"Supercycle" gets thrown around every time a chart goes up and to the right. So it's worth asking plainly: is the AI build-out actually a multi-year cycle, or a spending spike that's already peaking? The numbers from this summer say the cycle is still accelerating. They also say the winners are changing.
The numbers
| Signal | Latest | Why it matters |
|---|---|---|
| Big 4 hyperscaler capex, 2026 | ~$700B+ guided, vs ~$410B in 2025 | Demand is still being funded |
| Nvidia data center revenue | $89.0B last quarter, +117% YoY | Spend is converting to orders |
| Nvidia next-quarter guide | $108B total revenue | No sign of a slowdown yet |
| DRAM contract prices, Q1 2026 | Up ~90-95% QoQ | Memory has become a bottleneck |
| US data center power demand | 31 GW (2025) to ~66 GW (2027E) | Power is the next constraint |
Spending. Amazon, Alphabet, Meta and Microsoft entered 2026 guiding to roughly $700 billion or more in combined capital spending, up about three-quarters from last year's record. Then most of them raised it again at July earnings: Amazon to around $220 billion, Alphabet to $195-205 billion, Meta to $130-145 billion. Guidance going up mid-year is the opposite of what a peaking cycle looks like.
Compute. Nvidia reported $96.2 billion of revenue for the quarter ending in July, more than double a year ago, with data center at $89.0 billion and gross margins holding at 75%. It guided the next quarter to $108 billion, with no China data center revenue assumed. The detail I find most important: its non-hyperscaler segment (AI clouds, enterprise, sovereign buyers) grew 138%, faster than the hyperscalers themselves. The customer base is widening, not narrowing.
Memory. High-bandwidth memory (HBM) for AI accelerators uses far more wafer capacity per bit than normal DRAM. As Samsung, SK hynix and Micron shifted capacity to HBM, conventional DRAM got squeezed. TrendForce measured DRAM contract prices up roughly 90-95% quarter over quarter in Q1 2026, the fastest move it has recorded, and suppliers describe 2026 HBM output as sold out. TrendForce expects tightness to continue through 2027.
Power. Goldman Sachs expects US data center power demand to roughly double, from 31 GW in 2025 to about 66 GW in 2027. The grid can't move that fast. Large power transformers now carry lead times of well over two years, and interconnection queues in major data center markets run four to seven years. Of the new capacity targeted for 2026, only about a third was reported as actually under construction early in the year.
The thesis
The supercycle is real, but the pricing power is moving down the stack.
In 2023 and 2024, the scarce input was the GPU, and that's where the margins went. In 2026, GPUs are shipping in huge volume, and the binding constraints are now memory and electricity. When a resource is the bottleneck, its owner sets the price. That's what the DRAM contract data is showing, and it's the setup forming around power equipment and generation.
So the question I'm asking isn't "is AI spending going up?" It clearly is. It's "which layer is scarce next, and is the market still pricing it like a normal cyclical business?" Memory makers have historically traded as boom-bust stocks. If HBM demand keeps absorbing capacity through 2027, that cycle could run longer and steadier than the market is used to.
The risks
This is where most supercycle posts go quiet. Here's what would prove this thesis wrong:
- The spending outruns the cash. The four hyperscalers bought about $434 billion of property and equipment in the four quarters through March 2026, against roughly $149 billion of depreciation. That gap closes over time, and it will weigh on reported earnings. Amazon's trailing capex has already exceeded its operating cash flow, and several of these companies are raising debt to fund the build.
- Memory does what memory always does. Every past memory boom ended with too much capacity. If suppliers overbuild HBM and DRAM just as demand growth slows, prices can fall as fast as they rose.
- Power delays push revenue out. If data centers can't get connected, chips and memory sit in warehouses or orders get pushed back. A bottleneck helps whoever owns it but can hurt everyone upstream of it.
- AI revenue has to show up. Cloud growth is strong right now, but eventually the end customers of all this compute need to earn a return on it. If that stalls, 2027 capex budgets are where it will show first.
- Policy. Export controls already remove China from Nvidia's outlook, and tariffs and chip policy can change the math for any company in the chain.
What I'm watching
- First 2027 capex guidance from the hyperscalers, likely in their Q4 reports. A flat number would be the first real crack.
- Quarterly DRAM and HBM contract pricing. Price increases slowing matters more than prices falling.
- Capex as a share of operating cash flow. Rising debt funding is a late-cycle signal.
- Utility and grid equipment order books, the next place scarcity pricing could show up.
Sources
- Nvidia Q2 fiscal 2027 results (Aug 26, 2026)
- Hyperscaler 2026 capex tally (FT, via AI Weekly)
- Updated 2026 capex guidance after Q2 earnings
- Capex vs depreciation and cash flow (Silicon Analysts)
- TrendForce DRAM contract pricing, Q1 2026
- HBM supply and 2027 outlook
- US data center power demand forecasts (Goldman Sachs)
- Grid delays and transformer lead times
Not investment advice. I may hold positions in companies in this sector.