JerkyTreatsThe thin gruel of tokens on my $20 a month plan has me shook.
To be fair, I’ve been running $200 a month workflows on the $20/month plan, and I fully expected the leash to be yanked at some point. And here we are.
I have been hesitant to increase the plan- why would I when with some decent token management I can survive fine on 10x less cost.
The $20 plan usually gave me just enough credits from 5am-6am each morning to push the ball forward on the various projects I have on the go.
When I run out, I just switch to the work computer where I have effectively unlimited burn.
The only problem there is that I’m not working on my stuff. All the additional value I’m generating is wholly captured by the company. This is an important point, because to thrive under capitalism one must retain marginal value, especially when agentic augmentation multiplies your productivity.
I’m still hesitant to upgrade. The main reason being I don’t want to give Sam Altman $200 a month.
I can afford it, but I’m cognizant of those that can’t. There’s some real undertones of that permanent underclass. I remember the deep aversion to crypto bros when they started making a lot of money. I am not excited to see what happens as AI continues to exacerbate the real stresses in the job market.
I can also see how this treadmill goes. $200 today will not be the same token budget as $200 tomorrow. Enterprise is consuming the lion’s share of compute where $200 is a pittance. Real companies are increasing productivity while reducing workforce. Stonks moon, inference demand increases, and the $200 plan becomes $500.
But really its a matter of incentive.
I have a 128GB unified memory device- I’ve built out most of a homelab that allows me to use it for my personal token burn. Its $2400 a year to be a fully unfettered token goblin. A Strix Halo or Blackwell DG10 platform is what, $5000? The comparative advantage isn’t quite there yet, but there’s a “there” there.
The other problem with going full homelab is that its hard. Frontier out-of-the-box harness combined with bleeding edge models RL’d on the hardness is really compelling. Instant results.
The homegrown stuff is much more complex. Bootstrap the model, bootstrap an API, build out tools, build out code sandboxes, fiddle with the harness. And only then do you start to see the differences between frontier and opensource.
The jump from $20 to $200 is really an aversion to what the jump represents.
It’s an economic dislocation with sinister portents.
Because the dislocation is between the economy and society, a dislocation started in 2020 and the inflation that followed, now accelerated by a productivity boom where humans need not apply.
The only way Im wrong here is if the compute crunch leads to a compute glut.
It’s well known in oil markets that every shortage is followed by a glut. My base case is that given a few years, compute will be sufficiently deployed.
If there is enough competition in frontier labs at that time, prices should be competitive at the consumer level.
I don’t really see demand going down. I do expect the anti-AI and pro-AI camps to have sharper lines. Really we need more Level1Tech’s out there trying to communicate that a middle ground does exist:
The consumer-grade private AI cloud’s aren’t quite there yet. But demand generates supply, and supply distills down to the consumer. And the foundational elements are already there- the OpenClaw’s and Nemotron’s of the world are the building blocks of truly personal AI. There’s an entire class of data I will not share with Sam, Dario, or Elon; that I absolutely would want an AI to grind on.
So I don’t know. I probably end up paying because Captialism, Baby ™️
But it’d probably pay dividends to do the harder, more expensive work of going full personal AI cloud.