Compute planner·Data, runs, vendors, rent or buy

Compute planner

From data hours to GPU-hours to dollars. Size the training runs, the post-training passes and the annotation jobs, lay them out as periods, price them at every known vendor, and compare renting with owning under the buy gate from the Compute Management Policy.

Periods

Month 0 is

One row per line of compute. Rows can overlap: a rendering node runs alongside the training nodes. Start is the month the row begins; a new row starts when the previous one ends. In Compare, mark rows Rent or Own to build a mixed plan.

Name	GPU	Nodes (8 GPU)	Start	Months	Util %	Sourcing	Vendor	$/GPU-h

Data hours to GPU-hours to dollars

Hours of video to tokens to GPU-hours to dollars for one run, and the nodes that finish it in the window.

Data hoursvideo in the training set

Cameras1 for ego and web video, 2 for UMI

Frames / ssampled for training

Tokens / frameafter the tokenizer

Model, B params

Epochs

Overhead %vision encoder, action head

MFU %40 is a good multi-node run

GPU

Finish this run in, dayssets the node count, not the cost

$ / GPU-h

Scaling runs on a growing corpus

One row per scaling run, on the hours that exist at its date. Model, fps, tokens, MFU, GPU and price come from the fields above.

Run	Start month	Pretrain h ego + web	Epochs	Post h UMI	Epochs	Finish in, days	GPU-h	Nodes	Cost

Cameras per episode:pretraining post-training

Resume from the previous checkpoint: each run trains only on the hours added since the run before it

Preprocessing: decode, pose, tokenize

CPU for decode, resize and pose; GPU for the tokenizer. Once per hour of data.

Hours

Cameras

CPU core-h per video-hper camera; UMI with SLAM about 1.5, plain video about 0.3

Training fpsframes the tokenizer sees

Tokenizer, B paramsVAE or video tokenizer

Tokens / frame

Passesreprocesses over the project

$ / core-hModal CPU about 0.05

GPU

$ / GPU-h

Annotation and enhancement with a VLM

Captions, labels and filtering with a VLM: inference over sampled frames.

Hours to annotateego, web and UMI

Cameras

Frames / s sampled1 fps labels every second

Tokens / framethe VLM's image tokens

Output tok / samplecaption or label length

VLM, B params

Passesmodels or prompts run over the data

MFU %batched inference, 25 to 35

GPU

Finish in, days

$ / GPU-h

Token API, $ / M tokens0 to run on own GPUs; 7B-class VLMs are about 0.3, 72B about 1.0

Sourcing

Policy default is rent. Buy only when every line of the gate passes in writing, with CEO approval.

Presets from the plan

Scenario

GPU class for the vendor table and new rows

Rental term

Rental vendor

$ per GPU-hour

override

Reserved discount

% if no quote

Only vendors that pass the data bar Needs InfiniBand (multi-node) Show vendors with no stock or reference prices

Owned hardware assumptions

Node price and power, per class in use

GPU	USD / node	kW / node

8-GPU HGX server estimates, editable

Network fabric capex

$/node, once

InfiniBand switches, transceivers, cables; about $200k per 64-GPU NDR fabric. Set 0 for single-node Ethernet.

Site setup, once

$ fixed

Racks, PDUs, management and login nodes, OOB switch, install, shipping

Per-node setup, once

$/node

Tax and duty

Hosting, all-in

$/kW-month

Node network and storage

$/node-month

Support, spares, insurance

% of capex / yr

Internet and cross-connects

$/month

Build-out labor, once

$ fixed

Colo contract and onboarding, racking, cabling, burn-in, firmware, OS images, scheduler install: 2 to 4 engineer-weeks

Ongoing ops labor

$/month

About 0.4 FTE fully loaded: patching, DCGM, failures and RMAs, on-call, spares. Policy requires a named ops owner. This engineer is not building the data pipeline.

Remote hands and RMA shipping

$/month

Colo smart-hands at $150 to $250 an hour plus freight for returned parts

Cost of capital

% / yr on capex

Simple interest on the full capex while owned; 0 to compare on cash alone

Owned-node availability

% of hours

GPU and HBM failures, RMA turnaround, firmware and maintenance windows. Lost hours are backfilled from on-demand rent.

Warranty

months

Support after warranty

% of capex / yr

Extended support or self-insured spares once the OEM warranty lapses

Pay for hardware

Financing APR

% / yr

Financing term

months

Level payments; interest replaces the cost-of-capital line. Payments still owed at the horizon count against the net.

Depreciation

months

Residual value

% at end of life

Burst share rented

% of peak

Lead time

months rented

Storage, both cases

Hot shared FS

Hot rate

$/TB-month

Object store

Object rate

$/TB-month

Save, share, reset

Settings stay in this browser. Copy the JSON to share a scenario or paste one to load it.

Buy gate

Policy section 5.1. All lines must pass, in writing, with CEO approval. Owned nodes are depreciated over 3 years and hosting is charged to the same envelopes.

Cumulative cost over the horizon

Rent, cumulative Own, cash out (capex up front) Own, depreciated (P&L view)

Where the money goes

Per row, then run rates. Effective cost divides what is paid by what is used, the policy's unit metric.

Period	GPU	Nodes (8 GPU)	Start	Months	Util	GPU-h used	Rent	Own: burst rent	Plan	Plan cost

Rent

Own

Known neoclouds for this scenario

Vendor list from "Romoya Compute Planning - Vendors"; list prices re-checked on each vendor site on Oct 7, 2026. Click a row to price the scenario there. Edit a price to override it; overrides persist in this browser.

Vendor	Region	Fabric · orchestration	Status · commitment terms	Other GPUs, $/GPU-h	$/GPU-h on-demand	$/GPU-h reserved	$/H100-eq h	Node-month	Horizon

Assumptions. 730 hours a month. H100-equivalent factors from the Infrastructure Notes: A100 0.45, H200 1.1, B200 2.2, B300 2.5 (bf16 training; FP8 favors Hopper and Blackwell). On-demand is billed on used hours because interactive instances auto-stop; reserved and owned capacity bill the full month. Owned-node prices, fabric, site setup, build-out and ongoing ops labor, remote hands, power draw, hosting and cost of capital are planning estimates, not quotes; the buy case charges all of them, and the policy's 1.2x test uses the same lines. Owned nodes are assumed up 96% of hours with lost hours rented back, support rises after the warranty, and a financed purchase swaps capex for level payments with interest replacing cost of capital. Reserved prices come from the sheet where a quote exists, otherwise the reserved discount applies to the on-demand price. Marketplace vendors (vast.ai and similar) are restricted to public data under policy section 5.2. Location: US and Americas vendors are acceptable (decision Oct 7, 2026); mainland China is excluded.
Sources. Compute Management Policy v1.0 (Sep 29, 2026), Infra Plan M3 and M4 (Sep 2026), Romoya Compute Planning (Oct 6, 2026) and its vendor sheet.