ElevenLabs just doubled its valuation to $22 billion in a $300 million employee tender offer co-led by Wellington Management and T. Rowe Price. That’s a remarkable climb for a voice AI company that’s only been around since 2022, and it says something about where investors think the audio piece of the AI stack is headed.

The company makes text-to-speech and voice cloning tools that have become popular with content creators, game developers, and anyone building conversational AI products. The fact that two major institutional investors are willing to write checks at this valuation suggests the market for AI-generated audio is bigger than background tech for chatbots.

This is secondary money, so ElevenLabs isn’t getting the cash directly. But employee tender offers at doubled valuations typically mean the company is doing well enough that early employees want liquidity and investors want in badly enough to pay up for it. It’s also a signal that an IPO probably isn’t imminent. If it were, those employees would just wait.

Hardware design gets its AI moment

Flow Engineering closed a funding round at a $750 million valuation, backed by Valor Equity Partners, Atreides Capital, and Sequoia Capital. The company is building AI agents for hardware design, which is one of those unglamorous but enormous markets that doesn’t get much attention until someone raises at three quarters of a billion dollars.

Sequoia’s Roelof Botha joined as both an angel investor and board member, which tells you how seriously they’re taking this. Hardware design has been slow to adopt AI compared to software engineering, partly because the tooling is more complex and partly because mistakes are more expensive. You can roll back a bad code commit. You can’t un-fabricate a chip.

If Flow Engineering can make AI agents useful for hardware engineers, the market is massive. Every physical product with electronics inside it requires hardware design work, and the process is still heavily manual. The company hasn’t disclosed much about how its agents actually work, but getting Sequoia’s chairman on your board suggests they’ve figured something out.

Google’s Gemini 4 launch raises more questions than it answers

Google released Gemini 4 Argon and immediately restricted access to what it’s calling a “set of trusted cyber defenders.” The company says the model delivers frontier performance in software engineering, enterprise knowledge work, and cybersecurity defense. But if you’re not on Google’s approved list, you can’t use it yet.

This is a weird rollout strategy. Google says it’s “actively engaged in the U.S. government” on the release, which suggests either the model is powerful enough to raise security concerns or Google is trying to position itself as the responsible AI company ahead of regulatory fights. Maybe both.

The restricted access could be legitimate caution about a model that’s genuinely better at finding and exploiting security vulnerabilities. Or it could be marketing. Telling people they can’t have something is a good way to make them want it, and Google has been losing the AI narrative to OpenAI and Anthropic for the better part of two years.

Chief AI architect Koray Kavukcuoglu says the model is designed for “complex workflows” in software engineering and cybersecurity. That’s vague enough to mean almost anything, but if the government is involved in the release strategy, the cybersecurity piece is probably the real story. A model that’s meaningfully better at offensive security work would be worth restricting, at least initially.

The “trusted cyber defenders” framing is also interesting. It positions Google as the company that builds powerful AI for the good guys, which is good branding if you’re trying to avoid being regulated into oblivion. But it also means we won’t know how good Gemini 4 actually is until more people can use it. Benchmark scores are easy to game. Real-world performance is what matters, and that requires access.

Google says it’s marketing Argon as a “workhorse for coding and cybersecurity work,” which suggests it’s optimized for long, complex tasks rather than quick interactions. That would make sense if the company is trying to go after GitHub Copilot and other developer tools. But again, without broader access, it’s hard to know if this is a real breakthrough or just good positioning.

The AI agent hardware push is also heating up. OpenAI and Meta are both working on physical devices for their AI assistants, Dots and Muse respectively. Both companies are betting that people will want dedicated hardware for AI agents, which seems optimistic given how badly previous AI devices like the Humane Pin and Friend have flopped. But if you have infinite money and need to find the next platform before smartphones stop being the default interface for everything, you take some swings.

The throughline across all of this is that AI companies are hunting for the next big market. Voice, hardware design, cybersecurity, physical devices. Everyone’s looking for the application that breaks out beyond chatbots and code completion. ElevenLabs’ $22 billion valuation suggests audio might be one of them. Flow Engineering’s $750 million raise suggests hardware design could be another. And Google’s restricted Gemini 4 release suggests someone thinks cybersecurity is important enough to involve the government.

We’ll see which bets pay off. But right now, the money is moving fast and the valuations are climbing faster.