You speak into your phone and the transcript appears without a wait. You drive onto the highway and the lane lines move with you on the screen. The camera at your door recognizes who is coming home. None of this sends the data very far.
It relies on a very small chip nearby, closest to the data and farthest from the cloud. In the first week of November, two checks landed on this side: one for a chip, one for a model. The amounts are not startling, but the direction is clear.
Compute is splitting in two. One part sits in data centers, growing larger, powered by electricity and water. The other sits beside people, growing smaller, powered by batteries and heat dissipation. The second part used to be easy to miss. Now people are lining up to fund it.
Two Checks on November 4
On November 4, Jiemian News learned from people familiar with the matter that Houmo AI, a compute-in-memory AI chip company, had completed a B round of several billion yuan. The report came from those sources; the company did not publish the details.
Investors include industrial capital from upstream and downstream, state capital, and international investors, with several existing shareholders adding more. The money will mainly expand deliveries of the M50, advance the next generation of products, and complete the software toolchain.
On the same day, DenseAI, a company working on physical-world models, completed its third round. Within half a year, it has raised several hundred million yuan in total. Note that this is the sum of three rounds, not one, and the distinction matters.
Neither company is listed. There is no market value to calculate and no valuation to cite. What can be confirmed is the round, the type of investors, and the use of funds. Everything else is guesswork, and one more sentence would be speaking for the company.
Tens of Billions, No Finer Number
"Several billion yuan" is a range. It does not tell you the exact figure, only that it falls in the billions. Public reports carry no precise number, and this article will not supply one. Anything invented would be false, so it is better to leave the space blank.
Houmo AI was founded in 2020 and is headquartered in Beijing. It follows the compute-in-memory route. The name sounds convoluted, but the idea is simple: move computation into memory so the data travels less.
The company says the M50, less than a year into mass production, has won more than 150 product design wins and serves more than 100 customers. A design win comes earlier than an order: it means a customer has decided to build a product around your chip.
A design win is not a sale; it is a start. It weighs less than an order and more than a sample. It means the chip has entered someone else's product list and is waiting for mass production. For a chip company, that list carries more weight than a launch event.
What Compute-in-Memory Saves
In today's chips, the computing unit and the memory unit are separate, like a kitchen and a warehouse with a corridor between them. For every dish, the ingredients have to be carried back and forth, and the farther the trip, the more effort it takes.
The carrying itself produces nothing but consumes power. The more data, the more congested the corridor, and the higher the electricity bill. Compute-in-memory puts the stove inside the warehouse, saving those few steps. On the edge, that bill is especially heavy.
According to the company, this structure cuts the area of the core computing unit to about 30 percent of a conventional architecture and running power to about 10 percent. That is the vendor's own figure, not an independent measurement.
The benefit is direct: the same kilowatt-hour can run more inference, and the same board can fit into a smaller device. What the edge lacks most is never peak compute. It is power and heat, and whether the device can stay on for a long time.
The Model Moves to the Edge Too
What DenseAI does sits on a different layer from the chip. It trains a "world model": given the current frame and an action, predict what happens in the next frame, giving a machine a bit of common sense.
The capability sounds abstract, but its uses are concrete. A robot reaching for a cup, a car judging whether the vehicle ahead will brake hard, a drone flying through trees: all of them have to guess the next step first. A wrong guess is expensive.
Its model is called DenseWorld. The public description lists three keywords: next-state prediction, object-centric representation, and geometric-semantic self-supervision. In plain language, it teaches a model where things are and how they will move.
Models like this have to run on robots, cars, and drones, which puts them on the edge. The larger the model, the more it needs a power-efficient chip; the more efficient the chip, the more boldly the model can be placed at the edge instead of calling the cloud for everything.
The edge is not a data center shrunk and moved closer. It makes trade-offs within limited power, limited cooling, and limited space, cutting everything that can be cut and keeping only what cannot be avoided.
Another Curve Outside the Data Center
The activity at the edge is not the same story as the change inside data centers, but the two confirm each other. IDC data shows that in the first half of this year, China's data-center AI accelerator shipments reached about 2 million units, and the pool is still growing slowly.
Non-GPU accelerators accounted for 1,016,000 units, up nearly 80 percent year on year, surpassing GPUs at 991,000 units for the first time. GPU shipments fell over the same period, down more than 20 percent. One up, one down, and not a small fluctuation.
That figure describes data centers, not the edge. Treating it as an edge result would be a misreading. But it shows one thing: the shape of compute is changing, and no single type of chip can do every job.
Inside the data center, more work is being handed to specialized chips. At the edge, some work is being taken back from the data center. Both directions end at the same question: how much work can be done per unit of electricity. The logic is the same.
Closer to People, Harder the Math
The hard part of an edge chip is not speed but stability. Phones throttle when they heat up, automotive chips must pass certification, and cameras stay on for years. Any one of these is enough trouble, and none of it shows on a spec sheet. Users only feel lag.
Software is harder still. Whether developers are willing to rewrite code for a new architecture determines whether the chip sits in a showroom or ships in a million devices. This is what chip companies ultimately compete on.
That is why Houmo put part of this round into the toolchain and the ecosystem. Hardware can be taped out once; an ecosystem has to be moved person by person. The sentence is not new, but it is true every time, and it is never quick.
DenseAI faces the same thing. A world model has to be sold to robotics companies, automakers, and drone makers. What is sold is not a disc but a whole training and testing process, adjusted with each customer. That is slower than selling hardware.
Money flows to the chip closest to you for a simple reason: however strong the cloud is, its capability has to travel through networks, latency, and electricity bills before it reaches your hand. Some things cannot wait that long, and cannot afford to.
You will not see this chip at a launch event. It hides behind a car light, inside a camera, along the arm of a pair of glasses. One day the device will respond a little faster, and you will probably not think of it. That is its place.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.