CITIC Securities: Plummeting Inference Costs Set to Ignite New Wave of AI Computing Demand

Two Major Models Launch Simultaneously, Inference Cost Becomes the Core Battlefield

CITIC Securities recently published its latest research report, offering an in-depth assessment of the latest developments in the AI large language model industry. The report notes that OpenAI launched GPT-5.5 while DeepSeek simultaneously released its DeepSeek-V4 preview. Both heavyweight models have independently positioned "extreme compression of inference costs" as their core selling point, signaling that large model competition has officially entered a new phase where efficiency reigns supreme.

GPT-5.5 has dramatically reduced API call costs while maintaining powerful capabilities, while DeepSeek-V4 continues its signature "high cost-performance" approach, further lowering the inference barrier within the open-source ecosystem. Although the two camps differ in their technical approaches, their goals are highly aligned — making AI cheaper and more accessible.

Falling Costs ≠ Peak Computing Demand: The Jevons Paradox Returns

CITIC Securities specifically emphasized in the report that declining inference costs do not mean computing demand has peaked. On the contrary, this trend is likely to trigger the classic "Jevons Paradox" from economics — when resource efficiency improves and unit costs decrease, total consumption actually surges significantly.

Applied to the AI sector, the logic chain is remarkably clear:

• Lower per-call costs → More developers and enterprises willing to integrate AI capabilities
• Lower barriers to entry → AI use cases rapidly penetrate from high-value domains into long-tail scenarios
• Explosion of use cases → Token consumption grows exponentially
• Surging total demand → Computing infrastructure demand continues to climb

The report argues that the AI industry is currently at a critical inflection point, shifting from a "model capability race" to an "application scale race." Over the past two years, the industry's focus has been on who has the strongest model; going forward, the focus will shift to who can get more users to use AI more frequently.

AI Application Boom Imminent: Three Key Directions to Watch

CITIC Securities is bullish on a comprehensive boom on the AI application side and highlighted the following key trends:

First, Agent applications are accelerating toward deployment. With the sharp decline in inference costs, Agent application scenarios such as multi-step reasoning and long-chain task processing — previously too expensive to commercialize — will enter a genuine window for commercial adoption. From enterprise office automation to intelligent customer service, from code generation to research assistance, Agents are moving from proof-of-concept to large-scale deployment.

Second, edge AI and cloud collaboration are deepening. Lower inference costs make cloud-based large model calls more economically viable, further optimizing the edge-cloud collaborative AI experience. This will drive the flourishing of hardware ecosystems including smart terminals and wearable devices.

Third, vertical industry applications are accelerating adoption. AI solutions for healthcare, finance, education, legal, and other industries will see accelerated adoption due to cost reductions, transforming from "a luxury for a few leading enterprises" into "a standard tool for small and medium-sized businesses."

Dual-Track Strategy: North American Computing Leaders and Domestic Challengers

Regarding investment strategy, CITIC Securities maintains its long-term bullish outlook on two main tracks:

First, the North American computing power supply chain. The North American AI computing ecosystem centered on NVIDIA will continue to benefit from infrastructure demand growth driven by the global AI application boom. Cloud providers' capital expenditures are expected to remain elevated, with definitive opportunities across data center construction, high-end GPU demand, and network equipment upgrades.

Second, domestic AI computing power challengers. Against the backdrop of persistent international supply chain uncertainties, the demand for self-reliance in domestic AI chips, servers, optical modules, and other components is becoming increasingly urgent. The rise of domestic models such as DeepSeek-V4 is also providing strong demand-side support for the domestic computing ecosystem.

Outlook: The Efficiency Revolution Ushers in an Era of Universal AI Access

From a broader perspective, the inference cost revolution represented by GPT-5.5 and DeepSeek-V4 is essentially clearing the final economic barriers to the mass adoption of AI. Just as cloud computing's pay-as-you-go model gave rise to the mobile internet boom, the continued decline in AI inference costs will spawn a massive application ecosystem far beyond current imagination.

The core message from this CITIC Securities report is clear: The second half of the AI industry is not about shrinking computing demand, but rather a structural upgrade in computing demand driven by an explosion of applications. For investors and industry practitioners alike, now may be the optimal time to position for the dual-engine opportunity of AI applications and computing power.