Trillion-scale computing power is bottlenecked by a single tantalum capacitor.

Hundreds of servers reached the final assembly stage, only to be stuck on the production line due to a handful of tantalum capacitors.

Zhou Zihang, a supply chain executive at a leading domestic AI server OEM, told the Economic Observer that a small tantalum capacitor part holding back an entire batch of finished machines has occurred several times this year.

A business executive in charge of the tantalum capacitor division at a domestic listed company explained that along this computing power delivery chain, shortages of chips and copper foil halt production startup, and shortages of transformers stop power supply after servers are transported into industrial parks. But a shortage of tantalum capacitors blocks the very last step before machines roll off the production line for delivery. These components are not particularly noticeable, yet they can leave hundreds of finished machines stranded on the production line.

Tantalum capacitors in servers perform microsecond-level power compensation. As soon as an AI server is powered on, its graphics processor with a power consumption of over 1 kilowatt instantly draws hundreds of amperes of current. Tantalum capacitors must replenish power and stabilize voltage within microseconds; any delay of even half a beat can cause the entire machine to crash. As a result, the tantalum capacitor consumption per unit has risen from dozens of pieces in traditional servers to two to three thousand pieces in mainstream models.

Two hard constraints on the supply side are continuously widening the relevant computing power gap.

Both of these constraints are controlled by overseas parties. The first is resources: most tantalum mines are located in Africa, and high-grade tantalum powder required for capacitor production also relies on imports. The second is production capacity: over 70% of high-end production capacity is held by leading overseas industry players.

The aforementioned tantalum capacitor division executive stated that for domestic computing power enterprises, a transformer supply shortage is a hassle of queuing for goods, while a tantalum capacitor shortage brings the anxiety of having core lifelines controlled by others.

Under multi-layered ripple effects, the window for domestic substitution is opening up.

The aforementioned tantalum capacitor division executive told the Economic Observer that the localization rate of tantalum capacitors for AI computing power is gradually rising. The industry remains in the small-batch delivery phase in 2026, and is expected to see noticeable capacity ramp-up in 2027, with large-scale mass delivery not arriving until late 2027 to 2028.

01

A Supply Gap Spreading From Mines to Capacitors

Tantalum concentrate was the first to face supply disruptions.

Over 80% of global tantalum mines are concentrated in three African countries, most of which rely on small-scale manual mining with extremely low supply elasticity. After a landslide halted production at core mining areas in the Democratic Republic of the Congo in early 2026, roughly 15% of global tantalum raw materials vanished instantly. Resumption of production at such mines does not depend on capital expenditure, but on local security and governance environments, which are exactly the factors hardest to change in the short term.

Tantalum powder followed closely behind. Spot supplies in the midstream smelting sector are highly dependent on upstream mineral materials. Once the mining end tightens, both supply availability and bargaining space shrink simultaneously. The technical and production capacity thresholds for capacitor-grade high-capacity tantalum powder are even higher, with fewer alternative sources, further amplifying the supply crunch.

The most severe bottleneck lies in finished capacitors at the most downstream end of the industrial chain.

Kemet (under Yageo), the global leader in tantalum capacitors, has raised prices multiple times since June 2025. Its latest price adjustment was announced in March 2026 and took formal effect on April 1, with a noticeable price hike specifically for AI-dedicated polymer tantalum capacitors. Panasonic and AVX followed suit one after another, and the supply of high-end models remains persistently tight.

Server OEMs are the first to feel the pain of this shortage.

Zhou Zihang said: "The regular lead time for overseas tantalum capacitor manufacturers is extending, and prices have risen significantly compared to the same period last year."

To secure supplies, the company where Zhou Zihang works has started locking in long-term framework orders, forecasting demand 6 to 12 months in advance, and signing framework agreements with tantalum capacitor manufacturers to reserve production capacity. Even with locked long-term orders, capacity allocation from overseas manufacturers such as Kemet, AVX and Panasonic remains tight, making it impossible to fully avoid supply disruption risks.

The safety stock approach is also restricted by product specifications.

Zhou Zihang added that tantalum capacitors have extremely subdivided specifications: minor differences in parameters including voltage, capacitance value and equivalent series resistance make products incompatible for mixed use. It is impossible to stock all specifications, as neither the factory's capital nor warehouse space can bear the cost, so the company can only maintain small inventory buffers for high-frequency mandatory models.

What troubles server OEMs even more is that these components are so small and easy to overlook, yet carry non-negligible risks.

Zhou Zihang analyzed that in terms of BOM (Bill of Materials) cost proportion, the total cost of all tantalum capacitors in one AI server usually accounts for less than 1%, which is almost negligible compared to GPUs and CPUs. However, they are critical components in the power supply chain, and a single failed tantalum capacitor will trigger a single-board fault, making the entire machine unable to power on and operate.

Server contract manufacturers are also rearranging their production schedules.

Zhao Mingxuan, a supply chain project executive at a server contract manufacturer, also told the Economic Observer that the delivery lead time for tantalum capacitors has extended to 40 to 52 weeks, which exceeds the standard stocking cycle for regular server projects. Semi-finished machines can only be stacked in line-side warehouses, waiting for missing components to be replenished before being sent back to the production line. Disordered work orders directly reduce the production line utilization rate.

Zheng Guodong, a business executive at a leading tantalum powder enterprise, told the Economic Observer that the prices of tantalum powder and tantalum wire have risen in step with raw material prices, but shipment volume has not seen explosive growth. Product selling prices rise along with tantalum concentrate market conditions, instead of surging unilaterally, and there have been periods of price fluctuations driven by downstream wait-and-see attitudes and bargaining games. Fluctuations in mineral prices on the cost side significantly squeeze the enterprise's processing profit, and the raw material segment is not a "easy profit" sector as many outsiders imagine.

In Zheng Guodong's view, the incremental demand brought by the computing power track is real. Currently, orders directly supplied to tantalum capacitor customers related to computing power do not account for a particularly high proportion, but their growth rate is extremely prominent, with year-on-year increases significantly higher than those of traditional sectors such as consumer electronics, military industry and aerospace.

02

Strong Demand Meets Inelastic Supply

In January 2026, heavy rain triggered a landslide at the Rubaya mining area in the eastern Democratic Republic of the Congo. This mining area, which accounts for roughly 15% of global tantalum supply, was forced to halt production. Its resumption involves safety rectification and large-scale infrastructure construction, and it will take 6 to 12 months to return to normal mining status.

Mining stopped, but demand did not.

Zheng Guodong told the Economic Observer that the tantalum capacitor consumption per high-power AI server has multiplied, and the ripple effect on the upstream side has driven a very noticeable increase in demand for high-capacity tantalum powder.

There is actually not much the supply side can do. After the Rubaya mining area shut down, the spot price of tantalum concentrate rose rapidly, spot products in market circulation shrank, and a rush to purchase goods emerged.

Zheng Guodong introduced that his company does not rely solely on supplies from the Democratic Republic of the Congo. Leveraging the group's resource layout, it has built a global procurement system to allocate supplies through multiple channels, and its production has not faced supply disruptions. Meanwhile, the company has signed medium- and long-term procurement contracts for ferrotantalum-niobium ore sources with relevant overseas mining enterprises as stable raw material supplements, and has also ramped up recycling efforts to realize closed-loop recovery of tantalum waste inside factories, as well as purchase industrial tantalum scrap and dismantled materials from scrapped capacitors from external sources.

However, overseas mineral resources and recycled materials can only cover part of the demand.

A source from a state-owned metal mining enterprise told the Economic Observer that tantalum is a niche scattered rare metal, mostly associated with other ores. It is impossible to quickly build new mines and expand production just because tantalum prices rise. Mining plants in Africa mainly adopt manual and small-scale mining operations, which are affected by multiple factors including security, geopolitics and policies, leading to great uncertainties in production resumption. The release cycle for new primary ore production capacity generally takes several years.

Beyond the mining end, downstream production capacity is locked to a certain extent. The four leading overseas giants Kemet, AVX, Vishay and Panasonic focus their capacity expansion on automotive-grade and industrial-grade tantalum capacitors, with limited incremental capacity for AI-grade products. According to relevant supply chain sources, major overseas cloud vendors such as Microsoft, Google, Amazon and Meta, as end-demand parties for AI servers, often bypass server OEMs to sign long-term framework agreements directly with capacitor manufacturers to lock in more than 70% of the aforementioned high-end production capacity, in order to support their own computing power expansion plans.

The construction to mass production cycle for a tantalum capacitor production line is 18 to 24 months, meaning there will be no new supply added before 2028.

Where does this supply-demand mismatch lie?

The aforementioned source from the state-owned metal mining enterprise explained that first, demand is fragmented: different AI server models have different requirements for tantalum powder in terms of particle size, purity and porosity. Downstream capacitor enterprises need many subdivided customized specifications, but upstream smelting enterprises prefer to produce large-batch standardized products, as small-batch customized orders have high costs and large R&D investment, leading to low willingness to actively accept such orders. Second, information is not fully connected: information barriers exist between server OEMs, capacitor manufacturers and tantalum powder enterprises, making it difficult for end-user working condition requirements to be fully transmitted to the upstream. It is common to see situations where sample parameters meet standards but batch production consistency fails to meet requirements. Third, the verification chain is too long: any change to a process parameter by upstream material suppliers requires downstream parties to complete a full round of reliability verification, which often takes more than a year, leading to high iteration and trial-and-error costs.

Precisely because the supply side cannot catch up quickly and the mismatch cannot be resolved in the short term, this round of tantalum price hikes is unlikely to reverse soon.

The aforementioned source from the state-owned metal mining enterprise judged that this round of price increases for tantalum as a computing power metal is not entirely short-term speculation, but a structural market trend driven by rising demand superimposed on supply disruptions. He predicted: "The possibility of a simple short-term pullback is low, but prices will not keep surging unilaterally forever."

03

Domestic Tantalum Capacitors Break Through Bottlenecks

A domestic tantalum capacitor enterprise with accumulated R&D experience in high-reliability electronic components is spilling its mature reliability technology into the computing power track.

A business representative from this enterprise told the Economic Observer that its tantalum capacitor products have realized small-batch supply for domestic AI servers. Capacitors in computing power scenarios need to operate 7×24 hours under high temperature for a long time and withstand repeated shocks from large ripple currents, bringing great pressure to mass production consistency.

He Junfeng works at another domestic tantalum capacitor manufacturer that has already delivered small batches of products. He told the Economic Observer that the company spent 14 to 18 months from sample submission to realizing small-batch delivery. In the early stage, the company had to complete laboratory parameter comparison, board-level testing, high and low temperature cycling, long-term full-load aging, then move on to small-batch trial production and joint debugging with finished machines. Customers need to install the capacitors on real AI machines to run 7×24 hours stress tests for several months.

Consistency is the hardest hurdle for domestic components to pass.

He Junfeng said that a single tantalum capacitor sample can perform very well, but in mass production of tens of thousands of pieces, the parameter dispersion between different batches will be higher than that of leading overseas manufacturers. What customers are most strict with is not the static nominal parameters, but the long-term stability under working conditions, including leakage current drift under high-temperature load, and equivalent series resistance rise after thousands of hours of aging.

He Junfeng said: "Even if all single tantalum capacitor samples pass tests, if individual abnormal samples appear in multiple rounds of small-batch verification, customers will directly suspend the component onboarding process, which means pausing the certification procedure for this model to enter the official BOM."

Powder material is an even more upstream hurdle.

Zheng Guodong added to the Economic Observer that AI servers have harsh operating conditions, and their requirements for tantalum powder are far higher than those for ordinary consumer electronics. They do not only pursue high specific capacitance (specific capacitance refers to the capacitance value provided per unit volume or per unit mass; high specific capacitance means components can be made smaller under the same capacitance value, which is a key indicator for high-density integration of AI servers), but also attach great importance to highly consistent performance between batches, with highly stable impurity content, particle size distribution and porosity. Tiny fluctuations in impurity content will cause abnormal leakage current of tantalum capacitors, and directly trigger single-board faults of servers.

He Junfeng also pointed out that powder material is the ultimate ceiling. He said that the domestic supply chain for mid- and low-end tantalum powder is already mature, but ultra-high specific capacitance tantalum powder adapted for polymer tantalum capacitors of AI servers has extremely high requirements for impurity control and batch consistency of particles. Domestic powder products are still in the iteration and ramp-up phase, and the company still needs to match some imported tantalum powder for its high-end models. Even if equipment, processes and customer certifications are all in place, without a stable supply of high-end tantalum powder, the production capacity of computing power tantalum capacitors cannot be released on a large scale.

Policy support is focused on both technology R&D and verification segments.

A source close to the Ministry of Industry and Information Technology told the Economic Observer that high-reliability tantalum capacitors are mainly developed through programs such as the Industrial Foundation Reconstruction Project, with a focus on key upstream materials including ultra-high specific capacitance tantalum powder and tantalum wire, as well as polymer tantalum capacitor body processes and reliability verification systems. Meanwhile, the government is promoting the improvement of public verification platforms for domestic electronic components, to reduce the testing costs for domestic component sample submission and iteration.

But the mass delivery of domestic components cannot be accelerated quickly.

He Junfeng believes that there are no shortcuts in the certification of AI server components. The full process from sample submission, multiple rounds of reliability testing, small-batch trial production to entering the official BOM generally takes 1 to 2 years. Even if production capacity is built, if customer certifications are not completed, the production line can only remain idle. Domestic substitution is a gradual process, and cannot be achieved overnight.

This article is from WeChat Official Account "Economic Observer", author: Wang Yajie, authorized for release by 36Kr.