Show Me Where to Put the Fulcrum

THE NUMBER: $60 billion — what SpaceX paid for Cursor this week, in stock already up better than 30% since the biggest IPO in history, which means the deal got cheaper the higher the currency climbed. And the most valuable thing that $60 billion bought isn’t a model. It’s the index of how the world writes code — the one asset a free, commoditized model can’t copy. Everybody’s going to argue about the price. The price is the least interesting number in the story.

Archimedes had the whole thing figured out two thousand years ago, and he said it in one sentence: give me a lever long enough and a place to stand, and I will move the world. Every business-school deck that ever quoted him stopped at the lever. The lever is sexy. The lever is the machine, the technology, the force multiplier. Nobody frames their keynote around the fulcrum — the boring, fixed, unglamorous spot you brace your feet against so the lever has something to push off of.

Which is a shame, because this week the AI business handed us the cleanest proof in years that the fulcrum is the entire game. The lever just became free. And the people who own a place to stand are the only ones who are going to move anything.

Two stories, same week, almost nobody put them in the same sentence. SpaceX bought Cursor for sixty billion dollars. And Anthropic, on the same day Wall Street was still doing the SpaceX math, quietly published a study of four hundred thousand of its own coding sessions that explains why that sixty billion makes sense. One story is the trade. The other is the thesis behind the trade. Let’s take them in order.

🚀 What Sixty Billion Dollars Actually Buys

Start with the financial engineering, because it’s the tell, and because finance-brain never fully switches off around here.

Musk paid for Cursor in SpaceX stock — days after the largest IPO in history, with the shares already up better than thirty percent off the open. Sit with the mechanics. When you pay a fixed dollar price in a currency that’s climbing, the deal gets cheaper by the hour. Sixty billion dollars of an appreciating stock means fewer shares out the door, less dilution to existing holders, same trophy on the shelf. He paid full sticker and got a discount at the register. That’s not luck. That’s the oldest move in the Musk playbook: build a currency on a story that keeps enlarging faster than reality can catch it, then spend that currency to buy real things. He did it with Tesla for a decade. He’s doing it with SpaceX now, and Benedict Evans called the setup forty-eight hours early — “imagine what SpaceX would be worth if it had an AI business.” It went and bought one.

But the currency is the easy half. The hard question, the one that actually matters: what did one of the most valuable companies on earth decide was worth sixty billion?

Not a model. That’s the thing people keep missing. Cursor is the biggest coding shop on the planet that isn’t a frontier lab — roughly $2.6 billion in annualized revenue, mostly enterprise, growing fast. What makes it worth sixty billion isn’t a clever model. It’s that Cursor sees more cutting-edge code, written by more real developers solving more real problems, than anyone alive. He didn’t buy a tool. He bought a sensor — the single best-positioned sensor in existence for watching how software actually gets made.

🛰️ Everything Musk Builds Is a Sensor

Here’s the pattern, and once you see it you can’t unsee it. Every company Elon Musk has ever built is, underneath the product, a data-collection machine that feeds itself.

Every Tesla on the road is a camera rig collecting training miles for self-driving. Every Starship that flies is a telemetry stream mapped back into the next design. Every post on X — for him, against him, pro-capitalism, anti, doesn’t matter — is a labeled signal feeding the recommendation engine and, increasingly, Grok. The man builds products that happen to be sensors, and the sensor is always the actual asset. The car is a means. The data is the end.

This week even put the darkest version of that pattern on the table. A Pentagon filing — a single sworn statement from a defense official, filed to defend xAI in a lawsuit over data-center pollution — claims Grok helped target “2,000 munitions at 2,000 targets within 96 hours” in Iran. Treat it with the caution it deserves: it’s single-sourced, it’s motivated, it’s unverified, and the strikes it references reportedly killed hundreds of civilians. None of that gets trivialized. But strip it to the mechanism and it’s the same story: another sensor, another stream of engagements mapped back into the system. War is a fact of life, and if it’s going to happen, the side whose models learn from every engagement compounds an edge the other side can’t match. The flywheel reaches the battlefield too. That’s not a comfortable sentence. It’s a true one.

Now put Cursor in that lineage. It’s the one sensor Musk was missing — pointed not at roads or orbit or the timeline, but at the act of writing software itself. And here’s the wrinkle most of the coverage will skip: he doesn’t need the contents of your code. He’s not raiding your repo for your secret sauce. The gold is the process — the write, debug, deploy, recompile loop, run millions of times a day, the largest record of how software actually comes into being that exists anywhere on earth.

And it gets better, from his seat. Cursor is model-agnostic. It runs on every frontier model: OpenAI’s, Anthropic’s, Google’s, plus Cursor’s own. Which means Musk didn’t just buy a flywheel for Grok. He bought an observatory pointed at every competitor he has. He now gets to watch, at enormous scale and in live production, exactly how well OpenAI and Anthropic and Google actually perform on real work — not on a benchmark, on the job. We wrote two weeks ago that the public benchmark died as a decision tool and that private evals were the only test left that discriminates. Musk just bought the biggest private eval harness ever assembled, and it’s running on all of his rivals at once. Cursor is how Grok gets to recursively improve. Say that phrase out loud, recursively improve, and notice it’s the exact thing the doom crowd stays up at night about. He just bought the on-ramp to it for sixty billion in inflated stock, and the market cheered.

⚖️ The Lever Got Free. The Fulcrum Didn’t.

Back to Archimedes, because this is where the week’s two stories fuse.

The lever is the model — raw capability, the last ten percent of intelligence the whole industry spent a year treating as the crown jewels. We were as guilty of that as anyone. And then the floor fell out, exactly the way we said it would in Buy Wins, Not Players and again in The Good, the Bad and the Ugly: Washington pulled the best American model off the market on a Friday, and by Sunday a Chinese open-weights model was sitting at the top of the leaderboard at a tenth the price, three hundred tokens a second, no filter. Capability stopped being something you could defend. The lever became a commodity you rent from whoever’s cheapest this quarter, and the rental company might answer to people you didn’t vote for.

So if the lever is free and infinite and available to everyone — and it now is — then the lever is no longer where the value lives. The value moved to the fulcrum. To the place you stand. To the one thing you own that a free model can’t reproduce.

Cursor is a fulcrum. Its place to stand is the proprietary index of how the world writes code — the retrieval layer, the context, the data flywheel that no amount of model horsepower recreates from scratch. That’s why it’s worth sixty billion while the model it competes with is collapsing toward free. Same logic, inverted: the model is cheap, the context wrapped around the model is priceless.

🚀 Anthropic Published the Receipt

Which brings us to the second story, and it’s the one I’d frame the whole issue around if I had to pick one.

The same day, Anthropic dropped an economic study of roughly four hundred thousand Claude Code sessions, run between October and April. Strip out the academic language and here’s what it found. In a typical session, the human makes most of the planning decisions (what to build, what matters, what good looks like) and the model makes most of the execution decisions, the how. And the kicker: the more domain expertise the human brings, the more work the model does per instruction. Expertise doesn’t get replaced. It gets amplified. Over the seven months, the share of sessions spent debugging fell by nearly half, work shifted toward genuine end-to-end delegation, and the value of the typical task — measured against what the same job would cost on the freelance market — rose about twenty-five percent.

I’ll be straight about the one number that looks like it cuts the other way, because it’s actually the most important number in the study. The gap in success rate between an intermediate user and an expert is modest. Every major occupation succeeds at nearly the rate software engineers do. Read quickly, that says “expertise doesn’t matter much” — the opposite of everything above. Read correctly, it says something far more interesting.

Here’s the resolution, and it’s the whole issue. In almost every field, nobody has found the fulcrum yet. The technology is eighteen months old in any practical sense. Expert and intermediate are both standing the wedge in roughly the same default spot, the obvious one, so they get roughly the same result. The gap is modest because the discovery hasn’t happened — not because expertise doesn’t matter, but because the optimal placement is still hidden, for almost everyone, in almost every domain.

With one exception. Coding. And coding is the exception for a precise, mechanical reason: it has the tightest feedback loop in the working world. You place the wedge, you run the test, and in seconds reality tells you whether you were right. That loop, fast and cheap and objective, is a machine for discovering optimal placement. So in coding the people who found the right spot have already pulled away from the people who didn’t. That’s the “1000X coder” you keep hearing about — a figure of speech, not a measured constant, but the direction is real. There is no 1000X lawyer yet, and the reason isn’t that lawyers are slower or law is softer. It’s that law has no unit test. The verdict on a brief arrives slowly, fuzzily, sometimes never. Slow loop, hidden fulcrum, modest gap.

This gives you a lens sharper than any benchmark. The fulcrum moves first in whatever domain you can verify cheaply and fast. It already moved in coding. It moved years ago in high-frequency trading, where the P&L marks you in milliseconds — which is exactly why Jane Street and Citadel became machine-learning monasteries that pay like nowhere else on earth. It’s moving now in ad auctions, in recommendation feeds, in sports betting, in anything you can turn into a scoreboard. It has not moved in medicine, in strategy, in venture capital, in law — the places where the answer takes years and nobody can grade it cleanly. Watch the feedback loops and you can call which profession gets its 1000X next.

And here’s the entrepreneurial version, the one worth more than the observation: in a fuzzy field, the edge goes to whoever builds the loop that doesn’t exist yet. That’s exactly what Harvey did in law — it constructed private benchmarks so it could iterate, manufacturing a scoreboard where the profession never had one. The meta-skill of the next decade isn’t using AI. It’s turning unverifiable work into verifiable work, because the moment you can score it, you can find the fulcrum, and whoever finds it first owns the domain.

We made a version of this argument the week before last, in The Dr. House of AI: that the new literacy isn’t running the model, it’s being qualified to grade it. Anthropic just put four hundred thousand sessions of data underneath that claim. And the convergence kept coming. The very same day, Satya Nadella shipped Microsoft’s Copilot Cowork to general availability — long-running agents on multi-step work — and the headline feature was multi-model support. Route across models. Don’t marry one. The exact play we’ve been describing for a month, now standard-issue from the biggest enterprise software company in the world. When Redmond ships your thesis as a product feature, you’re either right or about to be copied. Probably both.

🌐 And If You’re Not Musk

Here’s where it lands for the reader who doesn’t have sixty billion in rocket stock to go shopping with.

Bots now outnumber humans on the web — actual measured traffic, more machines than people. Search is leaking value by the quarter; SEO, the discipline a generation of businesses built their demand on, is fading into irrelevance as customers stop clicking ten blue links and start asking one chat box. The self-help book market, Tim Ferriss says flatly, is already gone. The open web that everyone published into for twenty years is being replaced by an interface that answers the question and never sends the customer to your door.

So your fulcrum, if you run a business, is two things you have to own outright. First, an audience — a direct line to the people who buy from you that doesn’t depend on renting reach from a platform that’s quietly turning off the tap. Second, findability by machines. If an agent can’t reach you, retrieve you, and cite you, you are functionally invisible in the only storefront that’s going to matter. The revolution isn’t coming. It’s here, and its first casualty is rented distribution. Owned distribution is a place to stand. Rented distribution is the lever you don’t actually control.

One honest caveat on the Cursor trade, since we’re being clear-eyed. The bull case I just laid out — the process data, the observatory on every rival — is also a perception risk for the thing Musk bought. A chief security officer at a Fortune 500 doesn’t parse the subtle distinction between “he wants your IP” and “he only wants your process telemetry.” The headline that lands on their desk is our code now routes through Elon Musk’s company. Whether Cursor’s enterprise base holds or flinches is the open question sitting under the sixty-billion-dollar number. The flywheel that makes the deal brilliant is the same flywheel that could spook the customers who make it worth the money. That tension doesn’t break the thesis. It’s the thing to watch.

The Three Fulcrums

So here’s the whole issue in one frame. Three stories, one shape.

Cursor is worth sixty billion because it owns a place to stand: the index of how the world writes code. The expert in Anthropic’s data succeeds because she owns one too — domain judgment the model amplifies instead of replacing. And the business that survives the death of the open web will own a place the machines can actually find: an audience, and a front door built for agents. Proprietary data. Proprietary judgment. Proprietary distribution. Three fulcrums, one principle: in a world where the lever is free, the only thing worth owning is the ground you brace it against.

If you can name, in one clean sentence, the thing your business owns that a free model can’t copy — you’re fine. If the line comes up blank, that’s not a small problem. That’s the whole problem, and it’s the one worth spending this week on.

The lever is free. Knowing where to place the fulcrum is what matters.

Sources & Further Reading

• Elon Musk Doesn’t Run Six Companies. He Runs One Router — on organizational design as data routing
• Buy Wins, Not Players — on capability commoditization
• The Good, the Bad and the Ugly — on China’s response to export restrictions
• The Dr. House of AI — on expertise amplification and hiring models
• EP 07: The Talent Gap — Who Becomes a 100x Engineer When AI Does Everything?
• Anthropic Shipped Claude Channels. Your AI Agent Can Now Text You Back. — on agentic multi-channel deployment
• SpaceX–Cursor $60B acquisition (CNBC/Reuters, Jun 16, 2026)
• Anthropic, “Agentic coding and persistent returns to expertise” (Anthropic Economic Research, Jun 17, 2026)
• “How Cursor Indexes Your Codebase” (Jun 2026)
• “Pentagon used Grok to fire 2,000 missiles at Iran, official says” (The Independent, Jun 16, 2026)
• “Bots Now Outnumber Humans on the Internet” (2026 Bad Bot Report coverage)
• “The Web We Know Is Going to Disappear” (minid, Jun 15, 2026)
• Microsoft Copilot Cowork GA, multi-model support (@satyanadella, Jun 16, 2026)
• Benedict’s Newsletter No. 647 (Jun 14, 2026)