← IndexA month ago, Satya Nadella published an essay on the future of the firm with one line everyone quoted: “the real opportunity is not in picking the best model but instead in building a learning loop on top of models.” (A frontier without an ecosystem is not stable, June 14, 2026.) Everybody nodded. Then everybody went back to benchmarking GPT against Claude against Gemini, re-running the bake-off every quarter, treating the leaderboard like the thing that decides who wins.
A month later, I still haven’t seen anyone notice that the idea Nadella described already has a name, a fifty-year history, and a body of theory that tells you exactly how to build one.
So here’s the whole essay up front. Boyd taught us that fights are won by orientation, the internal model that turns what you see into what you do. Nadella says firms will be won by learning loops. Same claim. What’s new is that, for the first time, orientation is something a company can own instead of something that walks out the door at five o’clock. Advantage is tempo times retention.
The model is rented. The loop is owned.
The rest of this essay is why, and why the lab only sells you the engine, never the plane.
The loop already has a name
Observe, Orient, Decide, Act. Fifty years ago John Boyd, a U.S. Air Force colonel, wanted to understand why some fighter pilots won and others didn’t. What he landed on was bigger than any equation: a decision cycle, Observe, Orient, Decide, Act, that turns out to govern almost any complex contest, in a cockpit or out of one.
You’ve probably heard the pop version: decide faster than the other guy and you win. Boyd meant something deeper. Speed lives at the edges of his loop. Orientation sits at the center.
Orientation is the synthesized model you carry of how the world works, built from experience, training, culture, and the constant work of breaking down what you see and reassembling it into something useful. Two people can observe the same thing and act completely differently, because their orientation differs. Whoever holds the richer model converts the same raw information into the better move. Observe collects the information and Act spends it, but the fight gets decided in the middle, in the part nobody can see.
Boyd had a second insight that matters even more here. When your orientation is good enough, you stop deliberating. You skip the explicit Decide step and go straight from Orient to Act, because the model is rich enough that the right move is obvious. He called it implicit guidance. The better your internal model, the less you have to stop and think.
Hold onto that. It’s about to describe every agentic system your company is going to build.
Nadella’s loop is Boyd’s loop, externalized
Read Nadella’s essay against that backdrop and the mapping is exact. He talks about “human capital and token capital” compounding. He talks about a system that retains the “company veteran” expertise even when you swap out the underlying model. He talks about institutional memory that’s queryable, private evals that measure whether a model is improving against outcomes that matter to your business, a “hill climbing machine” that gets better with every use.
Strip the vocabulary and he’s describing Orientation. The learning loop is the firm’s orientation, lifted out of individual human heads and written down somewhere durable. The “company veteran who survives a model swap” is orientation that’s been decoupled from the substrate that happened to be carrying it.
That decoupling is the whole game, and nothing like it existed in Boyd’s world. His orientation lived in people. It was perishable. It walked out the door when your best operator quit, and it degraded every time someone retired.
Now a firm can externalize orientation, persist it, query it, and improve it on purpose. The thing that used to live only in your most experienced people can become an asset the company owns. Orientation becomes capital. That’s what Nadella means by token capital, even if he didn’t reach for Boyd to say it.
Where the loop breaks (and why that’s the interesting part)
I’ll flag where this stops being a clean fit, because the seam is where the real idea lives.
Boyd’s loop was adversarial. The entire payoff was to get inside your opponent’s loop and generate so much confusion that they fold. He wanted to create instability in the other guy. Nadella wants the opposite: a “stable equilibrium,” a “frontier ecosystem,” with value flowing broadly.
Boyd also optimized for tempo, the velocity of the loop in real-time conflict. His loop adapts, but it doesn’t obviously accumulate. Nadella is after a ratchet. Every cycle leaves a deposit. Each improved workflow generates a better training signal, which makes the next cycle smarter, which throws off a better signal again. Boyd’s loop gets faster with practice. Nadella is suggesting we can get both faster and smarter.
The theory: advantage is tempo times retention
Here’s my take, and I’ll label it a theory rather than a fact, because I’m extending Boyd past anything he wrote.
A firm’s competitive advantage is the product of two independent levers: the tempo of its loop and how much orientation each cycle retains. Tempo times retention.
Boyd maxed out tempo with retention effectively capped, because orientation lived in human memory and human memory leaks. AI adds a retention substrate, so each turn of the loop deposits durable, transferable orientation instead of letting it evaporate. Put the two together and you get the object Nadella calls a hill climbing machine. I’d call it what it is: a self-improving OODA loop. It adapts in real time, Boyd’s half, and compounds what it learns, Nadella’s half. Tempo and retention stop being a tradeoff and start multiplying.
This theory makes predictions, which is the only reason it’s worth naming:
- Firms will stop competing on model choice and start competing on orientation retention rate — how much of every interaction gets captured as reusable institutional memory versus lost to the void.
- AI-native companies will mature into fewer humans in the loop per outcome. As encoded orientation deepens, the approval steps drop out, just as Boyd predicted with implicit guidance.
- The moat won’t be how fast your loop runs on any given Tuesday. It’ll be the integral of retained orientation over time. Which is why Nadella says the advantage holds “regardless of any new individual model capability.”
Retention rate sounds abstract. You can measure it. Every unit of it gets paid for in tokens, and each token you buy either deposits durable orientation or evaporates on contact. Your AI bill is secretly a report on which one is happening. Most companies have never read it that way.
The pilot needs a plane
There’s a piece of Boyd’s world we keep leaving out. The pilot never ran the loop alone. He ran it with the plane. Boyd himself spent years on the aircraft-design math that shaped the F-15 and F-16, because he understood that a great pilot in a mediocre jet loses to a good pilot in a great one. The machine sets the ceiling on what the human’s orientation can actually express.
AI works the same way. The model is the engine, and it’s a remarkable one, but an engine bolted to nothing doesn’t fly. What turns a model into a loop is the harness around it: the tool access, the memory, the context, the evals, the orchestration that let a human and a model observe, orient, and act together at full speed. The harness is the plane. And right now almost everyone is still arguing about engines.
So what’s the F-35 of this generation? Worth remembering why the F-35 actually wins. In a knife fight, an F-16 can out-turn it. The F-35 wins anyway, on sensor fusion: it takes every input, radar, infrared, electronic warfare, data from other aircraft, and fuses it into a single coherent picture, so the pilot sees first, understands first, and acts first. It wins on orientation. It is, almost literally, an OODA loop with a cockpit.
That’s the harness worth building: the one with the best fusion, the system that pulls your organization’s context, history, and judgment into one picture and hands it to the human and the model at the moment of decision. The F-35 of the AI generation is whatever harness gives its crew the best orientation. The lab builds the engine. You build the plane.
The sovereignty question
This is where Nadella’s ecosystem argument stops being a policy aside and becomes the whole point. He warns about a world where a few models “eat everything they see” and entire industries find their knowledge commoditized out from under them. He’s stopped being subtle about it, too. At Stanford this month he put it flatly: “If you’re just a consumer of a foundation model, then I’m not sure how you can retain enterprise value, let alone create it.”
The orientation lens explains the fear and points at the fix.
If orientation lives only in the model, the model owner owns you. Every interaction you have feeds a loop you don’t control, and your hard-won institutional judgment becomes someone else’s training signal. That’s the hollowing-out he’s describing. But if orientation is capital you can externalize and own, the dependency inverts. You rent the raw capability and keep the model of how your business actually works. Nadella calls switching out a generalist model without losing your company-veteran expertise “the key test of your control and sovereignty in the era ahead.” He’s right, and the OODA frame is why: sovereignty in this era is ownership of your own orientation.
He’s no longer alone, and the volume is rising. Alex Karp spent twenty minutes on national television last week channeling what he called “the voice of American business” — enterprises that are “livid,” paying for “tokens that create no value” — and Palantir followed with a white paper on institutional sovereignty. Set the theatrics aside and the orientation lens diagnoses his complaint precisely: a token that creates no value is a loop cycle with zero retention. You paid for tempo and kept nothing. That diagnosis, and where I think Karp’s prescription goes wrong, deserves its own essay. It’s the next one I’m writing.
We’ve seen this movie one layer down. The companies that thrived in cloud built systems to understand their own operations no matter what the provider changed underneath them. The ones who didn’t became permanent tenants, re-architecting every time a vendor moved the floor. AI is the same bet with bigger numbers and a faster clock, and this time the stakes have moved well past your infrastructure bill. What’s on the table is whether your accumulated judgment belongs to you or to whoever rents you the model.
Steal the vocabulary
A frontier ecosystem stays stable only when value flows to the firms that build the loops, and that happens one company at a time, each deciding to own its orientation instead of renting it. Yours included.
So steal the vocabulary. The next time a meeting opens with model comparisons, ask the questions that actually decide who wins: What’s our retention rate? What did our loop learn this week that we own? Is this token spend depositing orientation, or evaporating? You’ll know the frame has done its job when the leaderboard conversation starts sounding like what it is — a debate about which engine to rent for a plane you haven’t built.
The model is rented. The loop is owned.