From TechRadar · · 2m
The AMD rig beats ChatGPT on cost after 3.5 hours a week
A dual-GPU AMD workstation priced against GPT-5.6 and Gemini, hour by hour, and the break-even lands in strange places.
In brief
A dual-GPU AMD workstation priced against GPT-5.6 and Gemini, hour by hour, and the break-even lands in strange places. Owned AI hardware only wins if you burn expensive tokens in volume. Against cheap cloud models the math never catches up, and the local model still reasons worse. Originally reported by TechRadar.
An $18,775 bet on local AI
A reviewer dropped two AMD Radeon AI PRO R9700 cards, 32GB each, into a workstation costing about $18,775 as tested. The measurements: power draw, token throughput, and amortized hardware against cloud tiers.
Cloud pricing sets the bar
Cloud output runs from $1.20 per million tokens for GPT-5.6 Luna up to $30 for Sol, with Claude Sonnet 5 at $10 and Opus 5 at $25. The pricier your model, the faster hardware pays back.
“The more expensive the cloud model a team would otherwise use, the sooner owned hardware pays for itself.”
A speed trick doubles output
Both cards together served eight simultaneous users at 156.2 tokens per second. Multi-Token Prediction nearly doubled that to 320.2 tokens per second with no loss in output quality.
Break-even in hours per week
At 320.2 tokens per second, the machine needs just 3.5 hours of weekly use to undercut GPT-5.6 Sol, 4.2 hours to beat Claude Opus 5, and 8.8 hours to beat Gemini 3.1 Pro.
Where the real money shows up
Past 20 million tokens a month at Sol pricing, the rig costs about $6,262 a year in electricity and amortized hardware against roughly $18,000 in cloud fees.
“That $11,738 yearly gap is the actual evidence behind the claim that heavy monthly usage makes AMD's rig worthwhile.”
Cheap models break the math
Compare against GPT-5.6 Luna at $1.20 per million tokens and you need 94.3 hours of use every week to break even. A week holds 168 hours, so the box would need to run nearly saturated.
“that math flips entirely in the other direction”
One card, lower bar
A single R9700 ran an eight billion parameter model at 34.5 tokens per second, pulling 221 to 283 watts versus 310 to 510 for the pair. That makes an $1,880 card easier to justify than $3,760.
“Smaller models therefore offer a second path into positive economics”
You pay in reasoning
On one independent intelligence index, the 27 billion parameter local model scored 37 against Gemini 3.1 Pro's 46. The cheaper tokens are measurably weaker on hard problems.
“a team chasing the cheaper token count may be trading away real reasoning quality, not just cloud subscription fees.”
The key point
Owned AI hardware only wins if you burn expensive tokens in volume. Against cheap cloud models the math never catches up, and the local model still reasons worse.
