Technical Deep Dive
The convergence of AI compute and energy materials is not a coincidence but a systemic outcome of two deep technical trends: the scaling laws of large language models (LLMs) and the electro-chemical physics of battery systems.
AI Compute Chain Architecture: The explosion in AI server revenue—companies like Inspur Information and Sugon reported AI server revenue growth exceeding 70%—is directly tied to the shift from traditional CPU-centric servers to GPU-accelerated, high-density clusters. These clusters rely on NVIDIA H100/H200 and domestic alternatives like Huawei Ascend 910B, interconnected via high-speed optical modules (800G/1.6T) from Zhongji Innolight and Eoptolink, whose revenues grew 80-120%. The thermal management challenge is acute: a single NVIDIA DGX H100 rack consumes up to 40kW, requiring advanced liquid cooling solutions from companies like Shenling Environment, whose data center liquid cooling revenue tripled. On GitHub, the open-source project ColossalAI (27k+ stars) has become a reference for distributed training optimization, reducing GPU memory usage by up to 80% through techniques like ZeRO optimization and tensor parallelism, directly lowering the cost of scaling.
Lithium Battery Materials Physics: The margin recovery in lithium carbonate and cathode materials (e.g., Tianqi Lithium, Ganfeng Lithium, Ningde Times' upstream partners) stems from a technical inflection point: the shift from LFP (Lithium Iron Phosphate) to high-nickel NCM (Nickel Cobalt Manganese) and solid-state batteries. High-nickel cathodes require ultra-pure lithium hydroxide, which commands a premium. The capacity consolidation phase (2023-2024) saw lithium carbonate prices crash from 600,000 RMB/ton to 80,000 RMB/ton, forcing high-cost brine and lepidolite producers out. Now, with demand from 4680 cylindrical cells and semi-solid batteries ramping, only producers with integrated refining and low-cost spodumene sources (like Tianqi's Greenbushes mine) are profitable. The open-source battery simulation tool PyBaMM (Python Battery Mathematical Modelling, 1.5k+ stars) is increasingly used by Chinese battery manufacturers to optimize electrode design and predict degradation, reducing R&D cycles by 30%.
Data Table: AI Compute vs. Lithium Materials Performance (2025 Annual Reports)
| Metric | AI Compute Chain (Avg. of Top 5 Firms) | Lithium Battery Materials (Avg. of Top 5 Firms) |
|---|---|---|
| Revenue Growth (YoY) | 65% | 45% |
| Gross Margin Change (YoY) | +8 ppts | +15 ppts |
| R&D Spend Growth | 55% | 25% |
| Capex Growth | 90% | 35% |
| Net Profit Margin | 12% | 18% |
Data Takeaway: While AI compute shows higher revenue growth and capex intensity, lithium materials demonstrate superior margin recovery and profitability, reflecting the different stages of their cycles—AI is in heavy investment phase, while lithium is in harvesting phase after consolidation.
Key Players & Case Studies
AI Compute Chain:
- Inspur Information (000977.SZ): The dominant AI server maker in China, with a 40% market share. Its 2025 report showed AI server revenue of ¥45B, up 75%, driven by orders from Baidu, Alibaba, and ByteDance for LLM training clusters. The gross margin improved to 14% from 11% as it shifted to higher-value liquid-cooled servers.
- Zhongji Innolight (300308.SZ): The leading optical module supplier, with 800G modules now accounting for 60% of revenue. Its net profit surged 120% to ¥3.2B, as it secured contracts for 1.6T modules with a major US hyperscaler (likely Google or Meta).
- Shenling Environment (301018.SZ): A thermal management specialist, its liquid cooling revenue hit ¥1.8B, up 200%. It supplies cooling for the new Alibaba Zhangbei data center, which uses 100% renewable energy.
Lithium Battery Materials:
- Tianqi Lithium (002466.SZ): Reported a net profit of ¥12B, up 300%, as its Greenbushes mine (51% stake) produced lithium spodumene at $400/ton, while spot prices recovered to $1,200/ton. The company is investing ¥5B in a lithium hydroxide plant in Western Australia to serve the solid-state battery market.
- Ningde Times (300750.SZ): While primarily a battery maker, its upstream subsidiary, Ningde Times Lithium, reported ¥8B profit, up 200%. It is deploying AI-based battery management systems (BMS) using reinforcement learning to optimize charging cycles, extending battery life by 15%.
Data Table: Competitive Landscape in AI Servers vs. Lithium Refining
| Company | Market Cap (¥B) | Key Product | Gross Margin | Key Risk |
|---|---|---|---|---|
| Inspur Information | 180 | AI Server | 14% | US export controls on GPUs |
| Sugon (603019.SH) | 120 | AI Server (Huawei-based) | 12% | Lower performance vs. NVIDIA |
| Tianqi Lithium | 150 | Lithium Spodumene | 70% | Lithium price volatility |
| Ganfeng Lithium (002460.SZ) | 100 | Lithium Hydroxide | 55% | High-cost Argentine brine |
Data Takeaway: Inspur and Sugon face an existential risk from US chip bans, while Tianqi's high-margin spodumene gives it a structural cost advantage over Ganfeng's brine operations.
Industry Impact & Market Dynamics
The dual-engine pattern is reshaping China's industrial landscape in three ways:
1. Capital Reallocation: In 2025, total capex from China's top 10 tech firms (Alibaba, Tencent, Baidu, ByteDance, etc.) reached ¥450B, with 70% directed to AI infrastructure (up from 40% in 2023). This is crowding out investment in traditional real estate and consumer internet. Meanwhile, lithium miners are raising ¥80B in equity and debt to expand refining capacity, with a focus on high-nickel and solid-state materials.
2. Supply Chain Localization: The AI compute chain is driving a domestic semiconductor ecosystem. Huawei's Ascend 910B chip, manufactured by SMIC using a 7nm-equivalent process, now powers 30% of new AI servers in China, up from 10% in 2024. This is reducing reliance on NVIDIA, but at a 40% performance penalty. The open-source CANN (Compute Architecture for Neural Networks) toolkit from Huawei (GitHub, 8k+ stars) is gaining traction as a compiler for Ascend chips, enabling PyTorch model deployment.
3. Energy-Compute Co-location: A new business model is emerging: data centers are being built adjacent to lithium battery recycling plants and solar farms. For example, a 500MW solar farm in Qinghai now directly powers a 100MW AI training cluster, with excess energy stored in lithium iron phosphate batteries from BYD. This reduces energy costs by 30% and carbon footprint by 50%.
Data Table: Market Size and Growth Projections
| Segment | 2025 Market Size (¥B) | 2028 Projected Size (¥B) | CAGR |
|---|---|---|---|
| AI Server (China) | 180 | 450 | 25% |
| Optical Modules (800G+) | 60 | 180 | 32% |
| Liquid Cooling Systems | 25 | 80 | 34% |
| Lithium Hydroxide (Battery Grade) | 120 | 200 | 12% |
| Solid-State Battery Materials | 5 | 40 | 68% |
Data Takeaway: The fastest growth is in liquid cooling and solid-state battery materials, both driven by the need for higher efficiency and energy density in AI and EV applications.
Risks, Limitations & Open Questions
Despite the bullish narrative, several risks loom:
- Geopolitical Overhang: US export controls on advanced GPUs (H100, B200) could cripple Inspur and Sugon's ability to deliver high-performance servers. The recent US CHIPS Act expansion may further restrict Huawei's access to EDA tools. If the 7nm process is blocked, the domestic AI compute chain could stall.
- Lithium Price Volatility: While lithium prices have recovered to $1,200/ton, they remain 80% below the 2022 peak of $6,000/ton. If EV demand slows (e.g., due to EU tariffs on Chinese EVs), prices could crash again, wiping out margins for high-cost producers like Ganfeng's Argentine operations.
- Energy Constraints: China's data center electricity consumption is projected to reach 400 TWh by 2028, equivalent to 5% of total national generation. Without massive grid upgrades and renewable integration, AI compute growth could be throttled by power shortages.
- Technical Debt: The open-source ecosystem for domestic AI chips (Ascend, Cambricon) is still immature. The PaddlePaddle framework (Baidu, 20k+ stars) is the most popular, but it lags behind PyTorch in community support and model availability. This could slow adoption of domestic hardware.
AINews Verdict & Predictions
The A-share 2025 annual reports confirm that the AI compute and green materials convergence is not a fad but the defining structural trend of the decade. Our editorial judgment is clear:
1. The AI compute chain will bifurcate: Companies with access to NVIDIA GPUs (via gray channels or partnerships) will outperform those relying solely on domestic chips. Expect Inspur to form a strategic alliance with a US cloud provider to secure chip supply, while Sugon doubles down on Huawei.
2. Lithium materials will see a 'super-cycle' within a cycle: The shift to solid-state batteries (expected commercial deployment by 2027) will create a new demand wave for ultra-pure lithium. Tianqi Lithium, with its low-cost spodumene, is best positioned. Ganfeng must acquire brine assets in Chile or Argentina to survive.
3. The 'digital-green' synergy will create a new asset class: We predict the emergence of 'AI-Energy' REITs in China by 2027, where data centers and renewable energy farms are bundled as yield-generating assets. This will attract pension funds and insurance capital, further accelerating the buildout.
4. Investors should watch for the 'battery-as-a-service' model: Companies like Ningde Times are already piloting battery leasing for data centers, where they own the lithium iron phosphate batteries and charge for usage. This could decouple energy storage costs from upfront capex, making AI compute more scalable.
What to watch next: The Q2 2025 earnings calls of Inspur, Tianqi, and Zhongji Innolight for guidance on capex plans. Also, monitor the progress of Huawei's Ascend 920 chip (expected 2026), which aims to match NVIDIA's H200 performance. If successful, it will trigger a re-rating of the entire domestic AI supply chain.