Apple's Huxe Acqui-Hire Leaves 1 to 6 Engineers in Play

On October 9, 2026, 9to5Mac surfaced a Digital Markets Act filing Apple quietly logged with the European Commission on June 9. The filing confirms Apple is acqui-hiring Huxe, the AI audio startup from three ex-NotebookLM developers, but the operative word is "certain": Apple will hire only some of the staff and take a non-exclusive license to the IP.

For anyone sourcing AI audio or personalized-assistant engineers, that single word opens a two to three week window on one of the smallest talent pools in AI.

What the DMA filing actually said

Apple reserved the right to hire a subset of Huxe's team and left the rest in play. The exact language, per the filing surfaced by AI Weekly, is that Apple will have "the right to make employment offers to and hire certain employees of Huxe AI, Inc." and "receive a non-exclusive license to Huxe's intellectual property rights."

Two things to notice:

• "Certain employees" is M&A counsel's term of art for "we picked some, not all." On a company of this size, that is a short list of names, not a rounding error.
• Non-exclusive license means Huxe's founders retain the right to use or relicense the underlying tech. The IP is not locked up. The people are not locked up either, unless they sign.

This is Apple's fourth disclosed AI deal of 2026, following a talent-and-IP arrangement with virtual avatar startup Animato in May. The pattern is consistent: small team, narrow IP license, no press release.

The Huxe cap table in plain numbers

Huxe is a 5 to 9 person company that raised $4.6M, shut down on May 28, 2026, and now has a partial Apple offer landing on part of the team. The arithmetic for recruiters is unusually clean.

Huxe was founded in December 2024 by three ex-NotebookLM staff: Raiza Martin (project lead, now ex-Gemini Notebook since Google's rebrand), Jason Spielman (designer), and Stephen Hughes (engineer). The round was led by Conviction and Genius Ventures, with angels including Figma CEO Dylan Field, Google Research chief scientist Jeff Dean, and Vercel CEO Guillermo Rauch (named in Japanese pre-seed reporting).

Headcount sources diverge, which matters:

• The open professional graph (April 30, 2026 snapshot): 5 employees
• note.com reporting mid-development: 9 people

Call it 5 to 9. Apple takes "certain." That leaves between 1 and 6 people quietly in the market, plus a non-trivial NotebookLM alumni diaspora in adjacent orbits.

47

Why this is a once-a-year sourcing event

Fewer than fifty people on earth currently hold a title that reads "AI audio engineer" in any flavor, and Huxe just put a measurable fraction of them back in motion. In Refolk's index of professional profiles, exactly 47 people globally carry a current title matching "Voice AI," "Audio ML," or "AI Audio Engineer." Top employers in that slice include pyannoteAI, Hugging Face, Nimble AI, and Iyuno.

That is the entire self-identifying pool. For context, there are more staff engineers at a single mid-size SaaS company than there are people in the world who publicly label themselves this way.

The title is unstandardized, and that is a sourcing tax

The Huxe-adjacent pool fractures across three near-identical labels. Any Boolean search on just one of them misses most of the market.

If you only search LinkedIn for "Voice AI Engineer," you miss roughly 79% of the already tiny pool. The fragmentation is the kind of friction where plain-English search beats Boolean: describe the person ("someone building generative audio or personalized voice assistants, ex-Google or ex-ElevenLabs, San Francisco or Paris") and let the retrieval do the title normalization. That is the exact gap Refolk closes.

Three reasons LinkedIn will mislead you for two more weeks

LinkedIn is a lagging indicator in acqui-hires, and this one is a textbook case. As of October 10, all three Huxe founders still listed Huxe as their current employer. Any recruiter filtering on "current company does not equal Apple" will miss them until the offer letters are signed and press-cleared, which typically trails the DMA disclosure by weeks.

Three faster signals:

1. GitHub commit history. Watch for a drop-off on Huxe-affiliated repos and a corresponding uptick on personal or new-org repos. Ex-NotebookLM engineers tend to keep public commits live.
2. Twitter/X bio changes. Bios flip before LinkedIn flips, because they do not trigger the same network-wide notification.
3. Domain and ENS changes. Personal domains, mail-forwarders, and ENS handles get updated when someone is quietly setting up a next thing.

The investor overlap is a hiring moat

Jeff Dean and Dylan Field personally backed Huxe, which means any ex-Huxe engineer taking a head-on competing offer risks burning the two most valuable AI references in the Valley. Jeff Dean is still at Google Research. Dylan Field runs Figma. Both are reference calls that get picked up on the first ring, and both are weighed heavily by future employers and investors.

The practical read for sourcers: do not pitch ex-Huxe staff into direct competitors. Pitch them into adjacencies where the reference call stays warm.

• Awkward: ElevenLabs, Spotify generative audio, Adobe audio, Meta audio research.
• Clean: Granola, Descript, pyannoteAI, Nimble AI, Hugging Face audio, or a stealth founder-mode play.
• Clean and under-pitched: personalized assistant startups that are not voice-first but want voice, and consumer podcast tooling with a research-and-summarization angle (the actual Huxe thesis).

The founder trio also retains rights to the Huxe IP under the non-exclusive license. That makes a second startup, with the same three names, a plausible 2027 event. Early relationship-building now is the entire game.

Who to pitch against Apple, with comp framing

Apple's comp ladder caps and visa constraints are typically what drive the "certain employees" exclusion in the first place, so the leftover 1 to 6 Huxe staff are usually the ones who either cannot or will not fit a FAANG band. Price accordingly.

Good landing-spot pitches, roughly in order of fit:

• pyannoteAI (France). Voice AI research shop. Overlaps culturally with the NotebookLM research ethos. Paris anchor matches a chunk of the global pool.
• Granola / Descript. Audio product shops that would pay a premium for anyone who shipped the Huxe personalized-news experience.
• Nimble AI, Hugging Face audio. Both show up as top employers in the Refolk slice. Known quantities, warm references.
• Stealth founder-mode startups. The investors on Huxe's cap table are exactly who you call to seed a new round. Conviction (Sarah Guo) is the obvious first ring.

The DMA filing is now a free sourcing feed

The EU's Digital Markets Act transparency obligations are the only reason this deal became public, and the EC's filings list is a free, recurring deal-flow pipeline for acqui-hire targets. Absent the DMA, Apple would have closed Huxe silently, the way most acqui-hires of 5 to 9 person startups vanish.

Set a recurring watch on the EC transparency list. Every "certain employees" phrasing in a filing is a short, exploitable sourcing window on a small, defined pool. Combine the filing (who got acquired, who got left) with a people-graph query (who worked there, who left recently, who is in orbit) and you have a repeatable playbook.

Every "certain employees" clause in a DMA filing is a sourcing window with a name attached.

A 14-day plan for the Huxe window

You have roughly two to three weeks before Apple's offers are signed, press-cleared, and LinkedIn starts flipping. Here is the sequence that uses the window.

1. Day 0 to 2. Pull the full Huxe roster from the open professional graph and cross-reference with GitHub contributors on any public Huxe repos. Target list: 5 to 9 names.
2. Day 2 to 4. Build a second ring: NotebookLM and Gemini Notebook alumni who left Google in the last 18 months. Many will not label themselves "AI audio."
3. Day 4 to 7. Reach out through warm paths first (Conviction portfolio founders, ex-NotebookLM teammates still at Google) rather than cold InMail. This pool reads InMail at single-digit rates.
4. Day 7 to 14. Pitch adjacencies, not head-on competitors. Keep Jeff Dean and Dylan Field's reference calls warm by respecting the obvious conflicts.
5. Day 14 onward. Set a DMA watch. The next Apple, Meta, or Google acqui-hire will surface on the same four-month lag, and the "certain employees" phrase is the trigger to look for.

$920K

The Huxe acqui-hire is small in dollars and large in signal. The ex-Google AI audio talent pool is one of the narrowest in the market, the Huxe cap table and NotebookLM diaspora are briefly addressable, and the DMA filing told you exactly when the clock started. Most recruiters will read the 9to5Mac story and move on. The ones who treat "certain employees" as a shortlist, not a footnote, will have the entire leftover pool mapped before Apple's offer letters go out.

FAQ

How many Huxe employees did Apple actually hire?

Apple did not disclose a number. The DMA filing uses the phrase "certain employees," which in M&A language signals a deliberately partial pickup. Published Huxe headcount ranges from 5 (the open professional graph, April 2026) to 9 (earlier reporting). The arithmetic therefore puts the leftover pool at roughly 1 to 6 people, plus the founders if they decline, plus the broader NotebookLM alumni network that was never part of the deal.

Why is "ex-NotebookLM" a legacy search term now?

Google rebranded NotebookLM to Gemini Notebook earlier in 2026, so current Google staff titles reference Gemini Notebook, not NotebookLM. Alumni who left before the rebrand still use the old name on LinkedIn and in bios, which actually helps sourcers: the "ex-NotebookLM" string on a current profile is a fairly reliable alumni filter. Searches should include both strings.

Can ex-Huxe engineers really take competing offers given the Jeff Dean and Dylan Field investor overlap?

Legally, yes. Practically, the reference-call risk pushes them toward adjacencies rather than head-on competitors. Pitching an ex-Huxe engineer into ElevenLabs or Spotify generative audio is awkward and reads badly to the Valley references they will need for the next round. Granola, Descript, pyannoteAI, Nimble AI, Hugging Face audio, or a stealth startup funded by Conviction or Genius Ventures are the clean pitches.

What is the fastest way to find the 47 people in the AI audio title pool?

Boolean search on a single title string misses 60 to 70% of the pool because the title fragments across "Voice AI Engineer," "Audio ML Engineer," "AI Audio Engineer," and long-tail variants. The faster path is plain-English retrieval against a people graph that normalizes titles for you. Describe the person you want (ex-Google or ex-ElevenLabs, generative audio or voice agents, San Francisco or Paris) and let the index handle the three-way title split.