Before diving into regulatory details, let’s take a moment to explore the technologies that power AI and how they differ from traditional computer processing.
To fully appreciate the rising importance of companies like NVIDIA and QScale, especially here in Québec, it's essential to understand the fundamental difference between a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). You hear these terms often, but what do they really mean?
The General-Purpose Nature of the CPU
Let us pause to appreciate the marvel that is the modern computer. At the heart of every machine lies a piece of silicon etched with billions of nanometer-scale circuits. These circuits operate at the boundary between classical and quantum physics. Thanks to them, a computer can execute arbitrary sequences of instructions, called programs, at billions of operations per second. This engineering wonder is what we call a CPU.
Despite its complexity, a CPU operates on relatively simple principles. It processes a sequence of instructions that define operations to be performed on data. For example, for a series of multiplications, the CPU handles each operation individually and executes them sequentially.
This architecture makes the CPU a general-purpose tool. Most imaginable computational operations can be carried out by a CPU. So what’s the point of a GPU? Aren’t modern CPUs powerful enough to handle all necessary additions and multiplications? As it turns out, they are not!
The GPU: A Specialist for Parallel Workloads
It was the development of video games that exposed the inherent limitations of CPU architecture. When simulating a 3D scene, a computer must perform a vast number of vector geometry calculations. These geometric transformations translate into matrix multiplications. To render this simulation in real time, as required by video games, the CPU must perform all of these operations within milliseconds, while also managing the rest of the computer's functions.
Fortunately for fans of smooth graphics, these vector geometry algorithms share a useful trait: they are parallelizable. In other words, operation 2 does not depend on the result of operation 1. In theory, tens of thousands of operations could be run in parallel instead of queuing them one by one for the CPU.
To take advantage of this opportunity, the first GPUs were created. The term "GPU" was first coined by NVIDIA in 1999 with the GeForce 256 graphics card. Unlike the CPU, which processes instructions sequentially, the GPU executes the same instruction on thousands of cores simultaneously, handling many data batches in parallel and achieving high levels of computational throughput.
Over time, other industries leveraged this parallel processing capability—notably blockchain networks at the heart of the cryptocurrency explosion.
Then came the boom in generative artificial intelligence. Neural networks, the backbone of this innovation, also benefit greatly from parallelization. Their training and inference processes are based on matrix multiplications, much like those used in 3D graphics rendering.
This trend proved highly favorable to NVIDIA, initially focused on gaming hardware. Today, NVIDIA is the leading provider of parallel computing power, one reason it is now among the most valuable companies in the world.
Other terms are now entering the conversation, such as TPU (Tensor Processing Unit) and NPU (Neural Processing Unit). These are alternative implementations of the same idea: processors specialized for parallel computation.
Now the question is: will AI regulation slow down investment and, by extension, innovation?
The European Union's Tough AI Regulation
The European Union’s Artificial Intelligence Act is expected to be approved and signed by relevant authorities in 2024. It establishes a legal framework governing how AI is used by organizations that sell or operate within the EU, affecting all companies regardless of nationality or size.
This law defines four categories of risk associated with AI products:
Category 1: Unacceptable Risk
Products in this category will be outright banned from the EU market within six months of the law’s enactment. These include:
- Real-time biometric identification systems
- Tools that collect facial images from the internet or public video surveillance to create facial recognition databases
- Tools that analyze people’s emotions in the workplace, such as software that interprets facial expressions
Category 2: High Risk
Products in this category are subject to strict, often costly compliance requirements. These systems are considered potentially harmful to health, safety, fundamental rights, the environment, or democracy. Examples include AI used in:
- Biometric systems
- Critical infrastructure
- Education and employment
- Law enforcement
Companies will need to modify their systems to comply with technical and legal evaluations. Proving compliance will be the responsibility of software vendors and may require expensive risk management strategies. As a result, AI development may increasingly shift toward countries like the U.S. and Canada, where regulations are less prescriptive.
Category 3: Limited Risk
Products in this category—such as chatbots and deepfake generators—will face lighter obligations. Vendors must inform users that they are interacting with AI systems and ensure that all AI-generated audio, video, or images are clearly labeled.
Category 4: Minimal Risk
Minimal-risk products like spam filters and recommendation engines are exempt from regulation under the AI Act.
Conclusion
We are at an inflection point where technology and regulation are starting to collide. It will be interesting to see how markets and innovators respond to this new landscape.