There are two conversations happening in AI right now, and they are barely talking to each other. The first is the one everyone follows: which lab has the best model, who hired whom, which researcher moved where. That conversation is about the frontier and it carries real weight, because when the people who built the infrastructure the whole industry runs on decide where to go next, the direction of travel is worth reading closely. The second conversation is running in parallel, it is getting louder every month, and it is the one more likely to decide who actually captures the value. It is not about how smart the models are. It is about what gets fed into them, who controls that, and whether the people operating these systems at industrial scale have any say in how they get built.
Palantir’s CTO put the first half of that plainly this week, arguing that we’re listening too much to the inventors of AI. The framing was not a dismissal, and he was explicit that the inventors are geniuses. It was a reorientation about whose report from the field counts as evidence. The signal that actually matters, in his telling, is the factory floor worker who just added a third shift because AI made the throughput possible. The operator. The implementer. The person who is not on the conference stage and will never be quoted in a launch post, whose experience of the technology is the only one that has been tested against a payroll.
Databricks’ CEO arrived at the same destination from a different road, arguing that AI has a context problem, not an intelligence problem. Every chief executive on the planet, he said, is asking one question, and it is not how to get a better model. It is how to feed the model their own data. Put those two observations side by side and they stop being separate takes. One says the industry is listening to the wrong people, and the other says the industry is solving the wrong constraint, and both are describing the same misallocation of attention. The bottleneck already moved downstream while the coverage stayed upstream.
What This Means for the Karpathy Story
The read on Andrej Karpathy moving to Anthropic was about signal: follow the person, not the press release. That still holds, and it should. When someone with that track record makes a move, the choice encodes information that no announcement will state directly, and reading it carefully is a reasonable use of attention.
But the Palantir and Databricks framing adds a layer that the talent story on its own can’t see. The researchers building frontier models and the operators deploying AI at industrial scale are increasingly optimizing for different things, and the gap between those objectives is widening rather than closing. The lab race is one theater, with its own scoreboard: safety versus capability, talent concentration, benchmark position. The data infrastructure race is a different theater entirely, and its scoreboard is the context pipeline and the question of who controls the information the models are allowed to see. A researcher moving between labs is a decisive event in the first theater and nearly irrelevant in the second. The researchers are winning the first war convincingly. The second war has barely started, and the entrants are not the same companies.
The Context Builder Race
You can already see the shape of it in what the ecosystem is actually building rather than what it is announcing. Vector databases, retrieval pipelines, knowledge graph layers, and enterprise data connectors have become a category unto themselves, with real revenue and real competition. Startups are raising nine-figure rounds not on model quality, which they do not have and do not need, but on context retrieval speed and accuracy, which is a much narrower and more defensible claim. The implicit argument in every one of those rounds is the same: the model is becoming a commodity, and the context layer is where value settles once it does. Investors are not saying that out loud, because saying it out loud would require repricing a lot of other things, but they are funding it.
There’s a product truth underneath this that Matt LeMay put cleanly in Product Management in Practice, asking whether the user needs and goals articulated by a team actually reflect the needs and goals of users, or just what the business wants those needs and goals to be. That question lands hard here. The labs have been answering the second version of it for years, optimizing for benchmark users and researcher intuitions about what capability should look like, because those are the users who are legible from inside a lab. The factory floor worker on the third shift is not in that sample, has no way to enter it, and has needs that don’t resemble a benchmark at all. The operators are now correcting the sample from the outside, which is the least efficient way for that correction to happen and apparently the only one available.
Who Already Sits Next to the Data
What makes this more than a philosophical disagreement is who is positioned to win the second race, because it is not obviously the labs. The advantage sits with the companies that already occupy the space between an enterprise and its own data: Snowflake, Databricks, Salesforce, the firms that spent years on the unglamorous problem of organizational data being messy, duplicated, badly governed, and scattered across systems that were never designed to talk. That work bought them relevance but not much excitement. Frontier AI handed them a new reason to matter, and it did so without requiring them to win a single benchmark.
Gene Kim’s line cuts straight through what follows: when coding is no longer the bottleneck, the rest of your organization becomes the bottleneck. Swap “coding” for “model quality” and you have this moment exactly. The models got good enough for a very large number of real tasks, and the instant they did, every other constraint in the system became visible at once. Data that was never cleaned. Permissions nobody can explain. Processes that only worked because a human was silently reconciling them. None of those are solved by a better model, and all of them are now sitting directly in the path of deployment.
The Stakes Underneath
Here is the uncomfortable version of the argument. It’s possible that who has the best model in 2026 matters less than who controls what goes into it. A lab can hire the best researchers in the world, and that is a genuine advantage in the first theater. Databricks already has the enterprise data pipelines. Google has the search index. Microsoft has the enterprise software moat and the distribution that comes with it. The model layer may be competing ferociously for a surface that the infrastructure layer is quietly turning into a commodity underneath it, which is a difficult thing to notice from inside the competition.
That is not a prediction, and it would be a mistake to treat it as one. The model quality gap is still real and still consequential, and a large enough capability jump would rewrite this whole analysis. But the claim that intelligence is no longer the binding constraint, that context is, deserves to be taken seriously rather than filed as a vendor talking point from a company that sells context infrastructure. The fact that a data company benefits from the argument does not make the argument wrong. It just means somebody with an incentive noticed it first.
The inventors of AI built something remarkable, and nothing here diminishes that. The people deploying it are now starting to tell them what it needs to be in order to be useful, which is a feedback loop the field has never really had, because until recently there was not enough deployment to generate one. How the labs respond to that loop, whether they treat it as signal or as noise from people who don’t understand the technology, is probably a more consequential decision than the next benchmark drop. Benchmarks get beaten every few months. A field that learns to listen downstream only gets built once.
Leave a Reply