There’s a frame from John Covach’s history-of-rock lectures that’s been rattling around our head for a few weeks. Before 1955, American popular music was structured around the song, not the recording. Music publishers ran the business. Songwriters wrote songs. Performers were “song stylists” — people who put their personal mark on a song that had been written, often years earlier, by someone else. Sheet music was the unit of trade; pianos were the home delivery system; and any one song lived simultaneously across a dozen sanctioned versions by a dozen different singers, none of which was the version.

Then rock and roll arrived and broke the model. The performer became the author. The recording became the canonical work. Cover versions stopped being the point and started being a sign that someone hadn’t done the original. By the late 1960s, the song-as-product industry was effectively dead in America.

It is now coming back — quietly, accidentally, and in a form the pre-rock industry would barely recognise. We think the right way to read 2026’s music industry is as a permanent partial restoration of the pre-1955 model, and we think there’s a real product to build on top of that observation. This is a long way of saying: the song should be the unit of product again, the stack to make that work now exists, and somebody is going to build the platform that runs on it.

This is that thesis.

The four forces that quietly rebuilt the song-first model

It happened without anyone framing it as a return to anything. Four largely unrelated curves reached usable scale at the same time, and the result is a music economy in which the song is once again the thing being shared, modified, and monetized — not the canonical recording.

1. The cover-and-clip economy is structurally the largest part of music consumption now. Instagram in India alone has over 500 million users, and they spend most of their listening time on 15-second hooks attached to videos other people made. Reels and TikTok don’t pay per stream — TikTok pays approximately $0.03 per video that uses a song, regardless of how many people watch it. The royalty unit is the use, not the listen. Every Reel using “Raina” is, structurally, a tiny licensed cover of the song-as-asset. Smule has 52 million monthly active users globally with roughly 7-8 million in India; StarMaker has 50 million MAU with 30 million Indian users. The global karaoke app market is projected to hit $8.28 billion in 2026 and $19.82 billion by 2035. These are 80 million people whose primary musical activity is performing other people’s songs.

2. Generative AI made every user a potential song stylist. Suno’s v5.5 release in March 2026 introduced voice cloning, custom models, and a Covers feature that “takes an existing track and regenerates it through a new performance direction.” The platform has reached 2 million paid subscribers and 100 million total users, and is valued at $2.45 billion after a Series C in November 2025. BandLab — a free, browser-based collaborative DAW — has 100 million users and 16-17 million new songs created every month, nearly double the entire Spotify catalogue. Splice has 8 million producers buying samples and stems. The infrastructure for a non-musician to produce a competent personal version of any song now costs less than a Spotify subscription.

3. Stem separation went from research curiosity to commodity in 24 months. Demucs v4 hits 9.20 dB of SDR on the standard separation benchmark; MVSep offers 100+ AI model configurations through a browser; UVR5 is free and ensemble-mode-capable. Any released recording can be deconstructed into a vocal-free karaoke track, an isolated vocal, drums, bass, harmony, and “other” — for free, on a phone, in three minutes. The pre-1955 industry’s product was sheet music: a set of instructions that lets a non-author perform a song. The 2026 stack does this automatically and retroactively against the entire recorded music catalogue. Every released recording is now also its own sheet music, in a sense, available for re-performance.

4. The royalty-routing rails finally work, in patches. DDEX standards — ERN/MEAD message formats — have made distributor-to-DSP metadata transfer good enough that, when a song is properly registered with a PRO and a publisher, derivative works (covers on streaming, sync to TikTok, mechanical reproductions) actually route royalties back to the songwriter. Spotify paid $5 billion to publishers in the past two years, with publishing payouts up 2.5x over five years. The mechanical-license layer (Harry Fox, MLC, IPRS in India) has compulsory licensing baked in: in the US, anyone can release a cover without permission for a statutory $0.091 rate per copy. The legal scaffolding for paying songwriters across many sanctioned versions of one song already exists. It just isn’t the marketing message of any consumer-facing music platform.

Put together: the cover/clip economy gives you the demand side, GenAI plus stem separation gives you the supply side at zero marginal cost, and the royalty rails give you the monetization layer. What’s missing is the product.

What’s missing: a song-first launch platform

Here’s the structural gap. Every music platform built since 1980 has been organised around either the artist (Spotify, YouTube, label rosters) or the user (TikTok, Smule, Suno). None is organised around the song as the primary unit.

Spotify is an artist-centric streaming product where the canonical recording is what gets played and where the artist captures most of the master royalty share. TikTok and Reels are user-centric video platforms where music is incidental backing track. Suno and BandLab are creator-centric tool products where the user is the artist. The publishing side of the industry — IPRS, ASCAP, BMI, the publishers themselves — is the only part of the stack that thinks song-first, and it’s an invisible back-office layer with no consumer-facing surface.

The product that doesn’t exist: a platform where a song is launched into a community, the community produces and records many versions of it, those versions distribute everywhere, and every distribution routes attribution and royalties back to the original songwriter automatically. Think of it as “the song-first stack”: metadata-clean from inception, license-baked-in, community-as-A&R, GenAI-friendly, distribution-aware.

The historical analogues are instructive. Tin Pan Alley publishers in the 1920s and 1930s functioned exactly this way — they signed songwriters, plugged songs to as many performers as possible, sold sheet music, and collected mechanical royalties when records were pressed. The product was “the song,” and revenue was a percentage of every sanctioned use. The 2026 version uses the same logic but replaces sheet music with stems, the local pianist with Suno + UVR5, and the song-plugger with TikTok-style discovery.

The sketch

Here is what we’d want to build, in product terms, if we were running it.

Submission layer. Songwriters submit a song as a bundle: lead vocal stem, instrumental stems, chord/MIDI scaffolding, lyrics with timing, optional vocal-style guidance for AI cloning, and a mechanical-license statement (compulsory rate or custom). At submission, the platform does the unglamorous publishing-side work: ISRC, ISWC, IPI assignment, registration with the relevant PROs (IPRS in India, ASCAP/BMI/MLC in the US), DDEX-clean metadata. The song now has the full identifier scaffolding it needs to track derivatives.

Community layer. The platform exposes the song to a community of “song stylists” — singers, instrumentalists, producers, AI users — who can record, record-with-AI-assist, fully synthesize, or remix their own version using the provided stems and scaffolding. Smule’s user base proves this behaviour exists at scale; the gap is that Smule’s covers don’t route royalties to the original songwriter cleanly because the song-bundle wasn’t designed for it.

Discovery / A&R layer. Songs that produce the most versions, the most engagement, the most cover-volume in the community become signal. The platform A&Rs upward: the songs the community has voted on with their performances get promoted to broader DSP distribution, with the original songwriter as the credited author and a community of co-stylists. This is fundamentally different from a streaming platform’s editorial-curation A&R, and it’s a return to a Tin Pan Alley logic — the publisher knows a song is worth pushing because three different bandleaders are already playing it.

Distribution / royalty layer. Every sanctioned version, when it gets played on Spotify, used in a Reel, synced to a YouTube video, generates a royalty event. Because the original song-bundle was metadata-correct from launch, the platform automatically takes its publisher cut and routes the songwriter share back. The performer of that particular version takes the master share. This has been technically possible for two years; nobody has wired it into a consumer product.

AI integration. Generative AI isn’t an afterthought — it’s a first-class song stylist. A user can generate a Suno-style version of a community song using their own cloned voice and the platform’s stems-as-conditioning. The AI version is a sanctioned cover under the original song’s license, the AI user gets credit as the performer, and the original songwriter gets paid. The legal ambiguity that’s currently slowing AI music adoption — Suno’s stalemate with UMG and Sony over downloads — disappears inside this platform, because every AI cover starts from a song that was license-cleared at submission.

This is the song-first stack. The pre-rock industry tried to build it in 1925 with sheet music and didn’t have the technology. The 2026 version finally does.

Why India is the wedge market

There is a specific argument for India as the launch market that goes beyond TAM-counting.

The Indian mainstream music industry never went through a 1955-style rupture. Hindi film music — and Tamil, Telugu, Punjabi film music — has always run on a song-first, performer-as-stylist model. Lyricist (Sahir, Gulzar, Javed Akhtar, Amitabh Bhattacharya) and composer (R.D. Burman, A.R. Rahman, Pritam, Sachin-Jigar) write songs; playback singers (Lata, Kishore, Rafi, Arijit Singh) perform them. The same song is sung by different singers across decades — A.R. Rahman’s “Tu Hi Re” exists as a Hariharan-Kavita Krishnamurthy duet, a stage-show solo, multiple Coke Studio re-imaginings, and several thousand wedding renditions. Indian music’s cultural default is the pre-rock American model.

The numbers reflect this. T-Series — India’s largest music label, 35% market share, 311 million YouTube subscribers, the most-viewed channel on YouTube — has built a substantial portion of its catalogue around re-recording and re-releasing older songs. “Dilbar” (2018) was a re-imagining by Tanishk Bagchi of a 1999 Hindi film song; it became a massive hit and is now on T-Series’ channel because the company also owns the original. Saregama’s Carvaan — a hardware product playing classic film songs — has sold millions of units to a generation that wants the old recordings, while T-Series is selling new versions of the same songs to a different generation. India is the country where a 1968 song still has a real economic life as 2025 covers, 2026 sync placements, and 2027 AI re-renderings, all without anyone calling this anything unusual.

A song-first platform launched in India is not asking the audience to learn a new behaviour — it’s asking them to do, with new infrastructure, what they already do. Smule has 7-8 million Indian users performing each other’s covers; Suno has a meaningful Indian creator base; Coke Studio Bharat is fundamentally a “song-stylist takes new version of song” platform inside a content brand. The cultural muscle memory is there. The product layer that connects it to clean attribution and routed royalties is not.

The wedge customer is the Indian songwriter — composers, lyricists, indie singer-songwriters with EPs and bandcamp pages — who currently has no clean way to get publishing royalties out of the system when their work is covered, sampled, sung at weddings, and re-performed. IPRS exists; the metadata problem is real and well-documented. A platform that makes registration painless at submission and handles cross-DSP routing is solving a known problem with willing customers.

The business model, at VC altitude

We won’t build a full pitch deck here, but the unit economics intuition runs roughly like this.

Revenue layers. (1) Publishing administration take — the platform retains 10–15% of the songwriter share of mechanical and performance royalties on every distributed version (lower than Songtrust’s 15% because automation drops cost). (2) Creator subscription — $5–10/month for advanced features (AI rendering, premium stems, distribution to all DSPs, custom mechanical-license terms). (3) Licensing marketplace — sync placements for ads, films, sound design, with the platform taking 20–30%. (4) Distribution fee — flat per-release for songwriter-direct DSP launches, the way DistroKid charges.

Defensibility. Three real moats. (1) Network effects on the song side: more songs → more covers per song → more songwriter signal → more songwriters. (2) Metadata graph: the platform’s clean attribution data on derivative works is a uniquely valuable asset that improves with scale and is impossible to retro-build from streaming logs. (3) Songwriter retention: once a songwriter’s catalogue is registered through the platform with clean metadata and live royalty routing, switching costs are real.

TAM intuition. India alone has — using the karaoke-app proxy — 30 million Smule + StarMaker users actively performing covers, plus an indeterminately larger pool on Reels. The relevant universe of would-be songwriters is harder to size, but the 13,800+ artists earning $100K+/year on Spotify globally is a useful floor — at the long-tail end, there are likely 10–100x as many songwriters earning between $1K and $100K/year, mostly under-monetized because of the metadata/registration/cover-tracking gap.

Comparable business reference points. Songtrust runs a publishing-administration model and makes money. Splice’s marketplace model — 8 million users, sample sales, royalty-free model — works at scale. BandLab proves a free browser-based creation product can hit 100 million users. Smule proves karaoke-as-engagement is a real business. None of these has stitched the four together; the integration is the opportunity.

What a $5–10M seed funds. A six-engineer team (publishing-admin automation, GenAI integration, stems pipeline, mobile-first community product), a publishing operations lead with PRO/IPRS experience, an India-first launch focused on Punjabi and Hindi indie songwriters (who have shown the most willingness to release outside the film system), and a 12–18 month runway to prove the songwriter-side flywheel in one music vertical before expanding.

What kills it

We list the four risks worth taking seriously, and what we’d watch for.

(1) Major-label adversarial response. UMG, Sony, and Warner have an interest in keeping the canonical-recording model alive, because they own most canonical recordings. If the platform succeeds at making derivatives a real market, the majors can either compete (release their own song-first platform) or restrict licensing of major-controlled songs from being submitted. The mitigation: launch with songwriters who own their own works (indie singer-songwriters, regional pop) and let the major catalogue come on its own terms.

(2) AI legal ambiguity. Sony’s pending Suno fair-use ruling, expected summer 2026, will define what AI covers can legally do. If the ruling is restrictive, the AI-rendering layer of the platform is harder to ship. The mitigation: AI is a feature, not the foundation. Smule and karaoke models work without AI.

(3) Royalty infrastructure friction. PRO registration and metadata routing actually working in India — at the speed and accuracy a consumer product needs — is unproven. IPRS’s track record is improving but the system isn’t yet “API-call clean.” The mitigation: build the metadata/registration layer in-house and treat IPRS as a downstream partner that gets cleaner data than they’re used to receiving.

(4) Songwriter-supply problem. The product’s flywheel starts with songwriters submitting good songs. If they don’t, no covers happen, no flywheel. The mitigation: launch with a small set of credentialed songwriter partners (50–100), seed the community with high-quality material, and use the cover-volume data as an incentive for new songwriters to submit.

The biggest risk is none of those, though. It’s that the product needs to feel like a music platform users want, not a publishing-admin tool with a community bolted on. Smule is fun. BandLab is fun. Suno is fun. The song-first platform has to be fun first and infrastructure second, in that order. If it leads with metadata, it dies.

The 1955 inversion

Here’s the closing thought, and the reason this thesis matters beyond product. Rock and roll’s 1955 rupture was a one-time historical event — it took a particular set of artists, a particular technology (the 7-inch single), a particular generation, and a particular cultural moment to push the music industry from the song-as-product model to the artist-as-product model. The shift was real and lasting. But it was also partial. India’s mainstream music industry never went through it. The cover-version corners of every market never went through it. The publishing side of the industry never went through it.

What the 2026 stack lets us see clearly, for the first time, is that the artist-as-product model was a technology-shaped accident — a model that fit the constraints of physical singles, FM radio, and label-controlled distribution. Now that the constraints have moved (zero-cost AI rendering, stem-perfect separation, automated royalty routing, attention markets that reward many small uses over one big play), the song-as-product model is competitive again. Not as a nostalgic revival, but because the math finally works.

The product question is who builds the platform that runs on this math. The cultural question is whether listeners — particularly Indian listeners, for whom the song-first model is already native — find a product that lets them participate in the singing rather than just the listening. The answer, we think, is yes, but only if somebody builds it on purpose.

Tin Pan Alley had the right idea. They just didn’t have a server.