The spec sheet is not the machine: two checks before you spend on local-AI hardware

Most local-AI hardware advice treats the spec sheet as the machine. Two checks from our own fleet: do the fit arithmetic before the money moves, and measure the machine you received, not the one you ordered — two identical units differed by 4.7× until a full shutdown brought the slow one back in line. For readers at the buying decision now, the distilled toolkit and config pack are linked at the end.

Most local-AI hardware advice fails in the same place: it treats the spec sheet as the machine. Twice now, our own fleet has taught us the two ways that goes wrong — once on paper, before you buy, and once on the bench, after you do.

Check one: do the fit arithmetic before the money moves

Model weights weigh the same on every machine, which makes the buying decision the one part of this field you can settle exactly, in advance, with arithmetic instead of benchmarks. A 27B model at 4-bit quantization fits comfortably in 32 GB of memory; the same model at 8-bit does not. A 125B-class model needs 128 GB before it is even a conversation. Between those two facts sits most of the money wasted on local AI: machines bought one tier too small for the model the buyer actually wanted to run, and machines bought one tier too large for a workload that would never touch the headroom.

The order matters, and it is the order we use when we spec machines for small offices: workload first, then model class, then memory tier, then — and only then — the machine. Benchmarks come last, because of check two.

Check two: measure the machine you received, not the one you ordered

When we indexed 45 published local-inference benchmark results, the headline was not a winner. It was a gap: two Mac Studio M3 Ultra units — same chip, same memory, same model spec — differed by 4.7× on an 8-bit quantized matmul, 22.91 TFLOP/s against 4.83 TFLOP/s. On a 25,970-token prompt the healthy unit prefilled at 373.7 tokens per second; the degraded one managed 118.2. Nothing was wrong that the operating system could see: memory was free, thermals were fine, and the degraded unit's own power telemetry reported its GPU at 100% idle while it worked. A warm reboot did not fix it. A full shutdown did — the same box then measured 22.61 TFLOP/s, in line with its twin.

The lesson is not that one vendor's hardware is suspect. It is that a spec sheet is a claim, and the only version of the claim worth paying for is the one your own machine reproduces on your own bench, in the first week, while returning it is still easy.

What we'd hand a friend who is about to spend $1,000–$5,000 on this

The free field notes are the raw lab reports — one measurement or failure at a time, in the order we hit them, and they stay free. For readers who are at the buying decision now and don't want to assemble a weekend of notes into a decision, we distilled them into the downloads below. Same numbers, organized by the decision, with the worksheet and checklist filled in.

• Local-LLM Hardware Buyer's Toolkit — $49. The 16 / 32 / 64 / 128 GB fit map, the seven-step decision worksheet (workload, peak concurrency, model class, KV-cache arithmetic), sizing for a 5–20-person office, the real three-year cost against a cloud subscription — including the workloads where cloud is genuinely cheaper — the same-day setup guide, and the full benchmark data appendix, every figure sourced. https://robotnurse.gumroad.com/l/llm-buyers-toolkit
• Serving Config Pack — $29. Already have the machine? The fleet's actual launch scripts and settings for llama.cpp, Ollama, and vLLM, each annotated with the hardware tier it runs on, the context size it was tuned for, and the failure it prevents — plus the Failure-Modes Field Guide: eight production serving failures as symptom → real cause → the distinguishing test → the fix. https://robotnurse.gumroad.com/l/serving-config-pack
• Both together — $69 instead of $78 bought separately. https://robotnurse.gumroad.com/l/toolkit-config-bundle

Full disclosure, because the notes would want it in writing: the buyer checklist PDF inside the toolkit is the same one-page checklist you can download free at robotnurse.io/checklist — it's bundled for convenience, not sold as an extra. Every download carries free updates when new measurements change a recommendation, and a 30-day refund: if it didn't help, email us and you shouldn't pay for it.

One chapter of the toolkit is a guide to deciding local AI isn't right for you. If that's where the arithmetic lands, the honest answer is to not buy the machine — and not buy the toolkit either.