I’m listening to a book called 1929, and it’s fantastic.
I’ve learned about the stock market crash in history classes. Everyone has. You get the dates, the headlines, the fallout. What you don’t usually get is the texture—the incentives, the psychology, the mechanics of how people convinced themselves something was worth more than it actually was.
One detail stopped me cold: people were buying stocks on credit.
Not metaphorical credit. Literal leverage. Borrowed money, chasing rising prices, under the assumption that value would continue to exist because everyone agreed it did. When the crash came and there was a run on the banks, the money simply wasn’t there. The value people thought existed had never actually been backed by anything tangible.
That’s what a bubble is.
A bubble is when the assumed value and real value diverge so far that the system collapses the moment anyone asks to cash out.
And I can’t stop thinking about how familiar this all feels.
We’ve Seen This Movie Before
We’ve had bubbles before. The dot-com bubble. The 2008 housing crisis. Different industries, same pattern.
Assets were appraised for more than they were actually worth. Money moved faster than reality. Optimism replaced fundamentals.
Right now, we’re watching the same thing happen in AI.
I just finished a chapter describing an RCA merger during the rise of radio. Radio was brand new in 1929. Revolutionary. World-changing. And because of the bubble, people with capital were able to trade promises, restructure ownership, and walk away richer without necessarily creating proportional value.
Tell me this doesn’t sound familiar!
NVIDIA promises GPUs to OpenAI.
OpenAI promises future purchases from AMD.
Oracle promises data centers for OpenAI and buys GPUs from both.
Around and around it goes.
Everyone’s balance sheet improves, everyone’s valuation goes up, but very few actual assets change hands. That circular promise-making is the clearest evidence of a bubble I know.
Hopes, Dreams, and Missing ROI
Let me be clear: hopes and dreams matter. We need them to drive innovation.
But when you bet too much on hopes and dreams without being able to deliver tangible return on investment you create fragility.
I’ve been saying this for a long time. I’ve been in the software industry for seven years. For the entire time, as a pragmatist, I’ve said the problem with machine learning, AI, deep learning—whatever you want to call it—is not potential.
The potential is unlimited.
The problem is ROI.
How much real, monetary value can you create from the system you’re building?
That question never goes away. And right now, the industry is struggling to answer it honestly.
We’re Out of the Things That Matter
Here’s the part that doesn’t get talked about enough: We’re out of RAM.
Not metaphorically. Literally.
RAM prices are going up. You can’t just buy more indefinitely. And RAM is a tangible resource that actually delivers value. GPUs matter, yes, but GPUs don’t work without memory.
I have a 5090 GPU at home. I rarely saturate all 32 GB of VRAM. I am constantly out of my 64 GB of system RAM. Even with mixture-of-experts models.You load models into RAM before you load them into VRAM. Funny how that works.
This is where the bubble starts to show cracks: we are constrained by physical reality, not ideas.
Inflated Metrics, Internal Comparisons
Another red flag is how we talk about progress.
Gemini compares to the previous Gemini.
ChatGPT compares to the previous ChatGPT.
Opus compares to the previous Opus.
These are inflated internal metrics, not grounded benchmarks tied to business outcomes.
And in practice? Most companies aren’t even using the cutting-edge models. They’re using last generation.
Why?
Because it’s cheaper, and itAnd it performs almost the same.
That’s real ROI: performance per dollar. Not leaderboard screenshots.
You Can’t Dodge a Bubble — But You Can Prepare
So how do you protect yourself from an AI bubble?
Honestly? You probably don’t.
We don’t know what the fallout will look like. We don’t know which promises will fail, which companies will survive, or which narratives will quietly disappear.
But skills don’t vanish.
AI isn’t going away, butwhat will go away are impractical uses.
AGI? Impractical.
Automating workflows? Very practical.
Agents? Useful, but limited.
The real work is learning where deterministic control flow ends and agentic behavior begins. Software architecture, infrastructure, hosting models, and reliability all still matter.
AI doesn’t replace those skills—it amplifies the cost of not having them.
Coding Agents Will Plateau
Here’s another uncomfortable truth that I’ll tell you because I’m not afraid of hurting your feelings: coding agents will hit a plateau.
There will be a point where this is “all Cursor can do for you.” After that, improvement slows and when we start paying for coding agents in addition to salaries, the math changes.There will be fewer jobs.
Differentiation becomes mandatory.
Knowing how AI works becomes a career skill. System prompts, fine-tuning decisions, and when to use software versus when to use an agent are not optional anymore.
Why I Self-Host Everything
This is why everything I do on my side projects is locally hosted, becauseiIt forces me to understand how things actually work.
OpenAI doesn’t just use your system prompt, they just add their own. This means you don’t get to see it or control it, which is detrimental. If you want real leverage, you have to do it the hard way.
On your own time.
I know you don’t have a lot of time, but if you have 10% or 20% time at work, use it. Leverage the hell out of it. You have to learnLearn the things that persist.
Compilers and infrastructure will always matter, tool-calling may change, but what you learn from it will carry forward.
History Repeats, Whether We Like It or Not
This is all too familiar. Rich people getting richer. Everyone else hoping nothing happens to them.
But hope isn’t a strategy.
Doing something doesn’t mean “make more money.” It means investing in yourself. Learning. Building. Figuring out what you’ll work on until 2am because you want to, not because you’re told to.
If you’re a new grad and love software, you might end up working at Target for a while. Keep building anyway.
If you don’t love software like that, that’s okay. Go be a plumber. I will happily pay you to fix my house. Building real things is still building things.
Just don’t build fake things.
My Advice for 2026
Figure out how to invest in yourself.
Meetups.
Online communities.
Open source.
Building at night.
Find something adjacent to your job that you actually care about and pour time into it before the bubble pops.
Nobody loves work. Work pays bills, and skills carry you through downturns.
History doesn’t repeat exactly—but it rhymes. And this rhyme is getting louder.