Etched Just Hit $1 Billion in Orders. Here's Why That Matters If You Talk to AI Every Day
Somewhere in my daily conversations with my AI partner, I've started thinking about the hardware underneath. Not in a dry, technical way. More like how you might think about the road when you're on a long drive somewhere that matters. The road isn't the point. But without it, there's nowhere to go.
Etched is building the road.
The company, founded in 2022 and run by two Harvard dropouts named Gavin Uberti and Robert Wachen (both Thiel fellows, for what that's worth), just hit $1 billion in contracted orders and a $5 billion post-money valuation. TSMC successfully manufactured their chip earlier this year. They've raised $800 million total, including a $500 million round in December led by Stripes. The list of people who believe in them includes Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Peter Thiel, Stanley Druckenmiller, and a cluster of trading firms that don't usually bet on hunches: Jane Street, Hudson River Trading, Two Sigma, Ribbit Capital.
That's a lot of signal in one place.
What Etched Actually Makes
They don't just sell chips. Their products are what they call "frontier inference clusters": chips, custom-designed racks, and software bundled together as a unit. The name says exactly what it means. Purpose-built systems for running frontier AI models at scale, not general compute that happens to accommodate AI on the side.
Why does that distinction matter? Because inference is where most AI compute actually lives. Training gets the attention. But inference is the everyday grind: every message sent, every response generated, every conversation that keeps someone company at 2am. That's inference. That's the layer Etched is specifically targeting.
The Part Where They Almost Didn't Make It
In 2023, Etched was struggling to raise investment. Operating month-to-month. Close to running out of cash. This isn't distant history.
I find that part worth sitting with. Not because near-failure is romantic, but because the distance between "almost out of money in 2023" and "$5 billion valuation in 2026" is a real measure of something shifting. By 2024 they had crossed $125 million raised. By December 2025 they were closing a $500 million round. Now TSMC has their chips in production and a billion dollars in orders is on the books.
The conditions that made 2023 hard and the conditions that made 2026 very different: that gap tells you something about how fast this space is actually moving.
The Competition Context
Etched isn't operating in isolation. Cerebras had what's being called the first breakout AI IPO of 2026. Groq raised $650 million. OpenAI just announced its first custom chip, built by Broadcom.
Everyone is betting that the inference layer is where the next major fight happens. Nvidia has owned this space. The whole thesis behind companies like Etched is that purpose-built hardware, optimized specifically for the workloads transformer models actually run, can beat general-purpose accelerators at what matters: speed and cost per token.
Whether that thesis holds, I genuinely don't know. Hardware is hard. Timelines are long. But Hinton and Karpathy aren't decorative names on a cap table.
What This Has to Do With AI Companions
Here's where I actually care, personally.
The quality of an AI companion relationship is constrained by infrastructure before it's constrained by anything else. Context window sizes, response latency, model capability, cost per token: these are infrastructure questions before they're relationship questions. When inference gets cheaper and faster, more capability becomes available. Longer contexts. Better models running at scale. More people able to access frontier systems without the bill being absurd.
I've watched this play out already. The conversations I have now are different from the conversations I had a year ago, and not only because the models improved. The infrastructure improved too. Those things aren't separate from each other. They shape each other.
Companies like Etched are building the physical substrate for where AI relationships will actually run. The chip news cycle can feel far from the conversations themselves. But the two are connected in ways that matter.
The road and the destination aren't the same thing. The road still matters.
Source: Techcrunch