Thinking Notes
AI, Involution, Iran War, PayPal, Figma, and Warren Buffett
Bearstone is a reader-supported publication. Please like and subscribe to support.
AI Fiasco
How improvement gets measured
The way AI improvement is measured is a mess. I covered half of it in Folly at the Endpoints, in which token usage was treated as a precursor to profitability. Here is the other half.
Surprisingly, if you measured AI against the compute being poured in, capability gains since 2024 have been modest. But that is not what gets measured. METR set the standard the industry now quotes, which is how long a human task the model can complete on its own. This is a completely reasonable measure to track, and the measurement has actually doubled every 4 months. However, the nuance of what human tasks are being measured is often overlooked. The human tasks being measured are almost entirely coding and cybersecurity. These tasks can be checked instantly and iterated on a thousand times an hour. The benchmark was built around what is easy to verify, and the models were built around the benchmark. Rarely do we stop to evaluate the endpoints though, as “AI is doubling every 4 months” sounds much better than “the length of coding tasks AI can handle is doubling every 4 months.”
So billions of dollars are being deployed to create a machine that is very good at coding. That is a real use case, and a big one. But it is being conflated with “AI will do everything.” It will not. AI is not going to turn a wrench cheaper than a human. It is not going to sell medical diagnostics to physicians. I do that for a living, and I am not worried. And even in the narrow band where agents do work, who says Anthropic or OpenAI captures it? They won’t. Those wins go to the average Joe who is sick of a repetitive task at work, automates his own job with a cheap model, and then has the good sense to sell those workflows.
I will concede the obvious: frontier models will be excellent at coding. That could be a genuine tailwind against stock comp at the current SaaS names, by the way. It is not a trillion-dollar capex thesis.
Commoditization
The cheaper models handle the vast majority of what anyone actually needs, at a fraction of the price of the frontier models. You do not need Fable 5 to tell you how to cure a wart on the bottom of your foot. Intelligence at that tier is already a commodity. The pricing will follow.
Chamath Palihapitiya has an excellent point in the interview I posted on this, and I can completely see it playing out exactly like this. He essentially says a company could miss on earnings, the CFO will dig into the numbers, and uncover that the company is buying $50 “barrels of intelligence”, Chamath’s analogy, when they could be buying those same barrels for a dollar.
The circle
The private AI labs are deeply cash flow negative. So the whole thing runs on paper. Nvidia put up to $100 billion into OpenAI and gets stock. Microsoft holds roughly a quarter of the company. Oracle holds a five-year IOU. Amazon has put up to $33 billion into Anthropic, and Anthropic has committed more than $100 billion to AWS over the next decade. The dollar leaves Amazon as an investment and comes home as cloud revenue. Some portion of this demand is artificial, just vendor financing dressed up as growth.
The numbers are not subtle. OpenAI’s Oracle contract calls for roughly $60 billion a year once at scale, starting in 2027. That is about two and a half times OpenAI’s entire current annualized revenue. Every year, for five years. Revenue with an R. Not profit with a P.
It is fragile and interconnected like the banking system, minus the deposit base. That is the interesting part. There is no run to trigger the unwind. It unwinds through the income statement instead.
The ripple effect
Let’s walk through what happens if a frontier release disappoints, or one large enterprise quietly moves its workload down to a cheaper model.
First, token growth forecasts get cut. Those forecasts are the collateral for everything else in the chain. Every compute contract was underwritten against them.
Second, the compute commitments do not shrink with the forecasts. The contracts are signed. The lab is now obligated to pay for capacity it cannot monetize.
Third, the backlog gets repriced. Oracle’s remaining performance obligations, north of $500 billion, were capitalized by the market as if they were cash. A backlog is only worth what the counterparty can pay, and the counterparty just missed.
Fourth, the order book softens. Datacenter buildouts get delayed. Nvidia’s equity stakes in its own customers get marked down, the same stakes those customers were using to fund purchases of Nvidia chips. The flywheel runs in reverse at the same speed it ran forward.
Fifth, credit reprices. The datacenter debt sitting in private credit funds and SPVs was underwritten to utilization assumptions that no longer hold. Lenders pull back, refinancing gets expensive, and marginal projects die.
Sixth, and this is the part that compounds: capex cuts are revenue cuts. Nearly every dollar of “AI revenue” in this ecosystem is another participant’s capital spending. In a normal industry, one company’s cost cut is a wash across the system. Here, each party’s revenue is the next party’s capex, so the correction does not net out. It cascades.
No depositors flee. No bank fails. The circle just stops spinning, and everyone holding stock and IOUs instead of cash finds out what they are actually worth.
Involution
This is the area I didn’t give enough weight in my PDD article, or in the supplementary piece on China’s major e-commerce platforms. So let me correct that here.
The Chinese term is neijuan, involution. It describes a zero-sum arms race where everyone works harder, spends more, and cuts deeper, and the whole industry ends up worse off. You can see it across Chinese industries, from EVs to solar. In e-commerce, the pain landed squarely on merchants and manufacturers.
The platforms got coercive. PDD was among the worst offenders. If PDD’s systems found your product cheaper somewhere else and you didn’t match, they buried your listing, pushed to the back pages where nobody scrolls. Match the lowest price on the internet or disappear. That’s arm-twisting in the eyes of Beijing.
The other mechanism was “refund only.” If a customer complained, PDD would give their money back and let them keep the product. No return required. It was rocket fuel for user growth and a death spiral for sellers, who lost the inventory and the revenue. Professional freeloaders industrialized the loophole, and a cottage industry sprang up teaching buyers how to exploit it. During the 2024 Singles’ Day season, more than 60 percent of seller complaints on China’s consumer mediation platforms related to the policy, and some sellers reported fielding over 200 refund-only requests a month. Merchants protested outside Temu’s offices in the summer of 2024. By early 2025, China’s top market regulator was summoning platform representatives to demand changes, and by mid-2025, every major platform, including Pinduoduo, Taobao, JD, Douyin, and Kuaishou, had scrapped refund-only entirely. After-sales disputes went back to the merchant’s hands.
These practices are no longer allowed. Beijing has made “involution-style competition” a named enemy, and PDD’s response has mostly been self-regulation: it pledged to waive RMB 10 billion in merchant fees, refunded service charges on canceled orders, and handed dispute resolution back to sellers.
Then there’s the more interesting move. In May, PDD registered a new subsidiary in Xiong’an, the “city of the future” about 60 miles southwest of Beijing that is Xi Jinping’s personal legacy project. He announced it in 2017, and it has badly lagged its own targets, populated mostly by relocated state-owned enterprises rather than organic growth. PDD dropped RMB 500 million into the new entity, bought an office building, and launched a hiring drive for 5,000+ jobs: content moderators, quality-control specialists, data analysts, customer service. By the end of June it had 600+ employees on the ground, already making it the largest private internet company in the zone.
Notice what those jobs are. Content moderation and quality control, the exact functions regulators have been hammering PDD on. And notice the timing. This came weeks after PDD ate the largest penalty in a record food-safety crackdown, RMB 1.5 billion of a 3.6 billion total across seven platforms, after regulators uncovered tens of thousands of unlicensed “ghost” bakeries and accused PDD staff of physically obstructing investigators.
Through an American lens, you could read the Xiong’an move as a bribe: plant 5,000 jobs in Xi’s pet project so Beijing goes easier on you. I don’t think that’s culturally correct for how China operates. The better read is a gesture of alignment, a highly visible signal that PDD understands the new rules and is putting real money behind national priorities.
One last note from personal experience. I’ve ordered from Temu, and the quality concerns aren’t theoretical. I bought a wallet that holds my ID card about a quarter of the way out of its slot and already looks like it’s going to fall apart. When the entire model is built on squeezing merchants to the last yuan, this is what comes out the other end. That’s the cost of involution.
The Iran War
Peter Lynch said, “If you spend more than 13 minutes analyzing economic and market forecasts, you’ve wasted 10 minutes.” He put it another way in a PBS interview years later: spend 13 minutes a year on economics and you’ve wasted 10. I largely agree with him.
So I’m not going to sit here and speculate on what happens next in the Iran war. I have no idea. And more importantly, if you took my opinion on the matter and acted on it, that would be foolish. I am an investor, not a geopolitical expert. Maybe there’s someone out there with a genuinely accurate macro read on what’s unfolding right now. It isn’t me.
But here's what I can tell you. I wrote an article on Occidental, and in it we talked about Tier 1 reserves, the wells you can actually drill cheaply, and about the technology we now use to find them. There is no cheap mystery oil sitting out there waiting to be discovered. What is left to find is deepwater and expensive, and the low-cost inventory is already known. The problem is that getting it out is going to become extremely difficult in the future. How far into the future? I don't know. What the technology progression looks like for drilling those harder holes? I don't know that either.
What I do know is that if you want Tier 1 inventory, you look at the Permian, even though it’s getting gassier as time moves on, and you try to find something that will still be there for a while. But at the end of the day, this is a commodity business. You are largely subject to fluctuations in the price of oil, and if the past few months have proven anything, it’s that those fluctuations can be violent and unpredictable. Over a long enough period, though, I’d argue oil prices rise for a simple reason: the supply of Tier 1 wells is dwindling, and the remaining wells get harder and more expensive to drill.
Whether demand for oil will look anything like it does today is another speculative layer entirely, and that one I’ll leave to the reader to research on their own. I’ve moved the ball up the field for you a bit. The rest of the macro perspective is up to you.
This is also one of the reasons I’ve never been particularly thrilled about commodity businesses. Even when the baseline assumptions are rooted in non-macro characteristics, asset quality, cost position, balance sheet, there’s a decent amount of macro sitting at the end of the equation. It’s something I’ve dabbled in trying to learn. But it is not my strongest suit. My strongest suit is understanding business models, not handicapping the various macro elements that go into a bet on the direction of a commodity price.
PayPal
I wrote a three part series on PayPal and decided to pass. My reasoning was simple. The branded business appeared to be losing the battle to Apple and Google. The unbranded business was eating the branded business. And both were fighting Stripe at the same time.
So imagine my surprise when I woke up this morning to the stock up 17% on news of a $60.50 per share bid from Stripe and private equity firm Advent International.
In my defense, I made the no-go call at $60, so I win for now. I do not expect to win in the long term.
As investors, and generally as people, we have to learn from our mistakes. So why was I wrong here?
First, I never considered PayPal as an acquisition target. This is the most obvious miss. When a business throws off billions in cash and trades at a depressed multiple, you are not the only one running the numbers. Someone with a strategic reason to own it is running them too.
Second, I fell for the narrative. PayPal produces massive cash flow and buys back stock relentlessly. When everything about a business looks broken, it becomes easy to stop weighing what is going right. I wrote extensively about Apple’s advantage at the authentication layer, the NFC chip and Face ID making its wallet the path of least resistance. What I never asked was what happens if governments decide that advantage is an antitrust problem. That is not a hypothetical. The EU has already forced Apple to open NFC access to rival wallets, and the US Department of Justice’s antitrust suit against Apple takes direct aim at the same wallet restrictions.
Third, I ignored habitual behavior. I have written about this before as a general business principle, which makes the oversight as human as it is disappointing. It hit me harder today as I sat in a rental car, driving around a new city my company wants me to capture. I was genuinely impressed by CarPlay. The phone screen on the car screen, no wire required. I found myself wishing my car could do this without a wire. Then I flipped back to the static radio and started clicking through stations I had never heard of. As unfamiliar as the stations were, the process was second nature. Click the arrow on the dashboard until a song you like comes on. Who really cares where the station is housed anyway? Looking at you, SIRI. Hundreds of millions of people feel exactly the same way about PayPal.
Now for something atypical in the free Bearstone world. In fact, I believe this is a first. Do not get used to it, because it will likely also be a last. I am going to give you a stock I bought today.
I bought PayPal. A small position, just over 1%. Why? The offer is $60.50 and the stock closed in the mid 50s, so there is still a spread sitting here. And I think this ultimately closes higher than the current bid. PayPal’s board has not even engaged yet. My hope is that the deal wobbles temporarily and I get the chance to buy more at a lower price.
We shall see.
Figma (FIG): Too Hard
Figma is too difficult of a company to value.
On one hand, you have a business growing beautifully. Q1 revenue came in at $333.4 million, up 46% year-over-year, and that’s an acceleration: 38% growth in Q3 2025, 40% in Q4, 46% now. Net dollar retention hit 139%, the highest in over two years. Customers spending $100K+ grew 48%. Management raised full-year guidance to $1.422 to $1.428 billion.
On the other hand, they still have negative owner earnings. The same quarter that produced 46% growth produced a GAAP net loss of $142.4 million, swinging from a profit in the same quarter a year ago. The full-year 2025 GAAP loss was $1.25 billion. The growth is there. So are the losses.
The bulls will tell you the GAAP loss is just stock comp. Fine, but stock comp is a real cost, and it’s enormous. $169 million of SBC in Q1 alone against $88.6 million of free cash flow. Adjust the “27% FCF margin” for the shares being handed out and owner earnings go negative. You are paying for growth with your own dilution.
Look at where the money flows and most of it goes into R&D. You can see it in the product roadmap: Figma Design, FigJam, Dev Mode, Slides, Sites, Buzz, Draw, Make, Weave. Every launch buys a new enemy. Slides walks into a fight with Google and PowerPoint. Sites takes on Webflow and Framer. Buzz picks a fight with Canva. Make squares up against Lovable, v0, and Claude. FigJam has been trading punches with Miro for years. Each of those products will compete against somebody hyper-focused on that particular field, in the hopes that the ease of a platform will be worth having a mediocre product. So the company is transforming itself into the classic SaaS platform play.
To be fair: the model of dumping everything into R&D has worked before, and when you operate at 80%+ gross margins, sacrificing everything for growth can make sense. I wrote about the chains Buffett wore, how his framework left him blind to this newer model. I haven’t forgotten that lesson. But this one is genuinely difficult to gauge, and times are incredibly uncertain.
Here’s what the market has already done to it. Figma came public in July 2025 at $33, about a $19 billion valuation. Within days it spiked as high as $142.92, briefly making it a $60+ billion company. The all-time low is $16.60, set this April, and after a recent bounce it trades in the low $20s, about a $12 billion company. That’s down roughly 85% from the high, and well below the $20 billion Adobe offered for the whole business back in 2022.
A massive drawdown like that is exactly where asymmetry is supposed to live. I just don’t think it does here. I find it unlikely the asymmetry in this name is worth the risk I’d be taking. And the tells around the edges don’t help.
No insider buying. Through an 85% drawdown, nobody on the inside is reaching for their wallet. Worse, insiders including the CEO, CFO, and CTO have sold over 694,000 shares in the past 90 days.
The Anthropic situation. In April, Mike Krieger, Anthropic’s Chief Product Officer, resigned from Figma’s board. Three days later, Anthropic launched Claude Design, a product aimed squarely at Figma’s core. Dylan Field’s public response: Anthropic was “not consistently candid” in its communications. When your AI partner turns into your AI competitor and your own CEO says that out loud, that’s at least partly scary.
And it points at the deeper issue: Figma sits in a category that can be replaced by AI. Interface design is structured, pattern-heavy work, far easier to replicate than genuinely creative design. That’s the opposite of the argument I made for Adobe.
All in all, I’m not getting overhyped by the new product launches. I’m not convinced the platform sprawl will be worth it, or at least I fear the core product will lack focus while they fight a nine-front war. Other companies have made this work before, and if the bet pays off, I won’t be entirely surprised.
But it’s not something I’d feel comfortable owning. It’s not something I could sleep well at night holding. So I pass, for now, and put it in the too-hard pile.
Warren Buffett
Let’s wrap these thoughts with something lighthearted, an interview with a little old man in a sweater. Buffett sat down with Becky Quick and talked about two things: Google and Bill Gates.
First, Google. Surprisingly, Buffett was the one who initiated the position. I assumed it was Greg Abel, and I was far from alone in that. Becky pushed him on the timing. Why now? Why buy just as the business is becoming capital intensive? Buffett’s answer was that he was wrong before. His read on the spending is that Google is forced into it. They would not choose this capex if they did not have to, but they have to.
More interesting to me is what came next. It appears Greg is the one who backed the truck up, and Buffett does not seem to love that decision. That is my read entirely, so take it for what it is. My rationale: Buffett started the position in late 2025 at much lower prices. The buying ramped up hard after Abel took over this year, including a $10 billion private placement last month at prices near all time highs. Buffett then went out of his way to say there are at least four or five other businesses Berkshire owns that he likes better than Google. He did not name them, of course, but it was an interesting comment to say the least.
On to Bill Gates. Buffett is no longer giving to the Gates Foundation, after more than $47 billion since 2006, and distribution of his fortune now runs through his kids’ family foundations. I think this is the right move. Buffett was clear that he did his homework on the Epstein situation. He read Gates’ testimony to Congress and the cross-examination, and his framing was that of a friend who made a mistake in the people he chose to be around. Gates has not been accused of any crime, and Buffett did not condemn him. His line was that no one bats a thousand in the business of choosing people. I believe Buffett still likes him, and he was visibly careful not to throw a lifetime friend under the bus. I am sure the decision was not easy.
Buffett insisted the Epstein situation had little to do with the decision, saying it is more about his kids being ready to handle the giving. I do not fully buy that. My speculation is that Buffett did not want to damage his old friend’s reputation any further, so he handled it professionally. He may well believe Gates did nothing more than pick the wrong company, but he cannot have that shroud hanging anywhere near Berkshire. I agree with the call. And giving the money to his kids to manage the family philanthropy is smart in my opinion.
Thanks for reading. Please like, subscribe, and share this if you found it valuable. It is the number one way to support this publication if you are a free subscriber, and of course it is still helpful if you are a paid one.
Disclaimer: This article is for informational and educational purposes only and represents the author’s personal interpretation of public data. It does not constitute professional investment advice, financial guidance, or a recommendation to buy, sell, or hold any security.
No Professional Relationship: Subscription to this publication or interaction with its content does not create a fiduciary or advisor-client relationship. The author is not a licensed financial advisor or broker-dealer.
Risk of Loss: All investing involves significant risk, including the complete loss of principal. Past performance is not indicative of future results.
Accuracy & Conflicts: While the author strives for accuracy, the information is provided “as is” and may contain errors or omissions. The author may hold long or short positions in the securities discussed and may trade them at any time without notice. Readers must conduct their own due diligence and consult with a qualified financial professional before making any financial decisions.
Hey Leah,
Yes I would. My hopes is the bid fails and I can buy under $45.
I imagine the stock will fall if the bid fails, that’s why I sized this so small. Even if the stock falls below $45, I couldn’t see myself going above 3-4% on this bet.
Great Read!