You Put Money Into AI and Now You're Not Sure. Here's a Saner Way to Think About It.
You did the thing everyone said to do. You bought some exposure to the AI trade — a few names, maybe an index, maybe you just started paying for the tools and quietly filed that under "I'm betting on this." Then you read one headline about circular financing and another about a frontier model getting yanked offline by a government, and your stomach does a small flip. Is this 1999 again? Did I buy the top?
I build software for a living. I use these tools every day. I'm not going to tell you it's fine, and I'm not going to tell you to sell. I'm going to do something more useful: separate the two questions you've accidentally welded together.
Quick disclaimer, in plain English: this is opinion, not financial advice. I'm a developer, not your broker. Make your own calls.
The anxiety is rational. Name it.
Say it out loud instead of swallowing it. The capex numbers are genuinely strange. Hyperscalers are spending roughly $600–700 billion in 2026 alone — Amazon, Microsoft, Google, and Meta each clearing $100 billion individually — as part of an estimated ~$2.1 trillion of deployment across 2026–2028. That is not a normal amount of money.
Then there's the part that makes the back of your neck prickle: the financing has started to look circular. The poster child is the Nvidia–OpenAI tie-up. Nvidia's headline pledge to OpenAI was framed as up to $100 billion — a non-binding 2025 figure since reported scaled back toward roughly $30 billion — and critics call the structure a loop: a chip vendor funding the customer that buys its chips. OpenAI's CFO Sarah Friar has pushed back on the "circular financing" label, framing it instead as building and diversifying infrastructure. Either way, OpenAI is running roughly $25 billion in annualized revenue against about $14 billion in losses this year — and raised $110 billion in new funding anyway. Morgan Stanley's Todd Castagno notes the dot-com peak hit a capex-to-sales ratio of 32% in 2000; the AI boom is projected at around 34% in 2026, climbing toward 37% by 2028.
Above-dot-com spending ratios, financing that loops back on itself, a flagship lab losing money at scale. Your gut is right to flag it.
If the question is "could AI stocks get repriced hard," the honest answer is yes, obviously. That's how markets work. But that's only one of your two questions.
The other question: is the technology actually real?
Here's where the threads get knotted. People look at the financial weirdness and conclude the technology must be hollow. That doesn't follow — and as someone who ships code with these tools, I can tell you it isn't.
Concretely. Claude Opus 4.8 shipped on May 28, 2026, just 41 days after its predecessor, and sits at #1 on the Artificial Analysis Intelligence Index. Anthropic describes it as having "sharper judgement, more honesty about its progress, and the ability to work independently for longer than its predecessors," and in my daily use that tracks — it's roughly 4x less likely than the prior version to let flaws in its own code slip by unremarked. Not a miracle. Independent developer Simon Willison called it "a modest but tangible improvement," which is the right register. The point isn't one giant leap. It's that the leaps keep landing, roughly monthly.
We've reached the point where a frontier model is treated less like a product and more like a controlled substance.
Look at what happened to Fable 5, the first public model in Anthropic's Mythos-class line, released June 9. Three days later, a US export-control directive ordered Anthropic to suspend access for any foreign national, inside or outside the country — and to comply, Anthropic pulled Fable 5 and Mythos 5 globally for every customer. Commerce Secretary Howard Lutnick sent the letter to CEO Dario Amodei. The underlying Mythos model can find and exploit software vulnerabilities better than all but the most skilled humans; it has identified thousands of high-severity zero-days, including flaws in every major operating system and every major web browser.
The reported trigger was narrow — a jailbreak that amounted to asking the model to "read a specific codebase and fix any software flaws." Anthropic called the demonstrated weakness minor and noted that "perfect jailbreak resistance is not currently possible for any model provider." You can argue all day about whether the response was proportionate. You cannot argue that a model capable enough to rattle a national-security apparatus is fake. There's even a status tracker, isfableback.org, for people refreshing to see if it's back.
The dot-com lesson everyone half-remembers
People reach for "this is the dot-com bubble" as if it settles the argument. It doesn't, because the dot-com story has two halves and most people only quote the first one.
Half one: the crash was real. Fortunes evaporated. Plenty of companies deserved to die and did. Pets.com did.
Half two: the internet still ate the world. Amazon did not die. The technology was real and the stocks were overpriced — both true at once. The repricing of equities and the durability of the underlying shift were two separate events that happened to share a calendar.
That's the whole mental model. "Are AI stocks expensive and partly propped up by circular money?" and "Is this technology going to matter?" can both resolve as yes. Conflating them is how you either panic-sell something durable or hold something overvalued for the wrong reasons.
Government money is a different animal
One nuance worth holding onto, because it changes the risk profile. Most of the scary capex is private money chasing private returns, sometimes in those uncomfortable Nvidia-to-OpenAI loops. Sovereign investment is a different kind of demand. The EU's InvestAI initiative aims to mobilize around €200 billion toward AI by 2030, including a set of "AI Gigafactories," funded roughly 65–70% private and 30–35% public. An early call drew 76 submissions representing over €230 billion in proposed investment. That money is driven by strategic motives, not next quarter's earnings, and it doesn't evaporate the same way on a bad call.
I won't oversell it. Skeptics already call the gigafactories delayed, a "€20B sovereignty mirage." Fair. Government money is slower, more political, and not immune to overpromising. But the existence of motivated, non-speculative buyers — states that want this capacity for reasons that have nothing to do with a stock chart — is a structurally different demand signal than one chip vendor financing one lab to buy its own silicon. Different, not safe. Hold both.
What to actually do — as a builder, not a trader
Here's the part I'm qualified to say something firm about, because it's my actual job.
If you're a builder, a founder, an operator, your durable bet is not a ticker. It's your skills and the products you ship that solve a real problem for a real person. That asset survives a stock repricing completely intact. When the dot-com equities cratered, the people who had learned to build for the web didn't lose their ability to build for the web. They were the ones standing when it recovered.
So, concretely:
- Learn the tools for real. Not demos — actual work. Capability that improves monthly is capability you can compound on.
- Build something that solves a genuine problem. A product people pay for because it removes their pain does not care what Nvidia's market cap did this week.
- Treat model access as a dependency, not a given. Fable 5 vanished overnight on a single government letter. The emerging narrative that model risk is the new vendor lock-in is worth taking seriously, and the sane response is to stay portable — multi-provider gateways, no single point of failure in your stack.
- Size your financial exposure to what you can stomach. That part really is on you, and it really isn't my department.
The trader's question — "did I buy the top?" — I can't answer, and neither can anyone honest. The builder's question — "is the ground under this real enough to build a career on?" — I can. Yes. The capability is real and compounding, and the worst-case stock-market scenario doesn't touch the skills you build or the problems you solve.
The internet survived its bubble. The capability that scared a government this month is not going to un-happen because a stock chart turns red. Pull the two questions apart, and you'll find most of the panic was living in the gap between them.
One more time, because it matters: this is my opinion as a developer, not financial advice. Your money, your call.
Also published on Medium. I build all this solo, in public — andygarcia.pro.