The artificial intelligence infrastructure boom may be building digital cathedrals on shifting sands.
The unprecedented expansion in AI infrastructure - with global spending on data center systems already reaching $260 billion in 2024 and projected to exceed $700 billion by 2030 per BlackRock - faces significant technological disruption risks that warrant careful consideration.
A wave of technological breakthroughs in AI efficiency presents an opportunity to reconsider our assumptions about infrastructure scaling. Now is the moment for everyone to take a step back and examine the fundamental premises behind the AI infrastructure buildout.
💡 Key Points:
✅ Current infrastructure investments may benefit from lessons learned during the 2000s telecom expansion
✅ The Jevons paradox may have different implications for physical AI infrastructure versus software efficiency
✅ Emerging AI architectures could reshape infrastructure requirements
✅ Investment strategies should balance scale with technological adaptability
Is the AI Data Center Boom Repeating the Telecom Bubble?
When the dot-com bubble burst, the telecom sector faced significant challenges. Of the $7 trillion decline in stock market valuations between 2000 and 2002, about $2 trillion was attributed to telecom companies. In addition, 23 telecom companies – including Exodus -- went bankrupt. The FCC found that by 2007, 73.4 million KM of fiber optic cable had been laid but that 48 million KM was unused – a stranded asset dubbed “dark fiber.”
While the late 1990s are remembered for the dot-com boom, the amount raised and lost by internet IPOs was tiny compared to the trillions of dollars that poured into the telecom sector. Established giants like AT&T and WorldCom took on billions in debt while upstarts like Exodus Communications -- which built 44 data centers in just a few years – raised almost $10 billion.
This was all justified by a simple rationale: internet traffic was doubling every 100 days. WorldCom made this statement, and it became widely echoed even though the reality was closer to every year.
Two decades later, it's worth examining whether current infrastructure growth assumptions merit similar scrutiny.
Does Jevons Paradox Apply to AI Infrastructure?
The Jevons paradox likely applies to the software layer, but physical infrastructure faces different considerations. While efficiency gains in computing naturally translated to increased usage, the same may not apply uniformly to the physical infrastructure layer. Power systems, cooling architectures, and real estate configured for today’s GPU farms may require adaptation for emerging computing paradigms.
Today's dominant AI models demand vast arrays of power-hungry GPUs. The $500bn Stargate project exemplifies this approach.
Then came news of DeepSeek's innovation in AI architecture, which suggests a different and potentially more efficient approach to infrastructure. The Chinese start-up's innovation — which achieves comparable performance to current models while consuming just 30 percent of traditional computational resources — prompted significant market attention that affected valuations across the sector.
AI infrastructure investors responded with a compelling framework: the Jevons paradox, the principle that increased resource efficiency drives higher total consumption.
As Microsoft CEO Satya Nadella wrote: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of." In January 2025, Morgan Stanley released a report that reinforced this view, projecting that a 90% drop in computing costs will unleash unprecedented AI adoption.
This framework has merit: the less something costs, the more people will use it. However, when it comes to AI infrastructure, this principle warrants careful examination.
DeepSeek has introduced new possibilities in the relationship between compute resources and model performance. It suggests that future advances may come from architectural innovation as much as raw computing power.
If proven successful, the efficiencies demonstrated in DeepSeek's open-source model could influence how larger companies approach model training, potentially affecting data center demand for training infrastructure. While this could accelerate inference compute adoption as token prices decrease, the infrastructure requirements for inference differ from those of model training.
The industry finds itself at an inflection point reminiscent of the transition from mainframe computing to distributed systems. Just as that shift transformed computer center requirements, today's investments in GPU-optimized infrastructure may need to evolve with technological advances.
While future demand for AI infrastructure seems certain, the timing and technical requirements may evolve. As NVIDIA's Jensen Huang noted, "Our systems are progressing way faster than Moore's Law."