Stop selling GPUs. Start selling inference.
GPU hours compete on price alone; an inference API is a product. ARK Core turns your fleet, mixed card generations included, into an OpenAI- and Anthropic-compatible token service. You license it per GPU and keep the margin on every token. There is no per-token fee to ARK.
GPU clouds are adding managed inference on top of their GPUs. DigitalOcean reports about $22 million of annual recurring revenue per megawatt, about double the $9–11 million it estimates for its peers.
InfoWorld, 15 September 2026 (CoreWeave Inference, Nebius Token Factory, Crusoe Managed Inference, Nscale Serverless Inference). DigitalOcean CFO at Citi's Global TMT Conference, 10 September 2026 (blended Q2 2026 figure).
The tokens per second of vLLM at 64 users, on the same two RTX 5090, with the first token 7.0× sooner. vLLM is faster for a single user.
Read moreQwen3.6-35B-A3B on four RTX 4070 Ti SUPER, 16 GB cards, at 48 concurrent requests on ARK Cloud.
Read morePer million tokens on that same deployment with 16 or more users, about a fifth of the energy of serving one request at a time.
Read moreWith a second copy of each part of the model, a failed server doesn't take the service down. Requests in flight are replayed on another path, and servers join or leave without a restart.
Read moreWhy GPU hours are a hard business.
Hourly prices you don't set
Median H100 rental prices range from $1.40 to $4.27 an hour across marketplaces, and CME has announced GPU rental futures, set to list from 5 October 2026. Hourly rental is a commodity priced by others.
AI Mining Co. data for 13 marketplaces, via TechCrunch, May 2026. CME Group, 11 August 2026.
Idle GPUs eat the margin
A card rented to one customer waits between jobs. At 30% utilisation, every token costs about three times as much to serve.
CloudZero, inference cost guide, 2026.
Power is the ceiling
When megawatts, not GPUs, limit growth, the business is measured in tokens per watt.
An API is a product to build
Serving, routing, metering and SDK compatibility need a dedicated team before the first token is sold.
More sellable tokens from every card, and every watt.
Every request in flight shares each GPU's decoding step. More customers on a card means more output, not slower service.
Many customers per GPU
ARK 2.0 batches every request in flight. Every model in our release benchmark served 64 concurrent streams on four RTX 5090, with 9.3× the total throughput of ARK 1.0 (median of 15 models).
Ahead of vLLM where customers share GPUs
On the same two RTX 5090, with the same model, prompts and third-party benchmark and both engines at default settings, vLLM is faster for a single user. With 4,000-token prompts ARK serves more from 8 users upward. At 64 users it generates 2.5× the tokens per second, gets the first token out 7.0× sooner, and neither engine failed a request.
Energy that falls with traffic
Energy per token drops 1.8 to 4.7 times once 16 or more users share a deployment. On ARK Cloud, a million tokens take 21 Wh on Qwen3.6-35B-A3B.
Every generation, one deployment
ARK runs on every NVIDIA generation from Turing to Blackwell, from H100, L40S and A100 to GeForce cards, and splits a model across mixed cards and machines over standard Ethernet. In production, Qwen3.5-122B-A10B runs across three machines and two GPU generations: four RTX 5090 plus an RTX 4060 Ti for its image encoder. An 8-bit attention cache doubles the context each card holds.
Rides out failures, reshapes live
Hold a second copy of any stretch of layers, on any GPU, and a failed server doesn't take the model down: its requests are replayed on another path. Machines join or leave while customers are being served, with no restart, and ARK sends more traffic to whichever copy is faster and less busy.
Metering built in
Per-request token usage, with cached and reasoning tokens reported separately, goes to your billing system through HTTP hooks. Exact token counts are available before a request runs.
You own the customer. ARK is the engine underneath.
Your brand, your price list, your billing and your regions. ARK Core is licensed per GPU, with no per-token fee. GPUs financed by the hour need to earn every hour, and ARK is built to keep them busy.
Your customers keep their OpenAI and Anthropic SDKs: streaming with usage, tool calling, reasoning, vision, embeddings and transcription. Models appear in the API as soon as their servers come online, with no restart.
From fleet to API.
ARK Core Your service
The platform and its models on your fleet, licensed per GPU. Your team runs it as your own API. Our engineers help plan GPU and model placement for your traffic.
ARK CoreTest before you buildARK Cloud
Point your test clients at ARK Cloud to check the API and models before your fleet is ready.
ARK CloudSend us your GPU inventory.
List your cards and the models you want to sell. We benchmark ARK on comparable hardware, or on yours, and send you throughput, time to first token and energy per token at the concurrency you plan to sell.