The hobbyist robot compute budget: what a cheap robot brain really costs

A real hobby robot budget has three viable shapes in October 2026. Use the 12 GB card you already own, and the light policies train on it today. Buy a used 12 GB desktop card, about 295 in current eBay sold data, and add about 230 in parts for the leader and follower arm pair. Or rent by the hour from 0.015 per GPU-hour, which lands near 29 per month even if you ran a card nonstop. The famous 5,499 robot brain is real, but it is not where hobbyists start.

All amounts in this guide are US dollar amounts. Every price carries its source and read date, because this market moved twice in 2026 and stale numbers are the main way hobbyists overpay.

What you are actually paying for

Robot hobby spend is not a utility bill. It is a series of training bursts.

A robot policy is a small model that turns camera images into motor commands. It runs at 10 to 50 steps per second on whatever computer you give it. That part is cheap. The expensive part is training, and training comes in short bursts: you record an hour of demonstrations, train for an hour or two, test on the arm, record more data, and train again.

The Hugging Face team that maintains LeRobot publishes a hardware guide with concrete numbers. On a common hobby dataset, about 50 episodes and 45,000 frames, an ACT policy trains in about 30 to 60 minutes on a single RTX 3090 or 4090. A diffusion policy takes about 2 to 4 hours on the same card. You can read the full table in the LeRobot hardware guide.

That single fact reorganizes the whole budget. One rented GPU-hour is a complete training run for the light policies. Two to four hours covers a diffusion policy. You do not need a supercomputer. You need the right memory size, available at the moment you press train.

At hobby scale, robot compute is a training budget, not a utility bill. Plan for bursts, and every path below gets cheaper.

Match the policy to the memory before you spend

Memory, not speed, is the first filter. The LeRobot hardware guide groups policies by peak video memory at batch size 8:

Three practical readings fall out of that table.

First, the entry class is genuinely small. ACT, the policy most hobbyists train first, peaks at 2 to 6 GB. The guide names a laptop RTX 3060 as a starter card for it. An ACT policy fits in the same memory a small language model needs, which is why the used PC parts market serves robot learning better than any robot store.

Second, memory scales roughly linearly with batch size. If a card is close, you halve the batch and halve it again. Training slows, but it fits. The guide also notes the default optimizer adds 30 to 100 percent on top of a plain forward and backward pass, so leave headroom rather than cutting it exact.

Third, 24 GB is the real wall. The large VLA family wants 24 to 40 GB, and the guide calls 24 GB tight even at batch 1. That wall is why the used 3090 exists in this budget at all.

One more line from the guide deserves quoting, because people ask anyway. For CPU-only machines, its advice reads: "Don't train. Use Colab or rent a GPU." Free cloud notebooks or a rented hour beat a CPU that would need a day per run.

Path 1: the card you already own, 0 new dollars

Start here, seriously. If your desktop or laptop has an NVIDIA card with 12 GB of memory, you can train the light class today, tonight, for nothing.

A 3060-class card runs ACT end to end on a hobby dataset. If you own a Mac instead, the hardware guide lists an Apple M-series option: ACT at batch 4 in about 6 to 14 hours. Slow, but free, and it is the same policy.

The point of this path is to find out whether you like the hobby before the hobby costs anything. Record demonstrations from a cheap arm, train ACT on hardware you own, and watch it succeed or fail. Every dollar in the later paths is then spent on a want, not a guess.

Path 2: a used desktop card, the value path

If path 1 is closed, buy used. Sold prices from eBay, tracked by getpcparts through October 24 2026:

Two things to know about this market in October 2026.

The cheap-3090 era ended. A March 2026 roundup on compute-market.com still lists the used 3090 at 800 to 999. That window closed. October sold data shows about 1,093, and the tracker BestValueGPU puts the card 23 percent above its own 12-month average of 978. Asking prices on 3060 cards run higher than sold prices too, 339 to 430 across trackers this month. Anchor on sold data, not listings.

The 12 GB floor matters more than the brand. Diffusion peaks at 8 to 14 GB, so an 8 GB card already forces painful batch cuts, and the small VLA class starts at 10 GB. A 12 GB card clears everything except the large VLA family. The used 3090 at about 1,093 is the only common used card that reaches every open policy family, and it still costs about one fifth of a Jetson Thor kit.

Buying used has its own checklist. Ask whether the card was used for mining, test it on arrival, and check your power supply: a 3090 draws up to 350 watts on its own. None of that changes the math much. It just means buying from a source with returns.

Path 3: Jetson kits, priced after the July 2026 reprice

NVIDIA's Jetson line is what people mean when they say "robot brain": a power-sipping module that bolts onto the robot itself. In July 2026, NVIDIA raised every Jetson price without an announcement, as CNX-Software documented on July 22 2026. Current list prices from NVIDIA's developer FAQ, read October 26 2026:

The entry kit went from 249 to 399, and Thor went from 3,499 to 5,499, in the same silent update. Blame memory costs, which the same coverage links to the rises.

What a Jetson buys you is the robot-side computer: low power draw, no desktop tower, and a deployment target that matches what you trained for. The Orin Nano Super kit at 399 is enough to run light policies on the robot. At the top, the Thor kit carries 128 GB of memory, enough to hold a 70B-class model onboard, which is a research capability rather than a hobby one.

What a Jetson does not buy you is training throughput per dollar. A 399 Orin Nano Super kit trains the light class slowly on 8 GB. A used 295 desktop 3060 trains the same class faster, on more memory, for less money. The honest split: desktop cards train, Jetsons deploy, and the Jetson versus cloud guide works through that trade in depth.

One more desktop appliance belongs in this price band: NVIDIA's DGX Spark, a small desktop AI machine, moved from 3,999 to 4,699 in February 2026 per NVIDIA's own announcement, again citing memory supply. It appears in hobby forums as a trainer. Used cards still win on dollars per training hour.

Path 4: rent by the hour, the no-commitment path

You can rent the exact card class a training run needs, by the hour, from distributed GPU marketplaces. This path owns nothing, ships nothing, and scales from zero.

Salad's Community Cloud, on its own site read October 26 2026, advertises GPU-hour rates from 0.015, on a pool that runs from RTX 3060 class through RTX 5090. 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. Run the arithmetic at 730 hours a month, the always-on figure: about 29 per month for the 3060, about 30 for the 3090. Nobody actually trains 730 hours a month at hobby scale. In practice the bursts cost about 1 per training day in our runs, because a burst is one to four hours.

Vast.ai, on its pricing page read October 26 2026, sells the same idea with different tiers: interruptible instances at "50%+ cheaper," aimed squarely at batch training, alongside on-demand instances billed per second. It publishes no static per-card prices, so this guide quotes none for it.

Interruptible is the catch, and it is smaller than it sounds. An interruptible instance can be preempted, so it suits jobs that checkpoint and resume. Training runs qualify: they finish in an hour or two and restart from a checkpoint. Closed-loop robot control does not qualify. Salad's own page says the network is not a fit below 10 milliseconds of latency, and your arm wants single-digit milliseconds. Rent for training bursts. Keep the robot side local.

Renting turns the hobby's biggest purchase into about one dollar per training day, and the interruption risk lands on a job that finishes in an hour anyway.

The rest of the bill: arm, cameras, robot

Compute is half the budget. The other half is the robot.

The standard hobby path is a SO-101 class arm, the open-source design used with LeRobot. A parts tracker that follows the build, dreamingofrobots.com, estimates about 230 in parts for a leader plus follower pair, or about 122 for one follower arm on its own. You need the pair: the leader arm records your demonstrations, and the follower arm learns them. Assembled kits exist if you would rather skip the build: one European store listed an assembled kit at about 388 EUR at the time of reading. Expect to add cameras, a power supply, cables, and shipping. Plan on another 50 to 100 for that layer. The budget rows below use the 230 pair figure.

If you outgrow arms entirely, the LeRobot Humanoid biped platform, announced by Hugging Face on May 21 2026, runs about 2,500 in parts. That is the ceiling of tier-1 hobby spending, and it is still less than half a Thor kit.

Three worked budgets

Putting the paths together, with every number sourced above:

The monthly column is the honest one. Hobby robot months cost 10 to 30 in compute when you rent, or the price of electricity when you own. The 5,499 and 4,699 machines are not in this table because they are not tier-1 purchases. They are lab purchases.

Decision rules

• You own any 12 GB NVIDIA card: train ACT tonight. Spend nothing yet.
• You own nothing, and only the light class interests you: used 3060 12 GB, about 295 sold.
• You want SmolVLA, or pi0-class research at batch 1: used 3090, about 1,093 sold.
• You want the computer living on the robot: Orin Nano Super kit at 399 for light policies. Thor at 5,499 only if you specifically need a 70B-class model onboard.
• You train a few times a month and hate idle hardware: rent. About 1 per burst day, interruptible tier, checkpoint on.
• You have only a CPU: use a free cloud notebook tier or a rented hour. Do not train locally.

When buying beats renting

The raw arithmetic is blunt. A 295 card at about 0.04 per rented hour breaks even at 7,375 hours, several years of hobby bursts. On hours alone, renting always wins at hobby scale.

Buying still makes sense for three reasons. Availability: your burst lands at 9pm on a Sunday and the card is in the next room. Consistency: the same card reproduces last month's result exactly. And interruptibility: nobody preempts hardware you own.

The usual pattern among builders is rent first, buy the class you keep renting. Rent a 24 GB card for a couple of training days before spending 1,093 on one.

Frequently asked questions

What is the best budget GPU for robot learning? For the light policy class, a used RTX 3060 12 GB at about 295 in October 2026 sold data. For everything up to the large VLA families, a used RTX 3090 at about 1,093. Hold the 12 GB line: the diffusion class peaks at 8 to 14 GB, and 8 GB cards force batch cuts immediately.

Can I train a robot policy on a CPU? Not usefully. The LeRobot hardware guide's own line for CPU-only machines: "Don't train. Use Colab or rent a GPU." A free notebook tier or one rented hour beats a day of CPU per run.

Is a Jetson kit worth it for a hobbyist? As the robot-side computer, yes. The 399 Orin Nano Super kit runs light policies on the robot at low power. As your only trainer, no: a used desktop card trains faster for less money. See the Jetson versus cloud guide.

What does a cheap robot brain cost? The phrase tracks the Jetson AGX Thor in news coverage: 5,499 at list since the July 2026 reprice, 3,499 at its August 2025 launch. For a hobbyist, the working brain is usually a 295 used card plus open-source software that costs nothing.

Do I need 24 GB of memory? Only for the large VLA family, pi0, pi0.5, and X-VLA class, which peak at 24 to 40 GB and run tight at batch 1 on 24 GB. ACT, diffusion, and SmolVLA all fit in 16 GB or less.

Can I rent a GPU instead of buying one? Yes. Interruptible marketplaces rent 3060-class hours from about 0.015 to 0.04 and 3090-class hours around 0.041, read October 26 2026. Checkpoint your training runs, and keep live robot control on local hardware.

If you would rather not run models at all

There is a third shape beyond owning and renting: hosted robot model routes, where somebody else runs the model and you call it. Whether that fits depends on what you are building and what is currently offered, which changes. Our pricing page carries live rates if you want to compare, the bill shock guide explains how hosted costs run away, and the OpenAI-compatible API guide covers the client side. You can open an account at signup.

For tier-1 hobbyists, the default answer stays local: a used card or a rented hour, an arm on the desk, and open-source policies.

Last verified: 2026-10-26