Availability Observatory
Shadeform aggregates 19 GPU clouds into a single live catalog. This page surfaces what no single-source view can: where each chip is procurable right now, across providers and regions; the per-provider price spread; boot-time tradeoffs; config availability; long-tail chip catalog. Observatory data — not part of the CCIR rate methodology. Shadeform-reported pricing is third-party-attributed.
Provider × region grid
—
Each cell is one (provider × region) pair for the selected chip. Green
= available now, red = offline. Hover for $/GPU/hr. Cells are sorted
by provider (rows) and region (columns), so chronic outages on a single
provider show up as a row of red. per_gpu_dph = hourly_price ÷ 100 ÷ num_gpus.
Per-provider summary
—| Provider | SKUs | Distinct | Available | Median $/GPU | Min $/GPU | Boot min (s) | Boot max (s) |
|---|
Shadeform tracks 19 GPU clouds; not all carry the selected chip. Boot times are minimum/maximum provisioning seconds reported by the provider. Latitude commits to ~60s; some hyperscale-scale fleets need 180-240s.
Per-provider price spread
—Each provider's min, median, and max per-GPU $/hr for the selected chip. Wide bars mean broad SKU coverage; narrow bars mean a single configuration. Cross-provider gap = arbitrage signal. (Same chip can clear $1.95-$2.99/GPU/hr across nine venues for H100 SXM.)
Available configurations
—| Provider | Interconnect | NVLink | GPUs | vCPU envelope | RAM (GB) | $/GPU/hr | Available |
|---|
Per-SKU configuration matrix. NVLink ✓ means the provider exposes explicit NVLink topology — not just SXM form. Empty cells reflect Shadeform's reporting; the underlying provider may have richer config detail in their own API.
Provisioning seconds by provider
—Min / max provisioning seconds per provider. Matters for elasticity- sensitive workloads — a 4× boot-time gap between providers can be more material than a $0.10/GPU/hr price gap on short-duration jobs.
Region distribution
—| Region | SKUs | Available | % Avail | Providers | Min $/GPU |
|---|
Where on earth is the selected chip procurable. The same chip can be online in Iceland but offline in Virginia, and price-anchored to the cheapest available cell across all regions.
Availability % over 30 days
—
Per-cycle pct_available = n_available ÷ n_total_sku_region
for the selected chip. Drops indicate fleet-wide outages or supply
cliffs; sustained climbs indicate new capacity onboarding. This
is the unique CCIR view — Shadeform exposes current state
only; CCIR persists the cycle-by-cycle history.
Cross-provider redundancy
—| Interconnect × num_GPUs | Providers | SKUs | Available | Carriers |
|---|
For each (interconnect × num_GPUs) configuration, how many providers carry it. Single-provider rows mean concentration risk — capacity loss has nowhere to fail over to. Multi-provider rows mean cross- venue arbitrage is possible.
Cross-chip headline
—| Chip | SKU × region | Available | % Avail | Providers | Min $/GPU | Median $/GPU | Cheapest provider |
|---|
Every chip Shadeform reports across the headline + long-tail panel. B300, GH200, and RTX Pro 6000 give early visibility into chips before they hit the CCIR rate panel.
Refreshed every three hours alongside /marketplace and /vast. Shadeform aggregator API authenticated via CCIR research key — values are public list-asks as Shadeform reports them. Aggregator depth: 19 GPU clouds, 28 GPU types, 269 instance types, ~671 (type × region) availability cells per cycle.