Micro1 bids $12.5M to pry Spirit Airlines' data out of Google's hands
The AI data land-rush reached a bankruptcy courtroom this week, a 26-year-old audio editor finally got dragged into the modern era, and someone did the arithmetic on why your GPU is slower than the spec sheet promised.
An AI training-data startup is trying to outbid Google for a dead airline's archive — after the auction closed. Micro1 sent a $12.5 million offer to Spirit Airlines' legal team on Wednesday, two days before the bankruptcy court's September 9 hearing on whether to approve Google's $10 million winning bid, according to Bloomberg and Business Insider. Founder and CEO Ali Ansari called his offer "materially higher" and Google's bid "actually quite low" for a carrier that operated for decades — Micro1, which pays some companies up to $2 million for their data, sees decades of real operational records as undervalued. The court hasn't approved the auction result, and it already delayed the hearing once after former Spirit flight attendants objected over what happens to their data.
Whether a judge even entertains the offer is genuinely an open question. Bankruptcy scholars told Business Insider the answer turns less on the amount than on the integrity of the auction that preceded it: judges sometimes consider late bids that could generate significantly more money for creditors, but as Nancy Rapoport of UNLV put it, "a duly noticed, well-run auction generally won't get undone" — the higher bidder had its chance during the auction. Lindsey Simon at Emory was blunter: "The bankruptcy code isn't clear on this." What is clear is the market signal. Three AI companies — Google, Mercor, and now Micro1 — fought over a failed airline's emails, chats, and internal records, which tells you where the scarcity has moved now that public web text is picked over. We covered the first round of this fight in August — Google's Spirit Airlines data deal stalls as a bidder and attendants push back — and the unresolved question then is still the one now: employee records get far less legal protection than customer records, and they're the bulk of what's for sale.
Audacity 4 finally ships, and the rewrite is bigger than a coat of paint. The open-source audio editor that most of the internet has used at least once moved to non-destructive editing: trim a clip and you can drag its edge back out to recover the audio, and clips can now overlap instead of stopping dead against each other. Underneath is a Qt rebuild with HiDPI scaling, light and dark modes, movable panels saved as Workspaces, ASIO support on Windows, and a new .aup4 project format that converts older files one-way. It's the first real modernization since 2000, and it moves Audacity much closer to a conventional digital audio workstation.
The catch is that a rewrite this deep leaves things behind: time tracks, MIDI/Note tracks, the mixer, the macro manager and scripting pipe, and VAMP and LADSPA plugin hosting are all absent in 4.0, along with play-at-speed and some effects and export paths. Anyone whose workflow depends on those should stay on 3.x until they return. Note also that this one has almost nothing to do with AI — despite being filed under The Verge's AI feed, the release is a plain engineering story, and that's fine.
A LocalLLaMA post put a number on the question every local-model runner has asked: how fast should this thing actually go? The answer for text generation is that memory bandwidth, not compute, is the ceiling, because generating each token means streaming the entire set of model weights plus the growing KV cache out of VRAM. Divide your card's bandwidth by the model's weight size and you get a hard theoretical maximum; everything after that is software losses. Worked example: Qwen3.8 27B at Q4 is about 16.8 GB of weights, so on an AMD Radeon AI PRO R9700's 637 GB/s the ceiling is 38 tokens per second — and the poster measures 29 tokens per second in llama.cpp, roughly 76% of ideal. Stretch the context to 256K and a 5090 drops from about 106 tokens per second to about 53, because those cached bytes get re-read on every single token.
That 76% figure is the useful part. It means the gap between advertised bandwidth and what you actually get is a stack-quality problem, not physics, and it gives you a way to sanity-check any benchmark you read: if a claimed throughput exceeds bandwidth divided by weights, the claim is wrong, someone is offloading to system memory, or the number is prompt-processing rather than generation. We published a beginner-friendly walkthrough of the technique that sits on top of these numbers — AI 101 — What is speculative decoding? — and the rule from Friday's guide still applies: the base rate has to be honest before the speedup means anything.
What to watch: September 9, when the New York bankruptcy court decides whether Google's auction result stands — and whether a late bidder can reopen it.
If a dead company's internal records are worth $12.5 million to AI buyers, should the people who wrote those emails get a share? Tell us in the comments.
