AWS Physical AI Toolchain Prices a Robot Training Run at $79, but Its Edge Step Is Still Planned

The most useful number in AWS's new robotics toolchain is not in the announcement. It is in the repository README, which estimates that a full GR00T training run costs about $79 on an ml.g5.12xlarge instance over 11 hours. Amazon Web Services and Nvidia staff introduced the toolchain on Oct. 7, and The Robot Report covered the launch on Oct. 8.

The launch posts talk about stages. The README talks about dollars per hour, which is the unit a robotics startup actually budgets in.

What the toolchain is

The code lives in aws-samples/sample-the-physical-ai-toolchain-on-aws under Apache-2.0. As read on Oct. 11 it showed 36 stars, 14 forks, 85 commits on main, two open issues and four open pull requests. The blog post does not name a licence; the repository does.

It is a set of independent Terraform modules (Foundation, Isaac Sim, Isaac Lab, GR00T and OSMO) that share configuration through SSM parameters. The blog lists five stages: raw teleoperation recordings landing in S3, synthetic world generation with Nvidia Cosmos, policy training with Isaac Lab and GR00T, validation in Isaac Sim, and deployment of optimised models to Nvidia Jetson hardware through AWS IoT Greengrass.

It is robot-agnostic by design. Users bring their own URDF robot description, teleoperation data and task definition, and the GR00T module ships with 27 UR3 pick-and-place episodes as an example.

What a run costs, from the README

The README's estimates are the first public price list for the workflow. They are AWS's own estimates and the README tells users to confirm current pricing.

Component	Estimate	Instance
GR00T training, smoke test	about $2 (15 min)	ml.g5.12xlarge
GR00T training, full	about $79 (11 hrs)	ml.g5.12xlarge
DreamZero fine-tune, 1,000 steps	about $93 (4 h 11 m)	ml.g7e.24xlarge
Isaac Lab RL training	about $10 to $30 (2 to 4 hrs)	ml.g5.xlarge

The hourly rates are where the argument sits. The GR00T run works out to roughly $7.20 an hour and the DreamZero fine-tune to roughly $22 an hour, both derived here from the README's totals. Cosmos 3 Predict is listed at about $37 an hour on a p5.48xlarge Capacity Block, and the README puts a full OSMO deployment at about $5 an hour for the EKS, RDS and ElastiCache underneath it.

1. Cosmos 3 Predict (p5.48xlarge)37 USD per hour
2. Cosmos Transfer 2.5 (g6e.12xlarge Spot)8 USD per hour
3. OSMO full deployment5 USD per hour
4. Isaac Sim workstation (g6e.4xlarge)1.86 USD per hour

Source: AWS toolchain README on GitHub, accessed 2026-10-11

The number that matters here is not the $79 for training but the $37 an hour for generating the synthetic data that feeds it. A team that generates for a day before it trains spends more than ten times the $79 training bill on Cosmos alone ($888 at 24 hours), on those estimates.

The edge gap

The last stage is the one that puts a policy on a robot, and the two documents disagree about it. The blog lists deployment to Jetson through IoT Greengrass as stage five. The README, as read on Oct. 11, says edge deployment is planned, not available, and names Jetson Thor and AGX as planned targets.

The Robot Report's account has the same tension. Sri Elaprolu, Director of Frontier AI Science and Engineering at AWS, told it the toolchain is "intentionally staying neutral to that final step." On hardware that is a design choice. On the last stage of a pipeline sold as end to end, it is a missing piece.

The blog is plainer about limits than the marketing around it. It says the toolchain provides "infrastructure patterns and architectural guidance, not finished robot behaviors" and that it "does not abstract away Physical AI complexity." It also says the cloud should not be the control loop for safety-relevant tasks.

That last point is the one the site's coverage of why a 99 percent per-action success rate finishes 37 percent of jobs bears on. Training cheaply does not move the on-robot failure rate.

Who is using the pieces

Elaprolu named four users of components, per The Robot Report: RLWRLD in South Korea, which AWS says built a specialised model for five-finger dexterity tasks; Telexistence in Japan, which he said has deployed more than 300 humanoids in convenience stores; Bedrock Robotics, which trains construction-robot models on AWS; and Amazon itself, which operates more than 1 million robots. The Telexistence figure comes from Elaprolu and is not independently verified in the article.

None of those deployments is evidence that this toolchain works, since the sample repository is days old. They are evidence that the components exist in production separately. Data collection is the part the toolchain leaves to the user, a cost that Mecka AI's $60 million round is aimed at.

Elaprolu also said AWS RoboMaker, shut down in 2025, was one component of its robotics stack and that this toolchain is not a direct replacement.

What the prerequisites say

A user needs an AWS account with GPU quota approved for SageMaker and EC2, an Nvidia NGC API key for container pulls, a Hugging Face token for model weights, and Terraform 1.5 or later on the production path. The README says containers build in AWS CodeBuild, so no local Docker is needed.

The GPU quota line is the practical gate, since the README's most expensive entry runs on a p5.48xlarge, which it labels as H100 hardware.

What would change my read is the edge module landing with a tagged release and a named Jetson target, or a customer reporting a real end-to-end run cost. Neither the blog nor the README gives a date for the edge work.