Renting an RTX PRO 6000 by the hour beats buying one outright

The picture

Amazon Web Services' own G7e instance page states that each G7e instance is built around an 'NVIDIA RTX PRO 6000 Blackwell Server Edition' GPU with 96 GB of GDDR7 memory - the server counterpart of the workstation card priced elsewhere in this batch at $16,000.00 on NVIDIA's own marketplace listing. AWS does not publish a per-hour dollar figure on the instance page itself; the tracking service Vantage Instances, checked 16 September 2026, records the smallest single-GPU size, g7e.2xlarge, at $3.36312 per hour on-demand, or $1.853 per hour on the spot market.

What the documents show

Dividing the $16,000.00 purchase price by $3.36312 per hour gives roughly 4,758 hours of rental before the running cost would match the purchase price. At a single eight-hour shift, five days a week, that is a little under four and a half years; at continuous round-the-clock use, it is under seven months. The AWS figure is a tracked on-demand rate from a third-party pricing service, not a number AWS itself publishes on the page checked, and it excludes storage, networking and any reserved-instance or savings-plan discount that would lower it further for committed, continuous use.

What it is allowed to decide

A make-or-rent calculation is a commercial fact, not a technical one: it says what a studio can afford to run and how often, not whether a technical-execution render is correct. This supports class 3, commercial, at most. It holds Provenance for the specific rate and source cited on a specific date. It does not hold Auditability over any output produced on rented hardware - nothing about the rental addresses whether a rendered scene can be reopened and reproduced later on different infrastructure.

The disclosure label

A disclosure travelling with a make-or-rent decision would read: cost comparison based on a $16,000.00 workstation purchase price against a tracked $3.36312-per-hour on-demand cloud rate for the equivalent GPU, both as checked 16 September 2026; Provenance held for both cited figures and their sources, Auditability not established for outputs produced on either path; asserted by whoever prepared the comparison.

• Is the work intermittent enough that the roughly 4,760-hour break-even point favours renting over buying?
• Does a reserved-instance or savings-plan commitment change the on-demand rate used in this comparison?
• Does the studio's need for offline, air-gapped or long-term archival access rule out a cloud-rented GPU regardless of the hourly cost?

Editorial reading: the arithmetic here is specific to intermittent, single-shift previs work; a studio running GPUs continuously across multiple shifts should expect ownership to win well before the roughly 4,760-hour mark cited above.