How should
The pricing logic of the AI industry chain is splitting. Upstream components like storage, CPUs, and optical modules are rising in price, while downstream token prices are competitively dropping—both happening simultaneously, and both showing signs of acceleration.
Zhang Xia from China Merchants Securities wrote in the strategy weekly report on June 15: “Currently, the AI industry chain is experiencing rare structural pricing. Upstream computing resource prices keep inflating due to rigid supply constraints, while downstream token consumption prices continue to deflate due to leaps in model capabilities and intensified market competition.”
This isn’t a simple “Is AI still working?” question. From 2023 to 2025, the market mainly trades the certainty of capital expenditure expansion by top overseas and domestic companies; starting from 2026, especially after June, trading becomes differentiated: upstream hardware price rises support profit expectations, but downstream applications are questioned—after token prices drop, can computing power become stable cash flow and profit?
The issue is, tokens are cheaper, but will usage be high enough to offset the unit price drop?
Upstream: Supply Can't Keep Up, Price Increase is Inevitable
The direct trigger for upstream inflation is a wave of raw material price hikes.
Tungsten hexafluoride is a typical case. Korean suppliers have notified Samsung and SK Hynix that contract prices for tungsten hexafluoride in 2026 will be raised by 70%-90%. Earlier in April, Japan’s major suppliers like Kanto Denka also issued supply disruption warnings. According to related information from Financial Associated Press, the price of Chinese tungsten hexafluoride with 99.999% purity has reached 1670-1810 yuan/kg, up 232.7% from 523 yuan/kg in the same period last year.
Price increases in electronic cloth are also pushing upstream prices. Looking further, AI hardware like storage, CPUs, and optical modules continue the price-hike trend.
Price increases can be quickly traded by the market because they are directly linked to EPS. When valuations are already high, funds prefer chasing upward earnings revisions; and since price hike logic is clear and data is easy to track, it naturally becomes the entry point for funds screening upstream AI targets.
A more midterm logic is supply-demand mismatch. Cloud vendors’ capital expenditure is rapidly rising, bringing sustained demand for GPUs, storage, optical modules, PCBs, and power distribution. But upstream capacity expansion has a cycle, and supply can't immediately keep up.
Global hyperscaler CAPEX is estimated to rise from $162 billion in 2022 and $154 billion in 2023 to $448 billion in 2025, with forecasts of $757 billion in 2026 and $920 billion in 2027. The steep demand slope makes upstream prices hard to stabilize.
Samsung and SK Hynix both mentioned in their April-May 2026 earnings reports and communications that AI-driven memory shortages will persist until 2027 and beyond, and some core customers have already locked in 2027 production capacity ahead of time.
Downstream: Prices Down, Usage Exploding
Token prices downstream are indeed dropping, and dropping fast.
The Silicon Data LLM Token expenditure index, which measures total market token spending, saw its first significant pullback recently after a single-sided rise since February this year—companies believe pricing for mainstream large models will continue to fall, prompting them to delay purchases. OpenAI is considering significant lowering of token prices due to pricing pressure from Anthropic; Tencent Cloud has announced a 66.67% price reduction for some model inputs.
Meanwhile, call volume is exploding.
Economics has a “Jevons paradox”—when resource efficiency improves and costs drop, total demand often rises rather than falls, as lower usage barriers foster more new scenarios. This has happened in coal, electricity, and bandwidth, and now it’s tokens’ turn.
According to OpenRouter data, as of June 8, 2026, global weekly token calls have reached 36.1 trillion, and are still growing exponentially.
Chinese figures are more intuitive: generative AI users increased from 249 million at the end of 2024 to 515 million in mid-2025, about double; daily token consumption in the same period increased from about 100 billion to 1.4 quadrillion in Q1 2026—about 1,400-fold growth. Users doubled, but consumption increased 1400 times, showing growth is mainly driven not by new users but by deeper usage per user and rapid penetration of agent-type applications.
Chinese manufacturers' low-price strategies are proving effective in this competition. The models with the highest global call volume are DeepSeek and Tencent Hunyuan, rather than overseas flagships with higher capability scores. By late 2025, open-source model token call share will be nearly one-third of total usage, with many developers beginning to use DeepSeek and Tongyi Qianwen instead of some closed-source APIs.
From an investment perspective, the key variable that determines the AI industry’s revenue space is not the price of a single token, but the growth rate of total token consumption—as long as the latter consistently outpaces the rate at which price drops, the overall industry revenue scale will not shrink.
Market Trading Logic is Switching
From 2023-2025, markets traded on the certainty of CAPEX expansion—so long as big players kept investing, upstream hardware had orders.
By June 2026, this logic starts splitting. Upstream continues to rise, downstream becomes competitive, and the market starts asking a tougher question: Can computing power be turned into sustainable cash flow and profit?
Last week, A-shares gave negative feedback to downstream AI deflation. But from a total perspective, the key to industry growth is not single token price, but speed of growth in total token consumption. When demand growth consistently outpaces price decline, overall AI industry revenues have hope to keep growing.
Anthropic’s launch of Claude Fable 5 on June 9 is a typical case, showing the strategic focus of top AI companies is shifting.
Fable 5 already surpasses any previously publicly released model in capability, with even greater leads on longer, more complex tasks. Yet its pricing is more than half lower than the earlier Mythos Preview—input $10/1M tokens, output $50/1M tokens.
This company didn’t take the highest price at its strongest capability, but set pricing even more aggressively. The reason is simple: for enterprise customers, being first in benchmark tests is no longer most important—scalable, low-cost deployment is.
2 Trillion Yuan Computing Power Investment: Where Does the Money Go?
On the policy side, the “National Data Infrastructure Construction Guidance” issued in January 2025 clarified investment direction during the “15th Five-Year Plan”: initial industry estimates say data infrastructure attracts direct investment of about 400 billion yuan per year, driving a future five-year investment scale of about 2 trillion yuan. At the same time, this round of computing power infrastructure requires at least 80% of AI chips come from domestic suppliers such as Huawei and Cambricon, with NVIDIA and AMD directly excluded.
According to EPOCH AI estimates, a typical 1GW AI data center requires about $3.8 billion in upfront CAPEX. Breaking up the 2 trillion yuan investment, the benefit scale for each segment is:
- GPU chips: 39%, about 780 billion yuan—this is the most elastic part of the entire industry chain
- Electric power & supply/distribution: 21%, around 420 billion yuan (generators, transformers, UPS each about 100-120 billion yuan)
- Land, buildings & data center engineering: 11%, roughly 220 billion yuan
- Optical modules, switches, cooling facilities: each 4%, about 80 billion yuan each, total 12%
Based on 2 trillion yuan total investment, about 780 billion yuan goes to domestic computing power chips, split over 5 years, driving over 150 billion yuan annual demand for domestic AI chips.
Localization is progressing faster than many expect. Domestic chip market share will reach about 41% in 2025, break 52% in the first half of 2026, surpassing NVIDIA. Huawei Ascend is the biggest winner, with 812,000 shipments in 2025, nearly half of domestic chip shipments; Alibaba Pingtouge ranks second with 265,000; Baidu Kunlun and Cambricon each about 116,000.
In May 2026, the China Information Security Evaluation Center for the first time listed “AI training/inference chips” as a distinct evaluation category, with 9 products including Ascend 310/910 and Cambricon Zhenwu M530/M890 obtaining level I certification valid for three years. This certification directly guides procurement for government, state-owned enterprises, and telecom.
Electric Power: Underestimated Beneficiary
Of the 2 trillion yuan investment in computing power infrastructure, about 420 billion yuan will drive supporting construction in power, cooling, and networking equipment.
Data from China Academy of Information and Communications Technology shows that in 2025, national computing power center electricity consumption reached 196 billion kWh, growing far faster than the average for all social electricity usage. The National Energy Administration expects that during the “15th Five-Year Plan”, annual added computing power electricity will be over 100 billion kWh, reaching about 800 billion kWh by 2030, accounting for about 6% of all national power consumption.
“East Data, West Computing” and “Power-Computing Coordination” are now national policy directions. The logic is: demand is concentrated in the east, while wind, solar, and hydro resources are abundant in the west—moving computing power westward can absorb green electricity, while “west power to east” still requires efficient transmission equipment. AI high-density cabinets are extremely demanding for supply stability, so efficient and energy-saving transformers, and high-voltage DC power systems see directly boosted demand.
The logic for energy storage is also shifting. Data centers need 24/7 uninterrupted power; storage can both replace diesel generators as emergency backup and serve as a main power source connected to the grid for daily supply. By late 2025, total installed modern energy storage reached 136 million kW/351 million kWh, up 84% from late 2024—AI infrastructure is increasingly an important variable in added storage demand.
Three Things to Watch: Price, Usage, CAPEX
This round of AI industry chain differentiation cannot simply be summarized as “upstream good, downstream bad”.
Upstream price increase depends on whether it translates to profit. Materials, storage, optical modules, and power distribution equipment price hikes are more easily traded in the short term, but midterm depends on whether capacity expansions ease supply-demand constraints.
Downstream token price drops depend on usage. As long as total token consumption growth outpaces price declines, overall AI industry revenues stand a chance to rise. Conversely, if price drops don’t bring enough usage, pressure will be more apparent at the application end.
Computing power infrastructure depends on CAPEX implementation. 2 trillion yuan data infrastructure investment, domestic AI chip replacement, and power/energy storage support are the most solid anchors in this chain.
Risk Warning and DisclaimerThe market has risks, investment needs caution. This article does not constitute personal investment advice, nor does it consider individual users’ particular investment goals, financial circumstances, or needs. Users should assess whether any opinions, viewpoints, or conclusions in this article suit their own situation. If you invest accordingly, responsibility is yours.