AI spending can't grow forever with P Equity Research | EP 167
Summary
- P Equity Research expects memory to absorb roughly half of hyperscaler capex, although estimates vary wildly. Forecasts for 2027-28 range from 36% to 73%, while UBS’s $900 billion estimate for next year exceeds this year’s total hyperscaler spending. The investable point is directionally clearer than the exact number: “No one will argue that memory is expensive.”
- Today’s compute shortage is really a compound constraint spanning chips, memory, packaging, power, and construction. CLSA estimated a 73% gap between ASIC/GPU demand and supply through 2030, B200/B300 rental rates reportedly rose from roughly $5 to $7-8 per hour, and even decade-old Volta hardware remains useful. Yet the guest called energy “probably the worst bottleneck,” citing data-center delays and political resistance.
- AI infrastructure remains a boom-bust industry because spending cannot rise indefinitely, but its next downturn may establish a much higher floor. Even a 50% cut to an estimated $1 trillion of capex would leave $500 billion—three to four times the combined 2023 spending of Oracle, Google, Microsoft, Meta, and Amazon. P Equity Research said the key test is ROI; Logan added that until free cash flow changes, questions will continue.
- Long-term agreements can smooth memory earnings, but they cannot repeal cyclicality. Typical deals cover four to five years, SK Hynix reportedly has 50-70% of supply contracted, and Micron disclosed $22 billion of upfront payments; price floors may coexist with no ceiling. A specialist who had worked at Samsung, AMD, and Renesas nevertheless warned that customers can quietly escape commitments—or warehouse unwanted inventory until future demand “simply evaporates.”
- The less obvious supply constraints sit in ABF substrates, printed circuit boards, and possibly indium phosphide, while networking broadens the trade. ABF capacity may remain constrained beyond 2028, while networking capex could compound around 64% annually for five years. These constraints broaden the trade beyond GPUs, even as mature power-equipment names already carry order books extending past 2030.
- Copper and optics will coexist longer than optical bulls may expect. Near-package optics could scale in 2027 and co-packaged optics in 2028-29, but heat, yield, and cost may delay CPO dominance until after 2030. Hyperscalers facing soaring memory bills want to “use copper as much as possible,” although ultimately “nothing travels faster than light.”
- High Bandwidth Flash may complement rather than replace HBM as inference and agent workloads expand. Logan said Grok 4.7 was described at 2.1-2.3 trillion parameters versus 1.5 trillion for model 4.6. HBF could hold model weights or reusable prefill results, but unresolved problems—endurance, heat, throughput, and cost—make smaller private deployments more plausible than hyperscale adoption initially.
- China is unlikely to flood global DRAM immediately, but dismissing its longer-term threat would repeat the industry’s history. CXMT’s limited capacity, roughly 25% HBM3 yield, and technology lag constrain near-term supply, yet domestic demand and rapid equipment localization could support 20-25% global memory share eventually. Logan argued that slowing US model development when Chinese open source is perhaps six months behind would be “inviting more competitors,” not creating a durable safety advantage.
Deep dive
1. Compute scarcity is real, but the binding constraint keeps moving
P Equity Research divided AI infrastructure into “logic, memory, power, and network technologies.” Google and Amazon say cloud growth is compute-constrained, while CLSA estimates a 73% ASIC/GPU demand-supply gap lasting through 2030.
Market prices corroborate the shortage: B200/B300 rental rates reportedly rose from roughly $5 to $7-8 per hour. H100s still fetch around $25,000 used, and Nvidia’s nearly decade-old Volta architecture remains deployed well beyond its expected economic life.
P Equity Research’s portfolio-level evidence made the abstraction tangible: startups cannot pursue promising work because “without access to compute power, you’re stuck.” One small company celebrated securing a batch of chips with a cake.
The guest’s counterweight came from an AMD expert: enough chips may exist, but usable compute is limited by data-center power, advanced packaging, and memory. P Equity Research therefore called energy “probably the worst bottleneck,” particularly amid construction delays, moratoria, and political resistance.
2. Memory could consume half the AI buildout
The headline range is enormous. CLSA projected memory at 48% of 2027 capex, SemiAnalysis 36%, JPMorgan 49% in 2027 and 60% in 2028, Citrine 60% by 2028, and UBS first 63%, then 73%, for 2027.
Applied to estimated hyperscaler capex of $1.1-1.2 trillion, memory could receive $500-700 billion; UBS’s more aggressive estimate is $900 billion next year. P Equity Research’s honest conclusion: “I don’t think anyone knows exactly” because HBM, commodity DRAM, and NAND are mixed together.
Suppliers want HBM to reduce dependence on commodity pricing. SK Hynix could derive about 35% of DRAM revenue from HBM by 2030, nearly twice the current proportion, while specialized HBM may sustain better margins than conventional DRAM.
Agent workloads strengthen the whole hierarchy rather than one product alone. The guest cited bank research linking distributed agents to heavier CPU/GPU utilization and therefore more HBM, DDR5, and NAND, while cautiously allowing that flash “perhaps” benefits most.
3. LTAs soften the memory cycle rather than abolish it
P Equity Research rejected the permanent-supercycle premise: for memory to stop being cyclical, hyperscaler costs and spending would effectively need to rise indefinitely. Like PCs and smartphones, AI should eventually mature into maintenance spending unless enormous future free cash flow—possibly around 2028—changes the equation.
Long-term agreements contain four elements: duration, committed volume, pricing, and financial guarantees. Four-to-five-year terms are common, but Samsung reportedly received inquiries for 10-year contracts; the guest’s answer was blunt: “They can’t make predictions beyond 2 years.”
SK Hynix reportedly has 50-70% of supply locked into LTAs, while Micron disclosed $22 billion in upfront payments. SanDisk discussed 80% margins by the end of 2030, and a Micron employee said DRAM margins could theoretically reach 94-96% where floors exist without price ceilings—an outcome both considered unlikely. Uncontracted volume may carry even higher margins, but LTAs provide demand stability.
The sharpest dissent came from an expert who had worked at Samsung, AMD, and Renesas: LTAs may be “overrated” because parties quietly cancel them in downturns. Enforcing unwanted volumes merely makes customers warehouse memory, destroy the relationship, and emerge after expiry with years of stock and no new demand.
4. Even a severe capex correction leaves a larger earnings base
Oracle, Google, Microsoft, Meta, and Amazon spent about $150 billion combined in fiscal 2023. If an estimated $1 trillion next year were cut by 50%, the remaining $500 billion would still be three to four times that earlier level.
P Equity Research’s base case is not a 50% collapse—perhaps 30% is conceivable—but a slowdown as the industry matures. That creates a “new average” where memory margins remain healthier than in prior recessions, particularly with contracted volumes and financial guarantees.
P Equity Research pointed to rising OpenRouter token volumes and agents reaching accounting, operations, and other non-technical users. He explicitly separated token consumption from revenue. Separately, Logan cited an unverified estimate that only 2-3% of the world currently pays for AI.
5. Substrates and optics broaden the bottleneck trade
In the Vera Rubin materials list, ABF substrates and memory showed triple-digit growth. ABF shortages may persist beyond 2028—possibly to 2030—and Dell, HP, Nvidia, and Broadcom have all identified substrates alongside DRAM, NAND, and wafers as constraints.
Indium phosphide may be another pressure point because it feeds lasers and optics; AXT, Sumitomo Electric, and JX Advanced Metals are expanding supply. Networking itself could grow around 64% annually over five years, making lasers and interconnects central to system performance.
P Equity Research expects a hybrid copper-optical period. Near-package optics may enter mass use in 2027 and CPO grow in 2028-29, but yield, heat dissipation, and cost could postpone dominance until after 2030—perhaps somewhere in 2030-40.
6. Power is essential, but the obvious equipment trade may be late
SemiAnalysis reportedly expects advanced labs alone to add another 14 gigawatts next year, though P Equity Research declined to claim precision: delays and moratoria change the tally constantly. Securing land, power, and infrastructure may reveal demand better than headline announcements.
Energy remains the largest bottleneck, but Logan said long lead times and exhausted capacity make it probably not the area worth investing in now. Gas turbines illustrate the timing problem: Mitsubishi, Siemens, and especially GE Vernova already have orders extending beyond 2030; a turbine ordered today may not arrive until then. P Equity Research also singled out gas turbines for their long lead times and market concentration.
7. HBF is speculative, while China is the durable competitive variable
P Equity Research questioned High Bandwidth Flash because low endurance meets GPU heat—a “double whammy,” especially beside Chinese chips already described as overheating. A SemiAnalysis source instead saw HBF serving low-output, smaller private deployments; a Nintendo CTO expert call likewise suggested coexistence with HBM, not replacement.
Logan’s bull case was storage economics: Grok 4.7 reportedly carried 2.1-2.3 trillion parameters versus 1.5 trillion for model 4.6, with future models potentially much larger. Read-heavy flash could hold weights or precomputed prefill results as workloads shift toward inference and decoding.
Near term, CXMT cannot flood DRAM: its monthly wafer starts were put at 200,000-300,000, HBM3 usable-die yield around 25%, bit density three to four years behind, and HBM one to two years behind. Domestic Chinese demand should absorb substantial new capacity first.
Longer term, the guest sees 20-25% global memory share as achievable. China already consumes about 30% of world memory for PCs and smartphones, while domestic equipment shares reportedly rose from 30% to 40% in etching, 20% to 40% in deposition, and 5% to 15% in implantation.
P Equity Research illustrated the volatility of memory leadership: the United States controlled about 95% of the market in 1975, Japan later reached 85% while the US fell to 2% in 1990, and South Korea now has about 62%.
Export controls may accelerate that localization. P Equity Research argued that restrictions helped push Nvidia from roughly 90% of China’s market toward a projected 80% domestic-chip share by 2028—the same dynamic that could eventually strengthen CXMT and YMTC.
Logan argued that this competitive pressure also makes a voluntary AI slowdown implausible. With Chinese open-source models perhaps six months behind, pausing US development means “letting China catch up”; he favored holding labs accountable under existing cybersecurity and data-breach laws over broad government intervention.