Kairos Signal · Capacity Telemetry · 6–17 Aug 2026

Two compute networks, the same eleven days, opposite directions

io.net grew its registered device fleet by 23.9% and delivered no additional compute. Akash barely grew its fleet at all and nearly doubled GPU utilisation. Both facts are invisible from a node count — the metric the category is usually sold on.

The number that isn't published

Utilisation is a ratio. Almost every DePIN network publishes the numerator — devices active, GPUs hired, cores in use — because it grows. The denominator, capacity actually registered to the network, is what turns that into a rate, and it is rarely exposed and almost never archived. Without it you can watch supply climb forever and never learn whether anyone is renting it.

We hold both halves, timestamped, sampled roughly hourly since 6 August. That makes one question answerable: is this network adding capacity faster than demand absorbs it?

io.net · devices
+23.9%
Registered fleet grew 2,196 → 2,722 devices
io.net · compute delivered
+0.0%
28,185 → 28,195 compute-hours per day. Flat.
Akash · GPU utilisation
31.6→61.2%
Demand up 83% on a fleet that shrank 5%

io.net added 526 devices and produced no additional output. Compute-hours served per day sat at 28,185 across the first three days of the window and 28,195 across the last three — a change of one part in three thousand — while the fleet behind it grew by nearly a quarter. Active devices fell 3.4%. Running clusters fell 37.9%.


The crossing

Plotted as utilisation, the two networks trade places. io.net enters the window as the better-used network and leaves it as the worse-used one.

Capacity utilisation, daily mean
Share of registered capacity actually in use. Higher is tighter.
0% 20% 40% 60% 6 Aug 11 Aug 17 Aug 44.6% 61.2% 28.0%
io.net — devices Akash — GPUs Akash — CPU cores

Akash's series begins 9 August, when both halves of its capacity pair start arriving from the same endpoint. Pairing its earlier CPU reading against the later denominator would have shown 15.9% instead of 20.6% — the same direction, a different level — so only the provenance-matched pair is plotted.


Supply and demand, separated

The same io.net window, with the two halves indexed to 6 August. The gap between the lines is the overcapacity: fleet growth that demand never met.

io.net — registered fleet vs. active devices
Indexed to 100 at 6 August 2026.
85 100 115 130 6 Aug 17 Aug 123.9 96.6
Registered devices Active devices

All measured pairs

Divergence is supply growth minus demand growth, in percentage points. Positive means capacity is arriving faster than it is being used.

NetworkResourceUtilisation SupplyDemandDivergenceDays
io.netDevices57.3% → 44.6%+23.93%−3.43%+27.3611.0
AkashCPU cores20.6% → 28.0%−6.82%+26.93%−33.768.0
AkashMemory14.9% → 24.7%−7.19%+53.73%−60.928.0
AkashGPUs31.6% → 61.2%−5.20%+83.33%−88.548.0

Four pairs, two networks. That is the honest extent of it: 181 networks were scanned for this window and only these two publish both halves often enough to measure change. Coverage is limited by what networks disclose, not by what we collect.


What this is not

  • Eleven days is one window, not a trend. The series begin at a clean collection epoch of 6 August 2026. Nothing here is a forecast, and none of it is a trading signal.
  • Daily means, not endpoints. These series swing about 25% intraday — Akash's active CPU ranged 3.50M–4.49M millicpu on 17 August alone. Anchoring on single first and last observations produced materially different numbers, so each end is the mean of a full day.
  • Only same-quantity pairs. A numerator and denominator are compared only when both resolve to the same resource in the same unit through the canonical ontology. Akash reports CPU in millicpu; it is converted to cores before any comparison.
  • io.net's GPU fill rate is excluded on purpose. Its hired/total GPU ratio reads about 96%, but that denominator is the already-active pool rather than registered supply, so it measures fill on hot capacity and hides every idle device. Devices active/registered is the comparable figure.
  • One io.net counter is frozen. Total inferences has read 491,583 for all twelve days. It is excluded rather than reported as flat, because a stalled counter and a genuinely flat one are not the same claim.