This week Satya Nadella handed every CEO a clean way to think about AI, and most of them will act on only half of it.
His framework is simple. Every company now holds two kinds of capital. Human capital is the judgment, relationships, and taste of your people. Token capital is the AI capability you build from your own work and own outright. The two are joined by a learning loop, where your workflows and your corrections feed a system that improves with every use. Nadella calls that loop “the new IP of the firm,” and warns of a world where companies are “ceding value to a few models,” their hard-won expertise absorbed into a general model and sold back to their competitors, the way outsourcing once hollowed out manufacturing.
He is right. Building your organization’s knowledge into an asset that compounds, instead of renting it back from a model that learned on your work, is exactly the direction companies should be moving. This is the same argument we have made about the harness for months: at today’s model floor, the durable advantage is the layer you build around the model, not the model itself. Nadella’s contribution is the language. “Token capital” turns an engineering project into a line on the balance sheet, and that is the framing that survives a budget meeting.
The half most CEOs will skip is the word “own.”
The framework is right. The instruction is incomplete.
A learning loop is about to become consultant boilerplate. By next quarter every vendor, integrator, and advisor will tell you to build one. That advice is correct and nearly useless, because it stops at the point where the real decision begins. The question is not whether to build token capital. Everyone will tell you to. The question is whether you own it in a form you can carry out the door.
Read who is making the case. Microsoft is the largest investor in OpenAI, a stake worth around $135 billion, and the primary cloud the company runs on. The executive warning you against ceding value to a few models helps control one of the biggest of them, while selling you the platform to build your “owned” alternative on. That is not a reason to dismiss the framework. It is a reason to read the fine print on where the loop actually lives.
And the ground under that question keeps moving. In its October 2025 restructuring, OpenAI became a public benefit corporation and, in the same deal, won the right to serve its products beyond Azure alone. The detail matters less than the pattern. The terms binding even the largest players get renegotiated on their schedule, not yours, so the stack you build your loop on can change shape without your input. That is precisely why the parts that compound need to be yours to move.
The loop has three layers. You only rent one of them.
The mistake is treating the learning loop as a single thing that either belongs to you or does not. It has three layers, and they carry very different ownership stakes.
The first is the data layer: the corrections your team makes, the examples of good and bad output, the transcripts, the evaluation sets that define what “right” looks like for your business. This is the rawest form of token capital, and it is the easiest to lose without noticing, because it accumulates inside whatever tool your team happens to use. Keep it in open, exportable formats that you control. If your corrections and golden examples live only inside a vendor’s proprietary store, you are renting your own memory.
The second is the definition layer: the workflows, the prompts, the context files, the codified skills that tell the model how your company actually works. This is where the compounding happens, and it is portable by design if you treat it that way. A skill written as a plain, documented runbook survives a model swap and a vendor swap. A skill that exists only as buttons clicked inside one platform does not. Open formats for exactly this now exist, skills and prompts you can store as plain, portable files rather than as settings locked inside one product, so the layer can move with you, and they are worth insisting on.
The third is the runtime layer: the model and the compute it runs on. This one you rent, and renting it is fine. Frontier capability is genuinely a commodity you should buy from whoever leads this quarter, because the leader keeps changing. The model decision is a framework problem, not a procurement problem, and treating the runtime as rented is the correct posture. The danger is not renting the runtime. The danger is letting the first two layers fuse to it, so that switching the model means rebuilding the asset.
Owned token capital, in practice, means the data layer and the definition layer stay yours and portable, while the runtime stays rented and replaceable. Get that division right and a vendor change costs you a migration. Get it wrong and a vendor change costs you the asset.
The layer that has to stay human
There is a fourth piece that never belongs to the vendor at all: the review layer. The corrections that feed the loop have to come from people who can tell good output from confident nonsense, and that judgment is what keeps the loop improving instead of drifting. This is where the harness actually pays off, and it is also the part most exposed when a forced model change drops a different model into the same workflow. Token capital without a human review layer is not an asset that compounds. It is an error that compounds.
Name the owner
An asset with no owner on the org chart is not an asset. It is an accident waiting to be lost when one person leaves. If token capital is real, someone owns it: a named person accountable for the data layer, the definitions, and the review standard. In most companies that person already exists, quietly maintaining the workflow that now does what their old job description used to. Give the role a name, a budget, and a line in the AI plan, or accept that your most valuable AI asset is held together by goodwill.
The objection worth taking seriously
There is a real cost to all of this. Portability is not free, and the deepest value often comes from integrating tightly with one platform’s tools. A small or mid-sized company that insists on total vendor independence can spend so much on optionality that it never builds anything that compounds. That objection holds, and the answer is not purity. It is tiering. You do not need to run your own models or avoid every platform. You need the IP layer, your data and your definitions, to be exportable, while the runtime stays rented and deeply integrated. Keep the part that compounds portable. Rent the part that is genuinely a commodity. That is the balance, and it is achievable at 80 to 200 employees without an infrastructure team.
The test to run on Monday
Here is the one question that turns Nadella’s framework into a decision. If your company switched its main AI vendor next quarter, what fraction of what you have taught the system would move with you?
If the answer is “most of it,” you are building token capital. If the answer is “we would start over,” you do not have an asset. You have a subscription that feels like one. The work between those two answers, making the data exportable, the definitions portable, the review standard explicit, and the owner named, is the actual project. Nadella named the asset. Whether it ends up on your balance sheet or your vendor’s is the part he left for you to decide.
Questions this article gets
What does Satya Nadella mean by token capital?
Token capital is the AI capability a company builds from its own work and owns outright, as opposed to the model it rents. In Nadella's framing it sits alongside human capital, the judgment and relationships of your people, and the two are joined by a learning loop where your workflows and corrections feed a system that improves with every use. He calls that loop the new IP of the firm. The value of the phrase is that it reframes an engineering effort as an asset on the balance sheet, which is the language that wins a budget decision.
How does a company actually own its AI learning loop?
Treat the loop as three layers with different ownership stakes. The data layer is your corrections, examples, and evaluation sets, and it should be kept in open, exportable formats you control. The definition layer is your workflows, prompts, and codified skills, and it should be stored as plain, portable files rather than settings locked inside one product. The runtime layer is the model and the compute it runs on, and that you rent, because frontier capability is a commodity and the leader keeps changing. Own the first two, rent the third, and a vendor change costs you a migration instead of the asset.
Should a mid-sized company build its own AI models to avoid vendor lock-in?
No. The answer is tiering, not independence. Running your own models or avoiding every platform is expensive and usually self-defeating for a company of 80 to 200 people, because the deepest value comes from integrating tightly with a platform's tools. The discipline is narrower: keep the parts that compound, your data and your definitions, exportable, while the runtime stays rented and deeply integrated. A human review layer that catches drift stays in-house, and one named person owns the whole thing.