Midjourney compute and growth docket#

Status: Decision log Scope: Proprietary creative subscription; relaxed scheduling is not evidence of GPU title.

2026-10-04 — Paid entry and priority-segmented capacity#

Compute and procurement#

Historical Google Cloud account, April 2023: Midjourney trained version four on Cloud TPU v4 with JAX; a founder statement distinguishes training TPUs from inference GPUs. This establishes cloud infrastructure use at that date, not the provider or title of every current GPU. The captured structured date is April 5 while other page metadata crosses April 6; the claim is intentionally month-dated. Google Cloud disclosure.

Current scheduling: Fast uses included GPU time; Relax queues image work as capacity frees, with waits influenced by prior use and advertised up to 30 minutes. Turbo uses a faster pool and more Fast allowance, with version/availability limits. These are consumer scheduling terms, not bare-GPU rental rates or measured utilization. Speed/billing definitions.

Acquisition and monetization#

Monthly list tiers remain Basic $10, Standard $30, Pro $60 and Mega $120, with 3.3/15/30/60 Fast hours. Annual subscriptions have a 20% discount paid upfront. Standard and above include relaxed images; Pro/Mega include relaxed SD video and Stealth privacy. Extra Fast allowance is $4/hour of the product's defined time—not a disclosed physical GPU's rental price. Plan comparison.

There is no current trial on the website or Discord; a limited niji mobile-app trial is a separate surface. Image ranking/tasks can reward additional time. Inference: public creative discovery, community and repeated style/video work support acquisition and retention while paid priority/privacy segment willingness to pay. This is not a documented recurring free-compute funnel or measured conversion. Trial boundary, time rewards and queues.

Outcomes and implication#

No current audited revenue, paying-user cohort or legal GPU-title evidence was established. Older revenue estimates and Discord-member counts are not renewed as current facts. Relax illustrates demand shaping through latency rather than proof that ownership makes inference free. Our own paid-capacity protection policy remains unchanged; this research does not authorize a competing queue implementation.

Direct documentation downloads returned HTTP 403 despite readable web retrieval. No account, checkout or workload was tested. Earlier profile preserved; batch manifest, coverage index.