When the Garage Needs Three-Phase Power

There's this story we all know. Two kids in Los Altos, sometime in the mid-70s. One of them is soldering boards in a garage. The other is preaching to anyone who'll listen that computers should belong to regular people, not just IBM. They scrape together a circuit board, a Motorola 6502, a handful of chips most people couldn't identify on a bet, and they build something. A few years later they're worth a fortune and the world looks different.

We tell that story like scripture. The garage. The two guys. The breakthrough. It's the founding myth of the entire tech industry — that anyone, with enough vision and a soldering iron, can build the next thing.

And I'm here to tell you... yeah, that's done. Officially. The garage is closed. The locks have been changed. There's a sign on the door that says "REQUIRES THREE-PHASE POWER AND A SUBSTATION PERMIT."

Because the next thing isn't a personal computer. It's AI. And AI is a rich man's game.

Let's start with the obvious

Want to train a frontier model? Cool. Get your checkbook. Then throw your checkbook away because it's not big enough. You need a wire transfer.

GPT-4, depending on whose estimate you trust, cost somewhere north of $100 million to train. Just train. That doesn't include the salaries, the failed runs, the data acquisition, the lawyers fighting about the data acquisition, the second round of lawyers fighting about the first round of lawyers. Just the raw compute. A hundred mil. To train one model. Once.

The next generation? People are throwing around numbers like half a billion. A billion. For a single training run. And nobody's blinking because the numbers have gotten so big that they've stopped feeling real, like Pentagon budgets or Elon's tweets.

Meanwhile, Sam Altman is out there talking about needing seven trillion dollars for chip manufacturing. Seven. Trillion. That's not a company budget. That's a small country.

You, sitting at home with your good idea and your enthusiasm? You're not invited.

"Just run it locally"

Okay, you say. Fine. I don't need to train models. I just want to run them. Open source has come a long way. Llama. Mistral. DeepSeek. Qwen. Plenty of options. I'll just buy a GPU and run my own stuff. Stick it to the man.

I love this energy. I genuinely do. I'm right there with you. Let's check the math.

Llama 3.1 405B at full precision needs roughly 800GB of VRAM. That's not a typo. Eight hundred. Gigabytes. Of VRAM. Your gaming PC has, what, 16GB if you're lucky? 24GB if you sold a kidney for a 4090?

"Okay but quantization!" Sure. Quantize down to 4-bit and you can fit it in around 230GB. Still not happening on any single consumer card. You'd need eight 4090s in parallel, and now you're talking about a $15,000 box plus a custom power and cooling setup that'll make your circuit breakers cry. Your electrical panel needs an upgrade. Your partner wants to know what that noise is. Your neighbors think you're running a meth lab.

Even the "small" frontier-ish models — the 70B parameter range, Llama 3.3, Qwen 2.5 — need around 140GB at FP16 or 40-ish gigs heavily quantized. So you're looking at a multi-GPU rig minimum, or a maxed-out Mac Studio with 192GB of unified memory, which Apple will happily sell you for the low low price of seven grand give or take.

And the kicker? Memory bandwidth is the actual bottleneck. You can stuff all the parameters into VRAM you want — if your bandwidth can't shovel them through the compute units fast enough, your tokens per second drops into "watching paint dry" territory. The H100 has 3.35 TB/s of memory bandwidth. The 4090? Around 1 TB/s. Your Mac Studio? Roughly 800 GB/s. There's a reason datacenter cards cost what they do, and it's not just margin (although... ha, ha, it is also that).

"Fine," you say. "I'll just use the API."

The tokens are killing me

API pricing has come down, which is genuinely a good thing. You can get decent frontier-class output for a few bucks per million tokens. That sounds cheap! And for chatting, it is.

But here's the thing nobody tells you when you're just doing Q&A. Real applications burn tokens. Agentic workflows burn tokens like a punk drummer burns through snare heads. Every retry, every tool call, every "let me think about this step by step" — that's tokens. Every long context document you stuff in there — tokens. Every conversation you've been having for the last six hours that's now 80,000 tokens of history — yeah, you're paying for that whole context on every. Single. Turn.

I've watched my homelab AI experiments rack up bills that genuinely surprised me. Not bankruptcy surprised. But "huh, didn't expect that" surprised. And I'm a guy who works in this stuff for a living and understands what I'm doing.

The little tinkerer who just wants to play? The student trying to learn? The person with the genuinely cool idea but no funding? They get a $20/month subscription with rate limits, or they hit pay-as-you-go and watch their experiments evaporate their grocery money.

This isn't accidental. This is the business model. The labs need to recoup that hundred million per training run somehow, and "somehow" turns out to be you.

The picks and shovels

The only people who got truly rich in the gold rush were the ones selling shovels. Levi Strauss made jeans. Wells Fargo ran the stagecoaches. The actual miners mostly went broke and died of dysentery.

In our gold rush, Nvidia sells the shovels. And boy, are they selling shovels. An H100 runs $25,000 to $40,000 depending on how desperate you are and who you know. They cost something like three or four grand to manufacture. The margin is criminal. The margin is beautiful. The margin is why Jensen Huang now owns a small fleet of leather jackets that probably cost more than my home.

Microsoft, Meta, Google, Amazon — they're each spending tens of billions on Nvidia chips. Tens of billions. Per quarter. Nvidia's market cap blew past three trillion dollars and people just kind of nodded.

Meanwhile, if you're an academic researcher, a small startup, an independent who wants to do real work, you're waiting in line behind hyperscalers who are buying every H100 off the assembly line before it's cooled down. Good luck getting hardware. Good luck even getting cloud capacity some weeks.

The labs are no longer labs

Here's the thing that really gets me. The big AI labs — OpenAI, Anthropic, the rest — they all started with this scrappy research vibe. Idealistic. Mission-driven. We're going to figure out alignment, we're going to build AI safely, we're going to share what we learn.

And then the bills came due.

OpenAI is now functionally a Microsoft subsidiary with ten-plus billion in deals and counting. Anthropic raised billions from Amazon and Google. xAI, Google DeepMind, Meta AI — they're all either inside or financially tethered to the biggest companies on earth.

I don't think this is malice. I think this is physics. You can't run frontier AI research on idealism and ramen. You need datacenters. You need power. You need the kind of capital that only six or seven entities on planet earth can provide. So the choice is: become a department of Microsoft, or stop existing.

The result is that the cutting edge of one of the most consequential technologies in human history is being decided in board meetings at companies whose primary historical concern was selling office software or shipping packages. We've handed the future to people who measure success in quarterly earnings calls.

And then they sold it to the Pentagon

Here's where it gets dark.

When you're a lab that needs billions to keep the lights on, you take money where you can get it. And the deepest pockets in the world aren't tech companies. They're governments. Specifically, defense and intelligence agencies.

OpenAI quietly removed the "no military or warfare use" clause from its usage policies in early 2024. Not with a press release. Not with a town hall. They just edited the page. Anthropic has announced partnerships to put Claude into classified government environments. Palantir, which has been selling AI-flavored everything to ICE and the military for over a decade, gets called a respectable AI company now instead of the surveillance contractor it actually is. Anduril is building autonomous weapons and getting unicorn valuations for the privilege.

Remember Project Maven? Back in 2018, Google's employees famously revolted when they found out the company was helping the Pentagon analyze drone footage. Six years later, basically every major lab is doing some version of Maven, and nobody's even pretending to be conflicted about it. The employees who would have walked out either got with the program or got laid off in the last round of cost cuts.

We're also starting to learn what AI-assisted warfare actually looks like in practice. Israel's military reportedly used a system called Lavender to generate kill lists during the war in Gaza — flagging tens of thousands of people as Hamas targets with what investigative reporting described as a 10% false-positive rate. Ten percent. Of human lives. Run through a classifier. Reviewed by a human operator in something like twenty seconds per target. That's not a hypothetical. That's a system that was actually deployed.

And on the home front? Clearview AI scraped billions of faces off social media and sold the database to police departments. Predictive policing algorithms are baked into law enforcement workflows in major cities across the country, despite repeated studies showing they encode the racism of the historical arrest data they were trained on. Facial recognition is at the airport, on city streets, in the convenience store. Every license plate reader is feeding a database somewhere. The IRS is using AI to flag tax returns. DHS is using it to flag travelers.

I do security for a living. I spend my days thinking about threat models. And the threat model here isn't subtle: an unprecedented surveillance and targeting capability has been handed to whichever government happens to be in power, and governments have demonstrated, repeatedly and across every political stripe, that they will use whatever capability you give them on whoever they decide is the problem this week.

The labs all have very nice AI safety teams. They publish papers. They run red teams. They worry, very publicly, about whether the AI might turn evil on its own.

You know what they worry about a lot less publicly? Whether the AI will be turned evil by the people who paid for it. Because those are the same people writing the checks. And it's hard to bite the hand.

Power. Literally power.

Microsoft is restarting Three Mile Island. Not because they suddenly love nuclear. Because they need the gigawatts. AWS bought a nuclear-powered datacenter campus. Google's signing deals for small modular reactors. The grid in Virginia — my grid, the one running my homelab right now — is straining under datacenter demand to the point where utilities are asking regulators for the right to build new gas plants just to keep up.

The garage doesn't need three-phase power anymore. The garage needs a reactor.

When the cost of entry to your industry includes building or buying nuclear power infrastructure, you have officially exited the realm of "two kids and a dream." You're in the realm of nation-states and trillion-dollar corporations. That's it. Those are the players. There are no others.

So what?

Look, I'm not saying it's hopeless. There's still cool work happening at the edges. Open weights models keep getting better. Quantization techniques keep improving. There's a whole indie community squeezing surprisingly capable models onto Raspberry Pis and old gaming rigs, and I love them for it. Some of my favorite weekend projects are exactly this.

But let's not pretend it's the same. Let's not tell the kids that anyone with a soldering iron and a dream can build the next OpenAI. They can't. The next OpenAI will be founded by people who can raise a billion dollars before lunch, who have the right Stanford connections, who know which VCs to call. That's the actual game now.

And the rest of us? We get to play in the API sandbox. We get to fine-tune the leftovers. We get to build clever wrappers around the magic boxes the rich kids own. Which is fine, sort of. There's real work to do at that layer. I'm doing some of it. You probably are too.

Just don't mistake what we're doing for what Jobs and Wozniak did. They built the thing. We're decorating someone else's thing. The actual machinery, the actual frontier, the actual decisions about where this goes — that's happening in rooms we'll never be invited into, between people who own utilities and chip foundries and small armies of researchers commanding transfer fees that would make Premier League clubs blush.

The garage is closed. The two-guys-and-a-dream era is over. And whatever comes next, it's not going to come from someone like you or me. It's going to come from someone whose seed round was bigger than the GDP of a small island nation.

Anyway. I'm gonna go boot up my Mac mini and pretend I'm building the future.

It's cheaper than therapy.