Teraplex capacity digest https://teraplex.us/capacity-digest A dated read of public GPU capacity and price pages, for people who own a site with land and power. Every number carries a live source and the date it was read. en-us Capacity digest 03: Public GPU capacity listings https://teraplex.us/capacity-digest/2026-07-28-issue-03 https://teraplex.us/capacity-digest/2026-07-28-issue-03 Wed, 29 Jul 2026 02:52:02 GMT Fourteen dated public GPU listings from six providers, shown with pricing, scale, provider-stated status and missing infrastructure fields preserved. Every row below is a dated read of public provider pages. The pages can change after re-read, so the source log preserves the observation boundary.

Offers

Provider Public configurations Capacity per listing Status boundary Published price Commercial basis Source
Impossible Cloud H100 SXM; GB300; two B300 configurations 8–1,152 GPUs Two dated listings; two not displayed $2.41–$4.42/GPU-hour 12–60 months Impossible Cloud
Prime Intellect B300 SXM6 reserved example 512 GPUs Not displayed $5.00/GPU-hour 3 years Prime Intellect
Verda Four Instant Cluster configurations; one H200 VM; one B200 VM 1–144 GPUs Provider-stated self-service; current stock not verified $4.00–$8.62/GPU-hour PAYG; VM use billed in prepaid 10-minute increments; VAT applies Instant Clusters, H200 VM, B200 VM
Gcore One H100 80 GB virtual GPU machine 1 GPU Provider-stated self-service. Sines-3 is the named launch region. Current stock is unverified. €1.65/VM-hour Minimum term not published; prices exclude VAT Gcore
Google Cloud One a3-highgpu-8g Flex-start VM 8 GPUs Listed in us-central1-c. The zone is supported. Current stock is unverified. USD 38.32/VM-hour PAYG; runtime up to 7 days; reservations unsupported Google Cloud
Crusoe One h100-80gb-sxm-ib.8x VM 8 GPUs Provider-stated self-service; current stock not verified $3.90/GPU-hour On demand; no minimum commitment Crusoe

The dashboard preserves each configuration inside its provider row. The five added records appear only in the currently listed bucket. Provider-stated self-service paths, supported zones, published prices and documented create workflows do not establish current stock, launchability, reservability or successful provisioning.

These records differ in hardware generation, cluster size, region disclosure, availability, term and source type. They do not establish a market average or directly comparable quotes. Every derived normalization is labeled in the dashboard. No authenticated inventory query, reservation, provisioning request or billable launch was performed.

Direct source log

Electrical disclosure boundary

The provider pages do not display the utility, service voltage, committed megawatts, study status or grid-energization date for these records. Compute-fabric terms such as InfiniBand or bandwidth per node do not establish electrical interconnection.

FERC’s 18 June 2026 action addresses the rules governing how data centers and other large energy users connect to the grid. FERC large-load integration release.

Boundary

This is dated public-data curation. It is not an offer, availability confirmation, valuation, pricing recommendation, demand forecast or matching service. Availability, configuration, taxes, region and commercial terms require confirmation before procurement or scheduling.

]]> Capacity digest 02: One GPU model, three sets of quoted terms https://teraplex.us/capacity-digest/2026-07-28-issue-02 https://teraplex.us/capacity-digest/2026-07-28-issue-02 Tue, 28 Jul 2026 00:54:48 GMT The same B300 silicon at $3.86, $4.42 and $5.00 per GPU-hour, and what separates the three numbers. Both pages were read on July 27, 2026, and both can change after capture. No page-level effective date was captured from either page. “Not captured” marks a field Teraplex did not record from that page on that date; it is not a statement about what the publisher did or did not show.

One model, three sets of quoted terms

Public B300 record Region Availability Minimum term Listed rate Source
256 GPUs, 32 nodes, SXM US 2027-08-01 36 months $4.42/GPU-hour Impossible Cloud
1,024 GPUs, 128 nodes, SXM US Not captured 60 months $3.86/GPU-hour Impossible Cloud
512-GPU SXM6 reserved example Not captured Not captured Three years $5.00/GPU-hour Prime Intellect

These public records differ in cluster size, term, interconnect detail captured, region, availability and source type. Teraplex does not treat them as comparable quotes.

Availability as a schedule field

A GPU-hour rate can be public while no delivery date is captured alongside it. Teraplex’s reading method for an electrical-infrastructure schedule is to treat an uncaptured delivery date as unknown rather than as immediate availability.

The first record carries a dated availability field. For the second record, availability was not captured; for the reserved example, neither region nor availability was captured. Teraplex records those differences to keep a public price record separate from a delivery plan.

Boundary

This issue is public-data curation. It is not an offer, availability confirmation, valuation, pricing recommendation, demand forecast or matching service. It collects no reader data.

]]> Capacity digest 01: Six fields that decide whether a GPU-hour rate means anything https://teraplex.us/capacity-digest/2026-07-27-issue-01 https://teraplex.us/capacity-digest/2026-07-27-issue-01 Tue, 28 Jul 2026 00:36:47 GMT H100 index at $2.72, public listings from $2.41 to $5.00, and why the headline rate is the least useful number on the page. Every value below was read from a public page on July 27, 2026. Listings are live documents and can change after capture. “Not captured” marks a field Teraplex did not record from that page on that date; it is not a statement about what the publisher did or did not show.

Price index

Ornn’s H100 SXM daily index printed $2.72 per GPU-hour on July 27. Across its last seven daily observations, the index moved from $2.73 to $2.72, a 0.4% decline. Ornn public history.

Ornn defines this global index as a weighted average of verified compute transactions in US dollars per GPU-hour. The daily value is an aggregate of hourly observations. Ornn price-index methodology.

Public capacity watch

GPU Public capacity record Listed rate Source
H100 SXM 8 GPUs, 1 node, US, 12-month minimum; availability not captured $2.41/GPU-hour Impossible Cloud
GB300 1,152 GPUs, 16 nodes, US, available September 30, 2026, 36-month minimum $4.32/GPU-hour Impossible Cloud
B300 SXM6 512-GPU example, three-year reserved term; region and availability not captured $5.00/GPU-hour Prime Intellect

These rows cover different hardware generations, cluster sizes, availability fields, regions and terms. Ornn is a transaction index. The other figures are provider list or example prices. The table does not establish a market average or an achievable quote.

Site-owner read

This section is Teraplex’s own reading method for these records. No published standard or industry convention defines it. We treat a GPU-hour rate as usable in a powered-site plan only when the same record also preserves six fields:

  1. GPU model
  2. Cluster size
  3. Region
  4. Availability date
  5. Minimum term
  6. Interconnect

Recording those six fields next to the rate is what keeps an index, an on-page listing and a reserved-cluster example distinguishable from one another before any rate enters an infrastructure or revenue model.

Boundary

This issue is public-data curation. It is not an offer, valuation, pricing recommendation, availability confirmation, demand forecast or matching service. It collects no reader data.

]]>