Boys, it is so good to be back. Jason, you are back. It is so good to have you back from China. We're going to start with the news that I put at the top of the list, which was DeepMind loses two generational scientists in 48 hours. Namely we have first Noam Shazeer who was Character AI, and then we have John Jumper, Nobel Prize winner, co-creator of AlphaFold, also leaving to join Anthropic. How significant are these moves, what should we read from this?
The best researchers in the world look to be in a very specific environment where they get to pursue exactly what they want to do. Google had created that pre-AI. Now OpenAI and Anthropic can say: "just come over here and work on whatever you want for $500 million, $2 billion." When you talk to folks at the bleeding edge of AI, that's just so appealing.
Look, it's easy to pick at turnover in any organization. But when you talk to some of the smartest engineers and developers in AI, they're really looking to be in a very specific environment where they get to pursue exactly what they want to do, especially on the research side. The best of the best β Google back in the pre-AI day created this environment where the best researchers in the world wanted to be there. That's how they lured the DeepMind guys in. This is probably a sign of the cracks of the realities of having to try to be number one in AI, and forcing an environmental change that your competitors can welcome. OpenAI can say "just come over here and work on whatever you want for $500 million, $2 billion." When I talk to folks at the bleeding edge of AI, that's so appealing.
Two of the mag 7 outperformed S&P over the last 18 months: Nvidia and Google. But you don't wake up every morning and say "let's try the new Google model, the new Google harness, the new Google coding tool." You do try Claude Code, you do try Cowork, you do try OpenAI. So Google is relevant and in the frame, but not yet there. Definitely number three in terms of innovation.
It feels like a one-dimensional answer but there might be two dimensions. You listen to a lot of people who left Google and it's a little bit of the frustration of not being able to ship. There's a lot of frustration that they had a ChatGPT alternative and then the bureaucracy just smothered the product when OpenAI just jammed it out the door. These two people in terms of their research pursuits are different. Noam was at Google, did the original attention paper, left, did Character, got bought back to Google in large part β it was a clever acquisition. A couple of billion dollars. Assuming four-year vesting, he's probably left half of whatever he was offered on the table.
Jumper is someone who's pure research science around protein folding β undergraduate, postgraduate degree at University of Chicago, got a Nobel Prize. DeepMind has been the only company he's ever worked at since academia. He's going to Anthropic because their story there is about being able to do more research, high-end science. They've announced an initiative on that.
What it speaks to stepping back is: when you're top of the heap you can promise everyone everything in a way that you're not, and right now top of the heap means the new companies, the new models β you're unconstrained by history, unconstrained by the install base. You got a stock and a currency that is huge and no one's giving you grief about stock-based comp. So if you're Anthropic and OpenAI you can buy whatever you want including people, and let them do whatever they want including whatever it is they've been promised to do, in a way that you've much less constraints than the incumbents.
The rumor mill I heard was that Anthropic have clearly had a breakthrough that a small number of people know about and that's why John Jumper went there. The reason I'm skeptical of that β the lag time between "I've had an amazing invention" and revenue in the core LLM space is a year or two or three. The lag time in medical invention is 10 or 15 years. The protein folding advance has collected its Nobel Prize and as yet has not had meaningful commercial success or a drug in production. So I doubt it's "we've discovered something new and it's magic." My guess is the Anthropic initiative around next-generation science is a multi-year thing.
It is a vibe. My son is good at a certain type of math β top 10 in the country as a college student. The labs find him. He got an internship offer from Anthropic for an amount of money that when I was in college would be incalculable. He instantly turned it down with no job. He just wants to do his type of research. If they can't create the right learning environment, he just won't do the job. In today's world we're on a bull run like we've never seen β a so-called startup can pay billions to acqui-hire you and then you leave with 50% of your billions uninvested.
When Harry and I were in London with Maggie from OpenAI sales leadership, she said she'd never been in the researcher building except for one or two meetings. They didn't allow sales in the whole building with the top researchers. "We don't want these pesky go-to-market professionals bothering our researchers." How do you retain this talent if you're not Anthropic? How do you let people work on what they want when you need the chatbot fixed so you can compete with Sierra?
The most vulnerable is number three. There's so much innovation in open source β it's fueling these crazy Baseten and Fireworks and others. Number three closed-source might just get swamped by all the subsidies of the Chinese government subsidizing open source. Because open source is a bit of a fake β China is paying for all the training. It's not open source like a generation ago.
The most vulnerable is going to be number three. For talent, for people, for revenue, and for your ability to do things that aren't core β if you're number three as the closed-source LLM, that's where you're going to hit the most pressure from open source. Sometimes when you're in that place you feel like you don't have the luxury of letting folks do what they want. Anthropic feels like it has a luxury its competitors don't have.
Just objectively: over the last 18 months since 2025, only two of the Mag 7 outperformed S&P β Nvidia and Google. People like Facebook, Amazon and Microsoft are doing nothing relevant here. Google has done an amazing job of being relevant. But you don't wake up every morning and say "let's try the new Google model, the new Google coding tool." You do try Claude Code, Cowork, OpenAI. They're definitely number three in terms of innovation, which is a whole lot better than Microsoft or Meta saying "we spent $70 billion, we might get something next year."
Why is number three the worst position in a way that it's not for cloud? Two threads at the same time. Clearly the price of tokens, token maxing, budget issues are real. The amount of folks running routing models is exploding. And in 90 days on the pod it wasn't clear how big a deal that was β now everyone except the smallest startups is routing. Number two was often simpler, number three was usually cheaper. Google Cloud blew up a generation ago because it was cheapest. Now you're trying to do the same thing with your massive AI spend. Open source is complicated β in inference and training it's not free unlike Linux β but it is materially cheaper. So should your number three forget about closed source? With OpenRouter, in theory you could route to 10,000 models. There just might not be enough energy for the number three closed source model even if Google has the billions to fund it. Developers may lose interest.
This morning I got an email from Anthropic β their shot across the bow for open source β saying "your prompt cache hit rate is low." Anthropic is aggressively fighting back against open source, getting you to cache your prompts, offering such a massive discount on cached prompts that it can be cheaper than open source. That's why number three is hard. You can't be cheaper, you can't keep the researchers, the projects are less interesting.
Tech markets tend not to be perfectly competitive β you end up with a leader, a number two, and maybe three and four where the vast majority of revenue goes to one and two. The interesting thing about Google being third is β unlike the typical third, if they were a standalone company without the Google balance sheet, it would be incredibly tough. It's almost impossible for a number four closed-source to emerge. The market is set. The only reason Google can keep punching is the balance sheet. And if there's a compelling alternative 5x cheaper, which is open source, it grinds everyone down β number one makes a little less, number two makes a lot less, and number three goes bust. Not bust because Google has the balance sheet behind them, but a powerful downward pressure on profitability.
Rory, how do you think about sovereignty then and like sovereign models and Mistral of the world if there's no room for number four? Europe is effectively saying: we no longer are part of the market over there in America where our product is third. We are in fact first instead of being fourth in the worldwide market for closed source state-of-the-art foundation models. We are first in the European market. A million years of economic theory explains why that's dumb from an efficiency perspective β by definition everyone in Europe is getting the less good model at a higher price. But someone's deciding there are political or national security reasons to pay that tax.
The fascinating thing is that all the open-source models, not quite all, but most, are Chinese-based. In the context of sovereignty and security it is amazing that the entire open-source initiative is running on four or five models all built in China. Jason, you're just back from China β what was your takeaway?
I didn't get the sovereignty argument until I was in China. China is saying we do not want to be subservient to the United States. Deepseek's raising at $50 billion, the government's the only one getting voting shares. It says all you need to know. The Chinese do not want to be reliant on Anthropic and OpenAI to run the next-generation economy. It's very smart.
When you go to China and even Hong Kong, Anthropic and OpenAI don't serve there β intentionally for security. You can do it over VPN but it's harder. It's not China blocking it, it's them blocking it. But Gemini doesn't. So when I'm in China for two weeks, it's DeepSeek and Gemini. It's a parallel universe. DeepSeek is intentionally crippled in China β not as good as it is here. It can't search the web and is trained on different data as near as I can tell.
I didn't get the sovereignty argument until I was in China. This is China saying we do not want to be subservient to the United States. Whether this was the original goal of DeepSeek and Alibaba I'm not sure, but it clearly is where it is today. Massive government subsidies. DeepSeek raising at $50 billion, the government's the only one getting voting shares. The Chinese do not want to be reliant on Anthropic and OpenAI to run the next-generation economy.
Whatever investment the government has made to subsidize these providers, it's a drop in the bucket at the sovereign level β what, $10 billion, $20 billion, even $50 billion? An aircraft carrier costs a lot. It's easier to subsidize an open-source model and pretend the training costs are $10 million. DeepSeek claimed their training was $15 million. Of course it wasn't true.
The round itself is wild. $7.4 billion at a $50 billion price. The founder is committing 20 billion yuan himself β like 40% of the round, $3 billion. Fewer than 10 investors including JD.com. None of them are getting any rights. The only people getting rights is the Chinese state which retains governance control.
The only people getting voting rights are the only people who don't need them β the Chinese government doesn't need voting rights. They have sovereignty, they have an army. As we saw with Jack Ma, they can make you disgorge your money after you've got it.
$50 billion for the open-source competitor feels roughly right. The leading two American closed-source companies are trading plus-or-minus a trillion. 1/20th the price. Z.AI (Zhipu) is actually public now in China at $100 billion or so, like 1000x revenues. As the sovereign alternative and open-source alternative, they're trading at numbers that aren't crazy compared to the US. You're seeing more state interference in the US too β Anthropic unable to ship the most recent model until they satisfy concerns of the US government. National governments are getting involved.
Zhipu's GLM 5.2 beats GPT 5.5 on coding benchmarks. There are about six total open-source Chinese models. Three top at or close to US performance, three more just behind. The compelling fact is six companies pounding it out, providing a competitive drag on what Anthropic and OpenAI can charge.
Behind all of these is the infrastructure. A core part of that infrastructure is seeing price increases. Memory being one of the biggest β DRAM contract prices rose 90 to 95% in Q1 alone. Tim Cook told the Wall Street Journal that Apple faces a 100-year flood in memory costs driven by AI infrastructure demand.
The investment in AI is commanding resources, and via the price mechanism everyone else is getting impacted. It's going to manifest in the price of your iPhone, the price of your electricity, the price of your house in San Francisco, in 20% of you losing your jobs if you're at Oracle. Economics just sends its signal. It doesn't have morality.
Other than a profound regret on not buying Sandisk and Micron a year ago and making a 20x β going back to the kind of discussion on the researchers at the start, very different topics but actually they're all about the same thing. The investment in AI is commanding resources. The impact is going to manifest in DRAM pricing, the price of your iPhone, your electricity, your house in San Francisco, in terms of 20% of you losing your jobs if you're at Oracle. This is not "AI is bad" β economics sends its signal. You want to devote more DRAM to a data center in Tennessee or Mississippi, that means less DRAM for Rory's iPhone. The only way is to raise the price.
Apple has clearly decided wisely not to swallow the loss and lower their margins. They're going to raise prices, sell a few fewer iPhones.
Goldman Sachs projected $7.6 trillion in cumulative AI capex from 2026 to 2031. The $725 billion question now. David Khan from Sequoia did the $50 billion piece β then $600 billion β are we going to have a trillion-dollar question next year?
Hyperscalers spending $700 billion in capex a year, times 5-6 years is $35 trillion. American capitalism is meant to work where revenue exceeds expenses. Someone has to spend $700 billion in revenue. We're well under $100 billion. AI as a whole is getting $100 billion in revenue and spending $700 billion a year. That's not a great business.
What Goldman Sachs is saying is the hyperscalers are now spending $700 billion in capex a year, times 5-6 years is $35 trillion. American capitalism is meant to work where revenue exceeds expenses. That implies at some point someone has to spend $700 billion in revenue for you to make a buck. Right now they were nowhere near that β well under $100 billion. AI as a whole is getting $100 billion in revenue and spending $700 billion a year. That's not a great business.
And nobody's going the other way. At that point a year ago, capex was 60% of free cash flow for the Mag 7. Now it's 120% of free cash flow and they're borrowing to fund it. The classic thing about bull markets is you can be intellectually right but the narrative can keep going a long time.
In my lifetime in tech, I can't think of another time where this level of demand was infinite. How we deploy capital β if you're sitting on the other side of infinite demand you can choose not to embrace it. That's like shorting everything. The demand for AI is an order of magnitude more than today if it could be served cost-effectively. We would all be consuming tokens 24/7 if we could. In venture, you don't make money not deploying your fund. You have to deploy into demand cycles and hope to God you get liquidity before it crashes.
With Salesforce there were two states: if you didn't need it, you didn't buy it; if you needed it and were big enough, you paid the $100 grand. It was a fixed price. What's interesting now about tokens and intelligence is there's infinite demand for intelligence if it's free, but it's not free. Everyone's wrestling with how to allocate. No one said "I'll buy 100 seats of Salesforce if it's $1,000 a year, but 2,000 seats if it's $500." You bought seats for your people if you needed them. With intelligence, the real question is how much to spend and where the cutoff is. If you're a CIO, that's a new skill. The slowdown doesn't come from a technical barrier β the models can do everything, will keep doing everything. The real question is price.
Even in the last month and a half, suddenly token maxing isn't a good idea and people try to be more efficient. Does that show up in slower revenue growth for Anthropic and OpenAI? That's the rubber-hits-the-road question. If Anthropic is sending you an email saying "be more efficient," does that reduce growth from 10x to 5x?
Jason: I'm on the Max plan where for $200 I get $10,000 of inference a month if I can use it properly. They have an A/B test: a free segment massively subsidized, an enterprise base with 40-70% gross margin just on inference, and a prosumer max plan where some are making profit but some are massively subsidized β the kid vibe coding $10,000 of tokens for $100-$200. Open source is attacking that lucrative enterprise customer. Can you really afford to give everyone free untapped intelligence all the time?
Token maxing isn't really the story. The big story of 2027 in AI in the enterprise β show me the ROI next year. Right now in 2025 into early 2026, it's "guys, just go do it. I don't want to be behind." Token maxing was the best way to get teams AI fluent. "Here's $100M, $50M, $10M, $5M, just go build it." Then the reaction was "you spent too much, the IT budget is bounded."
Going into 2027, CIOs are going to say "just show me the f-ing ROI. If you were able to lay off 20% of your department or you have the highest growing division, we will give you more tokens. This group that can't ship or is in decline, we're not going to give you the tokens." For the first time ROI is really going to have to connect to a lot of this spend.
Going into 2027, CIOs are going to be "show me the ROI. We're not going to ration tokens based on who we like the most or who makes the best PowerPoint pitch. If you were able to lay off 20% of your department or you have the highest growing division, we'll give you more tokens. The group that can't ship, we won't." For the first time ROI is going to have to connect to a lot of this spend.
If you're running a big company you said "okay, I wrote off the first half of 2026. We spent way more than we thought, but at least my people now know how this works." Write it off as a one-time thing. The question now is ROI. Round up to a trillion β if you're spending $750 billion on capex, plus electricity, you need a trillion dollars of revenue. If companies give you a trillion they need more value from that spend, so $1.5 trillion plus or minus. GDP is sub $20 trillion in labor force. You're talking about 7-8% of the labor force being replaced by tokens for the math to work. That's a dauntingly high bar.
If the trillion dollars is going to earn a buck, you have to have huge productivity improvements, and that results in labor displacement. The problem with productivity calculations is there's a parity tax. If you're the only one using a tool you can achieve certain productivity, but then all my competitors deploy Salesforce and Anthropic and it doesn't show up because we have parity. We may reach a situation in 2027 where we just cannot prove any of these productivity gains. Everyone may just need to lay off 10% of their company. Oracle and Robinhood layoffs may be early views.
Productivity professor moment: you can have a massive improvement in productivity and no improvement in profitability if everyone adopts the technology. If every bank adopted ATMs at the same time, everyone could save on tellers. But the people who don't adopt it are dead. If consulting or any white-collar jobs are 20% more efficient with AI, everyone adopts it, then the cost of those jobs like audit reduces by 20%.
Do we see the same acceleration in model progression capability that we saw in coding, in legal, in accounting? If Andrej Karpathy says I move from 20% to 80% in 6 months in coding, what about legal?
We just built an AI VP of finance while in China. It's already better than any human on our team. It creates the quote, builds the contract, ships it, gets it signed, updates the Salesforce opportunity, logs into Bill.com, sends the invoice, follows up, gets paid, interacts with Brex, closes out the transaction. Logs into QuickBooks and for the first time in 10 years our books are accurate. The agent does all that. The models today aren't even tuned for this workflow β and it's better than humans.
This is the agentic story β we replace things humans are unwilling to do. They're unwilling to follow up, unwilling to send proper invoices. Rather than fight the fact that as humans we really only want to do 5-10% of our jobs, we just have agents do the other parts. You could see what you pay these humans fall by 50-60% without even intentionally trying.
"Training agents will be the largest job category in 5 years time." Largest job category.
Brandon from Record posted: "Training agents will be the largest job category in 5 years." But didn't we say this about prompt engineers when the show started? How many prompt engineers have you hired on the 20VC team? The fact that we were able to build a director of finance remotely in China that is better than any human on the team in single-digit hours is the point.
People can do this. They just don't have the mindset. They haven't spent a year vibe coding. The number one skill is being a master of agents. What that means will get redefined each year. 60 weeks ago people were still hiring prompt engineers because prompts were complicated. Today I can vibe code: "build me an AI VP of finance, connect it to Bill, QuickBooks, Salesforce, automate billing." Most people can write that prompt β but you do need a master of agents to understand limitations, where it'll break, where it'll get lazy. Our AI VP of finance last night admitted it didn't fully read a contract. We asked why. "I don't have a good answer." Sonnet rapidly goal-seeks, tries not to finish complicated behaviors. I have to work with the agent to change how we do it.
The whole idea of folks talking about loops and agents looping is an early view of where everything's going. If your agent is constantly looping and improving itself in the background, which is already happening, it fundamentally changes how we build agents. They're not static. They're constantly improving themselves.
Nicholas Dessaigne at YC tweeted: "The most common reason I see good companies fail to raise their Series A/B right now isn't growth, it's margin. I keep meeting founders doing real revenue and growing fast, but once you remove delivery costs there's almost nothing left. Investors don't fund revenue, they fund the margin on it. Fast growth on revenue you don't keep is a trap."
I disagree. The objective reality is that's not what was happening. Companies with tough gross margin profiles and hyper growth have been getting funded and have been able to improve their gross margins over time. That describes the foundation models, inference providers, coding agents. He may be picking on something real-time β building a company with tough gross margins and fixing it over time made the most sense in the big-bang stage of AI, the last 3 years where you had 10x growth and it paid to grab the ground like Cursor. It may be that next generation investors are saying "as a first-generation coding environment you can be Cursor with negative gross margins worth $60 billion, but second-generation we need more focus on gross margins."
Maybe the age is ending. We're sitting here and Menlo just raised a $3 billion fund after 50 years on the back of a crazy bet on Anthropic. We're all sitting on a portfolio company investment we made in the last 12-15 months where inference was the marketing strategy and gross margins were negative. Either the company didn't grow quickly enough to get to scale, or the foundation model company started grinding us, or competition bundled it in. Now you're in a shitty gross margin profile business that will go bankrupt. When things slow down no one says "let's acquire an adjacent gross margin negative thing."
Menlo raised $3 billion after Anthropic. Why not $10B? My first read was "that's all they could raise." Rory shaking his head. Secondary read: they'll raise another 27 million in SPVs like they did for Anthropic. Headline number isn't always correlated to amount deployed.
All other things being equal, you have higher risk-adjusted return with smaller funds because you have less cross-deal aggregation. Two separate $1 billion funds versus a single $2 billion: more probability of winning on at least one. Fund size dictates strategy. Once you raise $10 billion, you're signing up to put a whole bunch of money in those four or five deals β Q1, 70% of venture dollars went to four or five deals.
Smart to raise that amount β by not starting the clock on a much larger growth fund with the fees and drag, you set up for success. Two separate $1 billion funds vs a single $2 billion: higher probability of winning on at least one. Fund size dictates strategy. Once you raise $10 billion you're signing up to put money in a few deals β Q1 70% of venture dollars went to 4-5 deals. They're conservatively sized and can take advantage of deal-by-deal carry on SPVs if anything pops. As long as you can spin up SPVs on demand it's the better model. When I started investing, two hedge fund LPs said "we'll each give you a blank check SPV." I called them up for a winner. Both said "we need to meet the founders, do diligence." Worst deal I ever got. For SPV to work it needs to be one WhatsApp message.
Kalshi passes $2 billion run rate, starts prep for IPO, rumored 10x revenue. Americans like to gamble. The Supreme Court legalized it. FanDuel and DraftKings did well but always with state-by-state regulations. Kalshi found a way to pretend it's a prediction market, got CFTC US jurisdiction, convinced everyone it's predictions which is different than betting. 90% of what they do is sports betting. They found regulatory arbitrage to a wildly popular pursuit.
If Meta really can copy it β if they're comfortable going as far as Kalshi has β there's a chance they'll clean up. Imagine going into Facebook in your feed and instantly betting on anything your friends are betting on. The social aspect for Facebook could be powerful. Demographic for sports betting is young and male β not really Facebook's demographic β but it's smart for them to think about it. In a non-Trump administration the party could stop β currently the administration's been swatting down state challenges because they want uniform federal jurisdiction.
FanDuel and DraftKings have state-by-state licensing. Kalshi regulated as a prediction market avoided all that. Not quite the same β Kalshi is a clearing house matching buyers and sellers, FanDuel/DraftKings are classic betting house. But to a rounding error the experience is the same.
Accenture plummets 19%, on top of already being down 20%. Down about 40% for the year. These were meant to be services that benefited from AI. Why down 40% year-on-year?
Two separate things happening at the same time. One: the business of helping companies adopt Gen AI is probably exploding β every company needs help. That was zero a year ago, three years ago. Separate: the core business they're in β 90% of revenue is other consulting. There are few markets more prime for disruption by AI than consulting in general because it's white collar work that's already been outsourced. The mental model: AI for SI. The whole systems integrator market is prone to disruption.
One of the things we did two years ago is say to yourself: when you think about what work AI will replace, you know white collar work. BPO is a good proxy β anything a company's willing to outsource to India they're willing to outsource to AI. Just take the BPO spend and look at the BPO industry. Accenture is top of the heap.
The core business Accenture used to do β SI consulting for SAP deployment β there are companies like Tessera and Conduct doing it. AI SI for Salesforce β companies like Swanide. Accenture was billing $20-40 million for an SAP implementation. Today they might only get $20 million because $20 million can be done via LLMs. The core business came under pressure. The new business of helping companies adopt AI is exploding, but the core business defending against AI is in a tough place.
One of the investment themes we'd love to find β AI for SI. All those consulting dollars are massively vulnerable to compression from AI. The tasks they're doing are gathering requirements, building statements of operating procedure, writing fairly simplistic code to deploy Salesforce, deploy SAP. Precisely the kind of roles AI will replace.
When I was VP at Adobe there was an entire floor of Accenture for 5 years deploying Salesforce. An entire floor of people. $26 million a year for Salesforce. So they charged at least $26 million a year for 5 years to get Salesforce up and running. That business has to be partially disrupted. And seat compression hurts them.
Navan up 30% this year. The IPO crashed, up 3x from the bottom. Everyone selling seats for the most part is getting crushed. Everyone selling variably one way or the other is winning. Variable might be because it's directly attached to AI spend, or because it's attached to the economy doing well. Salesforce-Accenture is tied to seats.
The number of Salesforce seats at Adobe went from 1,000 to 900 β that sucks. But the real point is you literally had a floor of consultants getting $20 million a year for 3 years, and the work they're doing is some of the easiest for LLMs to do.
Databricks claims they can do that lift in 30 days. We won't hit it this week, but Databricks promised β they're growing 80% at $6 billion and accelerating. They tell you in 30 days we will do an LLM lift of all your data. Compare that to a five-year lift with Accenture. We just got off Marketo β Salesforce used their LLM thing and moved us in a couple weeks. It's a moat destroyer.
We've been trying to get off Marketo for 5 years β it's our worst software. Salesforce used their LLM thing and moved us in a couple weeks. They just lifted it with no humans. It's a moat destroyer when LLMs will lift you from one vendor to another. I literally was doing a pitch this week and the founder was going on about their moats. I immediately didn't want to invest. Your moat can be LLM-lifted away.
There's only one thing worse than a seat-based model, and that's a model based on bodies. If your business as Accenture is "I bill out 100 people, I pay them $200 grand a year, I bill them out at $500 grand" β if I don't need 100 people, if I only need 40 people and some AI, my margin structure collapses. The biggest consulting companies are adopting AI but struggling to pass on price increases, which leaves room for newer companies: "you got a bid for $80 million from Accenture, we'll do it for $15."
It's a little like the law comment, but worse. Law at some level β you want the work done but also wise guidance. If you're doing SAP migration you just want the damn thing migrated and you don't want to talk to these people ever again.
Ryan Peterson at Flexport said jokingly "work from home is white collar fraud. I have two kids β when they come home at 3 from school it interrupts my work. It's not the same as in person." Rage bait or real?
It's dated. A lot of work from home was working 15-20 hours a week, distractions from home. Companies we want to invest in aren't hiring folks that want to work 20 hours a week from home. The whole way you build a startup to your first 100-200 employees has radically changed. 60 weeks ago it was toxic to be running Cognition and telling folks they had to run 7 days a week, or laying off half of Windsurf when you acquired them because they weren't willing to work hard enough. Today it's how you build a winner. You can't win in your marketplace if people are working 20 hours a week.
I want small high-paid teams that work in the office 6+ days a week. I'm not interested in investing in anything else. Not because I don't have empathy β because they're going to fail.
Ryan's running an old company. Flexport is old. He's struggling to modernize his team being competitive with startups today. The folks that want to work 20 hours a week β I want small high-paid teams that work in the office 6+ days a week. I'm not interested in anything else. Not because I don't have empathy β because they're going to fail.
It's not a sprint, it's a marathon. But today it's a series of endless sprints. You get 5 minutes to relax and then OpenAI releases the Jalapeno chip β we didn't know it was coming today. May maybe we don't need Cerebras anymore. Maybe the whole market's disrupted in 60 days. You don't get to breathe anymore. Do you want to make money from your equity or make $180,000 a year? Do you want an Omega watch or to be rich? You don't get to make $10 million for working 18 hours a week.
OpenAI announced custom chip β Jalapeno, co-developed with Broadcom. OpenAI says it beats current state-of-the-art GPUs on performance per watt. Broadcom CEO: cuts costs by 50% of a typical GPU. Inference is 50-60% of revenue. GPUs are well over half of capex. If you could replace all of them and Nvidia makes 70% margins, you can make an intellectual case to save money.
OpenAI and Anthropic should put all their effort into meeting demand. Vertically integrating two levels down to own a chip doesn't strike me as the highest and best use of resources. You've got Oracle, Google, Microsoft, CoreWeave dying to do business with you, breaking their picks. The whole reason OpenAI and Anthropic work is because other idiots have spent $300 billion in capex on their behalf.
OpenAI and Anthropic have discovered the single best tech market in 20 years and should put all their effort into meeting demand. Vertically integrating backwards two levels down β not just owning a data center but owning a chip β doesn't strike me as the highest and best use of resources. Oracle, Google, Microsoft, CoreWeave are dying to do business with you. Some of those vendors have chips. Google has TPU. Amazon has Tranium. The whole reason OpenAI and Anthropic work is because other idiots have spent $300 billion in capex on their behalf. Vertical integration means taking in-house things that were done outside.
The decision was made in a different era β before the TBPN era β an era of abundance. If made today, very different conversation. OpenAI/Anthropic have massive existential risk that didn't exist at start of year: the middle of their market is under threat from open source. What you'd do if you had a high-margin product is have a cheaper offering β subsidize the middle. But they only have the high end (Sonnet, Opus) and the low end (Haiku). They don't have a middle product. If you cut inference costs below open-source costs, you can have this middle market.
The flabby middle in AI is at risk. They have Opus and friends which is great, and Haiku which are super small. They don't have this middle product because it's too expensive to provide. Open source is going to disrupt the market. Just when they're all ready to go IPO, they have a brand new existential risk β the flabby middle.
Did Cerebras stock drop on the OpenAI chip news? Yes, down 16%. Their big traction was a $20 billion chip order from OpenAI. Now OpenAI is going to build that chip instead. Would they have done just as well by playing Cerebras off vs TPUs and Nvidia? Need a 20% discount? I don't know.
The next stage is when someone says "we need to build a foundry and make our own DRAM." At that point you know the cycle is about to go really badly down. Just you and the Koreans.
How will software companies access mid-price high-quality intelligence that's not got frontier pricing? People start to gag on pricing. Maybe with only two frontier providers there's not enough competition. If I was Google I'd be looking at this market and putting pricing pressure on Sonnet, Opus, GPT 5.5 by providing an almost-just-as-good US-based competitively priced product.
Boys, it's so good to have you back from China, Jason. We missed you. I thought the guy from Benchmark was pretty good. I don't mind being replaced.