The Price–Performance Frontier of Local AI Hardware

The frontier

Each point is one machine. The filled points are the front, joined left to right; the faint ones are beaten by at least one front machine. Switch the output axis to writing speed if you care most about how fast an answer appears, or to reading speed if you feed long documents and agent contexts. Switch the cost axis to the 5-year cost to count the electricity a machine burns running all day.

How to read the chart

One setup beats another when it costs no more and delivers no less, and is strictly better on at least one of the two. A setup on the front is beaten by nothing: the area to its upper left, cheaper and better, is empty. Everything to its lower right it beats.

The line between two front setups is a guide, not a machine. Strictly, a budget that falls between them buys the cheaper one's output, a flat step; the straight segments simply make the shape of the trade-off easier to read. For the same reason, a setup off the front is judged against what its own price can actually buy: its output as a share of the best output any setup delivers at or below its cost.

How it is built

The data

Every number comes from the ModelDock Benchmark data for Qwen3.8 Flash Next. Reading (prefill) and writing (decode) speeds are medians of published source reports, listed with their conditions on Evidence, or a clearly marked scaled estimate where a machine has no report yet; estimates are the hollow points. Prices are the current market prices on Data. Machines that fit the model only at 3-bit or below are left out, as Benchmark and Ranking hide them by default, and a machine without a price or a power figure cannot be placed.

Cost

The system price of a graphics-card machine is the card's price plus a $1,200 host: processor, board, 64 GB of RAM, power supply, case and storage. Integrated systems such as Macs, the Ryzen AI Max+ 395 and DGX Spark count at their complete price. The 5-year cost adds the electricity of running all day at $0.15 per kWh, for a machine drawing \(W\) watts and costing \(P\) dollars:

Every 100 W therefore adds about $660 over five years of continuous use.

Output

Daily output is the Benchmark's default workload: the tokens a machine gets through in 24 hours when 90% of them are read and 10% written. With reading speed \(R\) and writing speed \(D\) in tokens per second:

It suits a machine that runs agents around the clock. Writing speed alone is closer to what one person waiting for an answer feels; reading speed decides how fast long inputs are taken in.

The front

Sort the machines by cost and walk from the cheapest up, keeping each one that delivers more than everything before it. The kept machines are the front. Cost per million tokens, the measure the Ranking sorts by, is not an axis here: it is the ratio of cost to output, so it answers a different question, "how cheap is each token", rather than "how much does my budget buy". It is listed in the table of every setup at the end.

What the frontier shows

Steep, then flat

The front rises fastest where money buys graphics memory. From the RTX 3060 at $1,560 through the 16 GB cards to the RTX 5080 at $2,800, daily output climbs from 14 to 58 million tokens: on average, each extra million tokens a day costs under $30 of purchase price. For a mixture-of-experts model this is where the expert cache grows: every gigabyte of VRAM keeps more of the model's experts off the CPU, as the Strata article explains.

Past about $3,000 the curve bends. From the RTX 5080 to the RTX 5090 at $8,450, through two RTX 5070 Ti cards, an RTX 4090 and the Mac Studio M5 Ultra 96 GB, output rises to 120 million tokens a day at about $90 per extra million: three times the rate of the first stretch. The last step, to an RTX PRO 6000 at $19,200, buys 6 million more for $10,750, about $1,760 per extra million. Its 96 GB pays off with larger models or many parallel users, which this single-user workload does not measure.

Cheapest per token is not cheapest to buy

The cheapest machine on the front, the RTX 3060, is also the most expensive per token on it: $0.140 per million tokens, because its low output carries the same $1,200 host. The lowest cost per token on the front belongs to the Mac Studio M5 Ultra 96 GB at $0.047, followed by the RTX 4090 at $0.058 and the RTX 5080 at $0.059. A buyer with a fixed budget and a buyer optimizing every token's price read the same chart in different places.

Electricity reorders the middle

Switch the cost axis to the 5-year cost and the front keeps 15 machines but changes its members. Low-power machines move left and join it: the Mac mini M5 Pro 64 GB, the Intel Arc Pro B60, the Ryzen AI Max+ 395 and the RTX 5070. Power-hungry ones move right and drop off: the pair of RTX 3090 cards draws 850 W, worth about $5,600 over five years, more than its purchase price.

Writing and reading tell different stories

On writing speed, ten machines form the front, and the RTX 4070 Ti SUPER at $2,120 already writes 72 tokens per second. The Mac Studio M5 Ultra 96 GB, at 140, is the step before the RTX 5090's 174; the RTX PRO 6000 falls just off the front, at 98% of the cheaper RTX 5090. On reading speed the front is all graphics cards: the RTX 5080 reads 3,591 tokens per second at $2,800, the RTX 5090 6,257 and the RTX PRO 6000 8,988, and no Mac is on it.

Where not to trust it

• Sources differ. A median combines reports with different quantizations, runtime versions, context lengths and host machines. The same graphics card can be up to twice as fast in another PC, depending on its memory and PCIe link.
• Estimates are estimates. A hollow point is scaled from a measured anchor on similar hardware, not measured; its place on the front is provisional.
• Close is a tie. A machine within about 10% of the front is, given the spread of the sources, practically on it; the chart lists them.
• Prices move. The front is drawn at the day's market prices; the Data page keeps their history.
• One workload. Single-stream use, a 90/10 reading and writing mix and 24 hours a day. A different mix moves every point: the Benchmark workbench recomputes the economics for yours.
• What it leaves out. Headroom for longer contexts or larger models, noise, size and software support are not on either axis.

Use it

Start from the budget, find the front machine at or just below it, and look at the near-front list for alternatives that the spread of the sources cannot separate. Then check the machine's own page for its sources and conditions. To see where your own machine would sit, run the ModelDock Speed Test: your agent benchmarks it in one line and returns a share card, and results you choose to submit are reviewed before they join the public data.

Sources and conditions

• ModelDock Benchmark data for Qwen3.8 Flash Next: rates from the medians on Evidence, prices and power from Data, economics from the shared Benchmark model; the same rows are downloadable as JSON and CSV. The chart's date is shown under it.
• Defaults: 90% reading and 10% writing, 24 hours a day, $0.15 per kWh, a 5-year service life, a $1,200 host for graphics-card machines. All of them can be changed on the Benchmark page.
• The numbers quoted in the text are from the data of 15 October 2026.