Robot startup infrastructure cost: the real monthly ladder

A robotics startup can run its inference infrastructure for about 3 a month at the bottom of the ladder. At the top, the same five robots and the same workload cost about 2,300 a month. The rung decides the bill: community cloud card hours, on-demand pods, reserved contracts, an onboard computer bought once per robot, or a hosted robot route. Every rate below is named, dated, and worked in the open, as of October 2026. The pages ranking for this query price the office and the printers. This page prices the compute.

All amounts are United States dollars, written as bare numerals. Every price carries its source and read date. Last verified: 2026-10-26.

What the ranking pages get wrong

Search for this question and you meet business-plan checklists. One June 2026 guide tells you to budget a minimum of 5,000 per year for computers, software, and printers. The rest walk through company formation, office leases, and import rules. None of them names a GPU rate, a Jetson price, or a robot-hour cost. The word compute barely appears.

That fails the reader twice. First, compute is the budget line a robotics startup can actually shop for. Second, this market moved twice in 2026, and stale numbers poison the plan. NVIDIA raised every Jetson list price in July. Used 24 GB desktop cards rose about 6 percent in the last 30 days.

The honest frame is a buyer ladder. The buyer research behind this site's robot cost guides, finished in October 2026, maps four tiers:

Builders convert first. They spend today, in the open, on exactly the rungs this page prices. The tier-1 budget has its own guide, the hobbyist and builder compute budget. The tier-3 question, whether to build or buy compute at product scale, has its own guide.

What robot builders pay today

Before the ladder, here is the record of what builders actually paid in 2026, from the same buyer research.

The LeRobot community trains open policies on open weights, which cost nothing. Teams bring their own desktop cards or rent hours, and they buy arm parts. A leader and follower pair runs about 230. The LeRobot Humanoid biped platform runs about 2,500 in parts. Hugging Face sells a 9 per month PRO plan that includes 2 of inference credits. That plan is the usual entry into hosted compute for this community.

NVIDIA Isaac developers buy hardware and cloud hours. The Jetson AGX Thor developer kit launched on August 25 2025 at 3,499. It carries 128 GB of memory, the first Jetson able to hold a 70B-class model onboard. Its list price is 5,499 today. The DGX Spark desktop machine went from 3,999 to 4,699 in February 2026, per NVIDIA's own price-change notice, which blamed memory supply.

ROS 2 teams run open-source stacks with no vendor, and buy cloud GPU hours for training. Cloud GPU buyers in 2026 shop a mature market: RunPod, Lambda, Modal, Together, CoreWeave, and Vast.ai all sell by the card-hour.

One structural fact from that record shapes every budget below. Robots run policies at 10 to 50 steps per second, and paid spend concentrates in training and simulation bursts, not deployment. The deployment side is a small computer on or beside the robot. The bursts are the bill.

The compute ladder, rung by rung

Rung 1: community cloud card hours

The floor for sold GPU-hours today is a consumer card in someone else's gaming PC. Salad's public pricing page, read October 26 2026, lists a 24 GB RTX 3090 at 0.090 per hour on the lowest priority tier. The highest tier is 0.170, and 3060-class cards in the portal start from 0.015. A 3090 left on for all 730 hours of a month costs 65.70 at the lowest tier.

Our own account billing runs below the top tiers. In our own October 2026 training runs, a mid-priority 3060-class hour billed about 0.04, and a 3090-class hour about 0.041. That is about 30 a month for an always-on 3090, and close to nothing for the burst pattern a robot lab actually has.

The catch is labor. You get a raw container and an interruptible card. You keep the training server, the checkpoints, and the restarts yourself. For a three-person team that is usually fine, because training runs checkpoint and resume.

Rung 2: on-demand pods

The conventional clouds sell the same cards at higher rates with less work. Named parties, rates read October 26 2026:

RunPod rates are on their pricing page, Lambda's on theirs, Modal's on theirs, and the Hugging Face hardware table on theirs. An H100 pod at 3.99 per hour is about 2,900 a month left running. Almost nobody should leave one running: the same burst logic as rung 1 applies, and fine-tunes finish in hours.

One more rung-2 shape is credits. AWS Activate offers startups up to 200,000 in AWS credits, per their own page read today. Many robotics startups run their first year on credits like these. Budget the real rate for year two, because credits end.

Rung 3: reserved and dedicated

When training becomes continuous, contracts replace pods. Lambda's 1-Click Clusters are their reserved product, for terms of two weeks to a year. H100 systems price from 5.54 per GPU-hour at 256 GPUs to 6.16 at 16. HGX B200 systems run 8.87 to 9.86. All four figures are on their pricing page today. Salad sells Dedicated, reserved bare metal and virtual machines on contract built around the RTX PRO 6000 Blackwell, quoted by sales rather than listed.

Those figures are per GPU on multi-node systems with the fast interconnects that large training runs need. A startup reaches this rung when a fine-tune stops fitting in an overnight burst, not before.

Rung 4: the onboard computer

The deployment side is bought once per robot. NVIDIA's developer FAQ, read October 26 2026, lists the Jetson line after its July 2026 reprice:

The famous 3,499 figure in robot cost coverage is the Thor kit's launch price. The slot it left is now held by the AGX Orin kit at the same 3,499. For the bench, a used desktop card still wins on training per dollar. A used RTX 3060 12 GB runs about 295. A used 3090 runs about 1,093. Both figures are verified eBay sold data through October 24 2026. The split is the same one the tier-1 guide draws: desktop cards train, Jetsons deploy, and the Jetson versus cloud guide works through that trade.

Rung 5: hosted robot routes

The newest rung: somebody else runs the model, and you call it. Three shapes exist in October 2026.

First, per-hour rate cards. Abstract, a new entrant, publishes a public card in robot-hours on its own page, read today. The pi0.5 model from Physical Intelligence costs 0.85 an hour at 30 Hz. MolmoAct2 from Ai2 costs 0.99. Their own top model costs 4.05. Robotic Inference, another new entrant, sells one API across robot foundation models and publishes no public rates.

Second, token-metered robot models. Google's Gemini Robotics ER 2 lists input at 1.00 per million tokens and output at 5.00 through December 31 2026. Both double on January 1 2027, per the Gemini API pricing page read today. A token price cannot be converted to a robot-hour without a token count per call, and nobody publishes that count.

Third, per-call action routes, the unit this platform's robot lane uses. The rate side lives in the platform's model data, and this page quotes it only through the fence:

The honest state, in plain words: those three rows sit in the platform's model data with their rates set. They are not on the public list as of 26 October 2026, and this page quotes no live rate for them. The models index is the live catalog, the pi05 model page describes that lane, and the pricing page carries the listed lanes. A written record is not a launch.

The arithmetic those rates produce: at 2,160 calls per robot-hour, the pi0.5 class lands at 6.48 per robot-hour and the small class at 2.16. The full derivation, with the duty math, is the per-robot-hour derivation guide. Physical Intelligence opened its own API to beta on October 13 2026, with no public pricing, per its announcement. The units question, per call against per robot-hour, gets its own guide.

One worked month: three people, five arms

The canonical shape: a funded lab, three people, five SO-101 class arms. Two robots run demos eight hours a day on 22 workdays. The team runs twenty fine-tune runs a month, three card-hours each, and pays for one team plan for data and teleop.

The serving side needs almost no GPU. A robot-hour of policy inference takes about 84 seconds of GPU time on the published 39 millisecond per-call basis. The formulas:

robot hours        =  R x H x D
                      2 x 8 x 22  =  352

serving GPU hours  =  robot hours x 84 / 3600
                      352 x 84 / 3600  =  8.2

training hours     =  runs x hours per run
                      20 x 3  =  60

total card hours   =  8.2 + 60  =  about 68

Price those 68 card-hours per rung:

Add the data plane. Foxglove Pro is 20 a month with five devices included, then 20 per device per month, per their pricing page. A lab with ten connected devices pays 120. Viam starts free, then the first 5 a month plus usage, with cloud compute at 0.0025 per second, per their pricing page.

Add the one-time side. Five Orin Nano Super kits at 399 each is 1,995. Five AGX Thor kits at 5,499 is 27,495, and only a lab that needs a 70B-class model onboard should write that check. Evaluation episodes are their own cost, worked in the evaluation cost report.

The sensitivity line: the rung trades money for labor. At rung 1 you keep the servers yourself. At rung 5 somebody else does. Everything between is a negotiation over that one exchange.

The outsourced alternative: rental rates

A startup can also let someone else own the machines. Robots as a Service rents the whole robot, brain included, by the month. Named parties, dated:

The Digit filing also lists purchase at 200,000 plus 20,000 to deploy, and 36,000 a year in software and maintenance. It models a 1.1-year breakeven against 30.50 per hour of labor.

The structural fact: no rental contract anywhere publishes an inference line item. The brain rides inside the monthly fee. A startup selling inference sells into that fee, or into the tier-2 labs this page serves. There is no third line on an operator's invoice to land on.

How to pick your rung

• Month one: credits plus rung 1. Up to 200,000 in AWS Activate credits, or a 0.090 card-hour, covers a year of bursts.
• Buy the used 3090, about 1,093, when training runs weekly and the team wants the card in the room.
• Buy the onboard computer when robots leave the bench. 399 per robot covers light policies. 5,499 only for 70B-class onboard.
• Move to pods when queueing on community clouds costs more than the rate gap.
• Move to contracts only when a single fine-tune outgrows an overnight.
• Price hosted routes against your labor cost, not against raw card time. Labor is what they sell.

The tier above you

Scale changes the buyer. Figure runs its Helix policy fully onboard and buys compute for training, with reported multi-billion-dollar training clusters. 1X trains its models on cloud B200 systems and deploys at the edge. Big operators like Amazon and Ocado self-host their stacks and buy data planes at 20 to 250 per robot per month. None of them buys deployment inference from anyone, and every one of them buys training compute. That build-versus-buy question is the tier-3 guide.

Frequently asked questions

How much compute do I need to start a robotics company?

Less than the search results imply. A three-person lab with five arms runs on about 68 card-hours a month, training included. That is about 3 at community cloud rates, about 34 on a 24 GB pod, or about 170 to 272 on data-center cards. The bigger lines are people and robots, not GPUs.

Is it cheaper to rent or buy GPUs for a robotics lab?

Rent first. A used 3090 at about 1,093 breaks even against renting at 0.041 per hour only after about 26,600 card-hours, more than three years of nonstop use. Buy when training is weekly, when you want the card in the room, or when reproducing last month's run exactly matters more than the money.

What does robot inference cost per robot per month?

Count robot-hours first. At eight hours a day over 22 workdays, one robot runs 176 robot-hours. At the written-record rates above, hosted routes cost about 380 to 1,140 per robot per month. Raw card time at community rates costs well under 1, if you run the server yourself. The derivation guide shows the full arithmetic.

Can we run the whole lab on cloud credits?

For a while. AWS Activate offers up to 200,000 in credits, and accelerator programs stack more on top. Credits end, and the second year bills at full rates. Use credits to find your real workload, then size the rung with the tables above.

When does renting robots beat building our own stack?

When the robot and the task are both standard. Warehouse picking at 1,990 to 2,000 per robot per month, or a humanoid workcell at 8,500 plus 25,000 to deploy, buys the whole machine and its brain. When your robot or your task is the product, rental cannot serve you. The ladder above is your budget.

If you would rather buy inference by the call than run any of this yourself, the pricing page carries the listed lanes with live rates. You can open an account at signup. The tier-1 budget is the hobbyist and builder guide. The tier-3 question is the build-versus-buy guide.

Last verified: 2026-10-26