Read Books Books

Earlier benchmark evidence: October 25 baseline white paper · Original interactive clip-review gallery. The original gallery covers the RTX 5070 / 32 GB RAM configuration and reports time to a finished file.

The investment · Daily capacity · Rental · APIs · Next production test

The goal is finished episodes

My goal is a scalable movie-making system that can finalize five to ten vertical short-drama episodes per day. At 90 seconds each, that is 7½ to 15 minutes of finished story every day. I want to direct the work while an AI production assistant helps plan shots, make stills, generate performances, organize assets and carry out edits.

That is an operating goal, not a result this computer has already demonstrated. The companion hardware paper reports 96 completed video test configurations and 18 matched before-and-after pairs. This paper uses those measurements to estimate the rendering portion of a possible studio day. It separates money already spent, measured clip times, and assumptions about future production.

The central question for an author is not simply whether a large GPU is fast. It is whether ownership, rental, an API, or a mixture delivers enough usable work at an acceptable total cost—and whether the author wants to operate that system.

What I actually invested

These are my reported purchase figures, supplied October 29, 2026. They are not independently audited invoices or current shopping quotes.

ItemAmountHow it is used here
Original computer, purchased on sale on Christmas Eve 2025$1,400Historical cash paid
New RTX PRO 5000 72 GB GPU$9,000Incremental upgrade cost
New 128 GB system RAM$2,750Incremental upgrade cost
GPU + RAM upgrade$11,750Primary ownership comparison
Original purchase + these upgrades$13,150Known cumulative hardware spending

I estimate that the original computer now retails around $2,300 at Best Buy, and that its original GPU alone is around $800–$900. Those are owner-supplied market estimates; this analysis has not verified an exact current retail SKU. They are not added to the $1,400 I actually paid. Similarly, $13,150 is not an all-inclusive studio budget: it excludes any unlisted components, software, electricity, backups, maintenance, financing and labor. No resale credit for replaced parts is assumed.

The machine now has 72 GB of GPU memory and 128 GB of system RAM. Those are separate pools. Capacity creates options, but it does not prove that every author needs this configuration. In our ten-second LTX tests, the highest sampled whole-system RAM use was about 42.3 GiB. A controlled 64-versus-128 GB test remains undone.

Start with the output: five to ten finished episodes

For a planning example, assume a 90-second episode is assembled from eighteen retained five-second segments. Actual editing will use different shot lengths, overlaps, stills and sound bridges. This is an accounting unit, not a rule for cinematic pacing.

The upgraded computer’s LTX 8+3 at 1080p took a three-scene median of 61.002 seconds from submission to finished five-second file. Eighteen such clips therefore represent about 18.3 minutes of serial clip production before retakes, trimming losses or the rest of the pipeline. Measured timing source

The useful production variable is generated seconds ÷ seconds retained in the final edit. A factor of 3 means generating 270 seconds of footage to retain 90 seconds. This can include rejected performances, alternate coverage, handles and trimmed material; it is not a measured rejection rate. We test factors of 1, 3 and 5 below. All final screen time is assumed to come from generated video; reusing footage or holding stills would change the requirement.

Estimated serial LTX clip-production hours per day: five episodes require 1.5, 4.6 or 7.6 hours at generation factors of 1, 3 or 5. Ten episodes require 3.1, 9.2 or 15.3 hours. Editing and other production work are additional.
Figure 1. A planning projection from the measured 1080p LTX 8+3 median, not an achieved daily output. Both series use the same zero-based scale. One GPU render at a time; 90 seconds per episode.
Generated / retained footageFive episodes / dayTen episodes / day
1×: every generated second is retained90 clips · 1.53 hours180 clips · 3.05 hours
3×: one third of generated seconds retained270 clips · 4.58 hours540 clips · 9.15 hours
5×: one fifth retained450 clips · 7.63 hours900 clips · 15.25 hours

These hours include the benchmark’s file preparation but exclude still generation, voice generation, local-agent work, creative review, editing and final export. The underlying tests were landscape; native vertical production has not been timed. Multiplying three-scene medians is an estimate, not a queue simulation or a sustained all-day endurance result.

Recipe choice matters. At the same 3× footage factor, five to ten episodes using native LTX’s 1080p five-second median would occupy roughly 6.6–13.3 hours of clip production. H3 Turbo at the same output size would occupy roughly 19.9–39.9 hours. The fast LTX figure should not be applied to every model. These recipes also produce different performances; the preferred shot can justify a slower route.

The implication is practical: the 5–10-episode target needs a repeatable shot system, a measured usable-footage yield, and a schedule for the whole machine. Ten episodes at the illustrative 3× factor already exceeds an eight-hour rendering window. Longer unattended queues or rented overflow may help, but directing and quality control remain work to plan.

Renting a GPU is a different kind of purchase

My initial rental estimate was $1.69/hour. I retain it as a sensitivity case, not a verified offer. My October 29 console screenshots show RTX PRO 6000, 96 GB VRAM, at $2.49/hour, and a workstation variant at $2.59. Runpod’s public pricing page also listed $2.49 when checked. This is a higher-memory GPU than my 72 GB card, not an identical rented twin. Host RAM, architecture, software setup and availability matter. We have not run our recipes on it. Runpod pricing

Here is the narrow capital comparison: how many billed rental hours equal the $11,750 upgrade outlay?

Rental rateHours equal to $11,750At 2,000 billed hours per year
$1.69/hour, unverified sensitivity case6,953 hours3.48 years
$2.49/hour, observed listing4,719 hours2.36 years
Rental compute charges accumulate to the 11,750-dollar upgrade outlay at about 4,719 hours for 2.49 dollars per hour and 6,953 hours for 1.69 dollars per hour. The horizontal ownership line excludes running costs.
Figure 2. Capital-only cost crossings. Equal billed dollars do not establish equal productive output, full ownership cost or investment payback.

The annual example assumes 8 billed hours on 250 production days. At $2.49/hour, that rental compute bill is $4,980 per year; at $1.69 it is $3,380. If a machine is left billable for a 24-hour period, those amounts are $59.76 and $40.56 respectively, before storage. A forgotten running instance is a cost even while nobody is directing it.

Rental also carries storage charges. As one illustrative option, 350 GB at the listed network-volume rate of $0.07/GB/month is $24.50/month; other disks or storage tiers add or replace charges. Stopping compute does not necessarily stop persistent-storage billing. Runpod documents separate running/stopped storage rates and recommends backing up critical work. Pod pricing and storage

Availability is a real planning variable: several options in my screenshots were out of capacity. Runpod also documents that a stopped Pod may lose access to its former GPU because another customer rents it; storage can remain while GPU access is unavailable. Restart and availability

I would measure rental cost per accepted clip, including setup, downloads, idle time, failures and repeated attempts. I would not multiply my local speed by a rental price and call it a measured cloud result. For example, 4.58–9.15 billed hours at $2.49 would cost $11.39–$22.78 in compute, but that is only an equal-hours illustration; cloud throughput has not been tested.

Ownership adds electricity, cooling, maintenance, downtime and eventual replacement. Electricity should use measured wall power, not the GPU’s advertised power limit: average whole-system kW × powered hours × electricity price per kWh. Wall power has not been measured here. A local machine can retain resale value and serve writing, image, audio and agent workloads, but those benefits and future resale proceeds are not credited as proven savings.

API seconds are not finished-episode minutes

The H3 image-to-video endpoint checked on fal lists $0.05 per output second at 480p. That is $3 per generated minute, or $4.50 for 90 generated seconds. It is a rate for this specific endpoint/resolution, not a universal video-AI price. Its hosted implementation is not established as identical to our local quantized and adapter-based recipes. Endpoint and rates

480p H3 output charges onlyOne 90-second episodeFive episodes / dayTen episodes / day
1× generated / retained footage$4.50$22.50$45.00
3× generated / retained footage$13.50$67.50$135.00
5× generated / retained footage$22.50$112.50$225.00

These are simple output-second scenarios. They exclude stills, audio, agents, editing, taxes and any endpoint-specific additional charges. They also deliver 480p, so finishing or upscaling to 1080p/2K would be additional work. Do not compare this price directly with native 1080p LTX as though model, quality and delivery were identical. The same fal page listed $0.13/second for its 2K option: $11.70 for 90 generated seconds before repeats. That hosted output path is not our local 2K benchmark.

I have previously experienced roughly $20–$60 per minute in API video spending. That is my reported experience, not a verified rate for the current 480p H3 endpoint. We have not reconciled the invoices, models, resolutions, rejected takes or definition of “minute” behind that range. At that historical effective cost, 7½–15 finished minutes would imply $150–$900 per day, if the historical rate applied per finished minute. That conditional illustration must not be presented as a measured daily bill or have retry costs applied a second time.

Cheap generated footage and cheap finished stories are different quantities. A $3 generated minute with a 3× footage factor becomes $9 in video generation per retained minute, before the rest of production. The best economic comparison uses the same creative brief, delivery requirement and acceptance criteria across routes.

Which route fits an author’s work?

RouteWhen it may fitTradeoffs to budget
Existing computerOccasional clips, smaller models, low-resolution auditions, learning the workflowMemory limits, longer waits, some recipes may not fit
Owned high-memory workstationSustained daily use, persistent local assets, control of versions and queues, several compatible creative workloadsLarge up-front outlay, upkeep, one-machine bottlenecks, idle capital, replacement risk
Rented GPUIntermittent intensive work, testing a hardware class before buying, burst capacity beyond the local queueAvailable capacity, environment setup, storage, idle billing, Internet dependence and backup discipline
Hosted model APILow setup burden, infrequent production, access to specific hosted capabilitiesPer-output charges, retries, service/version changes, limits and less infrastructure control
HybridLocal everyday work with selected API features or rental overflowMore integrations and careful asset/version tracking

My inference from these early tests is to develop the production process around the machine already purchased, while treating rental and APIs as possible additions. That is different from recommending that every author spend $11,750. For a new buyer, a modest rental test with a real project can reveal whether the extra capacity is useful before making a large purchase.

Local agents are part of the research goal

The ambition extends beyond image and video inference. I want a local model to help operate production and execute editing notes. I am using Astra 6 as the current assistant baseline, then plan to test open-source agents on comparable tasks. That statement identifies the research direction; no agent benchmark or local-agent production result is reported in these measurements.

A fair agent test needs a fixed script, asset set and edit instructions, followed by checks of shot selection, timing, instruction adherence, preservation of originals, recovery from failures and the number of corrections I have to make. “It can load into VRAM” would not establish that it can edit a watchable episode. The comparison should measure completion time, human minutes, model/API cost, local memory use and whether the final edit meets the brief.

Running an agent beside a video renderer could also contend for GPU memory and compute. Concurrent agent-plus-render workloads have not been tested. Initially, a scheduler can separate planning, still generation, video generation and editing stages; any claimed concurrency benefit needs its own test.

The test that will make this a studio business case

The next unit of evidence should be an approved 90-second episode, then a batch of five using the same process. Lock the script, cast references and delivery requirements. Review an ordered still-and-audio storyboard before generating every moving shot. Record all attempts, including rejected or trimmed material, and preserve edits as new versions.

For each finished episode, record:

  • Finished output: approved seconds, native vertical dimensions, sound, captions and successful export.
  • Useful yield: generated seconds, retained seconds, and why footage needed revision.
  • Time: wall-clock turnaround, machine time by stage and EAW’s hands-on directing/editing minutes.
  • Cost: API bills, rental hours/storage, measured local electricity and any recurring tools used.
  • Reliability: failed jobs, missing assets, edit corrections, recovery work and final approval.

The target is five to ten finalized episodes, not that many drafts waiting in a review queue. If rendering is fast but review or continuity repair dominates the day, buying more rendering capacity alone will not solve it. If accepted-shot yield is good and the queue is consistently full, rental overflow or additional local capacity becomes a concrete decision.

Evidence and calculation notes

Hardware prices and the historical API-cost range are owner-reported on October 29, 2026. The rental screenshot is a dated observation; $1.69 is explicitly a sensitivity assumption. Public rates were checked October 29 and may change. No cloud job was launched for this paper.

Download assumptions and formulas · download scenario rows, CSV · measured hardware study · watch selected test footage.

Prepared with AI assistance, with Elizabeth Ann West directing the research. All capacity and cost scenarios are planning estimates. Achieving five to ten finalized episodes a day remains to be demonstrated.