Over the past month, the market capitalization of leading chip manufacturers has shrunk by more than $1.3 trillion. The market is panicking, selling off stocks amid fears of a collapse in demand for semiconductors. But as an analyst, I see the problem lies in a completely different area — infrastructure.
The paradox of the situation is that amid the sector's collapse, Taiwanese giant TSMC reported a record quarter. The company's revenue reached $40.2 billion, the growth forecast was raised above 40%, and confidence in the artificial intelligence megatrend was described as "very high." TSMC's profit grew by 77%, hitting an all-time high, and margins exceeded all previous figures. The company also announced an additional $100 billion in investments for its plants in Arizona.
Why doesn't the sell-off match reality?
TSMC's numbers speak for themselves: chips are being shipped, demand for them is accelerating, not fading. The market seems to have confused cause and effect. The fact is that microchips themselves are no longer a scarce resource for AI. Now, the industry's development hinges on steel, copper, and most importantly — time.
The first bottleneck is chip packaging. The CoWoS technology, which connects the processor to memory, is already operating at its limit. Nvidia buys up about 60% of the capacity, and demand for this service has tripled in two years. TSMC is nearly doubling its production in this area, but the lines are still running at 100% capacity.
The second and far more difficult bottleneck to address is energy. A finished chip is useless without electricity, and the industry is facing a catastrophic shortage of it. Data center statistics confirm the lack of energy capacity. Companies have announced 16 gigawatts of new capacity in the US for 2026, but are only building 5 of them. The remaining projects are frozen, and 25% don't even have an energy supply plan.
Where is the real "bottleneck"?
The root of the problem lies in equipment that the market rarely thinks about. This refers to high-voltage transformers that convert grid energy into power for data centers. Currently, their manufacturing takes an average of 48–60 months, whereas before 2020, 12 months was sufficient.
The shortage will persist for a long time. Hitachi, Siemens Energy, GE Vernova, and ABB — just four manufacturers of such units — have taken orders years in advance. Siemens Energy alone has a portfolio volume reaching nearly €136 billion. The problem starts at the raw material level: special electrical steel for transformers is produced by only five companies worldwide, and they cannot quickly ramp up production.
Circumventing the shortage is also not possible. Some data center operators decided to supply their own energy and ordered gas turbines, but GE Vernova, Siemens, and Mitsubishi have already sold their capacity for roughly the next five years and beyond.
The timelines for each constraint speak for themselves: chips are produced without delays, packaging capacity is fully booked, new transformers will appear in 2029, and turbines in 2030. The largest construction project in economic history is being held back by heavy electrical equipment that is almost handmade.
My conclusion as an expert: the market counted chips, when it should have been counting substations. The world's most advanced companies are now forced to build their own power plants to run neural networks. The AI story is no longer about intelligence — it's about electricity. And while investors panic over an imaginary decline in semiconductor demand, the real beneficiaries are transformer and turbine manufacturers, as well as owners of sites with ready-made grid connections.