Our radar pulled 88 items today. Four of them, posted within hours of each other by people who have never met, tell one story better than any of them do alone.
At the top, somebody bought a 5090 prebuilt and posted the receipt: 5,849 dollars plus tax, 6,279 all in. Their own summary of the decision was less polite than ours would be. They wanted a sanity check on whether they had bought at the right time.
Shortly after, someone else pointed out that the DGX Spark has now joined the 5090 in going up rather than down, and asked the question everyone asks next. Is the future used hardware?
Then came the answer to that question, from somebody living it. Two of their five CMP 170HX cards died inside two weeks, one dropping off the bus the moment vLLM starts and another throwing CUDA errors, and the third arrived with defective tensor cores. That is the used enterprise shortcut failing in public, in real time, on the same afternoon somebody else was talking themselves into paying six thousand dollars to avoid exactly that.
And then the floor moved
Here is the part that makes the day interesting rather than depressing.
Somebody ran MiniMax H3 on integrated graphics. Not a graphics card. The Radeon 8060S built into a Ryzen AI Max chip, under Linux with ROCm, using the official eight step acceleration LoRA, producing a 5.17 second video with native stereo audio. Their own note is that they could not find this specific combination documented anywhere, which is usually the sign of somebody genuinely first rather than somebody repeating a tutorial.
Elsewhere the same day: a fused turbo build generating five seconds of video in about a minute at four steps, a 104 GB model coaxed onto a 48 GB Mac at roughly twelve tokens per second, and 280 tokens per second out of a pair of AMD R9700s. At the far end, somebody with four RTX PRO 6000s had a local model write a working Minecraft mod.
Every one of those numbers belongs to the person who posted it. We have not reproduced any of them, and hardware results are the single most setup dependent thing in this field, so read them as reports rather than specifications.
What it actually means
The ladder used to be simple. More money bought more VRAM which bought more capability, and the only question was where you got off.
Right now the top of that ladder is getting more expensive rather than less, the used shortcut underneath it is a lottery with real failure rates, and the bottom rung has quietly dropped to hardware that does not have its own fan. Those three things are not in tension. They are what a fast moving open ecosystem looks like when optimisation work outruns hardware pricing.
The practical read for anyone about to spend money: the interesting question stopped being which card is best and became how far down your particular job actually goes. Video generation on an integrated GPU was not a serious sentence three months ago. It is a link now.
The part that is ours
Everything this channel makes runs on one consumer graphics card. The presenter, the lip sync, the voice, the daily uploads, the brain that writes the scripts. One card, one box, and we have published the exact spec including the parts that do not matter.
The honest footnote is that we bought it as a gaming PC, before this run up, and the AI obsession came afterwards. We would be making a very different decision if we were buying today, and anybody telling you they would confidently spend six thousand dollars right now is not accounting for how fast the optimisation side is moving. Two days ago we wrote about a model getting four hundred community builds in five weeks. A meaningful share of those builds exist specifically to make it run on less.
If you are working out where your own machine sits on that ladder, the calculator will tell you in about thirty seconds, and the rig guide covers the tiers if you are actually shopping.
A daily note from our news radar. The full week gets the DIY AI Brief treatment every Monday.