A Man’s Got to Know His Limitations

THE NUMBER: $6 billion. That’s what Nvidia paid this week to license Poolside’s model factory and walk away without owning the company. Add a billion more for equity at a $12 billion mark, and 109 engineers who changed badges, and you have the shape of the whole deal: Nvidia bought the capability and left the corporation on the table. The most expensive thing on the tape was a decision not to own something. Hold that number, because four separate companies did a version of it this week, and once you see the pattern you can’t unsee it.

The cleanest trade is the one you don’t do

Start with Poolside, because it’s the purest version of the move.

Poolside builds coding models. It was also, by the account of people who cover it closely, starved of the capital it needed to keep buying access to Nvidia’s chips. Grant Janich put it flatly in his Sunday letter: “Starved of compute, Poolside sells itself to Nvidia.” So the setup was the oldest one in venture: a company with a good asset and an empty tank, and a strategic buyer with all the money in the world and a reason to want the asset alive.

Here’s where a normal cycle writes a normal ending. Nvidia buys Poolside. Cleans out the cap table, absorbs the team, folds the models into its stack. Instead Nvidia paid $6 billion for a non-exclusive license to Poolside’s “Model Factory” — the machine that builds the models, not the models themselves — put another billion into equity, moved 109 people over, and left three founders running an independent company that is now worth four times what it was a few months ago.

Read the trade like an operator. Nvidia does not want to own a model lab. Owning one puts it in competition with its own customers, drags a regulatory tail behind every earnings call, and asks a company that prints money selling GPUs to suddenly become good at a business it has no reason to be good at. Why buy the building when you can lease exactly the output you need and let the tenant handle the upkeep? Harry’s word for it is a triple-net lease, and it fits: Nvidia takes the capability, Poolside keeps operating and maintaining the asset, and the landlord — the one with the actual scarce thing, the compute — collects. Cleaner. Simpler. Optimal.

The financial press wrote it up as a “licensing playbook that sidesteps acquisition scrutiny,” and that’s true as far as it goes. Structure the deal as a license plus an acqui-hire and the FTC never gets a merger to review. But sidestepping the regulator is the byproduct, not the motive. This is Nvidia’s third deal in this shape in eight months, after roughly $20 billion for Groq last December and $900 million for Enfabrica before that. A company doesn’t build a repeatable playbook to dodge a rule it would probably beat in court anyway. It builds one because it has decided, with total clarity, what it is and what it isn’t. It’s the compute. It is not the model. It will fund, license, and feed every model company it can reach, and it will own none of them.

The debt that lives on nobody’s balance sheet

Now Broadcom, which ran the same logic through a bond desk instead of a licensing team.

Broadcom’s job this cycle is to design and supply custom AI chips, and one of its biggest customers is Anthropic, which needs an enormous number of them. The problem is that chips cost money before they earn money, and putting tens of billions of dollars of depreciating silicon on your own balance sheet is a great way to turn a chipmaker into a leasing company by accident. Broadcom’s answer, reported this week, is to raise as much as $80 to $100 billion through an off-balance-sheet vehicle run by Apollo and Blackstone — the private-credit giants — to finance the chips and lease them into the deal. Benzinga’s headline said the quiet part: the arrangement “turned AI’s chip bet into somebody else’s debt.” Broadcom’s credit-default swaps widened on the news, which tells you the market noticed the size of what’s being moved around, even if it can’t see all of it.

Here is the discipline underneath the financial engineering. Broadcom makes chips. Apollo and Blackstone exist to raise capital and hold exactly this kind of long-dated, asset-backed paper — it is the entire reason they get out of bed. So the chip risk goes to the chipmaker and the credit risk goes to the credit shops, each asset routed to the balance sheet built to hold it. That’s not a trick. That’s the division of labor working the way it’s supposed to.

But we owe you the other half, because we don’t sell fortresses that are really windows. Off-balance-sheet financing has a history, and it isn’t a warm one. The last time an entire industry got clever about moving obligations into special-purpose vehicles so they wouldn’t show up where investors were looking, it had names: Enron’s Raptors in 2001, the structured investment vehicles that vaporized in 2008. The structure is not fraud — Apollo and Blackstone are doing exactly what they’re built to do, in the open. But a machine with this many moving parts, financing an asset whose useful life depends on a demand curve nobody can prove yet, is a machine with a lot of places to jam. The discipline is real. So is the fragility. Both things are true, and the second one is why you read the structure before you admire it.

The cheapest door is the courthouse

Google’s deal is the one that’ll make your neighbor look up from the dinner table, because it’s about their email.

Google paid roughly $10 million for the data of bankrupt Spirit Airlines — hundreds of millions of emails, internal Teams chats, billions of flight-pricing records, and anonymized passenger data. Ten million dollars. For a nationwide airline’s entire digital exhaust. The price is absurd until you understand the door: this was a bankruptcy, and a bankruptcy court is the one place on earth where a company’s data gets sold as an asset, cleanly, cheaply, and with the consent question handled by a judge instead of a terms-of-service pop-up.

And notice what’s not surprising here, because it’s the tell. Google already has more flight data than almost anyone alive. It runs Search. It runs Google Flights, which it built on ITA Software, the airfare-pricing engine it bought back in 2010. Buying Spirit’s records isn’t Google wandering into a new business. It’s Google doing the one thing it has always done — organize the world’s information and make it useful — and finding, in a bankruptcy docket, an unusually clean and cheap way to learn what the most price-sensitive fliers on the planet actually do. Dead center of the lane. Optimal access, through the optimal door, for the company whose whole identity is data.

The uncomfortable part, and the reason this one has teeth, is what it says about the training-data well. If the richest company in the world is picking through a dead airline’s inbox for ten million dollars, it’s because the clean, consented, proprietary data that models are hungry for has gotten scarce enough to scavenge. That’s a signal, and we’ll come back to what you do with it.

Own your intelligence, rent the rest

The fourth deal isn’t a purchase at all. It’s a build — and it’s the one that shows the discipline from the other side.

Harvey, the legal-AI company last valued around $11 billion, shipped Tenet this week, its first post-trained model. It didn’t raise a few billion to stand up a frontier lab and go chip-to-chip with OpenAI. It took an open-weight base — Kimi K3 — post-trained it on legal data with Fireworks AI, and, by its own account and its investors’, hit state-of-the-art on legal benchmarks at a fraction of the cost of calling a frontier model. Sonya Huang at Sequoia called it “super cool.” David Sacks, the administration’s AI point man, held it up as a case study in “how American companies are building world-class specialized models.”

Sit with the Sacks quote for a second, because there’s a wrinkle folded inside the applause. The “open-source base” Harvey used, Kimi K3, is Chinese. So the model being celebrated as an example of American AI strength is an American company’s law-specific training run sitting on top of a Chinese open-weight foundation. That’s not a gotcha — it’s the actual texture of where this is going, and it rhymes with something we wrote in Money Train last week: open describes the weights, it does not describe the money. Harvey’s move is the smart one precisely because it refuses the scope creep of becoming a lab. Its job is to be the best legal-AI company in the world, not to win a benchmark war it has no business fighting. So it owns the one layer that is its circle — the legal intelligence — and rents the foundation from whoever will hand it over cheapest. Even if “whoever” is in Beijing.

Know thyself

Put the four deals on one table and the pattern stops being about finance.

Nvidia won’t own a model company. Broadcom won’t become a $100 billion credit fund. Google will only ever organize data. Harvey won’t pretend to be a frontier lab. Every one of these is a company staring hard at what it actually is, drawing the border, and refusing to cross it — even when crossing it would capture more of the upside. Nvidia gave up the equity value of owning Poolside outright. Broadcom gave up the financing spread it could have kept by warehousing the chips itself. Each one traded a little money for a lot of clarity.

There’s an older word for the border you don’t cross. Warren Buffett calls it the circle of competence, and he’s spent sixty years insisting that the size of your circle matters far less than knowing exactly where its edge is. The Greeks got there first and carved it over the entrance to the temple at Delphi: know thyself. We put that exact idea in front of you three weeks ago, in an issue we literally called “Know Thyself,” and the market just spent a week proving it in cash.

Here’s the frame that makes it click, because it connects the finance to the thing you actually work with every day. AI is one long exercise in optimization. You optimize the data you feed it. You optimize the harness around it, the token spend, the routing, the inputs and the outputs, all tuned and re-tuned at machine speed until every watt is doing the most work it can. What happened this week is that the industry pointed that same optimizing instinct at its own cap table. It stopped asking “who should own this?” and started asking “what is the optimal home for each piece of this?” — the chips to the chipmaker, the credit to the credit shop, the data to the data company, the model to the vertical that will actually use it. The exotic structures are just what optimization looks like when you run it on ownership instead of tokens.

And the co-sign came from an unlikely place this week. David Sacks, arguing the administration’s line on data centers, wrote that “done right, data centers actually lower prices by producing excess power and funding grid upgrades” — which is very close to the argument we made in Memo to the Governors and again in The Usual Suspects: make the builder pay for its own power soup to nuts, then route the surplus back to everyone else’s bill. We’re not going to run a victory lap on it — the AI czar landing near our framing in the same week isn’t the same as calling it a month early, and we hold that bar high. But it’s worth noting that the discipline we keep describing keeps showing up in other people’s mouths.

So why is the rule even there?

One more turn, because it’s the one that should stick with you longest.

Every one of these deals routes around a rule. The license routes around the antitrust review. The SPV routes around the balance sheet and the scrutiny that comes with it. The bankruptcy sale routes around the consent and copyright fight that scraping your emails would otherwise start. And here’s the thing nobody in Washington wants said plainly: none of it required much effort, because none of the rules were ever tight enough to be the actual wall. Nvidia could have fought the FTC on a straight acquisition and probably won; it just decided owning Poolside wasn’t its job. The discipline here is not being imposed by a regulator. It’s self-imposed, a choice about identity, and the rules are almost incidental to it.

Which raises Harry’s question, and it’s a serious one: if the smartest, best-advised players in the market can stroll around your rule for free, and it costs them nothing, and they’d have won anyway — what is the rule for? A rule that only inconveniences the people too small to hire the structuring desk isn’t a guardrail. It’s a toll on the naive. Either it’s binding on everyone or it’s theater, and this week was a pretty good demonstration of which one it is.

What it means for you

Three things fall out of this, and they touch you whether or not you own a single AI name.

Draw your circle before your next build-or-buy decision. The most valuable companies on earth spent this week refusing to own things outside their lane. Do the same. Take the capability you’re about to build in-house and ask the Nvidia question: is this actually my circle, or am I about to over-reach to own something I could rent for less than the cost of owning it? Own the one layer where you can be the best in your market. Rent the rest, and pay the premium for the privilege of staying pure. Scope creep is how good companies rot; the fix is a border you’re willing to defend.

Inventory the proprietary data you already sit on, and start guarding it like an asset. Google just paid millions for a dead airline’s records because clean, consented, first-party data has gotten scarce enough to buy out of bankruptcy. You almost certainly own a dataset nobody else has — your transactions, your support logs, your customers’ behavior. That is the thing the whole industry is now short of. Write down the one dataset only your business owns, and decide today whether you’re protecting it, or leaking it into someone else’s model for free.

Read the structure of every AI deal you touch, not just the price. When a vendor finances your compute or your chips through someone else’s balance sheet, the risk didn’t disappear. It moved, and under stress it can move back to you. On any AI contract in front of you, find who actually holds the paper, who operates the asset, and who eats the loss if the demand never shows up. The price is what they want you to look at. The structure is the information.

A man’s got to know his limitations. This week four of the most valuable companies on earth showed you exactly what it’s worth to know yours — and if you’re paying attention, they also showed you where the next opening is. Find the one lane you can be the best in the world in, defend its border, and rent everything else from whoever owns theirs. The great ones don’t just know their limitations. They get paid for knowing them.

---

The email edition — THE NUMBER, three moves, and the day’s five stories — is in your inbox. If a build-versus-rent call is forcing the question of what’s actually inside your circle, that’s the conversation we run at Outsider Labs.