Theo says local AI is a hobby, not liberation
A $18,000 four-GPU box for a 10-person startup set off an argument about what frontier subscriptions will cost in two years.
The post that started it came from David Hendrickson, who described a 10-person startup running its own 128GB-VRAM AI server for $18,000. One Threadripper box, four Radeon AI PRO R9700 cards at 32GB each, Qwen3.8 and DeepSeek V4 Flash, 16 concurrent sessions. At 48K context per user he reports 538.5 tps aggregate decode, dropping to 365 tps at 128K context per user, with the GPUs undervolted to under 210W each. His framing was that local AI is good enough for a startup.
Theo did the division and came out somewhere else. That money is roughly eight years of $200 Claude subscriptions, and what it buys instead is quantized DeepSeek v4 Flash at 30tps. He drew a line between two versions of the same setup. Doing it because it is fun and cool, knowing it is not the best use of money, gets a check mark. Telling people they are a slave to OpenAI and Anthropic and do not really need frontier intelligence gets three red crosses.
I use local AI because it’s fun and cool. I know it’s not the best use of money but I genuinely enjoy it
The actual disagreement
The replies went after the part of the argument that depends on the future. Mia listed three objections, the first being that the Claude max subscription will not stay at $200, and added that for most companies privacy is king.
Theo's answer was that AI is getting cheaper, and that the $200 plans will improve rather than disappear. He also pushed back on the privacy point by noting that the companies in question are on Bedrock or zero data retention deals, not a pile of GPUs in someone's closet. On throughput he conceded ground narrowly, saying that if you are fine with 15tps on decent models, sure.
For working engineers, nobody in this thread is arguing about tokens per dollar at today's prices, because that part is arithmetic and both sides can do it. The argument is a bet on pricing direction. Buy the box and you have fixed your cost and your model quality for the life of the hardware. Rent the subscription and you are betting that the per-dollar capability keeps climbing fast enough that the hardware looks silly in a year. Theo thinks it does. Mia thinks the subscription price is a promotional number. No evidence in the thread settles it, and no one claimed to have any.
The one thing both sides agree on is that the local box is slower today.
Claude max sub will not cost $200 for long.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.