Token Deflation: Why China's Cheaper AI Isn't Bearish for the Supercycle

💡Chinese open-weight models have narrowed the capability gap with US frontier labs to a matter of weeks by consistently delivering 90% to 95% of frontier performance across coding, reasoning, and mathematics benchmarks. This rapid convergence is attributed to several structural advantages, including cheaper electricity, architectural efficiency like sparse Mixture-of-Experts (MoE) designs, and advanced cluster-level engineering. For example, Moonshot AI's Kimi K3 climbed to third place globally on the Artificial Analysis leaderboard, just behind Claude Fable 5 and GPT-5.6, while being priced at a third to a fifth of its US rivals.

💡Chinese AI models possess key cost advantages over US counterparts primarily due to cheaper electricity, architectural efficiency, and sophisticated cluster-level engineering. Electricity in China's western computing hubs is roughly 40% to 60% cheaper than in the US, ranging from ¥0.28 to ¥0.35/kWh (3.9¢ to 4.9¢/kWh), which is the single largest recurring cost for inference at scale. Additionally, Chinese models often utilize sparse Mixture-of-Experts (MoE) architectures, activating only 3% to 5% of parameters per token, and employ cluster-level engineering to achieve system-level scale, compensating for individual chip limitations.

💡China's 'open-weight' distribution strategy is a deliberate tactic to commoditize the AI model layer and capture downstream monetization, effectively eroding the pricing power of US frontier labs like OpenAI and Anthropic. By allowing the global developer community to download, fine-tune, and stress-test their models for free, Chinese labs create a sticky dependency, as seen with Alibaba's Qwen family spawning over 160,000 derivative models on Hugging Face. This strategy aims to route users and developers towards their monetized layers, such as cloud compute rental (Alibaba Cloud, Tencent Cloud) and advertising (Douyin, Baidu), rather than direct model sales.

💡The Jevons paradox suggests that increased compute demand will occur despite falling token prices because as the cost per unit of intelligence decreases, new and previously uneconomical use cases become viable, expanding total usage. Historically, this expansion in usage has outpaced the price decline. For instance, agent workflows consuming 10x to 1,000x more tokens become affordable at lower price points. Usage-weighted token prices are down approximately 40% since late June, yet AI spend per employee among heavy-using enterprises rose close to 50% month-on-month in July, illustrating this market expansion.

💡China's industrial strategy for AI models mirrors its past approach in solar panel and electric vehicle manufacturing: identifying a cost advantage in a key input or process, aggressively scaling production with state-directed capital, and then exporting the resulting overcapacity into global markets at prices competitors cannot structurally match. This strategy aims to commoditize the model layer by open-sourcing weights and eroding the pricing power of competitors, thereby becoming the default starting point that routes users and developers toward China's monetized downstream layers and sovereign AI infrastructure exports.

💡Falling token prices do not indicate an AI bubble or correction for US hyperscalers because the Jevons paradox applies, where cheaper tokens expand the overall market by making new use cases economically viable, thus increasing aggregate compute demand. While model prices are collapsing, capital expenditure commitments keep rising because the expansion in usage outpaces the price decline. Demand for underlying compute is holding strong, as evidenced by H100 GPU rental prices rebounding off June lows while token prices continued to fall, indicating that the market for AI intelligence is widening rather than shrinking.

💡GPU rental prices indicate sustained demand for underlying AI compute by showing a rebound in H100 rental prices off their June lows, while token prices continued to fall. This dynamic suggests that even as the cost of AI intelligence decreases, the demand for the foundational hardware remains robust because cheaper tokens enable more extensive and novel AI applications. Furthermore, B200 GPU pricing has been even firmer, trading above its May peak, reinforcing that the market for physical compute infrastructure is not weakening but rather experiencing sustained high demand.

💡Chinese pure-play AI labs are experiencing structural net losses despite growth due to an aggressive domestic 'involutionary' price war, with API prices for models cut to as low as $0.20 to $0.95 per million tokens. This state-directed export strategy, which mirrors past deployments in other technology sectors, leads to individual market participants being unable to defend profit margins, even for technologically sophisticated services. For example, Zhipu AI reported first-half revenue growth of 404% YoY but still posted a loss of over 2 billion yuan, underscoring the severe profitability challenges.

💡US closed-source models maintain premium pricing despite lower token volume by offering deterministic reasoning, robust service-level agreements (SLAs), compliance certifications (SOC2, HIPAA, FedRAMP), and enterprise data privacy protections, which open-weight Chinese models struggle to clear in Western corporate environments. Enterprise customers are willing to pay premium rates for these assurances, as evidenced by Anthropic alone accounting for approximately half of all dollar spend on Vercel's AI Gateway while processing roughly 12% of total token volume. This indicates that token volume does not directly translate to revenue capture in the high-value enterprise segment.

💡Given the polarization of the AI model layer, recommended investment opportunities are primarily in US-aligned, supply-constrained hardware across Asia, insulated from the model price wars. These include semiconductors, advanced packaging, power, and thermal management. Investors should focus on physical bottlenecks where profits are heavily concentrated due to hard physical supply constraints. The analysis suggests avoiding Chinese LLMs as attractive investment opportunities, as China's rapid AI growth does not guarantee attractive margins, with price competition and US export controls limiting global monetization.