It barely shipped chips, but its valuation hit $21B: can Etched challenge Nvidia in inference?

The bottom line

A startup that has yet to sell chips to a large-scale market has been valued at US$21 billion.

Etched’s latest funding round of $700 million in August pushed its valuation from $10.3 billion to $21 billion in less than a month. Its first customer is trading firm Jane Street, and the first rack was only delivered last month, but the company already claims to have secured more than $1 billion in customer contracts.

This is not "Etched has defeated Nvidia", but investors are betting on one thing: what AI lacks most in the future may not be training models, but the inference compute to run out model answers every day.

Chips are still being delivered, and valuations have already completed one lap.

What exactly does Etched sell?

It is not intended to be a general-purpose GPU that can do everything, but to design a dedicated system specifically for AI inference, which is the stage where the model generates answers.

The logic is simple: training a large model may consume a huge amount of compute, but when it comes to actually providing services to users, reasoning resources will be repeatedly consumed for every question and answer, code completion, and image generation. The more ubiquitous the model becomes, the more inference becomes like a never-ending electricity bill.

Etched hopes to use dedicated chips to make the model run faster and cheaper. The core indicator is not the "number of parameters" at the press conference, but:

• How many Tokens can be transferred per dollar;
• How much inference can be completed per watt of power;
• How long to wait from request to answer;
• Whether a complete set of racks can be deployed stably.

This is why investors discuss it next to Nvidia: not because it has replaced Nvidia, but because the inference market is large enough and may accommodate different types of hardware.

The headline number: valuation doubled in one month

In July, Etched was valued at approximately $10.3 billion after Series C financing; in August, after a new round of funding of $700 million, the valuation rose to $21 billion.

Participants include Jane Street, Kleiner Perkins, Sequoia, Andreessen Horowitz and Tiger Global. The company employs more than 400 people and already has working chips.

What excites the market even more is the customer signal: Jane Street is not only an investor, but also Etched’s first customer, having already obtained the first set of racks and deployed them to her own workloads.

This changed the Etched story from "there was a cool chip in the lab" to "someone actually plugged it into a production system."

But between these two sentences, there is still a gap between large-scale delivery, stability, software ecosystem and customer renewal.

Why did Jane Street buy first?

Trading firms are extremely sensitive to delays and have the ability to build complex infrastructure themselves. As long as a dedicated system can run certain model tasks faster, even if the advantage only occurs in a few high-value scenarios, it may be worth deploying.

This is not the same thing as ordinary consumers buying graphics cards. Jane Street does not need a chip to "do everything", it is more concerned about whether it can stably earn back the cost on specific tasks.

Therefore, Jane Street's first set of racks is a strong customer signal, but it cannot be directly extrapolated to "all cloud vendors will migrate."

Why is the capital market suddenly so excited?

Because AI money is spreading from model companies to infrastructure companies.

Large model companies are still competing to see who is smarter, but cloud vendors are beginning to face another problem: If everyone is calling AI every day, will power, racks, networks and inference chips become the biggest bottlenecks?

Etched is betting on the "specialization" route: when the model structure is relatively stable, special hardware can sacrifice part of the versatility in exchange for speed, energy consumption and cost advantages.

The risks are equally clear: If the model architecture changes too quickly, specialized chips may fall behind before delivery; if Nvidia continues to optimize software and systems, customers may not be willing to bear the cost of migration.

Is NVIDIA really going to be overthrown? Do not rush to share that claim

There is currently no evidence that Etched has beaten Nvidia in overall performance, shipment size, or revenue.

Reuters quoted experts as reminding that there are many cases in the history of semiconductors where "the technology was beautiful, but the company failed to make a big business." Etched's $2.1 billion valuation includes early pricing for inference market growth, customer contracts and the possibility of replacing Nvidia.

In other words, investors are not buying today’s income, but the scarce alternatives that may arise in the future.

This is the tension at the center of the story: the company is still proving it can become a durable business, while the market has already priced it like a future giant.

My verdict: Inference chips will get hot, but there won’t be just one winner

The AI industry is likely to see a coexistence of training chips, inference chips, network equipment and dedicated systems. If Etched can get its chips, compilers, rack delivery, and customer service right, it could indeed become an important player in inference infrastructure.

But for it to truly challenge Nvidia, it must answer at least four questions:

• Can chips be scaled from the first rack to hundreds or thousands of units?
• Is the customer contract a signed revenue, or a conditional future order?
• How high is the cost of software migration when changing models or frameworks?
• Will customers keep Nvidia at the same time, with Etched only complementary?

Until those answers emerge, $21 billion looks more like a high-priced option than a report card.

What to watch

• Etched’s chip shipment numbers and real customer list.
• Whether the Jane Street deployment scales to more workloads.
• The pace of redemption of the company's more than $1 billion in customer contracts.
• Whether Nvidia, AMD and cloud vendors launch stronger inference-specific products.
• Will the new model architecture changes reshuffle the dedicated chip route?

The last sentence:

The next battle for AI chips may not be who can train the largest model, but who can get answers faster and cheaper to every question the world asks every day.

Why this matters to you

If the cost of inference does come down, it could ultimately translate into cheaper APIs, larger free credits, and faster AI applications. But in the short term, the high valuations of chip startups will not automatically turn into user benefits; what is really worth seeing is whether the costs are transmitted to the product price, rather than the size of the numbers in the financing news.