The token price index crashing almost 60% in a few months is the kind of headline that sounds scary until you actually sit down and think about it. Silicon Data's LLM Token Expenditure Index — a usage-weighted measure of what a million tokens actually costs you — dropped to $0.9665 by August 31st. In May it was sitting at $2.0651. That's not a rounding error. That's a market deciding, loudly, that intelligence is becoming a commodity.
From my perspective, this is the most underrated AI story of the quarter. The "$600 billion AI boom" narrative keeps churning in the business press, but the people actually buying tokens are behaving completely differently. Open-weight models grabbed a huge chunk of routed traffic — the share of open-model tokens on one major routing platform went from 34% in January to 65% in June. That's not a niche preference. That's infrastructure-level behavior shifting in half a year.
Simply put: companies stopped paying premium prices for frontier models to do routine work. Routing tools now shove each request at the cheapest model that can actually handle it, and DeepSeek and friends made that calculus laughably easy. I've been doing this myself — my side project routes summarization jobs to a small local model and only hits the big API when something genuinely hard shows up. The bill went down maybe 70% and nobody noticed any drop in quality. That's the whole story in miniature.
To be fair, a falling price index doesn't mean demand collapsed. It means the same intelligence got cheaper. Whether the hyperscalers' capex plans survive contact with that reality is a genuinely open question — margins shrink when your product's effective price halves. Something has to give, and it probably won't be the model quality.
Meanwhile, on the hardware side, Minisforum dropped something at IFA 2026 that made me stop scrolling.
The MS-S1 Max-P495 is a mini PC with a ridiculous amount of unified memory — 192GB — paired with AMD's Ryzen AI Max+ Pro 495, the "Gorgon Halo" chip. The entire pitch is local LLM inference on your desk without renting a cloud GPU. And honestly? It works for a specific kind of person. If you're running 70B-class models for real work, or fine-tuning experiments where you don't want your data leaving the building, this thing is genuinely compelling.
The catch, and there's always a catch: 192GB of unified memory on a mini PC costs a lot, and the memory bandwidth still doesn't touch a proper workstation GPU. Fast enough for chat and agents, not fast enough if you're trying to pretend you're a datacenter. Also, I've owned enough Minisforum boxes to know thermals are always the wildcard — a chip like this in a small chassis is going to get toasty under sustained inference. I'd wait for the first round of reviews before ordering.
And if you want proof that local AI has gone mainstream in the other direction, people are booting Raspberry Pis straight into local LLMs now.
XDA had a nice piece on running a local model off a USB drive on a Pi, no cloud, no subscription. It's slow, obviously — you're not getting GPT-4-class answers out of a $60 single-board computer. But as a self-contained private assistant for notes, drafts, and basic agents, it changes the math for hobbyists. I ran a small 7B model on one of my SBCs over a weekend. Tokens trickle out at maybe a tenth of the speed of a real machine, and I still found it useful precisely because it's private and always on. Zero cloud, zero cost, zero leaks. That tradeoff matters more to some people than speed.
Quick add-on note: this whole "local AI everywhere" wave is fun, but the software side is still catching up. My Pi setup took an afternoon of fiddling to get stable, and the memory management on the desktop box needs tuning out of the box. The hardware is ready; the user experience is not, and pretending otherwise does nobody any favors.
Then there's the human layer, which nobody benchmarks.
A piece going around made me stop and think — workers describing how skilled jobs feel less rewarding now that a big chunk of the "solving" is done by an LLM. One person put it memorably: most of the sport these days is pointing out where AI is wrong. Not doing the work — checking the machine's work. There's an archer analogy in there about rifles making years of practice feel pointless, and I get it. It's a real, uncomfortable feeling: the parts of a job people actually love — solving the puzzle, testing the idea, making the call — are exactly the parts software took over first.
I don't have a tidy answer, and I'd be suspicious of anyone who claims to. But I do think the best individual move is to get weird: lean into the stuff AI is bad at — taste, judgment, pushing back on bad prompts, knowing why the answer is wrong rather than just that it's wrong. That's a skill that, honestly, might appreciate in value.
Last thing: OpenAI's "Astra" rollout is stirring a familiar argument about opaque AI and safety theater, with commentators like Parmy Olson pushing back on the security framing. A lot of people are wondering whether the industry's loud safety pledges actually survive contact with shipping deadlines. My read: watch what they do, not what they say.
That's it for today's roundup. The theme seems to be: intelligence is getting cheaper, local, and more personal — and the humans in the loop have to figure out where they actually add value. No rush on answering that one. I'm still working on my own answer, one local model at a time.
Keep this in mind if you're sizing your next AI budget: the cheapest token isn't always the right token, but the cheapest model that does the job is almost always the right model. Shop accordingly. And if you're crunching numbers on what a local setup would actually cost you, 7x24planning is a handy little planning calculator I've been using for quick budget math — nothing fancy, just useful.
Top comments (0)