Ask what an HGX B300 8-GPU system costs and you can collect several honest answers in the same week. The maker's specification sheet says what the system contains. A reseller's page shows an asking price. A seller sends a named buyer a quote with terms and an expiry date. A broker mentions a figure on a call. A company reports that it bought systems. An invoice and a delivery record confirm a sale. Each statement is true about something different. Averaged together, they produce a number that describes nothing.
This note sets out six evidence classes for GPU system prices, sorts public examples from 2024 to 2026 into them, and explains why the separation matters for valuation, lending and price indexes.
Six classes of evidence
| Evidence class | What it establishes | Typical source |
|---|---|---|
| Specification | What the system contains. It carries no price. | Maker's specification sheet |
| Listing | What a seller asks in public, at a point in time | Pricing page or pricing API |
| Quote | What a seller offers a named buyer, with terms and an expiry | Written quote |
| Indicative price | An estimate with no commitment behind it | Interview, analyst note, broker message |
| Reported sale | That a sale took place, according to a party or a publication | Filing, press release, press report |
| Verified transaction | A completed sale, confirmed by documents | Invoice, payment and delivery records |
The classes run from what describes the system to what a buyer actually paid. Each observation carries exactly one class, set when it is captured.
What the public record shows
Public prices for GPUs and GPU capacity sort into these classes once you ask two questions: who said it, and what does the figure include.
Listings move by pricing type. In June 2025, AWS cut on-demand prices for its H100-based P5 instances by 44% and for its H200-based P5en instances by 25%, against prices on 31 May 2025 [5]. In January 2026, AWS raised prices for EC2 Capacity Blocks, which reserve GPU instances for a future start date, by about 15% for H200 instances in US East (Ohio) [6]. AWS describes Capacity Block prices as dynamic, changing with supply and demand [7]. Both are listings from one provider. One fell and one rose because they are different pricing types. A rental series that mixes them records a change that comes from the mix.
Indicative prices travel far. In March 2024, NVIDIA's chief executive told CNBC that a Blackwell GPU would cost between $30,000 and $40,000, and later added that the cost includes designing and integrating data centers [8]. In May 2024, Tom's Hardware reported an HSBC estimate of about $3 million for a GB200 NVL72 rack, relayed through a journalist's social media post [9]. Both figures are useful context. Neither is an offer to a buyer or a record of a sale.
Reported sales depend on scope. In July 2025, IREN announced the purchase of about 2,400 B200 and B300 GPUs for about $130 million, including servers, storage, labor, racks and cabling [10]. Its annual report for the year to June 2025 describes an agreement of the same date for about 1,300 B200 and 1,200 B300 GPUs with ancillary equipment, at about $123.6 million [11]. Per GPU, the press release implies about $54,200 and the filing about $49,400. One buyer, one date, two documents and a gap of almost 10% per GPU, because the count and the scope differ.
System figures are not chip figures. The same filing reports an agreement for 4,200 B200 GPUs at about $192.9 million, with no itemized scope, and one for 1,200 GB300 GPUs with ancillary equipment at about $96.2 million [11]. Divided by GPU count, that is about $45,900 and $80,200 per GPU. Both sit above the $30,000 to $40,000 chip range from 2024. The gap says nothing about price movement. One figure is an indicative chip price. The others are reported purchase agreements whose totals can include more than the chip.
Why mixing classes misleads
A value built from listings describes what sellers ask. A value built from reported and verified sales describes what buyers paid. Both answer real questions, and they are different questions.
Mixing also hides scope. A per-GPU figure from a rack purchase includes CPUs, memory, networking and integration. A per-GPU figure from a chip estimate includes none of them. Without a grade attached to each observation, the two look comparable when they are not.
Commitment differs too. A quote binds the seller for a named buyer until it expires. An indicative price binds nobody. A listing for cloud capacity is executable only while capacity is available, which is why availability belongs next to the price.
Finally, a sale between related parties is a weaker signal of market value. The EU Benchmarks Regulation defines transaction data as observable prices representing transactions "between unaffiliated counterparties in an active market" [2].
What benchmark standards ask for
Financial benchmark standards address the same problem. IOSCO's Principles for Financial Benchmarks, published in July 2013, ask that a benchmark be anchored in observable arm's-length transactions [1]. Principle 8 sets a general hierarchy of inputs: concluded arm's-length transactions in the underlying interest, then in related markets, then firm bids and offers, then other market information or expert judgment [1]. Principle 9 asks the administrator to publish, with each determination, indicative percentages of each type of market data used [1].
The EU Benchmarks Regulation of 2016 takes the same line. Input data "shall be transaction data, if available and appropriate," and the administrator publishes guidelines on the types of input data and their priority [2].
These rules were written for financial benchmarks. They are useful design background for any price series, including one for GPU compute.
What it means for valuation and lending
GPU fleets back debt. In April 2024, Lambda secured a GPU-backed financing vehicle of up to $500 million [14]. In July 2026, Nebius announced a senior secured facility of about $775 million, backed by deployed GPU infrastructure and contracted cash flows [16]. A lender sizing an advance against GPUs needs to know whether the value rests on asking prices, estimates or completed sales.
Book values rest on estimates. Effective 1 January 2025, Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five, citing the pace of AI and machine learning development [12]. Meta extended the useful life of certain servers and network assets to 5.5 years from the same fiscal year [13]. Both are management estimates. Neither is an observed price for used hardware.
A valuation should state which classes it rests on. In financing diligence, borrower and lender can review the same records, see each record's class and source, and agree on the evidence before they discuss the number.
In the Rillor Compute Index
The Rillor Compute Index follows the same rule. Its methodology v1.0 defines four series: RCI Rental, RCI Availability, RCI Hardware and RCI Rent-to-Value [17]. In RCI Hardware, each observation of a published listing, quote or reported sale is tagged with its evidence class, and the classes stay separate in calculation [17]. A series publishes values only when it meets the methodology's data rules [17].
Read the full report
The full report, with the method, the worked examples, the comparison with benchmark and accounting standards, and the numbered sources behind every figure here, is available as a PDF.