What an Nvidia GPU earns on Sogni in Sept 2026

A forecast for a month still unfolding

We’re optimistic about another record month. The pace of video creation is already rewriting the recent daily high, and demand is growing well beyond a single model. But a month that is still in progress cannot be reported as a finished result. Nor does a busier network guarantee that every GPU will earn more.

For an RTX 5090, the scenarios below work out to about $413 for September on the currently funded pool, $503 if that pool grows another 25%, or $594 with 50% further growth. An RTX PRO 6000 comes out at roughly $813, $957 or $1,102. Each assumes a card stays available around the clock, participates in subscription work and keeps the same share of work per connected hour seen in our sample.

The demand behind the excitement

On September 9, Sogni completed 42,521 videos, up 92% from September 1. MiniMax H3 and its variants accounted for 35,438 of those videos, more than twice the September 1 total. Across the complete days from August 28 through September 9, each day’s video total exceeded the previous day’s.

That is a remarkable recent run. It is also a precisely bounded observation, not a claim that growth will continue every day. Daily new user signups reached 6,114 on September 9, up 43% from September 1 and the highest daily count in this window. That is the number of new accounts created on that single day, not a running total for the month. Earlier days fluctuated; we’re excited about the direction without pretending every metric moves in a straight line.

The broader picture matters too. Worker-rendered image completions rose 80% over that same comparison. And demand extends to third-party partner models: Wan 3 Enhanced — now listed as Wan 3 Uncensored — completed 19,037 videos in September 3–9, reaching 3,401 on September 9, the highest daily total in the 13-day window we checked. Its September 1–9 increase was 22%, so growth is broad but differs substantially by model.

These partner-hosted jobs show the breadth of activity on Sogni. They require separate pay-as-you-go spending and are not rendered by the GPUs in this worker sample, so we do not add their sales to the earnings forecast.

Our interpretation is that the combination of fair-use subscriptions and compelling video tools is encouraging more creation. The numbers establish the increase; they do not isolate how much was caused by Unlimited, H3, new signups or existing creators using Sogni more.

The sample in a wider worker network

At the September 10 snapshot, September’s Fast Worker subscription leaderboard listed 287 worker NFTs. The account behind this sample has 140 listed Fast Worker NFTs in total, including the 133 ComfyUI workers analyzed here; the other 147 NFTs belong to other accounts. These are subscription-leaderboard participants, not a count of every worker on Sogni.

The sampled account’s Fast Worker NFTs represented about 57.6% of the September pool’s recorded work at capture, down from 58.2% at the September 9 snapshot. The 133-worker ComfyUI subset represented 57.3%, down from 57.8%. Participation has broadened a little — the previous snapshot listed 283 NFTs — but this is still a large, carefully managed contributor. Worker counts and share of completed work measure different things.

September earnings, under three pool assumptions

At the September 10 snapshot, the estimated worker subscription pool stood at $43,299, representing the worker’s 51% share of $84,900 in net subscription revenue allocated to September. This already includes the September coverage funded by subscriptions collected so far. It is not a partial-month pot to multiply by the number of days left.

We combine that pool with nine complete days of worker activity, then ask what one representative card might receive over September’s 720 hours. Pay-as-you-go work continues at the observed rate; only the subscription pool changes between columns. The larger-pool scenarios use approximately $54,124 and $64,948.

The sample is the fleet of one carefully managed account, not a random cross-section of the entire Supernet. It includes 21 RTX 5080s, 50 RTX 4090s, 37 RTX 5090s, 23 RTX PRO 6000s and two RTX 3090s, classified by their hardware at capture. The two-card 3090 result is especially tentative.

Observed uptime is included because it makes the extrapolation visible: these cards were not all online 24/7. A forecast for continuous availability asks more of the lower-uptime groups. It assumes their observed work rate transfers to the hours they were offline; it does not establish that those hours would have been equally productive.

Two revenue lines, one worker

Pay-as-you-go rendering pays for individual jobs. We use the worker’s 51% share of the recorded render value, exclude subscription work from this line, and extend the observed dollars per connected hour to 720 hours. For the RTX 5090 sample, that contributes about $50 in each scenario.

Unlimited subscription work earns a share of the monthly pool based on eligible render work. On the current pool, that contributes about $363 to the continuously available RTX 5090 scenario; with 25% pool growth it becomes about $454. Together with the pay-as-you-go line, that produces the roughly $503 middle scenario.

The subscription revenue-share rules govern eligibility and settlement. Open-month figures remain estimates until reconciliation. Free signups and trials are not paid subscription revenue, and a creator’s displayed retail value of Unlimited usage is not cash owed to a worker.

When you can claim, and which work you host

Pay-as-you-go earnings can be claimed the same day once the applicable earnings threshold is met. Completed Spark- and SOGNI-paid jobs accrue regular worker earnings. Once those earnings are available and meet the current claim minimum shown in your dashboard, you can claim them to your Sogni account balance without waiting for a monthly subscription pool to close. Claim processing and any subsequent wallet withdrawal are separate steps.

Unlimited subscription payouts are planned for approximately 45 days after the end of the UTC calendar month in which the work was performed. The delay allows settlement and reconciliation across card payments, Stripe, Google Play, Apple in-app purchases and related providers, whose reporting and settlement periods differ. For September work, that means roughly mid-November, subject to finalization and the current terms. It is not 45 days from each job, and an open-month estimate is not yet a claimable balance.

Subscription Earnings participation is currently optional. Review the terms and opt in through the Subscription Earnings dashboard to make eligible workers on your account available for Unlimited jobs alongside regular pay-as-you-go work. You can also leave participation disabled and host only pay-as-you-go jobs. The combined forecasts above assume participation in both streams; the observed pay-as-you-go component is not a forecast of what a pay-as-you-go-only setup would earn.

See the current worker revenue-share rules for participation, reconciliation and payout terms, and your dashboard for the claim thresholds that apply to your account.

Why the GPU — and the work it receives — matters

The RTX PRO 6000 group leads this sample, but the table is not a pure hardware speed test. Earnings also reflect model eligibility, workflow mix, scheduling, uptime and competing capacity. A GPU’s memory can determine which jobs it can accept, while newer, faster workflows can change that opportunity during the month.

One example: FastH3 Turbo’s current base workflows can run on 24 GB-class cards with at least 23 GB of usable VRAM; custom-LoRA jobs require at least 32 GB. Workflow support keeps evolving, which can change the mix of work available to each card. Check the current worker requirements and catalogue before choosing hardware.

What could change the result?

• Paid subscription growth. More paid September coverage can increase the pool. Slower growth, refunds or reconciliation adjustments can pull the result the other way.
• Demand relative to available workers. Demand changes by the day, week and month. The number of other worker nodes online relative to that demand affects each GPU’s share. Increased video counts do not translate one-for-one into increased dollars per card.
• Your actual availability and workload. Keeping machines online and sharing popular, supported models matters. Offline time, network reliability, model downloads, speed and workflow eligibility all affect the work a card receives.
• Cooling and equipment care. Keeping equipment well cooled helps it sustain performance and availability. Heat-related slowdowns or downtime can reduce the work it completes.
• Your costs. Electricity, cooling, hosting, bandwidth, the worker licence and hardware depreciation come out of these figures. Token-denominated earnings also carry conversion-price risk. Revenue is not profit.

We’re excited because the demand is real and the opportunity is expanding. We’re keeping the forecast conditional because the market is moving on both sides: more people creating, more workers joining, and better models arriving. Another record month is our outlook, not a result we can declare on September 10.

Sources and measurement window

Worker estimates use the public subscription leaderboard and the daily worker analytics behind the Sogni dashboard. We selected the 133 ComfyUI workers in the Nosana.network fleet visible at capture, then summed their nine complete daily records. Subscription shares use eligible work from that same complete-day window for both the sample and the whole network. Pay-as-you-go dollars are calculated per day before being summed, with the worker share applied once.

For each GPU class, we divide observed pay-as-you-go income by connected hours and extend that rate to 720 hours. For subscriptions, we allocate the full-month pool using the class’s observed network share, divide by the number of cards, then normalize for observed uptime. This is an illustrative rate for one representative card while the rest of the market holds steady; it cannot be multiplied across a whole expanding fleet without recalculating network share.

Demand and signup counts come from Sogni’s production telemetry aggregated in Datadog US3. Videos count successfully completed outputs, not projects, requests or seconds of footage. Signups count successful account-registration events, not paid subscribers. The comparison window is August 28–September 9; it does not establish all-time monthly records.

Pool captured September 10, 2026 at 1:57 p.m. PT (20:57 UTC). Worker activity covers September 1–9 UTC: August 31 at 5 p.m. through September 9 at 5 p.m. PT. UTC dates therefore differ from local dates at the window boundaries. September 10’s incomplete activity is excluded from the earnings and demand calculations. The network-context paragraph uses the live month-to-date leaderboard at capture, including the partial day; those shares are separate from the complete-day shares used in the forecast. Dollar figures are rounded; all assumptions are held fixed at this snapshot.

• September subscription-worker leaderboard · live values will change after this snapshot
• Worker dashboard and daily analytics
• Unlimited plans and worker revenue share
• Fast Worker setup and requirements

Update, September 22: three weeks in

Data through September 21 UTC · Pool captured September 21, 2026 at 11:49 p.m. PT

With 21 of September’s 30 days complete, we reran this forecast using the same method on the same account’s workers. Two parts of the outlook have played out: demand kept climbing, and the subscription pool grew by about the 50% our most optimistic case assumed. The per-card estimates moved much less, because many more workers joined to share that work.

The day-over-day streak we described above ended on September 10. Video completions held between 38,220 and 45,528 a day from September 10 through 19, then set new highs of 47,000 on September 20 and 50,640 on September 21. MiniMax H3 and worker-rendered images also reached their highest daily totals in this window on September 21. Wan 3 Uncensored more than doubled from September 1; as before, those partner-hosted jobs are not counted in the earnings figures. Signups moved the other way: after the September 9 peak they held near 6,000 a day through September 15, then settled between 4,270 and 4,944 a day from September 16 to 21.

The worker subscription pool now stands at $59,963, the worker’s 51% share of $117,574 in net subscription revenue allocated to September. That is 38% higher than on September 10. The pool’s own forecast, based on the past week of subscription payments, projects $64,488 by the end of the month. That is 49% above the September 10 pool, almost exactly the $64,948 our +50% case used.

The network grew too. The September Fast Worker leaderboard now lists 345 worker NFTs, up from 287 on September 10. Eligible subscription work across the network averaged about 3.0 million render Spark a day over September 1–9 and about 4.3 million over September 10–21. The sampled account’s ComfyUI workers did 52.0% of that work across the full 21 days, down from 57.7% over the first nine.

The pool reached our most optimistic case, but the assumption behind that column, that each card keeps the same share of the work, did not hold. With more workers splitting more work, a continuously available card’s estimate lands close to the original current-pool column. It is 5–6% higher for the RTX 5080, RTX 4090 and RTX PRO 6000. It is 6% lower for the RTX 5090, which at $388 on the current pool now sits below our lowest September 10 scenario. Pay-as-you-go work also contributes less per connected hour than in early September: about $38 a month for the RTX 5090, down from $50.

The sample grew with the account’s fleet, from 133 to 143 ComfyUI workers: 20 RTX 5080s, 57 RTX 4090s, 40 RTX 5090s, 24 RTX PRO 6000s and two RTX 3090s. The RTX 5080 and RTX PRO 6000 groups were online far more than in the first nine days (63% and 57% uptime, up from 37% and 44%), so their full-time estimates now rely less on extrapolation. The earnings disclaimer above applies to every figure here.

One for the books. When September closes, we’ll follow up with the full month: final daily demand, the worker pool as it stands at month end and what this sample actually earned, rather than a projection. Subscription figures stay estimates until reconciliation, and September’s subscription payouts are planned for roughly mid-November. The final numbers are below.

Worker activity for this update covers September 1–21 UTC: August 31 at 5 p.m. through September 21 at 5 p.m. PT. The pool and leaderboard were captured September 22 at 06:49 UTC (September 21 at 11:49 p.m. PT). The month-end pool forecast is the subscription earnings endpoint’s own projection from recent coverage payments, not an editorial assumption. Leaderboard counts include the partial current day; the work share uses complete days only. September 1 demand values are from the original snapshot.

Final numbers: September, closed

Data through September 30 UTC · Pool captured October 5, 2026 at 10:01 p.m. PT

September is complete, so here is the follow-up we promised: the full month’s demand, the worker pool as it stands at month end, and what this sample’s cards actually accrued rather than a projection. The short version: the subscription pool finished at $64,693, within 0.4% of the +50% case we set out on September 10. Demand kept climbing through the last day of the month. A continuously available card landed near the original current-pool column rather than the +50% column, because 371 worker NFTs shared September’s pool instead of 287.

Video completions averaged 55,616 a day over September 22–30, against 30,961 a day over the first nine days, and set the month’s high of 59,466 on September 29. MiniMax H3 peaked the same day at 49,043. Worker-rendered images passed 240,000 on September 27 and 28. Wan 3 Uncensored, the partner-hosted model that is not part of the earnings figures, rose another 37% from its September 21 level to a 7,777 peak on September 25. September 30 finished 161% above September 1 for video.

Signups did not follow. They held between 4,092 and 4,392 a day from September 22 to 27, then fell to 3,578 on September 28, 2,003 on September 29 and 1,761 on September 30, the lowest day of the month. September’s total was 137,463 new accounts. Signups count successful registrations, not paid subscribers; paid subscription coverage, which funds the pool, still grew 8% between the September 22 snapshot and month end.

The worker subscription pool closed September at $64,693, the worker’s 51% share of $126,849 in net subscription revenue allocated to the month. That is 49% above the $43,299 pool on September 10, 8% above the September 22 snapshot, and 0.3% above the $64,488 the pool’s own forecast projected on September 22. Our +50% case assumed $64,948; the pool landed $255 short of it. The figure stays an estimate until reconciliation, and September’s subscription payouts are planned for roughly mid-November.

The network kept growing to the end. September’s Fast Worker leaderboard closed with 371 worker NFTs, up from 287 on September 10 and 345 on September 22. Eligible subscription work across the network averaged about 3.0 million render Spark a day over September 1–9, 4.3 million over September 10–21 and 5.5 million over September 22–30, for 128.8 million in the month. The sampled account’s 143 ComfyUI workers did 49.6% of it, down from 57.7% over the first nine days and 52.0% over the first 21.

The pool reached our most optimistic case; the per-card result did not, and the reason is the share of work. Each card’s slice of a pool that grew 49% shrank as the network grew, so a continuously available card finished within 3% of the September 10 current-pool column for the RTX 5080 (−0.3%), RTX 4090 (+2%) and RTX PRO 6000 (+2%). The RTX 5090 finished 12% below it at $364, and the two-card RTX 3090 group 19% below at $113. Against the +50% column, every class came in 25–44% lower. Pay-as-you-go work also kept thinning per connected hour: it contributed about $32 a month to a continuously available RTX 5090, down from $50 in early September, and $181 to an RTX PRO 6000, down from $236.

The last column is the one no forecast can give: what the sample’s cards actually accrued, including every hour they were offline. Across all 143 workers, the sample accrued an estimated $32,106 from the September pool and $6,698 in pay-as-you-go earnings, $38,804 in total, about $271 per card at 73% average uptime. The gap between the two final columns is the cost of downtime. The RTX PRO 6000 group, online 66% of the month, averaged $549 against $830 for a card that never went offline; the RTX 5090 group, online 84% of the month, averaged $307 against $364. The earnings disclaimer above applies to every figure here, and open-month figures remain estimates until reconciliation.

Worker activity for this section covers all of September 1–30 UTC: August 31 at 5 p.m. through September 30 at 5 p.m. PT. The pool and leaderboard were captured October 6 at 05:01 UTC (October 5 at 10:01 p.m. PT) with full-month revenue and usage coverage; the pool is not yet finalized. Each card is classified by the GPU it ran during September, not by its hardware at capture: nine cards moved from RTX 5080 to RTX 4090 deployments on October 1 and are counted as RTX 5080s here, so the class counts match the September 22 update. Subscription amounts use each worker NFT’s pool-eligible work for the whole month, including work the public leaderboard lists under the account rather than the NFT. Demand values for September 1–21 come from the two earlier snapshots. Signups count successful registrations, not paid subscribers.