The robotics OEM compute decision: build, buy, and the edge-first law

No leading robotics OEM rents the brain it deploys. Figure runs every inference tier onboard its robots, and it buys training compute by the billion. 1X trains its models in the cloud. It ships a Jetson module on each robot. Tesla builds its own inference silicon. The build versus buy question at OEM scale is really three questions. The deployment brain is bought as silicon. It goes on the robot. The training compute is bought from cloud owners. Per call hosted routes serve the programs below this scale. The record, the prices, and the arithmetic follow.

All amounts are United States dollars written as bare numerals. Every price carries its source and read date. This market moved twice in 2026, once for edge modules and once for cloud parts. Stale numbers are how these decisions go wrong.

Last verified: 2026-10-26

What the leading robot makers actually decided

Start with the record, because the argument is settled there. Six parties. Every row is dated, and every claim comes from the party's own posts and pages.

Three patterns fall out. One, nobody in this table rents the control loop by the call. Two, the answers on training differ by scale. Figure signs cloud contracts. 1X rents cloud hours. Tesla and Amazon build their own. Three, every maker that disclosed anything buys the deployment brain as parts, not as a service.

The edge-first law, read from the record

The cleanest statement in the industry is Figure's own Helix post, February 20 2025. It calls Helix the first vision-language-action model that "runs entirely onboard embedded low-power-consumption GPUs". Not mostly onboard. Not onboard with a remote fallback. Entirely onboard.

The same post gives the split. System 2 is a 7-billion-parameter vision-language model. It runs at 7 to 9 Hz. System 1 is an 80-million-parameter visuomotor policy. It turns those plans into actions at 200 Hz. Each system gets its own embedded GPU. A year later came Helix 02, January 27 2026, with a third tier. System 0 is a 10-million-parameter controller. It issues joint commands at 1 kHz. The post shows a four-minute task: walking, unloading a dishwasher, stacking, loading it again. No resets, no human help, all from onboard sensors.

Then the training side, which is where the money went. On September 3 2026 Figure announced a partnership with Nscale, an AI cloud. The initial commitment is 3.5 billion US dollars of compute. It can scale past 6 billion, on up to 100,000 NVIDIA GPUs. The platform is Vera Rubin. First deployment comes in the second half of 2027, at Barstow, Texas. Nscale also took an equity stake in Figure. The founder's logic fits one line. Helix "becomes more capable the same way every learned system does: with more data and compute."

The scale is worth a pause. Figure's Series C, September 16 2025, raised just over 1 billion at a 39 billion valuation. The training commitment is roughly triple the raise. The most edge-first OEM on earth still buys its training compute. It buys it from a cloud.

1X makes the same split legible at a smaller scale. Its GTC post, March 17 2026, names both ends. Training runs on NVIDIA Blackwell HGX B200 parts. Deployment models run on Jetson Thor, "directly on the robot with minimal latency". Two companies, two scales, one shape. Cloud for training, robot for action.

The law, stated once:

1. Deployment inference belongs on the robot. Every tier stays onboard, from 1 kHz joint control to 7-9 Hz planning, on parts Figure has never even named.
2. Training compute is bought, at any scale. The maker that keeps inference most on-robot still signed a 3.5 billion US dollar cloud contract for training.
3. Nobody at the top rents the control loop by the call. Hosted per call routes are a rung below OEM scale, not an OEM architecture.

What is actually known about Helix onboard GPUs

The question gets asked constantly. The honest record is short. Figure's February 20 2025 post says the whole stack runs on "dual low-power-consumption embedded GPUs". One GPU is dedicated to System 2. The other is dedicated to System 1. That is the entire hardware disclosure. Through Helix 2.5, September 17 2026, no Figure post has named a vendor, a part number, or a watt figure. When a report tells you a specific chip sits inside Figure's robots, that is the writer's inference, not Figure's record. The knowable facts are the count, two, the class, low-power embedded, the loads, 7 billion and 80 million parameters, and the rates, 7-9 Hz and 200 Hz.

1X is the contrast. It shows that disclosure is a choice. Its own post names Jetson Thor as the deployment part. If your vendor will not name its part, that is a question for the contract, not a mystery to solve.

The decision axes: latency, locality, and scale

Latency decides the top half of the table below. Each control tier has a deadline. The deadlines are not negotiable with physics.

Locality comes second. Some buyers must keep processing local, whatever latency says. Defense programs require air-gapped processing. Data rules in China keep Chinese robot data in China. The EU Machinery Regulation 2023/1230 extends machine-safety law to digital safety functions. It applies from 14 January 2027 by its own Article 54. None of this forces training local. It forces the safety-relevant control local, which is where the latency tiers already put it.

Scale comes third. The arithmetic below works it. Hosted routes lose first on latency. They lose on money as soon as robots work real shifts.

The buy side: module and cloud prices read today

Edge modules first, from NVIDIA's developer FAQ, read October 26 2026:

The July 2026 reprice is its own planning fact. CNX-Software documented it on July 22 2026. Every Jetson line rose, by up to 101 percent, with no announcement. Memory costs were the cited cause. A parts budget built on 2025 prices is wrong today. The top part is serious hardware. AGX Thor carries 128 GB of memory, 273 GB per second of bandwidth, a 40 to 130 W envelope, and 2070 sparse FP4 TFLOPS per NVIDIA's product page.

Cloud training parts, read the same day at gpudiff, changelog dated October 26 2026: a B200 part with 180 GB rents at 7.99 per hour at RunPod, 10.25 at Vast, 14.24 bundled at AWS. One such part, around the clock for a month at the low read, is 730 hours. That is about 5,833.

Hosted robot routes, for completeness, price in tokens. Gemini Robotics ER 2 lists 1.00 per million input tokens and 5.00 per million output through December 31 2026. Both double on January 1 2027. Physical Intelligence opened an API beta on October 13 2026. It bundles fine-tuning with a hosted route, with no public pricing. A token rate cannot be compared to a robot hour without a token count per call. No vendor publishes one.

The arithmetic at deployment scale

This platform's worked basis comes from the robot-hour cost page. A 30 frame per second camera loop with action chunks of 50 makes 2,160 calls per robot hour. The written-record per-call rates for the pi05 and smolvla model ids give 6.48 and 2.16 per robot hour on that loop. Its full derivation is shown there. The rows sit in the platform's model data with their rates set, and they are not on the public list as of 26 October 2026. This page quotes no live rate for them. Only listed rows render in the rendered table, and the models index is the live catalog. A written record is not a launch.

Extend it to a deployment of N robots, symbolically:

monthly hosted cost  =  N x H x C x P

N:  robots in the program
H:  robot-hours per robot per month
C:  calls per robot-hour, set by the control loop
P:  price per call, set by the rate you use

Now numeric, at N = 500 robots working one shift, H = 160 hours a month, C = 2,160 calls:

The crossover is the number to keep. A 5,499 module against the 6.48 lane pays for itself at about 849 robot hours. Spread across two years, that is roughly 35 robot hours a month. A robot on any real shift clears it in a week. This is the arithmetic under the edge-first law. At scale, the hosted lane costs several times the spread-out module. That holds before latency, before steady timing, before any safety case.

Training is a different bill. It does not scale with robots on shift. A single B200 part for a month is about 5,833 at the low read above. The serious programs rent by the hundred-part cluster. At Figure scale, the Nscale deal is that shape. The startup infrastructure guide works the middle of that ladder. The Jetson versus cloud guide works the robot-side placement question in depth.

The build side: what first-party really takes

Build means two different things, and the record separates them cleanly.

Build the deployment part. Tesla does, per its own AI page. One inference chip line, shared by cars and the robot. Nobody else in the table does. That path takes automotive-scale volume and a silicon team. It is a funding decision before it is an engineering one.

Build the training plant. Figure's route is the instructive one, because it did not. It committed 3.5 billion to a cloud after raising just over 1 billion in its Series C. Amazon builds its own, but Amazon is a cloud. For everyone else, training is bought.

So the honest menu for an OEM has three dishes. Buy modules for the robots, at the list prices above. Buy cloud hours for training and testing, at the rates above. Reserve true first-party building for the two or three companies on earth with the volume to justify it.

The compute ladder: hobbyist, startup, OEM

This page is the top rung of three. The hobbyist budget guide is tier one. One robot on a desk, a used card or a 399 kit, bursts of training. The startup infrastructure guide is tier two. A funded team, monthly costs, the first real hardware. Tier three is this page. Silicon strategy, module budgets in the parts list, training contracts, and a control loop that never leaves the robot.

Decision rules

• Your loop runs at 200 Hz or above. The part goes on the robot. There is nothing to discuss.
• Your robots work real shifts, above roughly 35 robot hours each per month at these rates. Edge capex wins on arithmetic alone.
• You are piloting, hours are low, the hardware is not final. Hosted per call routes fit, and the VLA break-even guide carries the crossover math.
• You are training foundation models. You are buying cloud hours whatever you decide about deployment. The B200 reads are the anchor.
• You are integrating a humanoid program around a big partner. Apptronik's row is the shape: partner hardware, your own stack.
• You are Tesla. None of this page applies.

Frequently asked questions

What GPU does Figure Helix run on?

No part has ever been named. Figure's February 20 2025 post says the whole stack runs on dual low-power embedded GPUs, one dedicated to each system. No post through Helix 2.5 in September 2026 names a vendor, a part number, or a watt figure. Any specific chip claim you read elsewhere is inference, not record. 1X, unlike Figure, names its part: Jetson Thor, per its March 17 2026 post.

Do robot companies buy inference?

Not the kind that runs their control loops. The leading makers put deployment compute on silicon they buy as parts. Their spending concentrates in training, bought at cloud scale. Figure's 3.5 billion commitment with Nscale, September 3 2026, is the dated example. Hosted per call routes serve builders and pilot programs, the rungs below this page.

Should an OEM run its policy on a Jetson module or in the cloud?

On the module. The joint tier cannot leave the robot at 1 kHz deadlines. The policy tier wants steady timing that networks do not promise. Cloud fits training, test bursts, and routing, the pattern in every row of the record table above.

How does per robot hour cost change with deployment size?

It is linear: N times H times C times P. At 500 robots on one shift, the 6.48 lane reaches 518,400 a month hosted. A 5,499 module per robot spreads out to about 114,563 a month over two years. Below about 35 robot hours per robot per month, hosted wins. Above it, capex wins, and the gap widens with every shift.

Where hosted routes still fit

There is a real band where paying per call is the right answer: pilots, demos, low-duty robots, and programs whose hardware is not final. In that band, the robot-hour derivation and the VLA break-even guide carry the numbers. The pricing units guide explains the units vendors quote. The listed rates for every listed lane sit on the pricing page, metered per call against a prepaid wallet. You can open an account at signup. When the shifts get real, buy the module, and let the cloud carry the training.

Last verified: 2026-10-26