Powering an Unlimited AI Economy That Works

Making the math work

When a centralized AI company sells you unlimited generation, it is arbitraging a fixed, brutal cost: enterprise GPUs rented from clouds, or bought outright. An NVIDIA H100 runs roughly $30,000; the cloud rental meter never stops. Every marginal render is real marginal cost against that bill, so a flat-rate plan is a bet that users won't use it — and creative tools are iteration machines. The heaviest users are the most passionate customers. The plan works right up until the product succeeds, and then the GPU bill eats it. That's the story of every retreat above.

Sogni's cost structure is inverted. We render on the Sogni Supernet — a network of independent GPU operators, most running consumer cards like the RTX 4090 and 5090. We don't rent their hardware at a fixed rate. They stake their idle compute on the decentralized network and earn upwards of 51% of network revenue. There is no fixed cloud bill to underrun. When network activity grows, so does the earnings pool, which recruits more supply — the same demand-attracts-supply dynamic that lets a rideshare network absorb demand spikes without owning cars.

Consumer GPUs are the right tool for this specific job

The obvious objection: consumer cards can't compete with data-center hardware. For some workloads that's true. For this one, the economics run the other way:

• Price per unit of hardware: an RTX 5090 lists around $2,000 versus roughly $30,000 for an H100 — 15× — while a 2025 peer-reviewed ACM study measured consumer RTX 4090 clusters delivering 62–78% of H100 throughput for LLM inference at up to 75% lower cost per token. (source)
• The workload fits the hardware: the most popular open image and video models — the ones the ComfyUI community actually runs — are powerfully efficient FP8, FP16, and BF16 variants engineered to fit 24–32GB of consumer VRAM. Serving them on consumer silicon delivers surprising quality and speed at a decisively lower cost per output. Is there a noticeable quality trade-off? We don't believe so — not enough to matter.

  Turbo isn't the whole catalog: distillation and numerical precision are separate choices. Sogni offers distilled and turbo models alongside full-model variants in FP8, FP16, and BF16. Our higher-precision FP16/BF16 options include Z-Image, Z-Image Turbo, Dark Beast Z-Image Turbo v9, and One Obsession.

  See our recent Krea 2 Turbo comparison rendered on Sogni, or try it yourself and be the judge.

• Idle supply: operators run hardware they already own, earning in hours it would otherwise sit dark. The Supernet holds substantial spare render capacity today — unlimited plans are, in effect, selling compute that currently goes unused.

The open-model wave makes the timing

The second thing the retreats have in common: they were built on closed, per-render-licensed or self-trained frontier models — the most expensive possible substrate for an unlimited promise. Meanwhile, open-weight models crossed a quality threshold that most of the market hasn't priced in. Lightricks' LTX-2.3 — open weights, synchronized audio+video, up to 4K — was downloaded more than 2.1 million times on Hugging Face last month (source). Alibaba's Z-Image Turbo: nearly one million (source). Krea 2 went open-weights on June 22 and was live on Sogni within days (source).

A cleaner example is image generation on the same Sogni pricing reference. One 1024×1024 GPT Image 2 render at medium quality with no references is estimated at $0.0685 (13.70 Spark), while a 1024px Krea 2 Turbo render at 8 steps is estimated at $0.0122 (2.43 Spark) — making GPT Image 2 about 5.6× the cost for these example jobs (source). These are not interchangeable models, and output quality is subjective; the point is not that every job should use Krea 2 Turbo. It is that open models can make high-volume iteration economically viable, while external frontier models remain available when their specific capabilities justify the premium.

That's the design of Sogni Unlimited: $20/month, fair-use unlimited across the open models we host — image, video, music, and LLM chat — with the three frontier vendor models (GPT Image 2, Seedance 2.0, HappyHorse 1.1) as metered add-ons at a member discount.

"Unlimited with fair use" — what we mean, precisely

Skepticism is warranted; the word has been abused. So here's the mechanics, published rather than buried: covered renders spend no credits — there is no monthly meter, nothing you buy and burn down. Fair use governs scheduling: each tier has published concurrency limits, exceptionally heavy sustained use queues behind other jobs at peak, and heavy-usage throttles reset each UTC day. Queued work gets done. The difference between this and "unlimited" plans that quietly suspend their heaviest users is that we're telling you the mechanics up front — they're the reason the promise is sustainable.

Can demand outrun the network? Three answers, in order: today there is substantial headroom; the worker earnings pool recruits supply as demand grows (and we can provision additional support nodes through rental partners like Salad profitably, because pool economics cover it); and the priority ladder protects everyone — paid pay-as-you-go jobs first, then Unlimited by tier, then free-tier jobs.

The bet

Every company that retreated from unlimited was making the same wager against its own users: please don't use this too much. We're making the opposite wager: that if compute costs scale with revenue instead of against it, heavy use is the business model, not the threat to it. One year and 158 million renders into the Supernet's public life, that's a bet we're structurally built to make — and the people who own the GPUs earn upwards of 51% of network revenue to help us win it.

Iteration is where creative work actually happens. It shouldn't have a taxi meter running.

Sogni Unlimited is $20/month or $199/year, with a 3-day trial for eligible new subscribers at sogni.ai/unlimited. If you have a GPU, it can join the network.