Here’s a number that should reframe how you think about the entire AI trade: on June 13, a Chinese lab called Z.ai released an open-weight model, GLM-5.2, that lands within about one percentage point of Claude Opus 4.8 — the most expensive frontier model money can buy — on closely watched agentic coding benchmarks. The catch? It costs roughly one-sixth as much to run. On output tokens, GLM-5.2 is priced around 17.6% of Opus 4.8.
That is the AI cost collapse in a single data point, and it is the biggest wealth story of 2026 that almost nobody is playing correctly. While the financial media argues about whether Nvidia is a buy after its 12% June swoon, the ground underneath the whole industry is shifting: the price of raw intelligence is falling faster than the price of almost anything in economic history. And when the cost of something falls toward zero, the money doesn’t disappear — it moves. The entire game is figuring out where it moves to.
Intelligence is deflating about 10x a year
Step back from the day’s headlines and look at the trend line. In November 2022, running a model at the level of GPT-3.5 cost about $20 per million tokens. By October 2024, the same capability cost about $0.07 — a roughly 280x decrease in two years. Researchers at Epoch AI who track this call it one of the fastest cost declines of any technology they measure. Andreessen Horowitz gave it a catchier name: LLMflation. For a fixed level of capability, the price to deliver it has been falling somewhere between 10x and, at the extreme, 900x per year depending on the task.
That is not normal. Solar panels, the previous poster child for deflation, fell about 10% a year. AI intelligence is falling 10x a year. Compounded, that is the difference between a slow tailwind and a hurricane.
Two forces drive it. First, the models themselves get more efficient — better architectures, quantization, speculative decoding, smarter caching. Second, and more importantly for your money, competition. When Z.ai can ship an open-weight model that a Vercel executive described as “almost shocking,” and undercut the US labs by roughly 80% on price, the US labs cannot hold their margins. GLM-5.2’s own cheapest input price fell 35% in about 90 days. Nobody has pricing power when the thing you sell is being commoditized in real time.
The contrarian read: the model labs may be the worst place to put your money
Here is where most investors get it exactly backwards. The instinct is: AI is the future, so buy the companies making the AI. Buy the frontier labs, buy whoever has the smartest model.
But think about what the cost collapse actually does to that business. If your product — frontier intelligence — is being matched by an open-weight competitor at one-sixth the price within weeks of your release, you do not have a durable moat. You have a treadmill. You spend billions training the best model, enjoy your lead for a few months, and then a cheaper clone erases your premium. Some analyses suggest OpenAI is losing money on inference at current prices — precisely because the price it can charge keeps getting compressed from below.
This is the classic pattern of a deflating input. The people who make the commodity rarely capture the wealth. The people who use the commodity to build something else do. Nobody got rich selling electricity at cost in 1920 — they got rich building the appliances, factories, and businesses that electricity made possible. The oil didn’t make the fortunes; the refineries, logistics, and downstream products did.
Intelligence is becoming the new electricity: a cheap, abundant input. And that reshuffles the entire wealth map.
Where the money actually moves
If raw intelligence trends toward free, value accrues to three places, and this is the part that matters for your portfolio and your paycheck.
1. The picks-and-shovels that scale with usage, not price
Nvidia slid 12% in June and the bears declared the AI trade dead. But look at what the buyers are doing, not saying: hyperscalers are on track to spend about $650 billion on AI capex in 2026, and Nvidia expects that number to approach $1 trillion next year. Cheaper intelligence doesn’t reduce compute demand — it explodes it. This is Jevons’ paradox in action: when a resource gets cheaper, we use dramatically more of it. Every 10x price drop unlocks a hundred new use cases that were uneconomic before. The infrastructure layer — chips, power, data centers, networking — benefits from volume, and volume is going vertical even as per-token prices crater. We broke down which infrastructure names are quietly outrunning Nvidia in this look at the AI infrastructure stocks up triple digits.
2. The application layer that owns the customer
The companies that will compound wealth from AI are the ones sitting between cheap intelligence and a paying customer with a real problem — software with distribution, proprietary data, and workflow lock-in. Their input cost (intelligence) is falling 10x a year while the price they charge customers holds. That is expanding gross margin, quarter after quarter, handed to them by the cost collapse. It’s the same logic behind why the smartest money in 2026 is skipping Nvidia and going two layers down — the market is learning to pay for businesses that turn AI into cash flow, not promises.
3. You
This is the part Wealtharian cares about most, and the part the finance accounts miss. The AI cost collapse is not just a stock story — it is the single greatest leverage tool ever handed to an individual. Capability that cost a mid-size company a fortune two years ago now costs you the price of a coffee per month. A solo operator in 2026 can run a research analyst, a copywriter, a coder, and a customer-service team out of a laptop for less than the old cost of a single junior hire. That is a wealth-creation opportunity, not a threat — if you are on the building side of it. As we argued in why the AI selloff just handed wealth builders a gift, the winners won’t be the people who fear AI, but the ones who quietly compound its falling cost into their own output.
The human angle Wealtharian won’t skip
There’s a genuinely good story buried in the spreadsheet. When intelligence gets cheap, it doesn’t just make coders faster. It puts a competent tutor in front of a kid who couldn’t afford one, a legal explainer in front of someone facing an eviction notice, a diagnostic second opinion in front of a patient in a country with three doctors per 10,000 people. Deflating intelligence is deflating the cost of expertise itself — and expertise has always been one of the things that separated the wealthy from everyone else. That gap narrowing is good for humanity and it redraws who gets to build wealth. Both things are true, and Wealtharian is not embarrassed to say so.
What to actually do
Don’t chase the “best model” — that lead evaporates in weeks. Think in layers instead. Own exposure to the infrastructure that scales with usage. Favor application businesses whose costs fall while their prices hold. And most of all, put the cost collapse to work in your own income: whatever you do, there is now a version of it that cheap intelligence makes several times more productive.
The people who get rich from the AI cost collapse won’t be the ones who guessed the winning chatbot. They’ll be the ones who understood that when an input goes to zero, the wealth moves downstream — and stood exactly where it landed.
Track your own path through the AI wealth shift. The Wealtharian Wealth Tracker lets you monitor your net worth, FU money progress, and investment milestones in one place — so you can see whether the biggest tech deflation in history is actually showing up in your bottom line. Try it free →