margins because it grabbed the ground. The bet that “my margins are terrible but my growth will cover it” sounds stupid when you say it and has been 100% correct.
What’s probably changing is timing. That bet made the most sense in the big-bang phase of the last three years when you went from nothing to 10x growth and it paid to grab territory. Investors now seeing the next generation may be getting more disciplined on margin. And all three of us are sitting on a deal from the last 12 to 15 months where inference was the marketing strategy, the margins were negative, and we’re no longer sure we get right-side up. When things slow down, nobody acquires a gross-margin-negative adjacency for fun. Those companies just fail unless they hit escape-velocity growth.
6. Menlo Raised $3B After the Anthropic Win. The Surprise Is It Wasn’t More.
Menlo, 50 years old as a firm, raised a $3B fund on the back of one of the great venture bets of the era. My first instinct, which Rory rightly shook his head at, was “that’s all they could raise.” Wrong read. The real answer is they’ll raise another $10B or $20B in SPVs and sidecars, the way they did for Anthropic. The headline fund size isn’t correlated to lifetime deployment.
There’s a structural elegance here. Provided your check size lets you play in any round, two separate $1B funds give you higher risk-adjusted returns than a single $2B fund, because you get less cross-deal aggregation and more probability of a clean winner in at least one. More importantly, the number of companies that can absorb a $1B or $2B check is tiny even on a decade-by-decade basis. In Q1, something like 70% of venture dollars went to four or five deals. Raise a $10B fund and you’re signing up to jam money into those few deals and nothing else. A conservatively sized fund plus deal-by-deal SPVs on the anomalies is the better model, as long as you can spin the SPVs up on one WhatsApp message. The worst LP is the one who promises a blank check and then wants to meet the founders when you call.
7. Accenture Is Down 40%: AI Is Eating the Consulting Business
Accenture dropped 19% on top of an earlier 20% decline. Down about 40% on the year. And it was supposed to be an AI beneficiary.
Two things are happening at once. The business of helping companies adopt Gen AI is exploding, because every company in America needs help. But that’s a small slice. The core 90% is traditional consulting and systems integration, and few markets are more prime for AI disruption. It’s white-collar work that’s already been outsourced, which is the tell. Anything a company was willing to send to India, it’s willing to send to AI. Run down the BPO spend and you’ve got a map of where AI wins, and Accenture sits on top of it.
The mechanics are vicious. An SAP or Salesforce implementation that billed $20M to $40M now goes for half because LLMs do the requirements-gathering, the SOPs, and the simplistic deployment code. When I was a VP at Adobe, there was an entire floor of Accenture deploying Salesforce for five years. That whole motion is being compressed. Databricks claims it can do a data lift in 30 days that used to take Accenture five years. We’ve been stuck on Marketo for five years, our worst software, and Salesforce just lifted us off it in a couple of weeks with no humans.
That lift is a moat destroyer. When an LLM can migrate you from one vendor to another in days, switching costs collapse. I sat in a pitch this week where the founder went on about their moats and I lost interest immediately. Your moat can be LLM-lifted away.
The deepest problem for Accenture is the business model. There’s only one thing worse than a seat-based model, and that’s a model based on bodies. If you bill out 100 people at $500K and suddenly only need 40 people plus AI, your entire margin structure collapses and your position at the top of the heap goes with it. That’s the opening for AI-native SIs to walk in and say “you got a $80M bid from Accenture, we’ll do it for $15M.” That gets a CIO’s attention. AI for SI is one of the most interesting places to invest right now.
8. We Built an AI VP of Finance in China. “Master of Agents” Is the Only Job That Matters.
While we were in China, Amelia built an AI VP of Finance in a single-digit number of hours, and it’s already better than any human on our team. It creates the quote, builds and ships the contract, gets it signed, updates the opportunity in Salesforce, logs into Bill.com, sends and follows up on the invoice, gets it paid through Brex, then logs into QuickBooks and makes our books accurate for the first time in ten years. The models aren’t even tuned for this workflow.
The trigger was human. A person forgot to invoice $80K of revenue at SaaStr AI Annual, gave no reason, and we wrote it off. So we built the agent. For us, agents don’t replace humans on the team. They do the 90% of every job that humans are unwilling to do: the follow-up, the Salesforce update, the proper invoice. As a side effect, variable contractor spend can fall 50 to 60% without even trying to save money.
Harry pushed: if it’s this easy, why isn’t everyone doing it? They can. They just don’t have the mindset yet, because they haven’t spent a year vibe coding and don’t know what’s possible. Brandon at Rippling posted that training agents will be the largest job category in five years. Harry’s counter was sharp: we said the exact same thing about prompt engineers, who were making $150K out of college 60 weeks ago and whose skill is worthless today.
He’s right that the skill keeps getting redefined. You don’t need a prompt engineer anymore. You need a master of agents who knows what they’ll get, where it breaks, where it gets lazy. Last night our AI VP of Finance admitted it didn’t fully read a contract. We asked why. It said it didn’t have a good answer. Being a master of agents is not throwing your monitor out the window at that. It’s knowing Sonnet goal-seeks and tries to skip complicated steps, then rewriting the instructions so all contracts get read start to finish, and being comfortable that it still won’t always comply. In a year the harnesses may be good enough that you don’t have to do that. The rate of change is the whole point.
9. Work From Home Is White Collar Fraud
Ryan Petersen’s clip, half-jokingly, that work from home is white-collar fraud got 5.5 million views. My read: dated. Not wrong, dated.
A lot of the work-from-home era was working 15 to 20 hours a week with a lot of distraction, and I saw it on my own team. But the reason it’s dated is that the companies I want to invest in aren’t hiring those people anymore. The way you build a startup to your first 100 or 200 employees has radically changed and is under-discussed. Sixty weeks ago, telling people to work seven days a week was a toxic thing for a founder to say. Today it’s how you build a winner. You cannot win your market with people working 20 hours a week.
Let me be direct, because empathy is exactly the word I think about here. I want small, high-paid teams that work in the office six-plus days a week, paid top of market, with double the equity because the teams are smaller. I’m not interested in anything else, and it’s not because I lack empathy. It’s because the alternative fails.
It is an endless series of sprints, not a marathon. You get five minutes to breathe and then OpenAI ships an inference chip you didn’t know was coming. So pick your path. Go work at the old software company growing 8% a year, make your $180K to $220K, and wear a really nice watch. Or work at the Corgi Cafe six and a half days a week with a shot at eight figures. You want an Omega or you want to be rich. There’s nothing in the middle anymore.
10. OpenAI’s Jalapeño Chip and the Flabby Middle of the AI Market
OpenAI announced its own custom inference chip, Jalapeño, co-developed with Broadcom. Broadcom’s CEO says it cuts costs by 50% versus a typical GPU. Cerebras dropped 16% on the news, having ridden a $20B chip order from OpenAI.
Rory’s instinct was that this is misplaced effort. OpenAI found the best consumer demand market in 20 years and should pour everything into grabbing enterprise customers, not vertically integrate two steps backward into a chip that goes into a data center. There are three or four chip providers and five or six hyperscalers desperate to do business with you and eat the capital risk. You’re the largest buyer of compute on the planet. Just beat them up on price. The whole reason OpenAI and Anthropic work is that other people spent $300B in capex on their behalf, and vertical integration is the act of pulling that back in-house. He also flagged the tell to watch for: if someone announces they’re building a foundry to make their own DRAM, the cycle is over. Just you and the Koreans.
My disagreement is about the flabby middle. The existential risk that didn’t exist at the start of the year is that as cost pressure hits, the middle of the closed-source market gets hollowed out by open source. People always use the best models for frontier work, and at the bottom these labs are competitive. But the middle is wide open. Anthropic has Opus and friends at the top and Haiku at the bottom and no real middle product, because the middle is too expensive to provide. If you can cut inference below open-source cost, you can serve that entire middle yourself instead of watching it get hollowed out. For B2B software to work, we need mid-priced, high-quality intelligence that isn’t priced at the frontier, because not everyone can run every workflow on Opus. People are starting to gag on pricing. That’s the real game, and it’s bigger than a chip.
One honest caveat Rory and I agreed on: this chip decision was made in an era of abundance, probably early last year. If OpenAI were making it today, it would be a very different conversation.
What This Means for B2B Founders
Gross margin is no longer something you fix later by default. It was, for three years, and that bet mostly worked. The window where negative margins plus hypergrowth gets funded and figured out is narrowing. If your delivery costs eat your revenue, raising your next round just got harder. Fix the unit economics or grow so fast it stops mattering, and know which one you’re betting on.
Your moat is probably weaker than you think. If an LLM can migrate a customer off you in days, your switching costs are not a moat. Stop pitching them as one. The defensible thing now is being the AI-native option that does the same job at a fifth of the cost, not the incumbent protected by integration pain.
Agents replace the work humans won’t do, not the humans. The fastest ROI isn’t firing people. It’s handing agents the follow-up, the invoicing, the CRM hygiene, the 90% of every role nobody wants. Build a master-of-agents capability on your team before you need it. The skill will keep changing, but the people who’ve been building agents for a year see possibilities the rest of the market can’t.
Pick a lane on intensity and hire for it. Small, high-paid, in-person, high-equity teams are winning their markets right now. That model isn’t for everyone and doesn’t have to be. But pretending you can win a frontier market with a 20-hour-a-week distributed team is the actual fraud.
The Real Bet Underneath All of It
Every topic this week reduced to the same question. Capital and talent are being pulled into AI harder than anything in our lifetimes, and the bill is coming due. Either AI delivers productivity gains large enough to justify a trillion dollars of annual revenue, which means real labor displacement, or the last capex dollar doesn’t earn its return and the cycle corrects. Nobody knows when. You don’t make money in venture by sitting it out, and you don’t make money ignoring that the math is stretched. You deploy into the demand, build the most efficient company you can, and try like hell to get liquidity before the chain breaks.
Quotable Moments
Harry Stebbings
“Training agents will be the largest job category in five years. We said the exact same thing about prompt engineers, and how many have you hired?”
“Do we see the same acceleration in legal and accounting that we saw in coding? Because if we do, whoa.”
“Menlo will deliver billions back. Three billion is conservative. Why is it not more?”
Jason Lemkin
“We built an AI VP of Finance while we were in China, and it’s already better than any human on our team.”
“Your moat can be LLM-lifted away. The founder went on about their moats and I immediately didn’t want to invest.”
“You don’t get to make ten million for working 18 hours a week. You get a nice watch. You want an Omega or you want to be rich. Make your choice.”
Rory O’Driscoll
“The only people getting voting rights are the only people who don’t need them.”
“The number one guy makes a little less, the number two guy makes a lot less, and the number three guy goes bust.”
“If you vertically integrate back into memory, then you know it’s over. Just you and the Koreans.”
