Goldman Sachs: China AI Demand Is Real, Recommends Hardware Sector
Nashnova编辑部
Goldman's One-Delta desk head Rich Privorotsky says China's AI demand has shifted from model launches to production workloads, with falling inference costs unlocking more compute consumption — a direct case for the hardware sector.
Inference is cheaper — so why are GPUs selling more?
The bear case ran like this: unit price drops → GPU-hours shrink → capex stalls.
Privorotsky's core call is the opposite: falling inference costs are not compressing demand — they are unlocking larger-scale usage.
This means → cheaper does not mean less buying; it means more users can afford to run more workloads, and hardware demand is driven by volume.
What do MiniMax and SenseTime's numbers actually show?
MiniMax's H1 revenue grew 283% YoY to $116.6 million; enterprise and API revenue surged 703%, now 63% of total.
July token consumption hit 20× January's level; August ARR — annualized recurring revenue, the monthly run-rate projected over a year — reportedly topped $800 million, roughly 80% from enterprise clients. In plain terms = this is not a chatbot traffic story; it is metered, pay-per-use billing inside production environments. Privorotsky's line: "20× token consumption is not a research cluster — it's a factory."
SenseTime's H1 revenue rose 23% YoY; generative-AI revenue grew 28%, accounting for nearly 80% of group revenue. Recurring revenue jumped 124%, and gross margins expanded. This reflects generative AI at these companies moving from research line-item to real P&L.
What does Zhipu's ultra-low pricing signal?
Zhipu (Z.AI) confirmed that the model previously tested anonymously as "Ox Alpha" is GLM-5.3-Flash, priced at $0.15 per million input tokens and $0.50 per million output tokens — the same ultra-low band as DeepSeek.
An early-bird promotion halves those prices through early September.
This means → China's leading model labs are collectively pushing inference prices to rock-bottom, reinforcing the "volume-for-price" logic — the lower the unit cost, the more aggressively enterprises plug in.
What should you actually focus on in Nvidia's earnings?
Nvidia's Q2 revenue rose 106% YoY; data-center revenue grew 117%. Q3 guidance beat expectations; management projected roughly 70% revenue growth through FY2028.
Privorotsky argues the real story is not the quarterly print but Nvidia's newly disclosed ecosystem financing structure.
Specifically: over $500 billion in third-party financing facilities — leveraged through private-credit partners including Apollo, BlackRock, Blackstone, Brookfield, and Goldman Sachs — with residual-value support covering up to 25% of a single deal; forward commitments total roughly $366 billion.
Why does this financing structure matter?
In plain terms = Nvidia is not just selling chips; it is helping customers fund chip purchases — turning large one-time capex into installment-like arrangements via third-party platforms and residual-value guarantees.
This means → Nvidia's ecosystem leverage far exceeds what any single quarter's revenue suggests; the "pipeline" of latent demand is much larger than the headline number.
This reflects the core reason markets remain cautious on hardware valuations even after digesting the earnings beat: whether China's AI production-side demand can keep delivering is the key node that will validate — or break — this leverage thesis.
市场有风险，内容仅供研究参考，不构成投资建议。