A shortage of advanced AI chips sounds like a manufacturing problem. But what if the chips themselves are not the main bottleneck? In 2025, the four largest AI accelerator designers together accounted for an estimated 90% or more of global High Bandwidth Memory (HBM) supply and advanced CoWoS packaging capacity, yet used only around 12% of advanced logic-die production. Making the processor was only part of the challenge. Securing enough memory and packaging to turn it into a usable accelerator was just as hard.
So when we ask who controls AI, we need to look beyond the companies building models. We need to examine the engineering supply chain behind large-scale computing, and the influential players who rarely feature in conversations about artificial intelligence.
The processor is not the whole story

For years, processor performance was discussed in terms of clock speed, core count and transistor density. AI workloads have changed that conversation. Modern accelerators must perform enormous numbers of mathematical operations while constantly moving data between processing units and memory.
This matters most for large language models. During training, a model processes data in batches, calculates errors and updates its parameters over many iterations. During inference, when a trained model generates a response, it repeatedly applies its learned parameters to incoming tokens to produce the next ones.
Both stages need heavy computation, but they strain hardware differently. Training demands large-scale mathematics and fast communication between processors. Inference can depend heavily on memory capacity and bandwidth, especially when a model must keep vast numbers of parameters and intermediate values within reach.
That is why an accelerator cannot be judged by theoretical processing power alone. If data cannot reach the processor quickly enough, extra computing capability sits idle. Moving data has become one of the central engineering problems in AI hardware.
Why AI chips need a different kind of memory
A GPU can contain thousands of processing units, and each needs a steady supply of data. Conventional memory architectures struggle to deliver enough bandwidth for large AI workloads without consuming too much space and power.
HBM addresses this. Instead of memory modules sitting some distance from the processor, HBM stacks memory dies vertically and links them through thousands of tiny connections called through-silicon vias. The stacks sit right beside the processor in the same advanced package, creating a much wider path for data.
The benefit is bandwidth. The trade-off is manufacturing complexity: stacking dies, connecting them reliably and integrating them with a high-performance processor demand specialised production and packaging processes.
SK hynix, Samsung Electronics and Micron are the major HBM suppliers, and their role is pivotal. A shortage of memory stacks limits how many complete accelerators can ship, even when processor dies are plentiful. A processor may be manufactured, but it is not a high-performance AI accelerator until its memory and packaging requirements are met.
The manufacturing technology behind a modern AI chip

Most leading AI chip designers do not run their own factories. Companies such as NVIDIA and AMD create architectures and designs, then rely on specialist foundries to manufacture the silicon.
TSMC is among the most important of these. Its advanced processes have served several generations of high-performance chips, including NVIDIA's Blackwell GPUs, which use a customised TSMC 4NP process.
Manufacturing involves repeatedly building and patterning extremely small structures on silicon wafers, and lithography is among the most critical steps. It transfers circuit patterns onto a light-sensitive material called photoresist. At advanced nodes, extreme ultraviolet (EUV) lithography prints increasingly intricate patterns. ASML builds these systems, which use light with a 13.5-nanometre wavelength. A laser strikes tiny droplets of molten tin, creating plasma that emits EUV radiation, a process ASML says happens up to 50,000 times per second.
EUV is not the only technology involved, and a smaller node does not automatically mean a faster chip. Architecture, power consumption, memory access and manufacturing yield all shape the final product.
The key point is that cutting-edge chip production rests on a highly specialized equipment ecosystem. Buying a factory building is not enough. A company also needs lithography, deposition, etching, inspection and metrology systems, plus the expertise to run them at scale.
Chiplets and advanced packaging
As AI processors grow more powerful, fitting everything onto one piece of silicon gets harder. Designers increasingly turn to chiplets and advanced packaging. A chiplet design divides a processor into smaller dies connected inside a single package. Different dies can perform different functions and even use different process technologies.
Connecting them is not as simple as placing them side by side. They need reliable, high-density electrical links that move data quickly while managing power and heat.
TSMC's CoWoS (Chip on Wafer on Substrate) is one advanced packaging technology that integrates logic dies and HBM stacks into one package, placing memory close to the processor and shortening the distance data must travel. This is why packaging has become critical to AI hardware production. Epoch AI's analysis of 2025 supply constraints found HBM and CoWoS capacity to be especially significant bottlenecks for leading accelerators.
The shift is significant: semiconductor innovation is no longer only about shrinking transistors. It is increasingly about integrating specialised components into one high-performance system.
When 72 GPUs have to behave like one machine

A large model may not fit in the memory of a single accelerator. Even if it does, spreading work across several GPUs can speed up certain operations. But this creates a communication problem.
In distributed training, processors exchange information such as gradients and model parameters. If one GPU waits for another to finish sending data, the cluster wastes valuable time. As GPU counts grow, the speed and organisation of communication become as important as the speed of each processor.
That is why high-speed interconnects matter. NVIDIA's NVLink provides high-bandwidth communication between GPUs, and switching systems let processors across a rack exchange data. NVIDIA's GB200 NVL72, for example, combines 72 Blackwell GPUs and 36 Grace CPUs in one rack-scale design, with a specified 130 TB/s of aggregate NVLink bandwidth.
Operators are no longer buying individual servers and treating them as separate machines. They are building tightly integrated systems in which dozens of accelerators work as one resource. The challenge is keeping them busy. More GPUs do not automatically mean proportionally faster AI. Results depend on workload distribution, memory access, network topology and how much communication a task requires.
The data center is becoming part of the processor
A modern AI data center is more than a building full of servers. Computing, networking, power distribution and thermal management must work as one engineered environment.
High-performance AI racks draw enormous amounts of electricity, and almost all of it ends up as heat that must be removed to keep components within safe temperatures. As rack power density rises, air cooling struggles to move enough air. Liquid cooling is an alternative: in direct-to-chip systems, coolant flows through cold plates attached to processors and carries heat away more efficiently than air alone.
Power delivery is another constraint. Facilities need electrical infrastructure that can handle heavy loads, along with backup systems and reliable grid connections. A site with enough space and servers still cannot run at full capacity if power or cooling falls short.
The scale is already measurable. The International Energy Agency estimated that data centres used around 415 terawatt-hours of electricity in 2024, roughly 1.5% of global consumption, and projects that figure could more than double to about 945 TWh by 2030. For AI infrastructure, electricity is no longer just an operating cost. Its availability, price and reliability can shape where facilities are built and how fast they expand.
So, who holds the power?

The AI computing ecosystem is spread across companies that specialize in very different layers of the stack. NVIDIA and AMD design accelerators. TSMC and Samsung manufacture advanced silicon. ASML supplies critical lithography equipment. SK hynix, Samsung and Micron make memory. Networking companies build the technologies that connect processors, while cloud providers such as Amazon Web Services, Microsoft Azure and Google Cloud operate computing infrastructure at scale.
None of these companies control AI in the same way. A chip designer shapes processor architecture and software compatibility. A foundry controls access to manufacturing capacity. A memory supplier influences how many complete accelerators can be assembled. A cloud provider decides who gets computing resources among customers who cannot afford their own data centers.
This creates dependencies across the industry. A shortage in one layer can ripple through the whole chain, while an improvement in another can shift who holds the advantage.
It also shows why AI leadership cannot be measured only by how many models a company releases. The ability to manufacture, connect, power and provide access to computing is becoming an equally important source of influence. In the AI race, the quiet layers of the supply chain may matter as much as the loudest names.



