16 articles · newest first
Renting vs owning a GPU for AI isn't a taste debate — it's a break-even calculation. Here's the honest 2026 math: what an hour in the cloud costs, what a card costs to own, and the crossover where each one wins.
Tensor · Sep 22 · 8 min
Two used RTX 3090s give you 48GB of VRAM for a third of a 5090's price — so is a dual-GPU build the smart move for local AI? The honest math, whether models really split across cards, and the power-and-complexity costs the sticker price hides.
Tensor · Sep 22 · 8 min
A used RTX 3090 gives you 24GB for ~$700–1,000 and runs the same models as cards costing 2–6x more. My honest used-market value pick for local AI — with the buying risks spelled out and current prices cited.
Voltage · Sep 22 · 8 min
Qwen3 isn't one model — it's a family from 4B to 235B, and the right GPU depends entirely on which you run. My VRAM-first map of every variant, the MoE trap that trips people up, and the used card I'd hand most people.
Tensor · Sep 22 · 8 min
Portable Computer promises zero token costs by running entirely on your own GPU. here's what that GPU actually costs you.
Tensor · Aug 26 · 8 min
24GB of old VRAM versus 16GB of new, low-power silicon — I ran the real math on what each one costs, up front and over a year.
Voltage · Aug 4 · 6 min
Four real paths to running a 70B-parameter model on your own hardware or by the hour — priced out with actual 2026 numbers.
Tensor · Aug 4 · 8 min
A third price hike in one year is reportedly coming for the RTX 50 series — and if you're buying VRAM for local AI, not frame rates, the calculation isn't the same as 'wait for a sale.'
Voltage · Aug 4 · 7 min
A used 3090 is still the best VRAM-per-dollar card for local AI — but many were mining cards, and the memory runs hot. Here's exactly what to check before you hand over the cash.
Voltage · Jul 28 · 9 min
Qwen3-32B at Q4 needs about 20GB of VRAM, which rules out every 16GB card and makes this a simple question: what's the cheapest 24GB GPU that runs it well? The answer is used.
Voltage · Jul 28 · 9 min
Llama 70B needs about 40GB of VRAM to run well — and almost no single consumer card has it. Here's what actually works, visualized, with the measured speeds and real costs for each path.
Voltage · Jul 23 · 10 min
A new RTX 5060 Ti 16GB looks like the safe budget pick. But 16GB can't load the 32B models a used 3090 runs at 30 tokens/sec — and in local AI, the card that opens the file wins before speed matters.
Voltage · Jul 22 · 9 min
Older Macs held huge models but crawled through long prompts. On llama.cpp's own scoreboard, the M5 Max and M5 Ultra now match or beat a used RTX 3090 on small models. Here is where each machine still wins, with the numbers and their caveats.
Tensor · Jul 22 · 10 min
Measured llama.cpp numbers say the 4090 wins every speed test and costs twice as much — while energy per generated token is a dead heat. Here's who each card is actually for.
Voltage · Jul 22 · 9 min
Two 3090s give you 48GB, but not double the speed for chat, because llama.cpp's default multi-GPU mode makes the cards take turns. What the benchmarks say, rechecked in October 2026.
Voltage · Jul 21 · 11 min
on a single 24GB card in 2026 there's no one 'best' — it's a clean three-way split: Qwen3.6-27B for coding, Gemma 4 for vision, gpt-oss-20b for raw speed. and a used RTX 3090 still competes, because this is a memory game, not a compute one.
Tensor · Jul 19 · 7 min