Key takeaways
- TrendForce forecasts that combined capital expenditure by the nine largest global cloud service providers will exceed US$886.7 billion in 2026, up roughly 90% year on year. North American operators account for close to 90% of it, with most doubling their spend.
- The composition is shifting: money is flowing increasingly into liquid cooling, advanced packaging, high-speed interconnect, power infrastructure and memory upgrades rather than servers alone. AI server shipment growth forecasts were revised up from 28% to nearly 31%.
- Alibaba CEO Wu Yongming put the domestic shortage in concrete terms: 2026 demand for Chinese-made AI chips is around 4 million units against roughly 3 million deliveries — a gap measured in millions.
- Nvidia's FY27Q2 revenue was US$96.2 billion, up 106%, with data centre revenue of US$89 billion. Anthropic signed a multi-year US$35 billion compute contract with Lambda.
Where the money actually goes
Start with the number: US$886.7 billion. Roughly RMB 6.3 trillion.
The detail that gets missed is that this is no longer a GPU story. Per TrendForce, a large share now goes to the "non-server components": cooling, packaging, interconnect, electrical infrastructure, memory.
That is because the constraint has moved. It is not that there are not enough chips. It is that there is not enough power, not enough cooling, not enough floor space, and not enough memory.
Which is also why memory prices eventually reached your phone: AI data centres absorbed the capacity, and consumer electronics queued behind them.
How big the gap is at home
Wu Yongming's estimate is that AI compute shortage persists at least to 2030.
In numbers: China's demand for domestically produced AI chips in 2026 is roughly 4 million units; actual deliveries are roughly 3 million. A gap in the millions.
That is not pessimism, that is an order book. A gap means domestic substitution has a defined market, and it means whoever can actually bring capacity online gains pricing power.
Corroborating data from the same period: Nvidia's FY27Q2 revenue of US$96.2 billion, up 106% year on year, with US$89 billion from data centres; and Anthropic signing a US$35 billion multi-year compute deal with Lambda.
Three signals that warrant caution
Long-dated contracts cut both ways. A US$35 billion commitment spanning several years is a heavy obligation if demand undershoots. The signatories are betting on three to five more years of explosive growth.
Valuations rest on a permanent shortage. The whole sector's high multiples assume scarcity runs to 2030. If demand slows or capacity lands all at once, that logic reverses quickly.
Depreciation is the bill that arrives later. A GPU is not a building. It is obsolete in three to five years. US$886.7 billion of assets will show up as depreciation pressure in earnings over the next few years.
The two spillovers that reach ordinary people
Electricity. Data centres are enormous loads. Large-scale expansion raises pressure on regional grids, which is why China's "East Data, West Compute" programme and western clean-energy bases keep resurfacing in policy discussion. The compute contest is, at bottom, an electricity contest.
Jobs. Data centre construction, operations, liquid-cooling installation and network commissioning are hiring heavily, with educational requirements more forgiving than people assume. A single new data centre's operations team can run to dozens or hundreds.
If a compute centre is being built near your city, that is a concrete block of local employment.
How not to get trapped by the compute story
Do not read capital expenditure as profit. Money spent is an outflow; whether it converts into revenue and earnings is a separate question entirely.
Separate the shovel-sellers from the shovel-buyers. Sellers collect cash; buyers carry the payback risk. The two cannot be valued with the same framework.
And keep one discipline: for any purchase justified mainly by "how big the future market is," first ask "how much is it earning right now?"
Honest limitations
TrendForce's figure is a forecast, not an outcome, and capex plans get revised. Attributing consumer memory price rises solely to AI demand ignores normal industry cycles. The domestic supply gap figures are one executive's estimate rather than an audited statistic.
Sources: TrendForce research (September 2026); public remarks by Alibaba's CEO; Nvidia FY27Q2 results; reported Anthropic–Lambda contract. Information only — not investment advice.