This chip can only think one thought. That's why it's 100× faster.
A normal AI chip is a printer: it can print any page, slowly. This chip is a rubber stamp: one page only, instantly. Five steps to see why that works, and what it costs.
1What an AI chip actually does all day
An AI model is a giant list of fixed numbers. Llama 8B is a list of 8 billion of them, called weights. They never change while answering you.
To produce each word of an answer, the chip multiplies your input by those numbers and adds up the results. Billions of times. That's the whole job.
Everything an AI chip does: multiply by fixed numbers, add up, spit out a word. Repeat.
2Why GPUs are slow: the numbers live far away
A GPU can't fit 8 billion numbers inside itself. They sit in separate memory chips next door.
For every single word, the GPU must haul all 8 billion numbers across a bridge into the chip, use each once, throw it away, and haul them all again for the next word.
That hauling, not the math, eats most of the time and most of the electricity. This is the famous bottleneck of AI hardware.
The GPU's problem is a commute. The workers are fast; the highway is jammed.
3Taalas' move: print the numbers into the chip itself
The weights never change. So Taalas asks: why store them in memory at all?
They etch the 8 billion numbers into the chip's own wiring at the factory, right next to the calculators that use them.
No memory chips. No bridge. No hauling. The slowest part of a GPU simply doesn't exist here.
The catch is obvious and real: the numbers are frozen forever. One chip = one model, permanently.
Same math, no commute. The number and the calculator that uses it are fused together.
4How can a wire "multiply"? Watch one weight work
Chips get doubling for free: routing a wire one position over doubles the number it carries. So your input x instantly exists as x, 2x, 4x, 8x… on parallel wires. No work done yet.
Any multiplication is a sum of doublings: 5x = 4x + x. 6x = 4x + 2x. 13x = 8x + 4x + x.
So "multiply by 5" just means: connect the 4x wire and the x wire to an adder. The weight isn't stored anywhere. The weight is the choice of which wires get connected.
Each connection is a microscopic metal plug punched between chip layers at the factory. Plug punched = connected. No plug = not connected. 8 billion weights = 8 billion punch patterns.
Two metal plugs = the number 5. A different weight would be a different plug pattern on the same wires. This happens 8 billion times across the chip, all at once.
5How they build a new chip in 2 months, not 2 years
Normally a custom chip takes years: design everything, make ~80 photographic stencils (masks) that pattern each layer, ~$10M+ up front.
Taalas' shortcut: every layer except the plug layer is identical for every model. Calculators, wires, adders: all generic. Only the punch pattern changes.
So the factory pre-builds stacks of "blank" chips, and a new model only needs a few new stencils for the plug layer.
Their software reads a model's weights and computes the punch pattern automatically. Weights in → chips out, ~2 months, ~20× cheaper than rivals.
Like a book printer that keeps pre-bound blank books and only prints the final page of ink. Only the plug pattern is custom.
6Your remaining questions, straight answers
Can it be fine-tuned later?
Partly. The etched numbers can never change.
But the chip keeps a small patch of normal, rewritable memory for add-on corrections (LoRA: a thin extra layer, under 1% of the model's size, whose output gets added on top).
You train that patch on regular GPUs, then load it onto the card like a settings file.
Want a truly new model? New chip. No way around it.
What does the shrinking cost in quality?
To fit on one chip, each number is rounded from 16-bit precision down to 3 to 6 bits.
Expect a small but real quality drop vs. the original model, like a heavily compressed photo: fine normally, artifacts in hard cases.
On a GPU you could re-tune this. Here the rounding is etched in, permanently.
Could this run Kimi-K3-sized models?
One chip is already at the maximum size a factory can make, and holds 8 billion numbers.
A frontier model has ~1 trillion numbers: roughly 100+ chips, wired together with the cable ports on the card's edge.
Models like Kimi only use ~3% of their numbers per word, but every number must still physically exist on some chip.
Doable, and it's their late-2026 roadmap. But each big model becomes a 100-chip, multi-million-dollar bet.
What's the payoff and the risk?
Payoff: ~17,000 words/second per user, ~200 watts per card, ordinary air cooling, ~100× cheaper per word than GPUs.
Risk: the chip is unbeatable on its one model and worthless on every other.
If labs release better models faster than Taalas can re-print chips, stock dies on the shelf.
So their real product isn't the chip. It's the 2-month re-printing pipeline.
If you remember one thing
GPUs waste 90% of their effort hauling the model's numbers back and forth from memory. Taalas etches the numbers into the chip's wiring, so there is nothing to haul. Printer vs. rubber stamp: they gave up the ability to print any page, and got a 100× faster stamp.