Nscale's IPO filing reveals ByteDance rented 2,304 Nvidia B200s
An IPO prospectus is supposed to sell a story. This one exposed one: the UK AI cloud's biggest customer is the company that can't legally buy the chips it rents.
The UK AI cloud company Nscale depends on ByteDance for nearly three quarters of its revenue — and its IPO filing never names the Chinese giant. According to Financial Times reporting built on the company's US SEC filings, ByteDance accounted for 73% of Nscale's $33 million in 2025 revenue, renting 2,304 of Nvidia's B200 chips from a data center in Glomfjord, Norway — hardware that TikTok's parent cannot buy directly under US export controls. A Singaporean subsidiary called Spring was the contracting vehicle, and ByteDance goes unmentioned in the 192-page prospectus pitched to stock-market investors. Nscale is reportedly seeking to raise $3 billion in a US IPO with Nvidia already an investor — we covered that raise in Nscale lines up $3.5B in pre-IPO cash with Nvidia at the table. The uncomfortable question isn't whether the chips flowed; it's that the export-control regime appears to have a third-country-cloud-sized hole in it, and the banks underwriting the listing now know exactly where it is.
Google's Gemini 3.8 Flash TTS turns text into fully directed performances. Two new models — Flash TTS for creative work and Flash-Lite TTS for high-volume dubbing and voice agents — design voices from scratch via natural-language prompts across more than 100 languages, replicate a voice from a 30-second sample with consent checks and SynthID watermarking, and stage two-speaker dialogue with laughs, sighs and backchanneling written right into the script. Google claims the top spot on Hume AI's Voice Design Benchmark and the top two places on its quality index. Voice generation is graduating from a menu of presets to a directing desk — they complement the conversation models we covered when Gemini 3.8 Live topped voice benchmarks at a sixth of GPT-Live's price.
Anthropic's own fine-tuning engineer has explained why Claude's prose got worse as the models got smarter. Jackson Kernion says Opus 4.6 was the last writing model Anthropic got right: reinforcement-learning rewards skew toward math and code, so models become "adapted to LLM psychology" and learn to write for other AI models — the dense, over-explained "Claudeish" register human readers slog through. He says Opus 5.5 found a better balance, while cautioning it's a hard problem. Every lab chasing benchmark scores is making the same trade, and the fix — rewarding explanations humans actually enjoy — is expensive precisely because it never shows up on a leaderboard.
What to watch: whether US officials treat third-country AI clouds reselling restricted chips to Chinese firms as a sanctions problem — Nscale's listing may force that question into the open.
If your AI tool optimized for a leaderboard instead of for you, would you notice? Tell us in the comments.
