The model is a commodity.
The loop is the company.
And the loop needs a home.

Something quiet is happening inside some of the biggest companies on earth. According to recent reporting in the business press, names like Amazon, Walmart, Uber and Meta, companies that spent two years telling the world they were “all in on AI”, have started doing the opposite. They are rationing it. Setting quotas. Filtering who gets to use which model, for what, and how often. Not because the technology stopped working, but because the bill stopped making sense.

On one side, the loudest voices in tech now argue that every company must accumulate what they call token capital: its own AI capacity, its own learning loop, its own accumulated knowledge, or risk being hollowed out. On the other side, in the real world, finance teams are closing the tap because unmanaged AI costs more than it returns. Both things are true at once, and the gap between them is where the winners and losers of the next decade get sorted.

There is a third thing almost nobody talks about, and it is the one that matters most in the long run: where that loop actually lives, and who controls the ground it runs on.

What “token capital” actually means

Start with the thing most people get wrong: the battle is not about which model you pick.

The best model, whichever one it is this quarter, is available to you, to your neighbour and to your competitor, for the same price. It is rented, like electricity. You do not own it, and nor does anyone else. A capability everyone can buy is not an advantage. It is a utility.

The advantage lives above the model, in three things you can actually own.

Your evaluations

Stop judging AI on public benchmarks. Judge it on your results: does it make your business money on your tasks? A private scorecard, built on your own outcomes, is something no competitor can copy.

Your training ground

The model stops improving in a vacuum and starts improving on your traces: your real cases, your edge conditions, the messy reality of your operation. That data is yours and no one else's.

Your memory

The whole knowledge of the business, queryable, that no longer walks out the door when an employee quits. You can delegate a task. You can delegate a whole role. But you cannot delegate its learning, unless the learning is captured in a loop that stays.

Put simply: your data is the seed. It belongs to you and no one can take it. But a seed at the bottom of a bag grows nothing. To grow, it needs soil, water and sun, which here means compute, rented by the hour and the token. The seed is the asset. The loop is what turns it into a tree.

Why most AI spend is not capital at all

Consuming tokens does not create token capital. Plenty of companies are already burning through prompts, agents, copilots and coding assistants with no memory, no evaluation, no reusable traces, no intellectual property accumulating underneath. That is not capital; it is just spend.

Capital begins only when the usage feeds those same three things: when it enriches your memory, produces an evaluation signal, or turns a repeated task into accumulated knowledge you own. When the usage produces a measurable return. Once that loop is built, you can unplug one model, plug in the next, and keep your in-house veteran. The knowledge stays. That is the real test of whether you own anything.

The numbers suggest most companies are failing this test. Analysts have estimated that only around a quarter of AI infrastructure spending fully hits its return-on-investment targets. Other studies put roughly three-quarters of the economic value from AI in the hands of about a fifth of companies. And those leaders are not simply buying more AI. They are redesigning their business model around it. Everyone else is bolting AI onto the way they already worked, and paying for the privilege.

The bill only goes one direction

There is a comforting story going around that AI is getting cheaper. On the unit, it is true. Goldman Sachs has estimated the cost per token falling something like 60 to 70 percent a year. But the unit price is not your bill.

Two things break the comfort. First, when a cost of production falls faster than the price, that gap becomes the supplier's margin, not your saving. Second, and this is the one that gets everyone, when a resource gets cheaper you do not use less of it. You use vastly more. You put AI everywhere, all the time. Economists call it the rebound effect, or Jevons' paradox: cheaper per unit means total consumption explodes.

Now add agents. The best models no longer answer instantly. They reason, backtrack, re-read themselves and test several paths before replying. All of that is inference, and inference is tokens. A single agentic task can burn many times the tokens of a simple chatbot query. And agents do not sleep. For one human you might run ten, or ten thousand, chewing through compute around the clock. Gartner has estimated that inference could become the majority of a model's total lifetime cost: the brain thinking continuously, and you paying for every thought.

That is the shift. The old cost was building the brain, paid once, by the lab. The new cost is the brain thinking, forever, paid by you.

The part nobody costs in: where the loop lives

Here is the question that gets skipped. You have built a loop. Your evaluations, your training ground, your memory. Now, physically, where does it run?

For almost everyone, the answer is a handful of foreign hyperscalers. Which means your seed, the one asset that was supposed to be yours forever, is sitting on someone else's soil. Your customer records, your pricing logic, your operational know-how, the traces that make your loop valuable, all of it flows through infrastructure you do not own, under laws you did not write, priced in a currency you do not control.

Three risks compound quietly here, and none of them show up on the invoice.

Data security

Every trace you send out to improve your loop is a copy of your business leaving the building. The more valuable the loop, the more sensitive the data feeding it. A loop is only an asset if the data underneath it stays yours in practice, not just in the terms of service.

Sovereignty

When the loop runs abroad, you inherit the jurisdiction it runs in. Access can be throttled, prices can be reset, terms can change, and export rules written on another continent can reach into your business overnight. You optimised for owning the loop, then rented the ground it stands on from a landlord who can change the locks.

Concentration

The OECD has warned that when the compute layer — the cloud, the data centres, the chips — is controlled by a few players, those players can lock everyone else out. Arthur Mench put it bluntly at a French parliamentary hearing: whoever owns the chips, the electrons and the energy is the one who wins. Solve demand and you still have a supply problem, because once supply is monopolised by a few foreign actors, you simply cannot turn your electricity into tokens any more. That is not an AI problem. That has been true since markets existed. The only difference now is the speed at which it is happening.

VPS, local cloud, and local tokens: bringing the loop home

The good news is that owning the loop and hosting it yourself are no longer exotic. The building blocks are ordinary.

A VPS — a virtual private server — is the simplest one. A local provider rents you an isolated slice of a real machine in a real data centre, usually inside your own country. Your data sits there, under your jurisdiction, and you are not shipping every trace across a border to make your loop work. For a large share of business workloads — memory, retrieval, evaluation, most agent orchestration — a local VPS or a local cloud is more than enough, and often cheaper than the hyperscaler equivalent once you count egress fees and the premium you pay for a brand.

The model itself is the harder piece, because the strongest frontier models still live abroad. But this is changing fast. Open-weight models you can run on local hardware are closing the gap for a growing set of tasks. And even where you still want a top model, you can route inference through a local API provider — a local token supplier — rather than sending it straight to a foreign lab. The token is the unit of intelligence. It matters who mints it, where, and under whose law.

The practical stance is a mix, not a religion. Keep the seed and the loop local by default, on a VPS or local cloud you control. Use local tokens wherever they are good enough. Reach for a foreign frontier model only for the narrow slice of work that genuinely needs it, and never let it hold your memory. Own the ground. Rent only the ceiling.

Why a country should pay for this

Zoom out from the single company and the same logic decides the fate of a whole economy.

If every business must build token capital, and token capital is really just a meter of compute that never stops, then demand for compute is no longer a bet. It is structural. The only open question is where that demand lands. If it lands on foreign infrastructure, then a country like Malaysia spends the next decade converting its companies, its data and its productivity gains into rent paid to data centres somewhere else. The value leaks out by design.

This is exactly the kind of gap a government should close with incentives, because the market on its own will default to the biggest foreign provider every time. A few moves change the maths:

• Tax credits or grants for companies that host their AI workloads on local VPS and local cloud, so the cheaper-in-the-long-run choice is also the cheaper-today choice.
• Support for local API and local token providers, so a domestic supply of inference actually exists to buy.
• Public investment in sovereign compute — the data centres, the energy and the chips that let a country turn its own electricity into its own tokens.
• Data residency rules that keep sensitive sectors — health, finance, energy, government — on home soil, which creates guaranteed local demand and makes the whole thing bankable.

This is also why every country needs its own champion — a national provider that mints tokens on home soil. France has Mistral. Malaysia has ILMU, the sovereign model from YTL AI Labs, trained on Malaysian languages, data and context and run on Malaysian infrastructure. A local champion is what turns “use local tokens” from a slogan into a real option. Without one, there is no domestic supply to buy, and sovereignty stays theoretical.

None of this is protectionism for its own sake. It is the recognition that in an economy where intelligence is made from electricity and tokens, the country that hosts the loop keeps the compounding, and the country that only consumes it pays rent forever. A dollar of incentive that keeps a loop at home returns as local jobs, local data centres, local tax base, and a generation of companies that own their advantage instead of leasing it.

The uncomfortable conclusion

Follow the logic and you land somewhere uncomfortable. If value lives in the learning loop, and the loop runs on compute that never stops, then AI no longer just separates the companies that use it from the ones that don't. It separates the companies that turn tokens into advantage from the ones that only burn them. And underneath that, it separates the countries that host their own intelligence from the ones that rent it.

That second split is brutal, because advantage in a loop compounds. A small early lead, learning slightly better rather than spending slightly more, cumulates fast. The likely outcome is not that everyone rises together. It is that a minority of companies rebuild themselves around the loop and use the returns to absorb the competitors who never did. AI becomes a multiplier for the few and a divider for the many. The same is true nation to nation.

You do not need to believe every forecast to take the instruction seriously. Token capital does not get decreed. It exists only if the tokens you spend produce a return, and only stays yours if the loop that produces it lives somewhere you control. It is not a pile of prompts, an army of mis-routed agents, or a cloud bill growing faster than revenue. It is capital in the strict sense: something that produces a return, improves, and reinforces knowledge that is yours, on ground that is yours.

Where Trees OS stands

That is the whole reason Trees OS exists, so we will say it plainly. Most companies will treat AI as a subscription: rent a model, ask it questions, hand over their data, and quietly become clients of infrastructure they don't control. A few will build the loop, the evaluations, the training ground and the memory that stays, and they will keep it on ground they own. The second group owns something. The first rents everything.

We build the second kind, and sovereignty is not a footnote for us. It is a design rule. Three commitments follow from it.

You choose where your data lives

We suggest a VPS or a cloud based on where you want your data stored: your country, your jurisdiction, your rules, not ours and not a foreign provider's default.

We build model-agnostic

We are not married to any single lab. The model is the commodity, so we keep it swappable, which means you can unplug one and plug in the next without losing your loop.

We always propose a local option

Wherever it exists, we route inference through a local API so you consume local tokens, minted in your own country under your own law. In Malaysia that means a champion like ILMU from YTL AI Labs; in France it would be Mistral. A foreign frontier model stays available for the narrow slice of work that truly needs it, and it never holds your memory.

We proved it on ourselves first, in energy staffing, which is where Treelance came from. Our own data was the seed. The result is a company that gets sharper every week without adding headcount, on infrastructure we keep close to home.

The model you use this year will be obsolete next year. The loop you build around it, and the ground you build it on, is the only thing that stays. Plant the seed you own. Grow it on soil you control. Then keep the harvest.