For the past decade, artificial intelligence followed a simple rule: if you wanted smarter AI, you threw more chips at it.
More GPUs. Bigger data centres. More electricity. Bigger cooling systems. Rinse and repeat.
It worked…. until it didn’t.
Now the world has hit a wall. Not a technology wall, but a power wall.
In places like Texas, Virginia and parts of the US West Coast, data centres are lining up for grid connections and getting told to wait years. The electricity simply isn’t there, and the grid is effectively sold out.
When one gigawatt of AI computing power can generate tens of billions of dollars in revenue, waiting five years for a transformer upgrade isn’t an option.
So the biggest tech companies are doing something almost absurd: building their own power plants next to their data centres.
Gas turbines, batteries and private energy infrastructure.
That’s how desperate the power problem has become.
Which brings us to a South Korean engineer who decided the real solution wasn’t more electricity…. it was smarter chips.
And that decision might end up reshaping how AI runs at scale.
When power became a bottleneck
GPUs became the backbone of modern AI almost by accident.
They were originally built for graphics – millions of pixels updated in parallel, doing simple math extremely fast. When AI exploded, those same mathematical strengths made GPUs perfect for training neural networks.
The industry leaned into that hard.
But GPUs were never designed to be energy efficient. They were built to be flexible, fast and general-purpose.
That flexibility burns power, especially when massive amounts of data have to move constantly between memory and compute.
At small scale, nobody cared. Power was cheap, grids had spare capacity.
But at data- entre scale, power is now the limiting factor.
And once power becomes the bottleneck, the winning chip is no longer the fastest one, it’s the one that does the same work using far less energy.
That insight hit a former Samsung engineer named June Paik.
From Samsung engineer to startup founder
Paik had a solid career at Samsung. Safe trajectory with big company stability.
Then a soccer injury forced him to spend months off his feet. Bored and curious, he started studying AI deeply – how algorithms map onto silicon, how models really consume compute, how power flows through chips.
That’s when the light bulb went on. The future ceiling of AI wouldn’t be algorithms; it would be electricity.
If power stops scaling, brute force computing stops working. And if that’s true, AI needs a fundamentally different type of chip.
So Paik left Samsung and founded Furiosa AI, with one clear goal: build ultra-efficient chips designed specifically for running AI inference at scale.
Why inference is where money lives
Training a model is like building a factory.
Inference is when the factory actually produces goods.
Every time ChatGPT answers a question, an image generator creates a picture, or a computer vision system scans a camera feed – that’s inference. It happens constantly, millions of times per second across the world.
Inference workloads are repetitive math: multiply numbers, add them up, repeat billions of times.
Traditional CPUs are terrible at this. GPUs are much better, but still inefficient because they’re designed to handle many different tasks, not just this one.
Furiosa took a different approach.
Instead of building a flexible machine, they built a factory that only does one job extremely well.
They built a Neural Processing Unit (NPU) called Warboy, purpose-built for inference efficiency.
The simple idea
The biggest power drain in modern chips isn’t actually the calculations. It’s moving data back and forth between memory and compute.
Traditional chips follow a model where data is constantly fetched from memory, processed, then written back again. At large scale, that data movement burns enormous energy.
Furiosa’s chip uses a different architecture called a systolic array.
Instead of constantly pulling data from memory, data flows through the chip in a rhythmic pattern – like water moving through pipes – and gets reused multiple times before ever leaving the chip.
The numbers stay close to the compute units: Less traffic. Less heat. Less power.
Think of it like cooking.
A GPU is a fast chef who keeps running back to the pantry. An NPU lays everything on the bench and never leaves the workstation.
At data centre scale, that difference becomes millions of dollars in electricity savings.
Furiosa pushed this even further by giving the chip massive on-chip memory, conservative clock speeds, and hardware designed specifically around data reuse rather than raw speed.
The result: dramatically lower energy consumption.
The industry has taken notice
Furiosa publicly demonstrated its chip running Meta’s LLaMA model at Stanford’s Hot Chips conference.
The numbers raised eyebrows.
The chip operated at around 150 watts, compared to 350 watts or more for high-end GPUs – and modern GPUs are now pushing well beyond 1,000 watts in some configurations.
Performance per watt was roughly 40% better than GPU-based systems.
At data centre scale, that’s the difference between building new power infrastructure or not.
Meta reportedly attempted to acquire Furiosa for close to US$1 billion. Furiosa declined.
More importantly, real customers stepped in.
OpenAI used the chip for a public demo in Seoul. LG AI Research ran seven months of testing on real workloads and measured roughly 2.5x better performance per watt compared to GPU systems.
LG has now signed a commercial deal and the latest Furiosa chip has entered mass production using TSMC’s 5-nanometre process and advanced packaging.
This is no longer a lab experiment.
Why this matters for investors
This shift changes the economics of AI infrastructure.
If inference can run at dramatically lower power, companies can deploy more AI capacity without expanding power grids, cooling systems or real estate.
That lowers operating costs. It accelerates deployment speed. It unlocks AI in locations that previously couldn’t support massive power draw.
It also creates a new competitive battlefield.
GPUs will still dominate massive training workloads. They’re unmatched for raw compute.
But inference – the part of AI that runs 24/7 in production – is increasingly about efficiency, not brute force.
That opens the door for purpose-built chips like Furiosa’s.
The competitive landscape includes giants like Google (TPU), Amazon (Trainium), and specialised players like Cerebras. But there are very few independent companies globally that can design, manufacture and deploy advanced AI silicon at scale.
South Korea now has one of them.
If Furiosa continues executing, it positions Korea not just as a memory-chip powerhouse – but as a serious player in next-generation AI compute architecture.
Sometimes the smartest revolution isn’t louder or bigger.
It’s simply more efficient.
Now read: The next big thing in AI is small – inside Nota AI’s KOSDAQ debut
