AI investment shows no signs of slowing, prompting concern about what happens when it eventually does.

Victor Huang has described ChatGPT’s arrival as AI’s “iPhone moment,” when the technology’s potential to change the way we live became clear to all. Indeed, the publication of large language models (LLMs) is arguably the most powerful innovation we have seen so far, enabling widespread user engagement spanning personal and professional purposes. Resulting from this, acceptance has snowballed and more and more of us have come to trust and rely on AI assistants. In turn, mainstream use has driven further improvements, as successive updates have delivered more humanlike interactions and additional capabilities.

Within a short space of time, it seems, an AI-based solution has become available for almost everything. In some of the most contentious examples, AI is giving us original novels, including new works in the style of our favorite authors, and virtual film stars. At the same time, the Albanian government has even put an AI – complete with a traditionally dressed avatar – in charge of its national procurement department. On a more day-to-day level, many of us are discovering how much we need the skills AI can bring to our lives, for finding information on the Internet, organizing our digital content, taking better selfies and fulfilling job responsibilities.

On the one hand, edge AI and tiny machine-learning (ML) applications, while on the other, cloud data centers are where we find the heavy lifting for social media, business applications, and intensive analytics. The AI data center server market has become a huge opportunity for equipment vendors as operators upgrade their data centers to meet the explosion in end-user demand. Already worth almost $300 billion, currently representing 17% of the total server market by value, it’s predicted to grow another 20% in 2026.

The effects permeate the entire hardware supply chain, extending to materials needed for chip-packaging substrates and PCBs for high-performance accelerator cards and motherboards. This is advanced, high-end materials technology, including low-CTE quartz glass and ultra-low-profile copper foils needed for high-layer-count boards. The boards are made using advanced techniques such as semiadditive process (SAP) and modified SAP (mSAP) for precise control over circuit dimensions, allowing designers to achieve feature sizes as small as 30µm and traces of 15µm width. Some advanced mSAP and full SAP can achieve even finer features, down to 5-10µm. The opportunities for AI data center equipment present a powerful lure for materials suppliers and fabricators to focus more sharply on high-end products or even realign their businesses away from PCB fabrication and into IC packaging. As AI processors and high-bandwidth memory drive demand for high-performance substrates, this is a $20 billion slice of the action in a global semiconductor market poised to reach $1 trillion in revenues by 2030.

Attracted by the prospect of greater margins, some suppliers are moving higher up the value chain and out of their traditional markets for conventional standard materials. While not all are following this trend, shortages and price increases are likely for these products as the supply chain adjusts and other companies in different geographic areas take over.

As consumer excitement over AI rises to ever-higher levels, however, signs of dissent are appearing. Doubters question whether the current boom is a bubble that’s about to burst. While arguments rage on either side of that debate, we must consider at least one spectacular precedent: the telecom industry of the early 2000s. Some of the industry’s biggest and most successful hardware suppliers, which built for what they thought was limitless demand, suddenly found themselves holding vast unsold inventories and were forced to revise their forecasts. Multi-billion-dollar write-downs followed, and key players suffered enormous pain in the stock markets. While some are even now just regaining their pre-boom market valuation, others went bust in the debacle, along with many overhyped dotcoms. We are certainly experiencing a hardware boom today, as data centers invest in new AI servers. In the historical telecom example, spending stopped suddenly due to the dotcom bust as operators stopped spending on hardware and reverted to their normal investment cycle of about 5-10 years, which is the typical lifetime of carrier-grade switches.

Today’s top computing companies estimate the lifetime of an AI server at 3-6 years. Investment could slow, although the timing is very difficult to predict accurately. After the telecom bust, a major materials supplier asked me to carry out an analysis to learn whether similar issues in the future could be anticipated and avoided. Notionally, of course, it’s easy to predict a fall happening sometime because demand cannot continue rising forever. In reality, dynamic high-tech markets that are constantly creating new services and capabilities have many variables as well as many unknowns. With so many interrelated and abstract factors in play, accurately pinpointing the timing and extent of any downturn is practically impossible.

This time, will we learn the lessons history teaches us? Few can resist the excitement of a boom, and shareholders will always expect executives to follow the money. Right now, data centers are still investing in hardware, and the market is expected to continue growing strongly. Sooner or later, however, every bubble bursts and all will be affected in some way. The industry needs to combine dynamism to make the most of the market highs with the resilience to handle the inevitable lows.

Alun Morgan is technology ambassador at Ventec International Group (venteclaminates.com); alun.morgan@ventec-europe.com. His column runs monthly.