Rare earth elements erbium and yttrium are critical to AI infrastructure, with China holding supply dominance.

Beyond electricity, two other natural advantages China possesses have been recognized overseas amid this wave of AI data center expansion.

Recently, IEEE highlighted two names that had previously received little market attention—

Erbium and yttrium.

One is hidden within the long-haul optical communication networks between data centers, and the other is embedded in the gas turbines powering the data centers—each corresponding to the two most pressing concerns in today’s AI infrastructure:

Internet and power.

In simple terms, erbium enables data to travel farther, while yttrium is used in high-temperature materials for gas turbines, helping them operate stably at higher temperatures and improving power generation efficiency.

As AI data centers expand rapidly, these two uses are also becoming increasingly important.

On one hand, computing power clusters are growing larger, with an increasing amount of computing power being deployed across campuses and cities, requiring more high-speed, long-distance optical networks between data centers;

On the other hand, AI servers are driving electricity demand to hundreds of megawatts and even gigawatts, prompting a renewed surge in gas turbine orders in the United States to meet新增 grid capacity.

As AI capital expenditures expand from GPUs to optical networks and power systems, two previously obscure metals buried deep in the supply chain are now being brought back to the table.

More notably, according to IEEE, our country accounts for nearly 100% of the production of these two rare metals.

In other words, the benefits from the current expansion of AI data centers both domestically and internationally may not be limited to obvious sectors such as GPUs, optical modules, and power equipment.

Keep digging further up the supply chain—we still have cards up our sleeve.

So, how exactly do erbium and yttrium participate in the construction of AI data centers?

Next, let’s go through them one by one.

Erbium has the chemical symbol Er, atomic number 68, and is a heavy rare earth element in the lanthanide series.

△

Its most direct connection to AI data centers occurs in the long-haul fiber optic networks between data centers.　

It is well known that today's data centers are no longer simply about cramming servers into a single building.

As the scale of individual AI clusters continues to grow, training, inference, and storage resources are becoming distributed across different campuses, cities, and even countries, leading to a sharp increase in data exchange between data centers.

For example, Google has built a private wide-area network connecting its global data centers, using fiber-optic cables spanning land and sea as the backbone.

This year, the total length of fiber on this network has exceeded 10 million kilometers, connecting 43 Google Cloud regions.

And these fiber optic links, often spanning hundreds or even thousands of kilometers, raise the most fundamental question: the farther the light travels, the weaker the signal becomes.

To better understand erbium's role here, let’s take a brief look at the principles of optical communication.

After leaving the data center, the data is encoded into high-speed laser pulses that travel along optical fibers to the next node.

However, even though modern quartz optical fibers have minimized loss, the signal still continuously attenuates as the transmission distance increases.

After a data signal has been transmitted tens or hundreds of kilometers, it needs to be amplified again.

Previously, the standard approach was "optical-electrical-optical," meaning the optical signal was first converted into an electrical signal, amplified and processed, then converted back into an optical signal for transmission.

However, in modern DWDM (Dense Wavelength Division Multiplexing) optical networks, a single fiber can simultaneously transmit dozens or even hundreds of signals at different wavelengths.

If each channel requires conversion to electrical signals, processing, and then conversion back to optical signals, the equipment becomes more complex, and both cost and power consumption increase.

At this point, the erbium-doped fiber amplifier, or EDFA, comes into play.

Specifically, in an EDFA, engineers dope a short length of fiber with a small amount of Er³⁺ ions and then "pump" it with a laser at 980 nm or 1480 nm.

After erbium ions absorb energy and enter an excited state, when a communication optical signal passes through, they release new photons through stimulated emission, directly amplifying the original optical signal.

Clearly, this directly reduces the process of converting light into electricity.

Interestingly, the wavelength band where erbium performs best is around 1550 nanometers, which is also one of the windows with the lowest loss in modern silica optical fibers and is widely used for long-distance communication.

In addition, according to the International Telecommunication Union (ITU), EDFA can simultaneously amplify multiple optical signals of different wavelengths without requiring individual photonic-to-electronic conversions.

Therefore, in today's terrestrial backbone networks, submarine fiber-optic cables, and inter-data center connections across regions, EDFA has become a highly mature core device.

In other words, the connection between erbium and this round of AI data center expansion lies in the fact that as more data centers are built and computing power becomes more widely distributed, the demand for constructing and upgrading long-haul optical networks increases, driving up the need for EDFA.

However, it should be noted that erbium primarily benefits medium- to long-haul optical communications, not short-distance interconnects of tens or hundreds of meters within racks.

Moreover, EDFA is not absolutely irreplaceable.

According to ITU documentation, in addition to EDFA, Raman amplifiers, semiconductor optical amplifiers (SOAs), and O/E/O regeneration devices that require optical-to-electrical conversion can all perform various forms of link amplification or extension.

In real-world long-haul optical networks, hybrid amplification solutions combining Raman and EDFA are also used.

However, after decades of industrialization, EDFA has become the most mature technology pathway among them.

The element symbol for yttrium is Y, and its atomic number is 39.

Although it is not a lanthanide element, it is commonly classified within the rare earth system in industry due to its similar chemical properties to heavy rare earth elements and its frequent association with them.

Yttrium's relationship with data centers is primarily through the power systems that supply electricity to them.

No need to elaborate—everyone knows how much electricity data centers consume.

The U.S. Lawrence Berkeley National Laboratory estimates that by 2030, data centers could consume approximately 11.8% of electricity in the United States.

The U.S. Energy Information Administration also predicts that natural gas power generation will meet a significant portion of the growing electricity demand from data centers.

GE Vernova, a leading gas turbine manufacturer, explicitly identified data centers as a key driver of demand growth for gas turbines in its 2025 annual report.

Thus, the expansion of AI data centers began to propagate upstream:

After the shortage of GPUs came the shortage of transformers; after transformers, the grid capacity became insufficient. When the grid could no longer keep up with the pace of data center construction, companies began sourcing power directly from power generation facilities.

Gas turbines have therefore regained popularity.

In simple terms, a gas turbine burns natural gas to produce high-temperature gas that drives a turbine to rotate, which in turn drives a generator to produce electricity.

To improve efficiency, a gas turbine has a very straightforward method: burn hotter.

Thermodynamically, the higher the gas temperature, the greater the potential for increased turbine efficiency. However, the issue is that when temperatures rise, the turbine blades can no longer withstand the heat.

So, to make the turbine blades more heat-resistant, one of today’s main characters—yttrium—comes into play.

Currently, industry primarily uses yttria-stabilized zirconia (YSZ) as a thermal barrier coating.

Its principle involves adding approximately 6% to 8% yttria to zirconia.

Although pure zirconia has low thermal conductivity and is well-suited for thermal insulation, it undergoes phase transitions accompanied by volume changes when exposed to repeated temperature fluctuations, causing the coating to easily crack or peel over time.

After adding yttrium oxide, Y³⁺ ions enter the zirconia lattice and create an appropriate number of oxygen vacancies, thereby stabilizing the high-temperature crystal structure of zirconia and enhancing its stability during repeated thermal cycling.

The resulting YSZ provides both thermal insulation and resistance to prolonged high temperatures and repeated thermal cycling, effectively acting like a ceramic thermal shield for gas turbine blades.

This coating reduces the actual temperature experienced by the underlying metal blades and decreases the consumption of cooling air, allowing the gas turbine to operate at higher temperatures.

Early NASA comparative tests evaluated yttria, magnesia, and calcia-stabilized zirconia, ultimately concluding that yttria-stabilized zirconia was the optimal choice among the three, based on overall durability and processing cost.

So, to summarize, the logic behind Yi's data center is:

The greater the electricity shortage in data centers → the stronger the demand for new power generation → the higher the orders for gas turbines → leading to increased demand for high-temperature blades and thermal barrier coatings → further upstream to yttria.

Of course, this chain is longer than erbium.

Moreover, not every data center installs its own gas turbine, and gas-powered generation is just one of several new power sources being added to data centers.

Meanwhile, magnesium oxide, calcium oxide, and more advanced novel thermal barrier materials can be tried as alternatives to YSZ.

However, for now, yttrium oxide still presents a barrier for testing and engineering validation in the short term.

Finally, let’s look at the supply side.

To be honest, although our country possesses the vast majority of erbium and yttrium deposits worldwide, the real advantage lies primarily in our subsequent separation and processing capabilities.

Because rare earth elements are often mixed together and have extremely similar chemical properties, separating erbium and yttrium individually and achieving industrial-grade purity typically requires dozens to hundreds of solvent extraction steps, demanding high standards in production lines, processes, and expertise.

According to USGS data, between 2020 and 2023, 93% of the yttrium compounds imported by the United States came directly from China, resulting in a net import dependency of 100%.

According to IEEE, there are almost no companies outside of China capable of commercially scaling the production of erbium oxide.

Therefore, China's true advantage lies not only in mining but also in the industrial capability to consistently convert rare earth ores into high-purity materials such as erbium and yttrium through separation and purification.

As overseas demand grows, if local separation and purification capacity fails to keep pace, this supply chain advantage concentrated in the midstream sector could further translate into a price premium abroad.

However, it’s important to note that, as minor metals, erbium and yttrium are both very small markets.

Based on current oxide prices and publicly available production figures, the combined global upstream raw material market for erbium and yttrium is roughly in the range of $100 million to $200 million in a typical year.

In terms of production, the USGS estimates that global yttrium production in 2025, expressed as yttrium oxide, will be only about 10,000 to 15,000 metric tons; there is no authoritative global single-element statistic for erbium, but the overall market is only in the kiloton range.

In other words, this is not a commodity market worth hundreds of billions or trillions of dollars, but rather a small metals market with a limited size and highly concentrated supply.

Therefore, as AI data centers continue to drive demand for long-haul optical networks and gas turbines, even modest absolute increases in demand could create noticeable marginal changes within such a small market.

Of course, erbium and yttrium are not irreplaceable; it remains to be seen how much real incremental value AI will ultimately deliver.

Reference link

[1]https://spectrum.ieee.org/rare-earth-shortage?itm_source=homepage&itm_medium=hero&itm_campaign=hero-2026-08-27&itm_content=hero2

[2] https://ntrs.nasa.gov/citations/19760025104?utm_source=chatgpt.com

[3]https://www.gevernova.com/investors/annual-report/ceo-letter?utm_source=chatgpt.com

[4] https://www.itu.int/dms_pub/itu-t/opb/hdb/t-hdb-out.10-2009-1-pdf-e.pdf?utm_source=chatgpt.com

This article is from the WeChat official account "Quantum Bit," authored by: Focused on Frontier Technologies