Mneme Labs field guide

AI infrastructure I use and evaluate by workload

There is no honest single-stack answer. Local hardware, rented GPUs, voice APIs, virtual servers, and managed deployment solve different constraints. This page separates what is already part of the Mneme workflow from services still awaiting controlled testing.

01
Local compute

RTX 3090 workstation

Private iteration, predictable hardware, no per-minute compute meter, and direct access to local media.

02
On-demand GPU compute

RunPod

Burst workloads that do not justify buying another GPU

03
Voice generation and dubbing

ElevenLabs

Narration for tutorials, explainers, and product walkthroughs

04
Cloud servers and application infrastructure

Vultr

Hosting APIs, background workers, databases, and agent services

05
Approachable app and API deployment

DigitalOcean

First deployments of web applications, APIs, and background workers

Start with the constraint

Infrastructure is a response to a measured constraint. Buy local hardware when repeated private workloads justify ownership. Rent a GPU when demand is intermittent or a model temporarily needs more memory. Use a voice API when a controlled narration pipeline matters. Deploy a server only when another person or system needs reliable access.

Choosing vendors first produces architecture by shopping cart. Define data sensitivity, model size, runtime, traffic, latency, geography, budget, and failure cost before comparing products.

The RTX 3090 remains the local baseline

A local RTX 3090 is the current practical reference point for Mneme’s media and visual-memory work. It keeps source footage on the workstation, supports direct iteration, and turns repeated inference into a hardware cost instead of a running compute meter.

That does not make local hardware free. Power, cooling, storage, maintenance, queue time, and the operator’s attention still count. Local also stops scaling when the required model, deadline, or parallel workload exceeds the machine.

Use on-demand GPUs for bursts, not confusion

RunPod belongs in the stack when a defined job needs a different GPU or temporary parallel capacity. The job should be reproducible, checkpointed, measurable, and able to store its output away from disposable compute. Upload time and storage can matter as much as GPU runtime for video.

Mneme Labs will add measured RunPod results after running the same representative workload locally and remotely. Until then, the RunPod material is an evaluation framework—not a claim that one configuration beats the 3090.

Treat voice as a governed production stage

ElevenLabs can sit after scripting and before final editing. The pipeline needs consent, pronunciation controls, segment-level regeneration, listening review, and an archive connecting every audio file to its script and settings. Fluent output is not proof that names, numbers, or technical claims are correct.

Separate hosting from expensive inference

Vultr and DigitalOcean can host applications, APIs, workers, and supporting databases. They do not need to run every model inside the web request. A durable pattern is to accept a job, validate it, queue expensive work, process it on the appropriate compute, and expose status and results separately.

DigitalOcean gets the beginner-deployment angle in this cluster. Vultr gets the infrastructure-control angle. The difference is editorial positioning, not a claim that either platform can only serve one audience.

What will turn these guides into benchmarks

Each hands-on update should publish the workload, model and version, hardware, region, storage path, transfer size, runtime, failure behavior, and complete observed cost. Screenshots and workflow videos should document the run, not decorate a conclusion chosen in advance.

That evidence is what will eventually justify stronger recommendations and affiliate calls to action. Until then, the pages link to official vendor information and state the testing status plainly.