Intelligence Etched in Stone
Humanity is doing something insane right now. We are building data centers the size of cities, putting power plants next to them, launching satellite networks, and cooling rooms full of supercomputers that consume hundreds of kilowatts with liquid. To run AI. While believing this is the future.
But history tells a different story. Every technological revolution began with monstrous prototypes, and once a practical breakthrough appeared, those monsters disappeared in an instant. Do you remember ENIAC? A vacuum-tube monster that filled an entire room. It showed the magic of computing, but it was slow, expensive, and impossible to scale. Then the transistor arrived, and everything changed. It evolved into workstations, PCs, and smartphones. The world chose to move beyond ENIAC instead of building more of them. The GPU data centers we are building now are AI's ENIAC. They work. They show us magic. But this is not the end.
Before you keep reading, I recommend going to the site below and asking it anything. It takes 30 seconds. You have to feel it for yourself.
chatjimmy.ai
An LLM appeared whose answer was already there the moment you pressed enter. We have lived assuming that lag is simply part of AI. That is why this is an unfamiliar shock that numbers cannot explain.
General-purpose computing changed the world because it became fast, cheap, and easy to build. AI will take the same path. The problem is that today's AI is far from that path. When you ask AI a question, it rests its chin in its hand and thinks for a long time. A coding assistant stares blankly for several minutes before giving you an answer, breaking your concentration. Even when you need a response in an instant, all you get is a leisurely reply. Talking to AI still feels like making an international phone call. Speak, wait, and wait some more. This latency is the wall between humans and AI.
The cost problem is even more serious. Running today's AI requires enormous amounts of equipment and money. HBM stacks, complicated input and output, cables, liquid cooling, advanced packaging, 3D stacking. Why is all of this necessary? Because where memory is stored and where computation happens are separate.
Think of it this way. Your brain is in Seoul, and all your memories are in a warehouse in Busan. Every time you want to remember something, you have to take the KTX to Busan and bring it back. Modern AI hardware has exactly this structure. Memory (DRAM) has high capacity and is cheap, but because it sits outside the chip, accessing it takes thousands of times longer than accessing something inside the chip. But you cannot put DRAM inside the computational chip either. The manufacturing processes are different. This contradiction creates all the complexity in AI hardware. To reduce the Seoul–Busan round trip, we lay HBM high-speed rail, build skyscrapers through 3D stacking, and run air conditioning through liquid cooling. Naturally, it consumes a lot of electricity, and the cost shoots through the roof.
Taalas turned this upside down at its core. Instead of fetching memories from Busan, it embedded them all inside the brain. It combined memory and computation on a single chip at DRAM-level density. Then it went one step further and built dedicated silicon for each model. Not off-the-rack clothes, but a tailored suit. Looking at the history of computing, deep specialization has always been the surest path to extreme efficiency. Taalas pushed that principle all the way.
How is it possible? It directly etches the knowledge AI has learned—the weights—into the metal layers of silicon. Literally, it is etching intelligence into stone. A single transistor holds a weight while also carrying out the multiplication. It remembers and thinks at the same time. To borrow founder Ljubisa Bajic's words, this is "not nuclear physics, but a clever trick that no one saw because no one had gone down this path." The chip's underlying structure stays the same, and only two metal layers are changed to fit a specific model. The same body with a different tattoo. Using TSMC's 6-nanometer process, a card that actually works can be produced in two months.
The HC1 chip etched with Llama 3.1 8B processes about 17,000 tokens per second per user. Nvidia H200 does 230, B200 353, Groq 594, Sambanova 932, and Cerebras 1,981. Everyone else is riding bicycles while Taalas is flying a jet. One card uses 200 watts. Even with 10 cards in a server, it uses 2,500 watts. That's fan-level power, and it can be plugged straight into a 30-year-old data center. One-twentieth the manufacturing cost, one-tenth the electricity. No HBM, advanced packaging, 3D stacking, or liquid cooling needed.
Of course, nothing is free. If a general-purpose GPU is a speaker that can play any song, a Taalas chip is a music box that plays one song perfectly. It is not intelligent in a general sense, and when the model changes, you have to produce a new chip. But it can adjust context size and support LoRA fine-tuning.
And crucially, we are approaching the threshold where models sufficient for everyday work become possible. Once frontier models improve a little more, there will come a time when they can be used for quite a long time for routine everyday tasks. At that point, the economics of a dedicated music box will work.
With Nvidia acquiring Groq for $20 billion, SoftBank swallowing Graphcore, and Intel reaching out to SambaNova, a major wave toward inference-specific chips is rising. Taalas stands at the most radical point of that wave. Its first product begins with Llama etched into silicon, but a mid-sized inference model will follow in spring, and a frontier model in winter.
Very fast AI is an entirely different kind of AI. Once latency below one-thousandth of a second becomes possible, scenarios we only imagined become real. It will feel less like an international phone call and more like running alongside the person next to you while talking. Taalas opening chatjimmy.ai as a beta even though its first model is not frontier-level shows its confidence that you should experience for yourself what becomes possible at this speed.
An autonomous vehicle has to finish making a decision within 0.01 seconds. A car traveling at 100 kilometers per hour moves 28 centimeters in 0.01 seconds. A ruler. About a handspan. Within that space, it has to recognize a pedestrian, make a decision, and turn the wheel. Drones are even more demanding. They have to avoid obstacles while flying at several dozen meters per second. Fly-by-wire in fighter jets, surgical robots, factory robots. In environments like these, sending a question to the cloud and waiting for a response is like mailing a letter to headquarters from a battlefield and waiting for a reply. People die from the network round-trip time alone.
Today's GPU inference cannot solve this. Putting hardware that consumes hundreds of watts and has to be liquid-cooled into a drone is like putting a refrigerator on a bicycle. Inside robots, inside cars, inside drones. Without the cloud. Without latency. A music box truly shines when the song it plays can decide between life and death.
We are still living in AI's ENIAC era. Huge, inefficient monsters are filling data centers. But history repeats itself. The transistor moment will come. For AI to fit into our hands, our pockets, and every machine we ride in, it has to become fast, cheap, and easy to build. ENIAC never spread everywhere. AI's ENIAC-like spread will not happen either.
The AI ecosystem needs new approaches. And challenges in this direction are urgently needed.
