SheeplzSubscribe
Where AI takes the economy
By Dylan Siegel
This is my understanding and opinion, for education. Not personal financial advice.
New to this? Read first: What's next: AI, robotics and productivity
When a computer can do anything a person does on a computer, work stops being limited by how many people you can hire. That's the part I don't think is priced in.
My take, in one line: the AI buildout isn't about chatbots; it's about machine labor, and if it works, growth goes far beyond anything in today's forecasts. Nvidia sells the engine.
Step 1: AI does the whole job on the computer
Today people operate software: engineers drive CAD, programmers write code, analysts build models. The shift I'm watching is from AI that helps with a step to AI that does the whole job:
Design a more efficient electric motor. Produce the CAD files, simulate the heat, pick the materials, write the manufacturing instructions, estimate the cost, and check it against the spec.
Every part of that is digital. When an AI can run it start to finish, the cost of another design stops being specialist hours and becomes compute plus checking.
The catch is reliability. Long jobs have many steps, and errors stack:
| Chance each step is right | Steps | Chance the whole job is right |
|---|---|---|
| 95% | 50 | 7.7% |
| 99.9% | 50 | 95.1% |
(Real errors aren't independent, so this is an illustration, not a forecast.) Going from 95% to 99.9% per step looks small. It turns a toy into a worker. That's why I watch how long a task AI can finish on its own, more than benchmark scores.
Step 2: AI speeds up AI
Once AI writes much of the code and runs many of the experiments, it helps build the next AI: better models, better algorithms, better chip designs, better ways to make chips. Each round can make the next one faster. Nobody knows if that becomes an explosion or runs into limits (data, power, experiments that take real time), but it's the reason the upside is so wide.
Step 3: it reaches the physical world
A rough order, overlapping, with no guaranteed dates:
| Phase | What happens | What slows it down |
|---|---|---|
| Digital work (now to ~2029) | Code, engineering files, analysis, design, operations | Reliability, security |
| Discovery (~2027 to ~2031) | AI proposes and tests ideas in drugs, materials, energy | Real experiments, trials, approval |
| Physical deployment (~2028 onward) | Robots, factories, logistics, cars, surgery | Hardware, safety, construction, rules |
The big idea isn't that one industry gets automated. It's that the time to invent, design and test things shrinks everywhere at once, and the improvements feed each other: better chips make better AI, which designs better chips, faster drugs, better factories.
What that could do to GDP
Small changes in growth compound into huge numbers. Start from a $30 trillion US economy growing 2% a year:
| Extra growth from AI | US economy after 5 years | Extra output per year by then |
|---|---|---|
| None (2% a year) | $33.1T | — |
| +1 point (3%) | $34.8T | +$1.7T |
| +3 points (5%) | $38.3T | +$5.2T |
| +8 points (10%) | $48.3T | +$15.2T |
This isn't fringe anymore. In September 2026 Anthropic published three scenarios for the US by 2030: AI makes the economy 1.6% bigger (modest), 8.3% bigger (substantial), or 32.4% bigger with growth around 15% a year (extreme). They say plainly these are scenarios, not predictions, and the extreme one comes with heavy job losses.
I don't think the extreme case is the most likely. I think most investors still treat anything above "modest" as science fiction, and that's the gap.
Why this is bullish for Nvidia even if AI gets cheap
The worry: chips and models get more efficient, so each task needs less hardware. But cheaper work tends to create more work:
| Compute needed per task | ÷10 (90% cheaper) |
| Number of tasks worth doing | ×100 |
| Total compute | ×10 |
Economists call this the rebound effect (Jevons' paradox). Agents make it stronger: one chatbot answer is one model call, while one agent doing a real job can make hundreds of calls as it plans, codes, tests and fixes. Demand stops being set by how many people chat with AI and starts being set by how many AI workers it's worth running at once.
How much of it Nvidia keeps
The economy getting bigger doesn't mean Nvidia gets the money. A rough, made-up example to show the scale:
| Per year | |
|---|---|
| Extra US output from AI | $5,000B |
| If 5% of that is spent on AI infrastructure | $250B |
| If Nvidia gets 40% of that | $100B |
| At a 45% profit margin | $45B of profit |
Most of the value goes to the people and companies using AI, and to customers through lower prices. But a small slice of a multi-trillion-dollar shift is still enormous, and Nvidia is trying to be in every layer: the data centers, industrial simulation (Omniverse) and the chips inside robots.
The twist: AI could also help Nvidia's rivals
If AI gets good at designing chips and rewriting software, it helps Google, Amazon, AMD and Chinese chipmakers catch up too. Nvidia's moat is partly years of hard, human engineering. A world of cheap machine engineering is good for total demand and harder on any one company's share.
What would tell me it's really happening
- Agents finishing long, messy projects with little human help.
- AI-found drugs, materials and designs that hold up in the real world, not just on tests.
- Productivity numbers rising across the economy, not just at tech companies.
- Total demand for compute still climbing even as cost per task falls.
