Recently, I used AI to make a first-person shooter game.

It took me 10 and a half hours… 1.3 million tokens.

Working physics.

Crisp graphics.

Every art asset generated out of thin air.

Total cost: US$60

I’m telling you this for two reasons.

One, because I want you to understand how remarkable this is.

Second, because it has everything to do with why I jumped on Mark Jeffrey’s podcast Hash Rate recently.

First things first.

Why this is crazier than anyone realises

A “token” is about three quarters of a word.

So the AI thought its way through a million words to hand me a video game. That’s about the length of the complete Harry Potter series.

And it did it in about 10 hours.

I couldn’t read a million words in 10 hours. Nobody could. Never mind use them to program a working game from scratch.

And how it did it is even more remarkable.

I lifted the method from Matt Schumer. You tell the model to screenshot the real Call of Duty, build its own version, screenshot that, compare the two, split the work across agents working in parallel, and keep going until they match.

I’m a coder by background. I remember trying to make games 20 years ago. It would’ve been impossible to make anything like this by myself, let alone in 10 hours.

Even two years ago, this project was a team of programmers, six months of work, and a million-dollar budget. I would have signed the check.

Today, it cost me sixty bucks.

Everyone is arguing about inflation. Well, here’s a deflationary number you’re unlikely to find in anyone’s spreadsheet:

One job.

US$1 million to US$60.

In 24 months.

That’s the part of the AI story I think investors need to get their heads around. Because if something that cost US$1 million two years ago can suddenly be done for US$60, you have to start asking what happens to the businesses that used to charge you the US$1 million.

I got into all of this with Mark. And that conversation took us somewhere that could get very uncomfortable for a lot of software stocks.

So what’s left?

If I can clone a billion-dollar franchise in a day for the price of a decent dinner, then the code protects nothing.

If the software doesn’t protect a company anymore, something else has to.

Three things survive.

Data. An AI is only as smart as what it’s been shown. And there’s information it will never see. A bank’s fraud records. A hospital’s scans. A police department’s body cam footage. Nobody is emailing that to an AI lab.

That makes proprietary data incredibly valuable. Whoever holds the most useful data holds something all the cheap intelligence in the world can’t copy.

Distribution. Being where the customer already is. I can build a better product this afternoon and never reach a single person using the incumbent. Ask anyone who built a better search engine.

Trust. The one everybody underrates.

I could write my own tax software. I use H&R Block anyway. Mine would probably be better. But when the audit letter arrives, I want somebody else’s name on the box. Nobody ever got fired for buying IBM.

Now, here’s what happens when you put all three together.

There’s an AI project called Score that teaches machines to tell a forest fire from a guy with a flashlight on a camera feed.

Sounds boring. It’s enormously valuable.

It works the way I just described – train the AI on specialised footage that a general-purpose model will never be shown. That’s the data.

Then it reaches customers through PricewaterhouseCoopers, so it arrives wearing a suit clients already trust.

And with PwC standing behind that relationship, it gets the third ingredient too: trust.

But here’s where it gets interesting.

Score isn’t some AI startup in Palo Alto. It’s a subnet on a crypto network called Bittensor.

The trade

Bitcoin talked strangers all over the world into plugging in computers and burning electricity, and paid them in bitcoin for the trouble.

Though it’s down from all-time highs, few people can argue it hasn’t been a massive success.

Bittensor pointed the same trick at AI. Instead of rewarding computers for securing a monetary network, Bittensor rewards them for doing useful AI work.

That work happens through subnets, essentially small AI businesses operating on the Bittensor network. Run the machines, contribute useful intelligence, and get paid in a token called TAO. The better the work, the greater the reward.

Meanwhile, Kimi K3 released its open weights on Monday. And these two stories are more connected than they might appear.

Because the moat everyone assumed the frontier labs would enjoy for years evaporated in about three seconds. State-of-the-art intelligence is rapidly becoming a commodity.

Commodities have no brand loyalty and no network effects. The cheapest producer wins. Full stop.

So who’s cheapest? Bittensor.

There’s another subnet on Bittensor running frontier models on consumer graphics cards instead of million-dollar racks, on the cheapest power on Earth. Cheapest anywhere. Its market cap is about US$25 million.

Meanwhile OpenRouter carries a valuation around ten billion. And a Bittensor subnet called Chutes supplies a great deal of what routes through OpenRouter and books roughly a million in revenue.

Those are the kinds of valuation gaps that get my attention.

Bittensor, the network all of these subnets run on, has been stuck around US$200 for more than a year, waiting for a catalyst strong enough to break it out.

Cheap frontier intelligence could be that catalyst.

And here’s what makes TAO particularly interesting at its current price.

Bittensor isn’t one AI company. It’s exposure to 128 of them through one network.

Vision. Translation. Bank fraud detection. Specialised data. AI infrastructure. Several are selling raw AI horsepower at extraordinarily low prices.

Together – they’re valued at roughly US$2 billion – at a moment when individual AI companies can command valuations stretching into the hundreds of billions, and the largest technology companies exposed to AI are worth trillions.

That disconnect is what makes Bittensor interesting.

And that’s the opportunity.

And it’s also why I’ve been paying attention to what my colleague Sam Volkering has been building.

I’ve just spent this entire essay explaining how quickly AI is changing the economics of software. Intelligence is getting cheaper. Old competitive advantages are disappearing. Entire industries are being reshuffled.

That creates opportunities. But it also creates a problem for investors: there are far too many companies to watch at once.

Sam has spent the last three years building something designed specifically for that problem.

It’s called Hyperion.

Hyperion uses AI to scan and rank roughly 1,800 investable stocks across the US and UK markets, looking for the characteristics that have historically appeared ahead of some of the biggest moves higher.

It doesn’t replace Sam’s judgement. It gives him a much bigger set of eyes.

And this Thursday at 4pm, Sam is going to demonstrate Hyperion live for the first time and show you exactly what it’s finding in the market right now.

Given everything I’ve just shown you about what AI can already do, I think you’ll want to see what happens when someone turns that power directly onto the stock market.

[Click here to register for Sam’s Hyperion event on Thursday at 4pm. By clicking this link, you agree to receive additional emails about Hyperion.]

Best, 

James Altucher
Contributing Editor, Investor’s Daily