Node health — what the headline counts hide
Findings from first-party endpoints · generated 2026-08-21T03:55:01Z
Four findings were tested against the underlying tables before anything was
written here. Two survived, one survived only as a level, and one was refused.
The refusals are at the bottom and they are the point: any scraper can produce numbers, and a
competitor pointed at these same endpoints would have published all four.
1. Akash: the GPU market is tightening
Available GPUs fell -43.5% while
active rose +101.6%. Idle supply is being
absorbed.
| Date | Total GPUs | Available | Active |
Available share |
|---|
| 2026-08-09 | 423 | 294 | 127 | 69.5% |
| 2026-08-10 | 443 | 304 | 138 | 68.6% |
| 2026-08-11 | 429 | 249 | 179 | 58.0% |
| 2026-08-12 | 430 | 244 | 185 | 56.7% |
| 2026-08-13 | 425 | 227 | 197 | 53.4% |
| 2026-08-14 | 419 | 222 | 196 | 53.0% |
| 2026-08-15 | 417 | 167 | 249 | 40.0% |
| 2026-08-16 | 401 | 154 | 246 | 38.4% |
| 2026-08-17 | 412 | 163 | 248 | 39.6% |
| 2026-08-18 | 420 | 150 | 268 | 35.7% |
| 2026-08-19 | 422 | 160 | 261 | 37.9% |
| 2026-08-20 | 424 | 166 | 256 | 39.2% |
Caveat, stated plainly: 12 daily
observations is a direction, not a conclusion. A single large deployment would produce the same
shape. We publish it because the direction is unambiguous and every point is checkable at
console-api.akash.network
— not because 4 points establish a trend.
2. Storj: the node count is inflated 3.7x
| Bucket | Nodes | Share of total |
|---|
| Total (headline figure) | 123,505 | 100% |
| Disqualified | 80,575 | 65.2% |
| Active | 33,045 | 26.8% |
Quoting 123,505 nodes overstates the working network by
3.7x. Source:
stats.storjshare.io.
Two things we will not claim.
(a) The buckets are not a partition — active, disqualified, exited, offline and
suspended sum to 124,363 against a stated total of 123,505, so they
overlap by 858. We show only the two ratios that are
well-defined against the total.
(b) A tempting reading — “the network is accumulating dead nodes faster than live
ones” — is not supported by our own data. Over
2026-08-06 to 2026-08-20, disqualified grew +340 while active grew
+108. Active grew faster. We checked because the story was attractive.
3. ThreeFold: 83% of registered CPU is dark
438 nodes reporting up.
1,575 of 9,142 registered cores in use
(17.2%). Source:
gridproxy.grid.tf.
See the utilisation report for the full capacity picture.
Refused — and why
“DefiLlama understates GEODnet revenue by 91%.”
REFUSED.
GEODnet's API reports total_usage_revenue = 10,886,080;
DefiLlama reports a 30-day revenue of 920,140. The ratio is
11.83x and is stable across all 7 days — which looks compelling and is
meaningless: GEODnet's field is cumulative lifetime revenue and
DefiLlama's is a 30-day window. 11.8x is simply ~11.8 months
of revenue at the current run rate. The stability is expected, not evidence.
This is the same error class our ontology already refuses for Akash
(ACTIVE_LEASES / TOTAL_LEASES: a lifetime denominator yielding a meaningless
0.1%). Publishing it would have been a false public accusation against a third party, derived
from a unit error. Additionally, total_usage_revenue fell mid-window,
which a cumulative total cannot do — it is token-denominated and revalued daily, so it is
not a USD revenue series at all.
“Filecoin revenue per miner rose 26% — oversupply
is clearing.” REFUSED.
Active miners did fall (537→529). But the revenue increase behind that claim is a
single day of a 24-hour series. Over the same window Filecoin's 30-day revenue
fell 45.2% (2026-08-06 to 2026-08-20) — including a
17.6% drop in one day between 2026-08-07 and 2026-08-08, which is an upstream revision rather
than anything organic. A per-unit ratio built on a 24h numerator over a 4-day window, when the
30-day series moves the other way, is a story rather than a finding.
Cross-network revenue-per-unit league tables. REFUSED.
A GEODnet weather station is not an Akash datacentre. The spread between them is arithmetically
real and economically meaningless. Within-network direction is publishable; the
cross-network level is not.