AI chip supply chain splits as upstream costs surge 233% and token prices crash

The AI industry is experiencing a rare structural pricing split: upstream hardware costs are surging while downstream token prices are collapsing, with both trends accelerating in the first half of 2026.

"The AI industry chain is undergoing a structural pricing divergence — upstream computing resource prices face supply-driven inflation while downstream token consumption prices face competitive deflation," Zhang Xia, a strategist at China Merchants Securities, wrote in a June 15 report.

The divergence is stark. Tungsten hexafluoride, a key material for semiconductor etching, has surged to 1,670-1,810 yuan per kilogram, up 232.7% from 523 yuan a year earlier, after South Korean suppliers notified Samsung and SK Hynix of 70%-90% contract price increases for 2026. Storage, CPUs, and optical modules are also extending their upward trend as hyperscaler capital expenditure accelerates — projected to reach $757 billion in 2026 and $920 billion in 2027, up from $448 billion in 2025, according to industry estimates.

The question investors are now asking: can surging token consumption volume offset the price declines enough to sustain industry-wide revenue growth?

Supply constraints drive upstream inflation

The upstream price surge traces directly to a supply-demand mismatch that shows no sign of easing. Hyperscaler capital expenditure has more than quadrupled from $162 billion in 2022 to a projected $757 billion in 2026, but upstream production capacity takes years to bring online. Samsung and SK Hynix both said in their April-May 2026 earnings calls that AI-driven memory shortages will persist through 2027 and beyond, with core customers already locking in 2027 capacity.

China's 2 trillion yuan data infrastructure plan, announced in January 2025, adds another demand layer. The plan requires at least 80% of AI chips to come from domestic suppliers such as Huawei and Cambricon, excluding Nvidia and AMD. Of the total investment, about 780 billion yuan — or 39% — will flow to GPU chips, making it the largest single beneficiary. Power and power distribution equipment accounts for 21%, or 420 billion yuan, while optical modules, switches, and cooling each take about 4%.

Domestic chip makers are already gaining ground. Chinese AI chip market share surpassed 52% in the first half of 2026, up from 41% in 2025 and overtaking Nvidia for the first time. Huawei's Ascend series led with 812,000 units shipped in 2025, nearly half of all domestic chip shipments, followed by Alibaba's Pingtouge at 265,000 units and Baidu Kunlun and Cambricon at about 116,000 each.

Token deflation meets the Jevons paradox

Downstream, the pricing picture could not be more different. OpenAI is considering significant token price cuts to counter Anthropic's aggressive pricing, while Tencent Cloud has already reduced input prices by 66.67%. Anthropic's Claude Fable 5, released June 9, outperforms all previous public models on complex, long-context tasks yet prices at $10 per million input tokens and $50 per million output tokens — less than half the cost of its predecessor Mythos Preview.

The Silicon Data LLM Token Expenditure Index, which tracks total market token spending, showed its first notable decline since February as enterprises delayed purchases in anticipation of further price drops.

But total consumption is exploding. Global weekly token calls reached 36.1 trillion as of June 8, according to OpenRouter data, and continue growing exponentially. In China, generative AI users grew from 249 million at the end of 2024 to 515 million by mid-2025 — roughly doubling — while daily token consumption surged from about 100 billion to 1.4 trillion over the same period, a 14-fold increase. The gap between user growth and consumption growth suggests the driver is not new users but deeper per-user engagement and the rapid penetration of agent-based applications.

This is the Jevons paradox in action: as the cost of a resource falls, total demand rises — often by enough to increase total spending. It happened with coal, electricity, and bandwidth. The question is whether token economics will follow the same pattern.

Investment implications

For investors, the divergence creates two distinct trade opportunities. Upstream hardware suppliers — chip makers, memory producers, optical module manufacturers, and power equipment providers — benefit from structural supply constraints and policy-driven demand that are likely to persist through at least 2027. The 2 trillion yuan Chinese infrastructure plan favoring domestic chips over Nvidia and AMD is a multiyear secular trend that reshapes competitive dynamics in the world's second-largest AI market.

Downstream, the picture is more nuanced. Companies that can capture the exploding token volume — model providers like DeepSeek and Tencent's Hunyuan, which rank among the most-called models globally — may offset price compression through scale. Open-source models now account for nearly one-third of total token usage, with developers substituting DeepSeek and Alibaba's Qwen for some closed-source APIs.

Power and energy storage sectors also emerge as indirect beneficiaries. China's AI data centers consumed 196 billion kilowatt-hours in 2025, and the National Energy Administration projects that figure will reach 800 billion kilowatt-hours by 2030, or about 6% of total national electricity consumption. Installed energy storage capacity reached 136 million kilowatts by the end of 2025, up 84% year over year, with AI infrastructure becoming a significant new demand driver.

The key risk is whether downstream token demand growth can sustain its exponential trajectory. If the Jevons paradox holds, total industry revenue grows. If price declines outpace volume growth, application-layer margins face sustained pressure. The next two quarters of hyperscaler earnings will provide the first real test.

This article is for informational purposes only and does not constitute investment advice.