If you could hear only one block from AI Update Bangkok 2026, Productize would pick this one, because it is the only block that is not about things you buy and use. It is about where the money in this industry flows.
It is also the block shaped most by the read of the industry that Dr. Jim (Jimmy Tejasen), who hosts the event, has built up over a long time in it.
Following part two, the map of the whole day, this is the last block: which layer of the AI stack the money sits in.
Part 1The pipe is worth more than what flows through it
The image he opens with is oil. Somebody drills. Somebody lays pipe and never drills at all. And a household puts solar panels on the roof and stops buying electricity.
The middle layer in AI is the one laying pipe: hundreds of models from dozens of providers behind one door, routed by task and price. Not one graphics card. Not one datacentre. A cut of everything that passes.
Three figures arrive together: roughly 25 trillion tokens a week, an annual value around 900 million dollars, and a take of about 5%. What is never said is whether the 900 million is everything flowing through or the middle layer's own cut. Those two readings are about twenty times apart, so take the set as a scale rather than as a number.
Then he knocks down his own analogy, which is the thing that makes this block more credible rather than less. Oil runs out, and gets dearer as it gets scarce. Tokens cannot run out, and their price falls continuously.
What follows is that value does not accumulate in the commodity. As the commodity keeps getting cheaper, value migrates to the route, and to the far end where the user touches something.
What behaves like oil is the pipe, not what flows through it.
Part 2Three layers, and who does the most work
Laid out as a table, the conclusion is fairly brutal for whoever works hardest.
| Layer | How hard it works | What it keeps |
|---|---|---|
| Model and hardware providers | Hardest: they train it, buy the machines, carry the power bill | Margin squeezed by a price war |
| The middle layer holding the route | Lightest: no machines of its own | About 5% of everything routed |
| Running it on your own machine | Middling: you maintain it yourself | No share paid to anyone |
The bottom row is the one people skip, because it generates revenue for nobody. Running locally does not merely avoid the model fee. It removes the middle layer and the provider at the same time, the way a house with solar panels drops the power station and the grid from its bill together.
One thing has to be said plainly here. This was told two ways. One says local execution and calling over the network are separate markets that do not compete. The other says local execution cuts both layers out. Productize tells the second version because it was the one pressed hardest, which is not the same as the question being settled.
Part 3How an AI gateway differs from the old kind
Anyone who has built backend systems already knows the old kind: the thing that controls who may call, how often, and where the call goes next. The line he draws here is sharp and easy to keep.
| The old gateway | The AI gateway | |
|---|---|---|
| Treats a request as | One call over the network | One job that has to be decided before it runs |
| Has to decide | Allowed or not, then forward it | Which model, what the tokens cost, and whether the rules permit it |
His advice on choosing one ages well, because it ranks your situation rather than the products. Look at what you already run, at whether you want to host it yourself, and at what your real question is. If the real question is where the money went, the tool that answers that is the one to pick.
What makes this layer valuable is neutrality, and he points out that neutrality is more fragile than it looks.
Four things would destroy it: ordering providers, capping price, excluding somebody, and keeping the traffic data to use yourself. The last is the dangerous one, because the day usage data gets monetised, neutrality ends immediately, and every bit of value built on it evaporates with it.
Slow to build, quick to break.
One more thing worth knowing before picking up tools in this family. He attached a warning to a tool he otherwise recommends: anything that intercepts the connection of a subscription-bound program and spends that quota against a different backend may breach the terms of both sides.
Part 4The layer the user touches, and the caveat it must travel with
The largest claim of the whole day sits here. Whoever holds the layer that wraps the model holds this era's operating system.
The reasoning is sound. That layer is the one a person actually opens, actually types into, and actually grants access to files. The model and the middle layer stay behind the curtain, to the point where the user does not know which model they are using.
He defines the word hold more precisely than expected: it means leaving becomes expensive. Expensive because of accumulated context, expensive because of a way of working that has been tuned into place, and expensive because of permissions already granted.
The sharpest point in the block is where local execution sits. A downloaded model is an inert file that does nothing until something drives it. The wrapper is therefore a router on the user's side, choosing local or remote before the middle layer gets to choose anything at all. Default it to remote and you feed both the middle layer and the provider. Default it to local and you cut both.
So what comes home from this block is not who wins. It is the question he leaves at the end: what actually decides it. His answer is that it is not capability. It is trust.
Whoever solves trust first gets to be the operating system.
And that is a question you can answer for yourself without waiting for anyone: the tool you use every day, how far would you go in handing it the keys to the house.
Part 5Which part to read next
This is the end of the series. If you have not read the earlier parts, start at part two for the shape of the whole thing.
- Part 1 Physical AI, and whether a mistake can be undone
- Part 2 The map of the whole day, and four questions before believing a number
- Part 3 Compressing models, and the quality cost you cannot see
- Part 4 Training a model, and the line that says when it is worth it
Source
Everything here comes from AI Update Bangkok 2026, held on 21 August 2026 at The Cloud Bangkok by Jimmy Tejasen. This part covers the session on the business structure of AI.
- Event livestream: watch the recording
- Jimmy Tejasen's work: aiserver.in.th, news and reviews of AI server hardware, and agentic-press.com, a publishing house for AI books
The figures here come from what was said on stage, not from anybody's financial statements. Use them for ratios and for asking better questions, not as market numbers. Productize has not reproduced anything shown at the event.