DOS

Unlock the full potential
of AI compute

DOS (Decart Optimization Stack) pushes AI hardware further.

Maximizing throughput, reducing latency, and improving compute efficiency across inference and training workloads.

  • 8xTPS

    TPS in TPU vs. baseline

  • 30xFPS

    FPS in TRN vs. baseline

  • 1500TPS

    TPS LLM workloads (gpt-oss 120b)

  • 100FPS

    FPS in live video generation

Demand for compute
is outpacing supply

Chip production can’t keep up with the demand AI is creating, and securing more compute isn’t always an option. DOS gets more out of the compute you already have access to, cutting idle time and raising utilization across GPUs, TPUs and Trainium, so you unlock more capacity without waiting on new hardware.

  • AI compute demand
  • Chip supply
  • The compute gap
Relative compute (illustrative)202320242025202620272028

Most AI chips never reach their full potential.
DOS changes that

The Decart Optimization Stack (DOS) is a vertically integrated inference and training platform for LLM, agentic, video and world model workloads. It sits between the model and the chip, tuning the kernels, compilers and runtime to the hardware so every core stays busy. The result is the same model running faster and at higher utilization on any accelerator, integrated into your existing infrastructure.

AI WorkloadsHardwareDecart Optimization Stack (DOS)DOSOptimizing infrastructure for inference and training.

What DOS improves

  • Latency

    Higher utilization cuts time to first token and time per frame, enabling real-time experiences and products that weren’t possible before.

  • Throughput

    By keeping every core busy, each chip serves more users and requests, so you scale without scaling your hardware.

  • Cost

    When compute you pay for is fully used, bringing real-time video generation down to $0.01 per second and turning AI from an expense you limit into a product you can grow.

Built to optimize across architectures

DOS is hardware-agnostic from the ground up, tuning each model to every chip it runs on so it can move between accelerators. Run where compute is cheapest, most available and best suited for your workload.

Nvidia GPU
  • Market average
  • With DOS
010020030040050060070080090010001100120013001400150016001700010K20K30K40K50K60K70KInteractivity (tok/s/user)Token throughput per chip (tok/s)04008001,2001,700010K20K30K40K50K60K70KInteractivity (tok/s/user)Token throughput per chip (tok/s)

No trade-off between
speed and scale

Every AI deployment balances two things: how many users a chip can serve, and how fast each one gets a response. DOS moves the entire curve. At every level of responsiveness, the same chip delivers more tokens, and it reaches speeds the market average can’t, with up to 3× faster responses per user than leading systems.

Who DOS is built for

Everyone running AI faces the same limit: compute. DOS helps you get more from it.

Model developers

Train and serve frontier models faster, at lower compute cost.

Chip manufacturers

Bring out your silicon’s full performance with optimized workloads and benchmarks.

Inference platforms

Serve more users per chip, with lower latency and higher margins.

Enterprises

Get more from every token, with faster responses and lower costs.

DOS makes live AI possible

Live AI has no room for delay: every frame must be generated before the next one is due, continuously, for as long as the user is there. DOS is what makes that possible in Decart’s own models. By keeping every core busy and every operation tuned to the chip, it brings generation fast enough to happen in real time, letting Lucy edit live video in under 40 ms at up to 100 FPS, and Oasis generate interactive worlds that respond to every action in under 200 ms.