The real moat in AI coding sits above compute, on the model

The two loudest facts in AI coding tools right now point in opposite directions, and the gap between them is where your next bet lives.

The most money went to Cursor, the AI coding editor. In a deal reported this year, SpaceX bought it for $60 billion in stock and folded it toward Colossus, the giant computing cluster Elon Musk’s companies built to train and run models at scale. The most developer affection went somewhere else entirely. A July market-share analysis by andrew.ooo puts Claude Code, the coding tool built by the AI lab Anthropic, at a 46% developer satisfaction rating, against Cursor’s 19% and GitHub’s Copilot at 9%. Same category. The best-funded tool and the best-liked one are not the same tool.

If you skipped the tooling news this year, here’s the whole of it. “AI coding tools” are the assistants that write and edit code for working developers. The market consolidated fast, to about three that matter: Copilot, Cursor, and Claude Code. The JetBrains developer survey in January, with more than 10,000 responses, put them at 29%, 18%, and 18%. Adoption is near-total: getpanto.ai’s 2026 roundup reports 84% of professional developers using these tools and 51% using them every day. That’s most of the profession’s daily work.

Now the read most people took from that consolidation, and it’s a fair one. Once every tool could plug into the same outside services, the thing separating winners had to be something you couldn’t copy by wiring up an integration. That something looked like raw compute: owning cheap capacity at Colossus scale, so you weren’t at the mercy of a model provider’s pricing. SpaceX buying Cursor was that whole thesis made literal. Compute is the moat.

The satisfaction spread says the moat climbed one floor higher than compute. The durable edge in AI coding tools is the frontier model on top of the stack, the most capable class of AI that only a handful of labs can train, and that’s the one layer a compute deal structurally cannot buy.

Start with why a satisfaction gap is the number to watch, not this quarter’s revenue. When developers hit a usage cap, they switch tools. That’s the main trigger, more than any worry about safety or features. So the metric that predicts who survives a switching event isn’t how much money a tool booked. It’s how much its users want to keep it. Cursor has the money story: its own reported numbers put it around $2 billion in annual recurring revenue, roughly doubled in a quarter. Claude Code has the wanting story: 6× growth in nine months and a satisfaction rating more than double Cursor’s. A 46-to-9 gap between the model maker’s tool and Copilot is people saying the product is better, and the product is mostly the model.

Here’s the part that makes this sharper than a rich company getting out-liked by a scrappier one. The compute-is-the-moat view has the logic right. It just stops on the wrong floor.

The logic goes: as each layer of a stack becomes free to copy, the moat rebuilds on the layer above it. That already happened once. A shared standard called the Model Context Protocol, which lets any AI tool tap the same outside services, made integrations free to copy. So the moat moved up, to compute. Correct. The view just stopped climbing one floor too early. There’s a layer above compute, and it’s the frontier model itself. That’s the floor Claude Code is standing on, and it’s the one a compute acquisition can’t reach. SpaceX could hand Cursor a computing cluster. It could not hand Cursor a frontier model that Cursor controls. Cursor still rents its intelligence from Anthropic and OpenAI, the same labs whose own tools now out-satisfy it.

So Cursor sits in the middle of a sandwich. It commoditized the integration layer below it, and it rents the model layer above it. What it owns outright is the part in between: the editor and the workflow it wraps around the model. Useful, real, worth money. But the layer that developers are actually rating when they say one tool is better than another is the model, and that layer belongs to someone else.

I want to be honest about the strongest version of the other side. Compute bundling bought Cursor things that matter. It bought scale, a $2 billion revenue run, and relief from margin pressure, since owning capacity means not paying a model provider’s markup on every call. A company can run a long time on that. And these satisfaction figures come from blog-level market analyses, not audited numbers, so hold them as reported, not gospel. But even granting all of it, notice what the compute deal did not buy: the thing developers reward when they choose what to keep. Money bought Cursor a strong present. It didn’t buy the top floor.

Which turns the model-provider-owns-the-top-floor idea from a side risk into the main question. For most of this year that idea sat in the “things that might go wrong for Cursor” column. The satisfaction data just moved it to the front.

So if you’re a PM betting on this category, the sharp question isn’t which tool has the most funding or the biggest cluster. It’s this: is a compute-backed tool that doesn’t own its own frontier model a durable business, or a margin-improved tenant on a floor its landlord can repossess? And the harder version, the one worth actually trying to answer: does an independent app built on someone else’s model have a permanent seat here, or does every developer-facing surface eventually get absorbed by whoever makes the model it’s really selling? The integration layer already collapsed into the layer above it. The open question is whether the app layer is next. If it is, the first sign will be one we can already read today: which tool developers say they’d hate to lose.