Your Moat Isn't Intelligence. It's Context.

Monday morning. 40 AI newsletters in my inbox from the past few days.

OpenAI shipped GPT-5.6. Moonshot…

A Chinese lab most hadn’t heard of, dropped Kimi K3, a free, open-weight model that benchmarks in the same territory as the best American frontier models.

Google launched Gemini 3.5 Pro. Anthropic extended the deadline on pulling Fable 5 from subscription plans. Again.

And scattered between those: a wave of agent launches, funding rounds with a B in them, and at least a dozen threads explaining why each of these changes everything.

That’s not a backlog. That’s less than one week of AI news. By the time I get through reading them, they'll likely be out-of-date.

You can’t keep up with this.

I can’t keep up with this, and paying attention to it is a meaningful part of my job.

Every morning there’s a new shiny toy, and every morning the industry picks it up, shakes it, and asks the same question: is this the one that changes everything?

It’s the wrong question. The flood of shiny new toys isn’t just exhausting us. It’s pointing our attention at exactly the wrong layer and what matters most.

I’ve spent the past year arguing that intelligence is becoming a commodity. That intelligence is becoming free, and judgment isn’t. This week, of all weeks, made the case better than I ever have, and it’s time to say the next part plainly.

AI capability is now the least durable moat in technology. Not one of the least durable. The least. A few years ago, access to the best models was an advantage. Today it’s a temporary condition, measured in months, or maybe even days.

Let’s have a quick look at what was actually happening underneath this week’s subject lines and why none of it really matters.

The real story isn’t any single launch. It’s a war on three fronts, and one month tells it.

This June, the Commerce Department ordered that Anthopic’s Fable, the most capable model in the world, go offline on national-security grounds. In a couple of weeks, they reversed course and cleared it.

Anthropic has since tried three times to pull it from subscription plans because there isn’t enough compute to serve the demand; the current deadline is today. OpenAI locked its newest model behind vetted partners to stay ahead of a “voluntary” executive order that’s already been enforced with takedowns. Google’s model cleared for launch reportedly because it tested less capable. And the hyperscalers are pouring some $700 billion into data centers this year to keep it all fed.

A war over models. A war over compute. A regulatory fog nobody can read.

Meanwhile, Kimi K3 landed Thursday: the capability Washington spent June treating as a state secret, free to download from a lab in Beijing.

The asset everyone is fighting over is collapsing in value, in real time, by free download. January’s insurmountable lead is June’s table stakes.

A lead that refills every quarter isn’t a moat. It’s weather.

Model quality still gates the game: below the frontier, agents fail and users leave. Capability decides whether you get to compete. It just no longer decides who wins.

If intelligence doesn’t hold value, what does?

Watch what doesn’t reset when a new model drops. Or when a model gets ordered offline by the government.

Trust doesn’t reset.

Trust doesn’t reset.

The client who’s watched you handle 400 of their transactions doesn’t re-run the vendor evaluation because a lab shipped a better benchmark, and doesn’t fire you because a regulator benched your model for three weeks. You swap the engine. The relationship doesn’t notice.

Workflow doesn’t reset.

Workflow doesn’t reset. Once you’re embedded in how the work actually gets done, ripping you out costs more than any intelligence gain is worth.

Data doesn’t reset.

Data doesn’t reset. Every transaction teaches you the edge cases, the exceptions, the way this regulation collides with that deadline. That knowledge compounds, and nobody else is accumulating yours.

That’s context. And the whole thing comes down to one asymmetry:

You can rent intelligence by the token. You can’t rent context (at any price).

You earn it, transaction by transaction, at the speed of trust.

And if you doubt this is where the value went, watch what’s going on around us in AI. In the middle of the most expensive arms race in business history, Anthropic and OpenAI diverted five and a half billion dollars into services ventures whose real product is just a matter of getting a seat inside the customer’s building.

The one thing their trillion-dollar valuations can’t scrape off the internet is the texture of real work inside real companies. The people winning the intelligence war are spending billions to buy context. Follow the capital, not the keynotes.

I run a company in one of the most trust-dependent transactions there is: real estate closings. In all the years of selling into that world, no client has ever asked me which AI model we use. Not once.

They ask who carries the liability when something goes wrong. They ask what happens to their client’s wire at 4:58 on a Friday.

That question is the moat talking.

In the least glamorous place possible: the gap between demo and deployment.

I’ve watched agents book meetings, write code, and review contracts. Impressive work. I’ve also watched those same agents collide with a security review, a compliance officer, an audit requirement, enterprise procurement.

The demo is what the model can do. Deployment is what the organization will let it do. That distance remains enormous, and in a summer when the government itself couldn’t decide what models were allowed to do, it widened.

That gap isn’t friction waiting to be engineered away.

That gap is the market.

Every governance requirement, every audit trail, every procurement gauntlet is a place where context gets earned, and every review you pass becomes a wall the next vendor has to climb.

If a frontier model can vanish on a Thursday by government order, what an enterprise needs isn’t the smartest model. It’s a partner who absorbs that chaos: swaps engines without breaking the workflow, keeps the audit trail intact, carries the liability.

The next decade of enterprise value won’t come from smarter models. It will come from making them, and the agents around them, transparent and governable at scale, inside workflows someone earned the right to sit in.

Next Sunday, my inbox will refill. Forty more newsletters. Maybe fifty this time.

There will always be shinier toys, a new model wearing the crown for a quarter. The capex will get more absurd before it gets rational. None of that is yours to win.

What’s yours is the question underneath: who trusts you? Whose workflow do you live inside? What do you know about your corner of the world that no lab is positioned to learn?

If you’re building or leading though an AI transformation right now, even if it’s just an idea on a napkin, audit where your context lives.

Ignore the tools for now. List the workflows you’re embedded in, the data only you accumulate, the reviews you’ve passed that competitors haven’t. That list is your balance sheet. The benchmark comparison is a depreciating asset with a half-life of one news cycle.

Longer term, stop trying to be the company with the smartest model. Become the company that context accumulates in, not the one that intelligence passes through.

The newsletters track the toys. The toys will always depreciate.

Intelligence is becoming free, even while a trillion dollars fights over it.

Context is earned at full price.

Go earn it.