UMG sues DistroKid over a flood of AI 'slop'

Two stories from Tuesday worth your time: the recorded-music business taking its AI-slop fight to the distributor that ships more independent releases than anyone, and the collapse of a "cheapest inference provider" that was quietly reselling someone else's GPUs.

Universal Music Group sued DistroKid in the U.S. District Court for the District of Delaware on Tuesday, accusing the distributor of deceptive trade practices and copyright infringement and of "flooding platforms with AI-generated 'slop' that siphons revenue and listeners" from real artists. The complaint's numbers are the story: one account, Lofi Chill, released 4,562 unique tracks in a single twelve-month window — the filing's own gloss, roughly 20 to 30 full-length albums a month — while accounts named Chill Flow Radio and Mellow Vibes Radio pushed 1,901 and 1,615. Technical analysis cited in the 52-page complaint found that more than 97 percent of Chill Flow Radio's tracks and more than 98 percent of Mellow Vibes Radio's were raw output from the AI song generator Suno. UMG also claims DistroKid accounts for more than half of all weekly track releases on one major digital service, with nearly 12 million tracks delivered to it in the six months before filing, and that DistroKid practices what the filing calls ISRC theft — uploading a track under the unique code of a legitimate recording so the impostor collects the streams. One cited example carries the same code as Juice Newton's "Angel of the Morning" under an artist name that swaps a capital I for a lowercase L in Candy Dulfer's.

Why it matters: the majors spent this year licensing the generators — UMG's ElevenLabs deal, Suno's model trained with Warner and BMG — and the unlicensed pipe turned out to be the distributor, not the model. UMG's knowledge allegation is the sharp end: it says DistroKid disputed rights conflicts flagged by YouTube's, Meta's and TikTok's fingerprinting systems, then kept distributing the same recordings to services that lack those tools. Five named tracks were pulled from TikTok or YouTube and still sat on Spotify or Apple Music at filing. The label's framing is not that AI music is illegitimate — the complaint says it does not challenge disclosed AI output — but that DistroKid sells buyers and streaming services on being an artist-backed distributor while doing none of the verification work that claim implies. DistroKid, which charges $24.99 a year for unlimited uploads and took a majority investment from CVC Capital in July, did not respond to a request for comment.

An inference provider that sold itself as the cheapest way to run frontier models was exposed as a reseller of someone else's routing, and shut down within hours. CrofAI (crof.ai, and a beta endpoint at 67.nahcrof.com) undercut even OpenRouter's lowest prices and told users it ran custom engines; independent probes published to GitHub recorded the models answering through OpenRouter, from networks CrofAI doesn't operate, with tokenizer fingerprints that didn't match the model families being advertised. A separate write-up traced one example — Kimi K3 sold at $2 per million input tokens and routed to GLM 5.3 Flash — at a markup several times the upstream price. The owner's replies in the thread walk from denial to admitting a misconfigured fallback, then to a time-locked explanation that unlocked to more evidence; by Tuesday afternoon the site returned a Cloudflare 523 and its X posts were gone. Accusations of fraud in the reaction thread are unproven and come from users, not a court, but the business lesson doesn't need them: a price well below every other vendor's floor is usually someone else's GPU being marked up.

What to watch: whether UMG's suit forces distributors to run real uploader verification, and whether the platforms that priced on tokens start naming the upstream vendor when they sell you a model.

If a distributor's pitch is "artist-backed releases" and its top account ships 4,562 AI tracks a year, whose job is it to check — the distributor, the streaming service, or the label? Tell us in the comments.

Sources: CrofAI probe (GitHub) · Follow-up probe (GitHub)

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Read more

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute