Xing Ziqiang: AI Investment Enters Halftime, a 9·24-Style Reversal Is Unlikely
Miles Bennett
Why "halftime," not game over?
Big Tech's planned AI capex — roughly $800 billion this year, $1.2 trillion next — has not been cut, and supply-chain orders show no cracks.
But the market has already priced in the next one to two years of good news. This means → even if fundamentals hold, share prices may stall.
In plain terms = the shovels (chips, compute) are still selling, but the shovel-selling story is fully told. Now the market wants to see who actually digs up gold.
What is this sell-off really about?
Crowded positioning is the core pressure. It is not that investors stopped believing in AI — it is that too many believe. Xing cites South Korea, where leveraged ETFs offering multiple-times-long exposure have proliferated — a classic late-stage signal in a thematic bull run.
The second pressure is a funding drain: he estimates AI-related companies need to raise roughly $1 trillion in equity and debt over the next year, steadily pulling liquidity from stock and bond markets.
This means → the market has become hyper-sensitive to marginal shifts in inflation, rates, and central-bank policy — any tremor gets amplified.
How is China's AI path different?
The key number: the token cost of China's domestic large models — the price to process a small chunk of text — is roughly one-tenth of the U.S. equivalent.
Xing proposes a "dual-track compute" framework: frontier model training still requires cutting-edge chips, but inference and applications need far less per-chip performance. Domestic chips can gain ground through state-run compute centers, improving through real-world use.
In plain terms = training a large model is like building a rocket — you need the best parts. Running a large model is like driving a taxi — good enough will do. Domestic chips have an opening on the taxi side.
Why won't a 9·24-style reversal arrive soon?
Xing's report title states it outright: "The 9·24-style inflection point has not arrived." Although some Q2 economic indicators weakened, policy remains focused on tech, energy, and supply-chain security — not broad-based stimulus.
He estimates that unspent bond quotas and newly created financial instruments leave more than RMB 2 trillion of fiscal firepower available in H2 — but the money is earmarked for power grids, energy storage, AI, and compute infrastructure, not direct consumer handouts.
This means → the second half is more likely about accelerating deployment of money already budgeted at the start of the year, not a sudden deficit expansion or a new wave of special bonds.
How to bridge the K-shaped split?
The upward leg — AI, the "new three" in energy, hard tech — is pulling away from the downward leg: domestic demand, consumption, property, and jobs. The gap is hard to close quickly.
Xing offers three proposals: ① redirect part of the trade-in subsidies from durable goods toward dining, tourism, and services; ② modestly narrow fiscal support for hard-tech exporters already riding a global upcycle, freeing resources for service-sector tax cuts; ③ keep strengthening the social safety net to reduce households' precautionary saving.
This reflects a judgment: the upward leg already has its own momentum; fiscal resources should tilt toward the downward leg.
Is it time to buy the dip?
Xing says explicitly that he cannot answer whether the market is at a bottom.
From a longer-term view, he sees China holding a comparative advantage in two areas of the AI second half: energy transition and low-cost large models.
Whether that thesis pays off depends on two things: how fast domestic compute infrastructure is built out, and whether AI applications actually reach commercial viability.
Content is for reference only, not financial advice.