Dollars, not a ship date

The GPU is cents. The thing you want is the rest of the invoice.

A working pipeline for feasible lyric swaps over a frozen mix, with a menu of licensed singers, is a funded product: roughly $400k–$3M to build, then tens of thousands a month to run if it has users. The YouTube cover is two orders of magnitude cheaper. Matching Suno, or shipping unlicensed likeness, is not a software quote.

To get what you described, plan on mid-six to low-seven figures before famous names, and eight figures the moment you want those names.

Price the slice, not the fantasy

The YouTube cover

Sinatra timbre on Careless Whisper. Same words, same sax, same clock.

Build $0–$40kRun pennies / track

This is RVC / So-VITS on isolated vocals. A hobbyist does it on a 4090. A small product (upload, split, convert, mix, download) is a normal web app plus a GPU worker. Inference is seconds. You already heard this genre. It does not change lyrics.

The website

Prompt, library, player, credits. Suno’s product loop.

Build $30k–$150kRun $50–$500 / month at toy scale

Auth, object storage, CDN audio, a queue UI. No model breakthrough. This repo is the cheap end of that slice.

Lyric lock, same throat

New words, original accompaniment, original singer, sample-locked.

Build $250k–$1.5MRun $0.05–$0.40 / edit

Stems, singing alignment, duration-budget phoneme packing, infill (the MeloDISinger problem), mix-match of reverb and doubles. A demo verse is glue around open models. A tool that survives a stacked chorus and a wet pop mix is labeled data, a specialist, and a lot of failed takes. Arbitrary fan lyrics that do not fit the melody still fail — that is physics, not budget.

That, plus licensed voices

Pick a singer from a menu. Not a living star you do not own.

Build $400k–$3MRun voice deals + GPU

Record or license voicebanks (session singers, Synth V / ACE / Kits-class contracts). Custom bank from a willing singer: typically mid four to low five figures to record, plus training. A catalog of twenty decent singers is a production budget, not an API key. This is the path that actually ships.

Anyone, including famous people

That slider on a hit, in a named artist’s style.

Build not an engineering invoiceRun $10M–$500M+

Likeness and catalog. Suno’s Warner settlement was reported at $500 million. A single AI-artist record deal already cleared millions. Unlicensed celebrity conversion is a complaint, not a line item. Budget for lawyers and licenses, or do not build this.

Your own Suno

Closed weights, eight-minute songs, consumer brand, GPU fleet.

Build $20M–$100M+Run a GPU fleet + a label relationship

Foundation-model training, data, serving, moderation, mobile, feed. Compute for a serious music model is millions to tens of millions before the product exists. This is a company. Open weights (ACE-Step, YuE) skip the training bill and still leave you short of v5 polish and of lyric-lock.

What you actually pay per song

Raw GPU

A100-class cards are about $1.20–$1.60/hr on public clouds. ACE-Step-class generation can be a few seconds, so the silicon for one full-band generate is well under a cent. Nobody’s invoice looks like that.

A real song job

Stem split + align + lyric infill + optional voice convert + loudness/true-peak + retries + idle workers. Budget $0.05–$0.50 per successful edit, more if you keep GPUs hot for latency.

Storage and CDN

WAV/FLAC masters are large. A million tracks is a storage bill, not a rounding error. Cheap compared with models. Not free.

People

The spend is ML + audio engineers, not React. One person who has shipped singing models is worth more than the GPU month. A lab that can touch mix-quality vocals is a payroll, not a prototype credit.

Spend it on this roadmap Why lyric lock is the expensive slice