The Darkbloom network

Everything below is aggregate, public coordinator data — the whole fleet, not your box.

Active providers

1.1K

1.1K code-attested

Requests / 24h

4.92M

3.1K tokens per request

Tokens / 24h

8.94B

950.19M generated

Utilisation

8.0%

23.0K tok/s capacity · 0 queued

GPU cores

41.7K

fleet-wide

Memory

106.6K GB

unified, fleet-wide

Memory bandwidth

572.7K GB/s

aggregate

Power draw

185.8K W

~185.8 kW while serving

Hardware mix

Providers by chip family and tier.

M1 Max169

M5 Max149

M4 Max147

M4 Pro142

M3 Ultra136

M5 Pro78

M2 Max58

M4 Base52

M3 Max49

M5 Ultra44

M1 Ultra39

M2 Ultra29

Trust & state

Only hardware-trusted providers receive traffic on a hardware-trust network.

Trust level

hardware638

self_signed502

none7

Status

online881

serving260

untrusted6

Live demand by model

How many providers are actively serving each model right now, versus how many merely list it. The coordinator publishes no per-model request counts, so this is derived from each provider's reported current_model and sampled over time.

Model	Serving now	Share	Loaded	Listed	Contention	Trend
Qwen3.5-9B	116	45%	180	285	2.5x	collecting…
gpt-oss-20b	67	26%	153	458	6.8x	collecting…
gemma-4-26b-qat-4bit	49	19%	326	555	11.3x	collecting…
nvidia-nemotron-3.5-super	7	3%	20	30	4.3x	collecting…
EigenLabs/Qwen3.8-27B-4bit-mtp	6	2%	40	111	18.5x	collecting…
ternary-bonsai-2-27b	3	1%	13	162	54.0x	collecting…
glm-5.3-flash	3	1%	5	5	1.7x	collecting…
mimo-v2.6-flash-mopd	3	1%	3	14	4.7x	collecting…
qwen3.6-35b-a3b-vl-mtp-mxfp8	2	1%	93	523	261.5x	collecting…
EigenLabs/Ternary-Bonsai-2-9B-MLX-2bit	2	1%	9	87	43.5x	collecting…
nvidia-nemotron-3.5-lightning	1	0%	16	152	152.0x	collecting…
qwen3.5-35b-a3b	1	0%	16	237	237.0x	collecting…
gemma-4-26b-8bit	0	—	0	16	no demand	collecting…

Contention is providers listing the model divided by providers actually serving it — lower is better for a provider choosing what to run. A model with zero serving providers has no measurable demand, however few competitors it has.

Model supply and price

How many providers serve each model, beside what it pays per 1M tokens.

Model	Providers	Input	Output
gemma-4-26b-qat-4bit	555	$0.042	$0.220
qwen3.6-35b-a3b-vl-mtp-mxfp8	523	$0.050	$0.700
gpt-oss-20b	458	$0.018	$0.090
Qwen3.5-9B	285	$0.080	$0.130
qwen3.5-35b-a3b	237	$0.065	$0.750
ternary-bonsai-2-27b	162	$0.075	$0.500
nvidia-nemotron-3.5-lightning	152	$0.025	$0.120
EigenLabs/Qwen3.8-27B-4bit-mtp	111	$0.035	$2.200
EigenLabs/Ternary-Bonsai-2-9B-MLX-2bit	87	$0.080	$0.130
nvidia-nemotron-3.5-super	30	$0.060	$0.350
gemma-4-26b-8bit	16	$0.042	$0.220
mimo-v2.6-flash-mopd	14	$0.100	$0.280
glm-5.3-flashbottleneck	5	$0.065	$0.180

Where the compute is

Top regions by provider count.

California, US96

Tokyo, JP41

England, GB36

Texas, US35

New York, US27

Florida, US23

Ontario, CA20

Taipei City, TW17

Illinois, US16

Colorado, US14

Where the demand is

Top regions by requests in the last 24h.

The coordinator is reporting no request-location rows right now. It suppresses regions below a minimum request count for privacy, so this blanks out when per-region volume dips.

Supply and demand barely overlap

These are two separate rankings, and comparing them is the point.

Countries serving

Countries requesting

Serve but don't use

host providers, send no traffic

Use but don't serve

send traffic, host nothing

Source: the coordinator's public /v1/stats and /v1/pricing. Location data is aggregated by the coordinator and suppressed below a minimum provider count, so small regions are absent by design rather than missing.