AI Infrastructure’s Hidden Chokepoints: Memory, Power, and Substrates
TREE NEWS reports: A veteran anonymous analyst known for deep supply-chain research on AI infrastructure has delivered a sobering assessment of the artificial intelligence buildout, arguing that hyperscaler claims of decade-long demand visibility are fundamentally unreliable and that the true constraints on AI expansion have shifted away from GPU supply toward memory costs, power generation, and advanced packaging materials.
Speaking on a widely followed technology podcast, the analyst behind the research outfit stated bluntly that “if a hyperscaler tells you they have 10 years of demand visibility, they are just fooling you. They cannot predict what will happen two years from now.” He maintained that memory remains a cyclical industry, driven by the nature of capital expenditure, and that the assumption of infinitely sustainable hyperscaler spending is just that — an assumption.
Memory: The Trillion-Dollar Line Item
The analyst estimates that next year’s hyperscaler capital expenditure will reach $1.1 trillion to $1.2 trillion, with memory alone accounting for 50% to 60% of that figure — roughly $500 billion to $700 billion. Some investment bank estimates put the memory spend even higher, at $900 billion, which would exceed this year’s total hyperscaler capex. The mix includes HBM, commodity DRAM, and NAND flash, with the exact breakdown varying widely by source.
To smooth out the boom-bust cycle, memory manufacturers are aggressively pursuing Long-Term Agreements (LTAs). Samsung has reportedly begun receiving requests for contracts extending up to 10 years. However, a former Samsung and AMD executive warned the analyst that LTAs may be overhyped: if demand collapses, these contracts could be quietly cancelled behind closed doors without penalties, because forcing customers to absorb unwanted product would lead to inventory gluts and a cliff-edge revenue decline.
The Real Bottleneck: Power and ABF Substrates
While GPU rental prices confirm acute compute scarcity — with used H100s still fetching around $25,000 and B-series GPU hourly rates rising from $5 to $7-8 — the analyst argues that raw chip supply is not the core constraint. An AMD expert confirmed that silicon supply is adequate; the bottlenecks lie in data center power, advanced packaging, and memory.
Breaking down the bill of materials for Nvidia’s next-generation Vera Rubin architecture, components seeing triple-digit price increases include ABF substrates, PCBs, and memory. ABF substrate capacity is extremely tight, with shortages likely persisting through 2028 or even beyond 2030. Dell, HP, Nvidia, and Broadcom all cited DRAM, NAND, ABF substrates, and wafers as capacity constraints in recent earnings calls.
On the energy side, gas turbine lead times are extraordinarily long. Mitsubishi, Siemens, and GE Vernova have backlogs extending past 2030. Ordering a gas turbine today might mean delivery in 2030. “Power is probably the worst bottleneck right now,” the analyst said.
Copper vs. Optics: A Long Coexistence
In data center networking, the analyst believes copper and optical will coexist well beyond 2027. While optics will ultimately win — “nothing travels faster than light” — co-packaged optics (CPO) face yield and thermal challenges and remain expensive. Near-package optics may see adoption in 2027, with CPO ramp in 2028-2029, but meaningful penetration likely won’t occur until after 2030. Hyperscalers are motivated to extend copper’s life as long as possible to save costs amid soaring memory expenses.
Market Implications
- Memory stocks: Cyclicality persists, but LTAs may soften downturns. Investors should watch for signs of contract renegotiation and inventory buildup.
- Power and industrials: Gas turbine makers have multi-year backlogs, limiting incremental upside but providing revenue visibility.
- Semiconductor materials: ABF substrate and PCB suppliers are key beneficiaries of the AI buildout’s hidden bottlenecks.
- Optics: Lumentum, Broadcom, and Coherent remain positioned for long-term growth, but the transition from copper will be slower than bulls expect.
- Hyperscalers: Free cash flow is trending toward zero as capex surges. The ROI question remains unanswered and will drive sentiment.
Key Takeaways for Investors
The AI infrastructure trade is not just about GPUs. The real constraints — and opportunities — lie in memory, power, and advanced substrates. Investors should scrutinize claims of long-term demand visibility and prepare for cyclicality in memory, even if LTAs provide some cushion. The copper-to-optics transition will be gradual, and power generation remains the most intractable bottleneck. As the analyst put it, “Every great technology goes through a period of massive investment, and then it starts to slow down. AI will mature too.”
