There are two conversations happening in AI right now and they’re 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 has real weight. When the people who built the infrastructure that the whole industry runs on decide where to go next, you pay attention.
But there’s a second conversation running in parallel, and it’s starting to get loud.
Palantir’s CTO said something pointed this week: “We’re listening too much to the inventors of AI.” Not as a dismissal — he acknowledged they’re geniuses — but as a reorientation. The signal he said actually matters is the factory floor worker who just added a third shift because AI made it possible. The operator. The implementer. The person not on the conference stage.
Databricks’ CEO framed the same thing differently: “AI doesn’t have an intelligence problem. AI has a context problem.” Every CEO on the planet, he argued, is asking the same question — how do we feed it our data? Not: how do we get a better model. The bottleneck has already moved downstream.
What This Means for the Karpathy Story
Yesterday’s read on Karpathy moving to Anthropic was about signal — follow the person, not the press release. That still holds. When someone with that track record moves, the direction is worth reading carefully.
But the Palantir and Databricks frame adds a second layer. The researchers building frontier models and the operators deploying AI at industrial scale are increasingly optimizing for different things. The lab race — safety versus capability, talent concentration, model benchmarks — is one theater. The data infrastructure race, the context pipeline, the question of who controls the information the models get fed — that’s another theater entirely.
The researchers are winning the first war. The second war hasn’t really started yet.
The Context Builder Race
This is already visible in what the ecosystem is building. Vector databases, RAG pipelines, knowledge graph layers, enterprise data connectors — the tooling for feeding context to models has become its own entire category. Startups are raising nine-figure rounds not on model quality but on context retrieval speed and accuracy. The implicit argument: the model is becoming a commodity; the context layer is where the value actually lives.
What’s interesting is who’s best positioned to win that race. It’s not necessarily the labs. It’s the companies that already sit between enterprises and their data — the Snowflakes, the Databricks, the Salesforces. The companies that have spent years solving the messiness of organizational data at scale. Frontier AI gave them a new reason to matter again.
There’s a product truth underneath all of this that Matt LeMay put cleanly in Product Management in Practice: “Do the user needs and goals articulated by my team actually reflect the needs and goals of our users, or just what the business wants those needs and goals to be?” Labs have been answering the wrong version of that question — optimizing for benchmark users and researcher intuitions instead of for the factory floor worker on the third shift. The operators are now correcting that.
The Stakes Underneath
Here’s the uncomfortable version: it’s possible that who has the best model in 2026 matters less than who controls what gets put into it. Anthropic can hire Karpathy. But Databricks already has the enterprise data pipelines. Google has the search index. Microsoft has the enterprise software moat. The model layer might be competing for a surface that the infrastructure layer is slowly turning into a commodity.
Gene Kim’s framing cuts right through it: “When coding is no longer the bottleneck, the rest of your organization becomes the bottleneck.” Swap “coding” for “model quality” and you have the current moment exactly. The models got good enough. Now every other constraint in the system is showing up.
That’s not a prediction. The model quality gap is still real and still consequential. But the argument that intelligence is no longer the bottleneck — that context is — is worth sitting with seriously.
The inventors of AI built something remarkable. The people deploying it are now starting to tell them what it actually needs to be useful. That feedback loop is new. How the labs respond to it — whether they treat it as signal or noise — might matter more than the next benchmark drop.
Leave a Reply