For the last three years, the AI story was an intelligence race. Benchmarks, release dates, which lab's model could out-reason the other's. That race just hit a speed bump — frontier releases are slowing while regulators review them, and open models keep closing the public gap — and the competition moved somewhere more interesting.
Look at what the major players actually shipped recently. Apple is rebuilding Siri around access to your messages, photos, calendar, and screen. Anthropic put Claude inside Slack, where teams grant it access to selected channels, tools, and codebases. OpenAI's internal research shows its employees converging on Codex by pointing it at the files that matter. Three different companies, three different product shapes, one identical thesis: the model doesn't need to be smarter — it needs to know what's going on.
Nate B. Jones calls this the context war, and I think the framing is exactly right. A moderately intelligent model that already knows your situation beats a brilliant model you have to brief for ten minutes. Anyone who works with AI daily has felt this: the model is plenty capable, but before it can help, you paste the email, explain who the client is, explain which version of the deck is current, explain that yesterday's Slack thread reversed the decision. The bottleneck isn't intelligence anymore. It's situation.
Intelligence, meanwhile, is becoming a commodity. Capable open models get cheaper every month. What doesn't commoditize — what compounds — is context. Which is why every lab is fighting for it.
Whose moat is it, though?
Here's the part of the story that gets less airtime. Look at the roster of combatants in the context war: Apple, Anthropic, OpenAI, Google, Slack. Now look at the territory they're fighting over. The email you triaged at 7 a.m. The decision buried in a channel you got tagged into. The promise you made on Tuesday. The reason you picked vendor B back in March. None of that belongs to any of them. You produced it. All of it.
Yet every offer on the table has the same shape: give us your context, and we'll give you utility. The utility is real — I'm not cynical about that. Permissions and audit logs and private clouds are genuine engineering, and the products are genuinely useful. But notice what accumulates, and where. Each assistant gets more valuable to you the more of your situation it holds. That's the point. That's also the trap.
Play the default timeline forward. Your personal context lives in Siri's private cloud. Your team context lives in whatever memory the Slack assistant built. Your project context lives in a coding agent's workspace. Each vendor holds a shard of you. The shards don't interoperate — that would defeat their purpose. And in three years, when a better model ships from a different lab, you'll discover a new kind of lock-in: not your files, which you can export, but your accumulated situation — the years of context an assistant has built up about how you work. Files move. Vendor memory doesn't.
When context is the moat, and the moat is built entirely from your work, it's worth asking who owns the moat.
The missing player
Every war story has combatants and territory. In this one, the territory is you — and the territory doesn't get a seat at the table. Nobody on the current roster is building for the person who generates the context.
What would fighting for yourself look like? I think it comes down to four properties:
It accumulates. A context layer only pays off if it compounds — across years, jobs, and tools. That immediately rules out living inside any single vendor's product, because vendors churn and jobs change. The layer has to outlive every tool that feeds it.
You hold it. Physically. On hardware you control, in formats you can read without anyone's permission. "Trust our privacy policy" is a fine promise, but "it never left your machine" isn't a promise at all — it's an architecture.
You permission it. Not every model should see everything. The messy, sensitive stuff — compensation threads, health appointments, the candid assessment of a colleague — might only ever touch a model running on your own machine. The hard technical problem can go to a frontier model with a narrower slice. You decide, per question, what leaves.
The intelligence is swappable. This is the piece that inverts the war. If your context lives in a layer you own — exposed through open interfaces like MCP that any assistant can consume — then models become interchangeable clients of your data, instead of you being an interchangeable user of their memory. Plug in a local model for private work. Plug in this year's frontier model for horsepower. Plug in next year's winner when it ships.
Notice what that last property does to the economics. If you own the context layer, the labs have to keep competing for you — on intelligence, on price, forever — because switching costs you nothing. If they own it, they win you once and keep you through accumulated memory. Commodity intelligence plus owned context is the version of this future where the user wins. It's also the version nobody with a cloud to fill is going to build for you.
Answer the quiet question deliberately
The intelligence race was fun to watch, but it was never really about us — we were spectators to benchmark charts. The context war is different. Its territory is your working life, and for the next few years every convenient new assistant will arrive with the same quiet question attached: where does what-this-thing-learns-about-you live?
Most people will answer by default, one convenient permission grant at a time. It's worth answering deliberately instead. Your context — everything you've seen, decided, and promised — is the most valuable thing you produce at work. The tools are finally smart enough to prove it. Make sure the value lands with you.
This is the principle behind Third Brain — a local-first engine that builds your context layer on your own machine: email, chat, meetings, commitments, notes. Local models by default, any model you choose, and none of it becomes someone else's moat.