AI chip startup Etched defies skeptics, hits $10.3B valuation from big-name investors

We have spent the last three years in a state of GPU worship. Nvidia became a multi-trillion dollar company because they owned the only shovels in town while everyone else was frantically digging for gold. But if you talk to hardware founders, there is a quiet, nagging realization: general-purpose chips are inherently inefficient. They are built to do everything, which means they do one specific thing—running Large Language Models—with a lot of wasted energy and overhead.

Etched, a startup founded by three Harvard dropouts, is betting their $10.3 billion valuation on the idea that the era of the general-purpose GPU is peaking. They aren't trying to build a better graphics card. They are building an application-specific integrated circuit, or ASIC, designed solely for inference. If they are right, the cost of running AI models could drop by an order of magnitude. If they are wrong, they are building a very expensive monument to a specific snapshot in architectural history.

The Specialist vs. The Generalist

To understand why investors are handing out decacorn valuations to a company that doesn't yet dominate the data center, you have to understand the bottleneck. Nvidia’s H100s and B200s are masterpieces of engineering, but they are designed to handle a massive variety of tasks. They handle training, they handle different types of math, and they maintain legacy compatibility.

Etched is taking the opposite approach. Their chip, Sohu, is hardwired for Transformers. This is the architecture that powers GPT-4, Claude, and Llama. By stripping away the hardware required to run other types of code, Etched claims they can run these specific models significantly faster and with much less power. In the hardware world, specialization is the ultimate form of optimization. When you know exactly what the math looks like, you don't need a flexible brain; you need a powerful, dedicated muscle.

The Risk of Being Right Too Early

As a founder, the biggest fear with an ASIC is the "hardwired" part. Software moves at light speed. Research papers come out every week proposing new ways to structure neural networks. If the industry suddenly pivots away from the Transformer architecture to something entirely new—like State Space Models or something we haven't named yet—an Etched chip becomes a very high-tech paperweight.

However, the bet here is that the Transformer is the "SQL" of our generation. It is the fundamental building block that is stable enough to build hardware around. Etched is banking on the idea that even if the models change, the core mathematical operations will remain consistent enough for their specialized silicon to stay relevant. It is a massive gamble on the technical direction of the entire industry.

What This Means for Builders

If you are building an AI-native startup today, your biggest line items are likely compute and talent. We have been living in a world of artificial scarcity where high inference costs dictate what products are viable. You might want to build a feature that requires 10,000 tokens of context for every user interaction, but the margins don't allow it because the GPU rental fees would kill you.

Specialized chips like Etched change the unit economics of the entire ecosystem. If inference becomes 10 times cheaper because we are no longer using general-purpose GPUs to do specialized work, then an entire class of currently "impossible" applications becomes profitable overnight. This is the moment where we move from "AI as a luxury" to "AI as a utility."

• Lower Entry Barriers: Cheaper inference means bootstrapped startups can compete with venture-backed giants on feature depth.
• Architecture Lock-in: Builders need to watch if hardware specialization forces us to stay with Transformers longer than we should, simply because the hardware is too good to ignore.
• Decentralized Compute: Efficiency gains could eventually move high-order inference out of massive data centers and closer to the edge.

The Infrastructure Pivot

The involvement of big-name investors in this $10.3 billion round isn't just about chasing the next Nvidia. It's about hedging against the current monopoly. Sophisticated investors know that the current margins on GPU rentals are unsustainable. The market is desperate for an alternative that doesn't require a blood oath to a single chip manufacturer.

Etched isn't just selling a chip; they are selling a world where the "Nvidia Tax" no longer exists. For founders, this is the signal to start thinking about what your product looks like when compute is no longer the primary constraint. We have spent so much time optimizing our prompts and our RAG pipelines to save on tokens. What happens when you have a surplus?

The Skeptic's Corner

I have to stay grounded here. We have seen many "Nvidia killers" come and go. Most of them fail because software is harder than hardware. Nvidia’s real moat isn't just the silicon; it's CUDA—the software layer that everyone uses to write their AI code. Etched has to convince developers that the performance gains are worth the headache of switching to a new stack.

Furthermore, hitting a $10.3 billion valuation puts a massive target on their back. They are no longer the scrappy underdog; they are a major player that has to deliver actual silicon that beats the world's most valuable company. The jump from a design on a screen to a chip in a server rack is where most hardware startups die.

The future of AI isn't just better code; it's smarter atoms. If we keep trying to run the world's most complex intelligence on general-purpose chips, we are just burning money to stay warm.

For those of us building in the trenches, the Etched story is a reminder that the layer below us is still shifting. We shouldn't get too comfortable with the current cost structures or the current leaders. The hardware wars are just getting started, and the real winners will be the founders who are ready to capitalize on the collapse of inference costs.

The Takeaway

Specialization is the inevitable next step for AI infrastructure. Etched is leading the charge by betting that Transformers are here to stay. For builders, this means the cost of intelligence is headed toward zero. Stop worrying about token counts and start worrying about what you'll build when compute is finally abundant and cheap.

Read the original at TechCrunch AI →