Excess Returns — Jason Hsu: China AI gap, capex arms race, S&P seven

Opening

China's AI models are treated as on-par with the US on software (open-source as strategy), energy/transmission is a Chinese structural edge, hardware captures rent until overcapacity, and the S&P 500 is a concentrated one-tree AI/CapEx bet — not the diversified passive sleeve retail thinks it is. Jack Forehand and Justin Carbonneau interview Jason Hsu (~55m), founder/CIO of Rayliant Global Advisors and Research Affiliates co-founder. Ground covered: Anthropic-CEO safety letter vs US–China race; DeepSeek→near-parity path; open-source vs closed rents; where AI profits accrue (hardware/energy → integrators/apps); NVDA/TSMC export controls as accelerant for domestic catch-up; Mag7 CapEx arms-race health; S&P seven-stock concentration; momentum crashes vs value; China retail-driven alpha. YouTube auto-captions (asr — "Rayliant" as "Railant," "TSMC" as "TFMC," "Kimmy"/Kimi as gap reference). Watch.

Key takes

US AI safety slowdown without China coordination is self-stunting: regulate US labs and Chinese firms "simply leapfrog." Hsu: assume China is on-par on model development today, perhaps edged on some real-world models; Stanford gap cited by host as maybe <5%; Hsu's earlier DeepSeek read was "~80% of the way… at like 1/20th the cost" (asr — host recounting). DeepSeek's origin as an HFT hedge fund noted as the non-Big-Tech exception; other Chinese model shops look like US Big Tech franchisees. [asr]

China open-sources models as hardware-subsidy strategy — closed US models hold temporary rent that open-source convergence erodes. In equilibrium most value is not in the model layer: SAS/integrators pick cheap open models, stitch pipelines, charge for solutions; users pay for energy/tokens. ~5% model-quality gaps become "indiscernible" when price gaps are large (hiring analogy). China would welcome mutual open-up "for safety" because it equalizes access they already grant outward. [asr]

Energy/transmission is a Chinese competitive advantage Hsu credits to (sometimes) effective central planning: multi-source (nuclear, world's-largest solar, Central Asia pipelines, Oman oil outside Hormuz), ultra-high-efficiency long-distance grid from Xinjiang solar. US DC power/politics struggle contrasted. [asr]

Hardware cycle: front-end infra (DRAM, GPUs, servers, DCs) captures rent while capacity is scarce — multi-year / ~five-year (asr) build cycles; overcapacity then shifts rents to application/SAS layers on cheap infra. Host: demand still unmet so overcapacity may be further out than classic tech cycles; Hsu doesn't disagree that the hardware rent window can run. China chip catch-up: "nowhere close" to NVDA GPU or TSMC foundry quality today, but closing speed "rapid and predictable… matter of years" — export controls accelerated domestic effort once framed as a weapon. [asr]

Mag7 CapEx has a healthy and an unhealthy reading. Unhealthy: arms race where you overspend so the other can't leapfrog you → declining marginal return, software "overcapacity" (feature-rich products that can't raise price), bad for today's CapEx giants. Healthy/solvency: not close to running out of money — generating large cash flow, spending it, and pulling future cash flow forward via borrowing that is not distress financing. [asr]

S&P 500 as "passive diversification" is a fiction under AI concentration — "one tree." Active investors making a conscious AI factor bet are fine; passive buyers told they're diversified are exposed to MSFT/Meta-class drawdowns without knowing it. Market beta is the AI CapEx theme today (analogy: pre-GFC beta was real estate/finance). Multifactor / low-vol / dividend sleeves are the actual diversifiers; China barbell (SOE value/dividends + speculative AI growth) contrasted with US "just one thing." Korea index called out as essentially two stocks (SK Hynix/Samsung) — "hard to beat two stocks," not efficient diversification. [asr]

Momentum crashes (semis reverse) are a known fat-left-tail feature that resets crowdedness; pair with value, which often pops on the other side. ML now lets Rayliant run 200+ factors with robustness/hyperparameter tools that once took months — nonlinear multifactor portfolios with fewer names than the benchmark because "S&P has 500 stocks but really it's got seven." [asr]

Key math

DeepSeek earlier read: ~80% capability at ~1/20th cost (asr — host citing Hsu) — catch-up baseline. [asr]

Stanford US–China gap ~<5% (asr — host) — near-parity claim Hsu broadly accepts. [asr]

Hardware/DC cycle ~five years before capacity online (asr) — rent-window framing. [asr]

US national wealth ~$17T → ~$170T over ~30y of China trade (asr) — macro interdependence aside. [asr]

US gov expenditure >30% of GDP; >80% of that welfare-oriented (asr) — macro frame, not AI math. [asr]

China equity ownership ~80–90% retail (asr) — alpha-source claim. [asr]

S&P "really… seven stocks" vs 400+ ballast (asr) — concentration. [asr]

Rayliant multifactor: 200+ factors (asr) — process density. [asr]

Quotes

"We should just assume today that they are on par with the US when it comes to model development." — Jason Hsu [asr]

"In equilibrium most of the value is not going to be in the models… you're really just paying for the energy token." — Jason Hsu [asr]

"When you have an arms race… you're basically spending money with increasingly lower return coming back from it because everyone's overspending." — Jason Hsu [asr]

"These firms are… pulling future cash flow to today… this is not these firms borrowing money because they don't have money." — Jason Hsu [asr]

"That is not what is in your S&P 500 portfolio today at all… you're really actively choosing to bet the farm on one thing." — Jason Hsu [asr]

"S&P has 500 stocks but really it's got seven stocks." — Jason Hsu [asr]

Variant perception

Priced in — China is closing the model gap; open-source threatens closed-lab rents; power is a bottleneck; Mag7 CapEx is enormous; index concentration is widely discussed.

What's new — Open-source-as-China-hardware-subsidy and equilibrium rent stack (hardware/energy → integrators, not models); export controls as accelerant of catch-up rather than durable moat; CapEx split into arms-race-unhealthy vs cash-flow-healthy; explicit "passive is active one-tree" framing from a factor CIO; Korea two-stock index as cautionary twin of S&P-seven.

Bear case — Near-parity may overstate production-grade reliability and chip dependence; hardware rent window could compress faster if neocloud/open silicon lands; arms-race framing could be wrong if ROI stays high; factor pitch is also product marketing for Rayliant.

Discount — Quant/factor shop talking its book on multifactor vs cap-weighted. China expertise is a Rayliant franchise — optimism on Chinese market structure and energy may be selection-biased. Safety-letter segment is geopolitical opinion, not portfolio math. ASR garbles firm/chip names.

Positioning

AI capex durability — MIXED / soft WEAKENS on arms-race, soft STRENGTHENS on solvency. Unhealthy overspend → lower returns and eventual overcapacity hurts today's CapEx cohort; cash-flow strength and unmet demand argue the build continues near-term — durability of spend ≠ durability of returns.

HBM supply binds — STRENGTHENS (soft). DRAM/GPU/server named as front-end rent collectors while capacity scarce; Korean memory as index-dominating bottleneck names.

Enterprise agent stall — NEUTRAL. Integrator/SAS layer as eventual rent home implies applications matter, but no F500 production-$ evidence.

Inference margin inversion — STRENGTHENS (soft). Open-source convergence + "pay for the energy token" + model-layer rent erosion is exactly the price-vs-cost pressure path on closed labs — without disclosing lab GM figures.