The race for AI chips and computing power: scarcity, geopolitics, and the impact on the Netherlands

Artificial intelligence is often colloquially described as a purely software-based breakthrough. Words like 'cloud', 'neural network', and 'algorithm' suggest an intangible, virtual world. The reality, however, is grittier and much more physical: the current AI revolution relies entirely on computing power, silicon, advanced cooling, and extreme amounts of energy. Those who do not have access to specialized hardware simply cannot compete in the highest echelon of model development.

In this article, we unravel the dynamics behind the global battle for AI chips. We look at the causes of the ongoing GPU scarcity, the key players in the market's crosshairs, the physical production bottlenecks, the geopolitical forces surrounding ASML and Taiwan, and finally, what all of this means in practice for Dutch developers and organizations.

1. Why traditional hardware falls short: from CPU to GPU

To understand why there is an acute scarcity of AI chips, we must look at how large language models (LLMs) perform calculations. A classic Central Processing Unit (CPU) is designed as a 'jack-of-all-trades': extremely fast at executing complex, diverse tasks one after the other (sequential processing). A CPU has a limited number of highly powerful processing cores.

AI models, and in particular the architecture of transformers and deep neural networks, work fundamentally differently. They require billions of simple mathematical operations — mainly matrix multiplications — to take place simultaneously. This is known as parallel processing.

The rise of the Graphics Processing Unit

Graphics Processing Units (GPUs), which were originally developed for rendering 3D graphics and video games, turned out to be ideally suited for this type of work. A GPU contains not dozens, but thousands of smaller processing cores that can perform the same type of mathematical task in parallel. When researchers discovered that neural networks trained exponentially faster on GPUs than on CPUs, the graphics card transformed from a consumer component into the most important raw material of the modern tech sector.

2. The bottlenecks: Why making chips is so extremely difficult

Building a modern AI accelerator is one of the most complex industrial processes on Earth. It has long ceased to be just about etching the smallest possible transistors onto a slice of silicon (the wafer). Today, the biggest friction points in the supply chain lie in two specific areas: memory bandwidth and advanced packaging.

High Bandwidth Memory (HBM)

While training or running a model with hundreds of billions of parameters, the model's weights must be constantly transferred back and forth between the computing memory and the processing cores. Traditional DDR memory is far too slow for this. The processing core is then left waiting for data — a phenomenon known as a 'memory wall'.

The solution is High Bandwidth Memory (HBM). This involves stacking multiple memory chips vertically and placing them directly next to the processor via microscopic connections. Global production capacity for HBM is extremely limited and almost entirely sold out by a handful of chip manufacturers.

Advanced packaging (CoWoS)

A second crucial bottleneck is what is known in the industry as Advanced Packaging. TSMC (Taiwan Semiconductor Manufacturing Company) uses a patented technology called CoWoS (Chip-on-Wafer-on-Substrate). This involves placing the processor and the HBM memory stacks on a thin intermediate layer (interposer), allowing them to function as a single hybrid chip.

Key Insight: The physical shortage of AI chips is currently not primarily caused by a lack of silicon wafers, but by the limited capacity of advanced packaging lines such as CoWoS and the scarcity of HBM memory.

3. Market dynamics: Monopolies, software walls, and challengers

The market for high-end AI hardware is currently dominated by a single player: Nvidia. This dominance is not just the result of superior hardware, but primarily reflects a strategic choice made nearly two decades ago.

The invisible wall of CUDA

In 2006, Nvidia launched CUDA (Compute Unified Device Architecture), a software platform that allows developers to program GPUs for general-purpose computing. Because virtually the entire academic and industrial AI world has been trained on CUDA over the past fifteen years, a deeply entrenched ecosystem has emerged. Almost all well-known AI frameworks (such as PyTorch and TensorFlow) are primarily optimized for CUDA. Switching to a different chip manufacturer often requires rewriting software and redesigning workflows.

The alternatives: AMD and custom ASICs

Nevertheless, several parties are trying to break this hegemony:

• AMD: With the Instinct range (such as the MI300 series), AMD offers a powerful hardware alternative. Through the open-source platform ROCm, AMD is trying to close the software gap with CUDA, although compatibility in practice remains a challenge at times.
• Big Tech and custom silicon (ASICs): Large cloud providers are rapidly building their own application-specific integrated circuits (ASICs). Google has been using its own TPUs (Tensor Processing Units) for years, Amazon has Trainium and Inferentia, Meta is developing MTIA, and Microsoft is rolling out its Maia chips. These chips are less flexible than general-purpose GPUs, but are significantly cheaper and more energy-efficient for specific tasks.

4. Geopolitics and the strategic supply chain: The role of the Netherlands

Hardware for artificial intelligence is no longer a purely commercial matter; it has become an issue of national strategic importance and geopolitical power dynamics. The value chains are extremely globalized, yet shockingly vulnerable due to monopolies in specific links.

ASML: The Dutch crown jewel

The Netherlands plays an absolute key role in this global chain thanks to ASML, based in Veldhoven. ASML has a de facto global monopoly on the production of EUV (Extreme Ultraviolet) lithography machines. Without these machines, which can print patterns at the nanometer scale onto silicon, it is technically impossible to produce the most advanced AI chips in the world.

Because of this key position, the Netherlands is under constant pressure from the US Department of Commerce. The United States has imposed stricter export restrictions to prevent advanced lithography technology and high-end AI chips from being supplied to China. This has led to political tensions and strict licensing regimes for the Dutch chip machine manufacturer.

The vulnerability of Taiwan and the European Chips Act

Another geopolitical pain point is the geographical concentration of chip manufacturing. The most advanced chips are physically produced in factories in Taiwan, primarily by TSMC. Any geopolitical escalation around the Taiwan Strait could bring the global supply of computing power to a standstill within days.

To reduce this dependency, both the United States and the European Union have launched subsidy programs (such as the European Chips Act). The goal is to build more chip fabrication plants (fabs) on home soil. While this may improve long-term security of supply, building an advanced fab takes many years and requires hundreds of billions of euros in investment.

5. What does this mean for the Dutch market and AI users?

The geopolitical battle and the physical scarcity of hardware are not a remote concern. They have direct consequences for Dutch companies, government institutions, and software developers.

1. High and variable cloud costs

Because AI chips are expensive and demand outstrips supply, cloud providers pass these costs directly on to their customers. Both the price per token for API-based models and the hourly rates for renting dedicated GPU instances remain high. This raises the barrier for startups and SMEs to train or fine-tune their own models at scale.

2. Physical limits: Energy and grid congestion

Running large GPU clusters requires massive amounts of electricity and cooling capacity. In the Netherlands, the expansion of data centers faces concrete physical limits. Grid congestion — the lack of capacity on the Dutch power grid — makes it difficult to connect new, large data centers. This directly impacts the regional availability of computing power. Read more about this in our analysis on AI and energy consumption and our overview of data centers and AI in the Netherlands.

3. Sovereignty and data location

Because the most powerful hardware clusters are located in US cloud infrastructure, Dutch organizations are faced with questions regarding data sovereignty and privacy. Especially within healthcare and government, there is a growing desire to run models locally or on European soil. If you want to weigh whether your own infrastructure or a cloud solution is the best choice, consult our decision guide on cloud versus on-premise hardware.

6. Practical strategies: Efficiency over brute force

Given the scarcity and high costs of hardware, waiting for 'unlimited cheap computing power' is not a realistic strategy. Successful Dutch organizations are therefore increasingly opting for smart, efficient alternatives instead of ever-larger computing clusters:

• Deployment of smaller models: For many specific business applications, a massive model with hundreds of billions of parameters is not necessary. By switching to optimized small language models (SLMs), organizations can get by with a fraction of the required GPU capacity.
• Quantization and optimization: Using techniques like quantization (reducing the precision of weights from, for example, 16-bit to 8-bit or 4-bit), powerful open-weights models can run locally on consumer hardware or lighter servers.
• Hybrid inference: Simple queries are handled by lightweight, local models, while only highly complex tasks are forwarded to expensive, external GPU clusters in the cloud.

Conclusion

The race for AI chips is a complex interplay of technical physics, market dominance, and global geopolitics. Sand and silicon have been transformed into the strategic gold of this century. Although Nvidia currently rules the roost with its CUDA ecosystem and high-end chips, the rise of custom ASICs from Big Tech and alternative chip designs ensures a dynamic market.

For Dutch organizations and developers, the key does not necessarily lie in joining the race for the largest GPU clusters. The real gains are to be made through smart architectural choices: applying quantization, utilizing specialized small models, and maintaining the balance between efficiency, privacy, and costs. Computing power remains scarce, but those who build efficiently can achieve maximum returns even with limited hardware.