To compare GPU server prices across configurations, put both builds on one basis first: the same kind of price, a stated date and the same parts, or a stated adjustment for each part that differs. On 8 November 2026, Exxact's configurator listed four eight-GPU HGX B300 servers at $529,958 to $588,841 [21, 22, 23, 24]. All four carry the same GPU board, the same three-year warranty and the same capture date, and the spread between them is $58,883, or 11.1% (derived) [21, 22, 23, 24]. Their host CPUs, host memory, storage and network hardware differ.
Statistical agencies meet the same problem with computers and microprocessors, where models turn over quickly [1, 11]. This note applies four of their tools to public GPU server listings: matched builds, adjustment with line-item prices, price per unit of performance and hedonic regression. Each one answers part of the question, and each stops at a point the listings make visible.
Why price per GPU misleads
Dividing a system price by eight gives a figure per GPU that still carries the CPUs, memory, chassis and everything else in the box. Epoch AI built its price-performance figures this way, from the cost of a DGX server divided by eight, which it says "accounts for server and interconnect hardware costs" [31].
The parts behind that figure vary by system type. An NVIDIA DGX B200 has two host CPUs with 112 cores for its eight GPUs [15]. A GB300 NVL72 rack has 36 Grace CPUs for 72 GPUs, one CPU per two GPUs and 36 CPU cores per GPU (derived) [16]. A per-GPU figure from a rack carries more host per GPU than one from a server.
Fleet purchases carry more again. In July 2025, IREN announced the purchase of about 2,400 B200 and B300 GPUs for about $130 million, including servers, storage, labor, InfiniBand, racks and cabling, or about $54,167 per GPU (derived) [29]. The evidence classes note explains why a fleet purchase, a listing and a quote stay apart.
National accountants measured the size of this effect for computers. The Bureau of Economic Analysis compared the average unit value of single-user computers with its quality-adjusted price index for 2001 to 2005. Unit values fell 4.9% a year and the index fell 16.4% a year [12]. A price per GPU is a unit value of the same kind (this note's reading). What a system price covers, part by part, is set out in what an HGX B200 or B300 server price includes.
Matched builds: the same configuration at two dates
A matched-model index prices the same products every period, which separates price change from quality change [4]. Exxact's TS4-126572371 page kept one default HGX B300 build across three captures: two AMD EPYC 9554 CPUs, 2,304 GB of DDR5-5600, 1.92 TB of U.2 storage and no add-in network card [21, 25, 26].
| Capture | Default price | 24 x 96 GB to 24 x 128 GB |
|---|---|---|
| 16 February 2026 | $439,982.40 | +$7,920 |
| 16 May 2026 | $513,073.00 | +$14,520 |
| 8 November 2026 | $535,513.00 | +$25,080 |
Sources: [21, 25, 26]. February and May from the Internet Archive's Wayback Machine.
The build's price rose 21.7% over the period, and the memory upgrade on it rose to 3.2 times its February price (derived) [21, 25, 26]. That is a price change for one build at one seller, read from listings. The report on memory prices and GPU server costs follows the same configurators across more dates. TrendForce projected on 30 September 2026 that conventional DRAM contract prices would rise 10% to 15% quarter on quarter in the fourth quarter of 2026 [35].
Matching breaks on turnover. The authors of a working paper of the U.S. Bureau of Labor Statistics (BLS) write that for microprocessors, whose prices seldom change during a product's life, a strict matched-model method "would yield constant price levels, no matter how drastically prices or product qualities differ for new models" [4]. Pakes found matched-model indexes for PCs showing no price fall on average, while the hedonic index showed steep declines in every year [8]. Labels move too: the board on that page read "2.3TB HBM3e" in February and May and "288 GB HBM3e" in November, with the build unchanged [21, 25, 26]. Match on the parts list, never on the label.
Adjusting with line-item prices
When two builds differ in a few parts, price the difference with the seller's own options. Eurostat's price index manual calls this option pricing and notes that it has been used for desktop computers [13]. Here the three Intel-based B300 builds are taken to two Xeon 6747P CPUs, 3,072 GB of DDR5-6400 and no local drives:
| Build | Listed | Adjustment | Adjusted |
|---|---|---|---|
| TS4-131188947 | $529,958 | Drives out, 32 x 96 GB in | $546,678 |
| TS4-198437339 | $564,718 | Drives out, CPUs to 6747P | $569,778 |
| TS4-104366747 | $588,841 | Drives out | $580,921 |
Sources: [22, 23, 24]. This note's arithmetic on each page's option prices.
The gap between TS4-131188947 and TS4-104366747 shrinks from $58,883 to $34,243, or 6.3% (derived) [22, 23]. What remains sits in parts no configurator offers as an option, such as the chassis, power supplies and integrated network adapters (this note's reading). The fourth build, on AMD EPYC CPUs, cannot reach the same basis: its page offers 3,072 GB of DDR5-6400 for $29,040 over the default memory but no Xeon CPUs [21].
Option prices are differences inside one configurator on one day, and they are not market prices for the parts. Adding 1,024 GB of DDR5-6400 cost $21,120 on one Exxact B300 page, moving from 64 GB to 96 GB modules, and $33,440 on another, moving from 96 GB to 128 GB, or $20.63 and $32.66 per added GB (derived) [22, 24]. On one page, the 86-core CPU upgrade cost less than the 64-core one [22]. Eurostat's method values a feature at a proportion of its market price that is "no larger than one" [13].
Price per unit of performance flips with the metric
A price per FLOP or per terabyte of GPU memory replaces the GPU count with a performance figure. Lenovo stated a "usual sale price" as of 15 June 2026 of $550,475.10 for an eight-GPU B200 server and $785,606.50 for an eight-GPU B300 server [27]. With NVIDIA's dense figures [14, 15]:
| Metric | B300 against B200 |
|---|---|
| Per GPU | +42.7% |
| Per dense FP8 PFLOP/s | +42.7% |
| Per dense FP4 PFLOP/s | -4.9% |
| Per TB of GPU memory | -2.1% |
| Per FP64 TFLOP/s | About 42 times |
Sources: [14, 15, 27]. This note's arithmetic.
The ranking depends on the metric. The two Lenovo servers are also different models, the B300 one designed for Intel Xeon 6 processors, so the premium mixes the GPU board with the host [27]. State precision and sparsity with every figure: NVIDIA prints sparse and dense FP4 values for HGX [14], and AMD's MI355X chart shows B200 at NVIDIA's sparse figure for FP8 and its dense figure for FP6 (derived) [14, 17]. Comparing generations has its own method, set out in how to compare GPU generations on price per FLOP. A price per unit of measured throughput is a derived metric, and MLCommons requires a comparison on such a metric to state its basis and never to present it as an official MLPerf® metric [34].
Hedonic regression needs more data than one seller shows
Hedonic regression estimates a price for each characteristic of a product from many observations, then removes the value of a quality change from a price comparison [5, 6]. BLS introduced a two-period time-dummy model for server microprocessors with its July 2018 data, with cores, cache per core, frequencies, thermal design power and a SPEC CPU2017 score among the candidate variables [1]. The time dummy's coefficient estimates the price change between two adjacent quarters [1].
The method needs volume. BLS's study of cloud prices used 128 to 236 observations per model for one provider and reported counterintuitive signs on some coefficients [3]. A BLS working paper found that error rises as samples shrink, and that the simple time-dummy model has lower variance on small samples, with a potential for more bias [4]. Eurostat's manual says hedonic methods need "a reasonably large and detailed data set" [13]. Four B300 builds carry more characteristics than observations, so this note runs no regression on them (this note's reading).
How to compare GPU server prices across configurations
- Compare like with like on evidence: a listing with a listing, a quote with a quote.
- Date both prices, and treat captures weeks apart as separate observations.
- Check scope. Warranty terms differ: three years of parts and labor on Exxact's pages [21], three years of labor and one of parts on one Wiredzone listing [28]. NVIDIA AI Enterprise is licensed per GPU and bought separately for Blackwell DGX systems [30].
- Adjust the parts that differ with line-item prices from the same page and date, and report the residual.
- Choose the unit after the builds match, and state it.
Rillor's GPU systems pricing models configurations by GPU family, GPU count, memory, CPU, interconnect, cooling, condition, location and seller type, and every observation keeps its capture time, source and evidence class, so a listing is read as a listing and a quote as a quote.
Read the full report
The full report adds the time-dummy model and its bias adjustment, the BLS evidence on sample size, the line-item steps for every build, the five metrics in dollars and the numbered sources.