I just reviewed TSMC’s quarterly report and realized that the shortage of computing power for AI is not temporary at all. Wei Zhejia gave the most honest answer on a live broadcast: the problem will hang around for at least until 2027, and there are no workarounds.

The numbers speak for themselves. Revenue for the first quarter came in at $35 billion—up more than 40% year over year. But the most interesting part is capital expenditures: TSMC has increased its budget to $56 billion for 2026. This is a historical figure—more than half of all investment over the past three years. The HPC business grew 20% quarter over quarter and now makes up 61% of revenue. Clearly, the company is going all in on AI.

But here’s the point: even with such investment, building a new plant is 2–3 years of work. TSMC can’t just snap its fingers and double output. So new 3nm fabs are on the horizon in Taiwan, Arizona, and Japan, but they won’t start operating until 2027–2028.

And the GPU rental market situation is even more acute. H100 prices have risen by roughly 40% over six months—from $1.70 to $2.35 per hour. In China, monthly H100 rentals jumped from 40–50 thousand yuan to 80–90 thousand. And this isn’t the end: the entire H series is essentially sold out for years ahead, and Blackwell has already been fully reserved through the end of 2026.

Tencent, Alibaba, and Baidu are all raising prices for cloud services at the same time—by 5–30% in April. That makes sense: if GPUs become more expensive, API tokens will be more expensive too. Demand is exceeding supply so much that providers can afford such jumps.

But the key observation is this: total capital expenditures by major cloud operators in 2026 are projected to reach $700 billion—60% higher than last year. Almost all of that money goes to NVIDIA chips that are produced by TSMC. On the one hand, demand is enormous; on the other hand, there are physical production constraints. This imbalance won’t disappear over the next couple of years.

Conclusion: the shortage of computing power is not a temporary setback, but a structural reality. TSMC, NVIDIA, and HBM memory manufacturers are in a period of high growth certainty. For the rest of the chain—cloud providers and model developers—challenges remain serious, and the trend of rising prices will continue at least until 2027.