Power semiconductors are seeing another round of price hikes: Why has the AI industry started scrambling even for mature chips?

What is the most scarce resource for AI?

The first reaction of many people is GPU.

Over the past few years, from NVIDIA to HBM, from advanced manufacturing processes to high-speed interconnection, as long as it is related to AI servers, prices, production capacity and supply chains have almost all been re-examined.

But now, price hikes are starting to spread to a seemingly not so "advanced" corner — power semiconductors.

According to reports from media including Securities Times, since the beginning of this year, manufacturers such as Infineon, Texas Instruments, STMicroelectronics, and ON Semiconductor have successively adjusted prices of some of their products. Some Taiwan-based power semiconductor manufacturers also plan to raise prices of some spot products by about 10%-15% starting from October.

What is more noteworthy is that the industrial chain has begun to point to AI data centers as one of the reasons.

Here comes the question:

AI servers need the most advanced GPUs, why are power devices largely produced with mature 8-inch manufacturing processes also starting to be in short supply?

The answer lies in a frequently overlooked fact:

AI servers do not only have GPUs.

To make several chips worth tens of thousands of yuan or even more run stably in one server, a huge set of power supply, conversion, protection and heat dissipation systems is required behind them.

The stronger the AI computing power, the higher the value of this "behind-the-scenes system".

This means that AI is transmitting the semiconductor boom from the most top advanced chips down layer by layer.

The more expensive the GPU is, the more important those "unremarkable" chips next to it become

In an ordinary computer, the power consumption of a CPU may only range from tens of watts to hundreds of watts.

But when it comes to AI servers, things are completely different.

High-performance GPUs and AI accelerators require larger power supply, and multiple high-power consumption chips may be deployed in one server cabinet.

After the computing power density rises, the first change is:

Power is getting harder to deliver.

After electricity enters the data center from the grid, it cannot be directly sent to the GPU.

It needs to go through multiple links such as power supply system, server power supply, voltage conversion, and power management, and finally convert the voltage into the state that the chips can actually use.

A large number of power semiconductors are used in these links.

What are they responsible for?

Simply put, it is to control and convert electricity.

For example, converting alternating current into direct current, converting higher voltage into the low voltage required by chips, while improving conversion efficiency, reducing losses, and protecting equipment under abnormal conditions.

Looking at one single device, it is far less expensive than a GPU.

But a large number of such devices are needed in one AI server.

The more servers there are, the greater the power, and the denser the data centers, the greater the demand.

Therefore, the expansion of AI computing power will eventually form a very long industrial chain:

Increase in GPUs → Rise in server power consumption → Upgrade of power supply system → Increase in demand for power devices.

In the past, when the market talked about AI hardware, more attention was paid to the most expensive computing chips.

But after the construction of AI data centers enters the large-scale stage, the industrial chain begins to find that:

What really determines whether a data center can work stably is never just whether there are enough GPUs.

It is also power.

Why is the "mature manufacturing process" starting to be tight in supply?

The semiconductor industry often gives people an impression that:

The more advanced the manufacturing process, the more valuable it is.

3nm is more advanced than 7nm, and 7nm is more advanced than 28nm.

Then here comes the question:

Why have the 8-inch production lines that have existed for many years also seen changes in supply and demand?

The reason is that not all chips need to pursue the most advanced manufacturing process.

GPUs pursue higher transistor density, lower power consumption and stronger computing power, so they keep moving towards advanced manufacturing processes.

But power devices are completely different.

Many power semiconductors pay more attention to high voltage resistance, large current, stability, cost and long-term reliability.

For these products, using mature processes for production is more economical.

This is also why a large number of analog chips, power devices, automotive chips and industrial chips around the world still rely on 8-inch wafer fabs.

They are not "advanced", but very important.

Problems also arise from this.

Manufacturers of advanced processes can invest huge sums of money to build new wafer fabs, because the price of high-end chips is high enough.

However, after years of development of the 8-inch mature process, many production line equipment have been used for many years, and the industry has long faced fierce price competition in the past, so manufacturers' willingness to expand production on a large scale is not always strong.

This has created a special situation:

The demand was originally very stable, and the supply was also relatively stable.

Once a new source of demand suddenly appears, the balance between supply and demand may change rapidly.

AI data centers have precisely become such a variable.

Reports show that industrial institutions have noticed that AI-related demand is taking up an increasing share of the production capacity of some 8-inch mature processes.

This means that the logic of the mature process in the past, which was "sufficient supply and continuous price cuts", is changing for some products.

Of course, this does not mean that the entire 8-inch market is in short supply, let alone that all mature process products will see price hikes.

The situation varies greatly for different chips, different wafer fabs and different customers.

But one trend is worth paying attention to:

Mature process no longer means permanent oversupply.

This round of price hikes is not exactly the same as the previous "chip shortage"

The semiconductor industry is no stranger to price hikes.

During the global "chip shortage" period from 2020 to 2022, automotive chips, MCUs, power devices and other products all experienced obvious supply shortages.

But if this round is only understood as a repetition of the last cycle, the most important changes may be missed.

The previous chip shortage was largely caused by the disturbance of the epidemic, logistics problems, mismatch of supply chain inventory, and the sudden recovery of demand for automobiles and consumer electronics.

The new variable that appears today is that AI has changed the power demand structure in the electronics industry.

In the past, data centers also needed power supplies.

But the power density of traditional servers is not comparable to that of today's AI servers.

When the power consumption of a single chip rises and the computing power that a cabinet can carry becomes higher and higher, the importance of power supply efficiency is also rising synchronously.

Assume that a data center consumes 100 units of electricity.

If 5% of the electricity is lost during the power conversion process, 5 units of electricity are converted into heat.

If the loss is reduced to 3% through power devices with higher efficiency, it seems that there is only a 2 percentage point difference.

But on the scale of the year-round operation of a large data center, these 2 percentage points represent quite considerable electricity bills, heat dissipation costs and infrastructure investment.

Therefore, in the AI era, what power semiconductors are facing is no longer just the problem of "selling a few more devices".

More importantly:

How much are customers willing to pay for higher efficiency?

This will directly affect the product structure.

The importance of high-efficiency power supplies, power modules and related semiconductor devices may all increase with the investment in AI infrastructure.

As a result, an interesting change has taken place in the semiconductor industry:

AI is pushing the most advanced chips to 3nm and 2nm;

On the other hand, it has pushed some mature process products back to the spotlight of the industrial chain.

10%-15% price hike cannot be simply understood as the entire industry is raising prices

For this kind of industry news, the most common misinterpretation is that after seeing the words "price hike", people directly equate it with the fact that the industry has fully entered a boom cycle.

In fact, it is not that simple.

In the currently disclosed information, some manufacturers involve specific products, specific customers or the spot market.

What some Taiwan-based manufacturers plan to adjust is also the price of some spot products, with a range of about 10%-15%.

Spot price is not the same as long-term contract price.

Large customers and small customers do not get the same price.

The supply and demand situation of different models, different delivery dates, and even different factories of the same enterprise may be different.

Therefore, to judge how strong the boom of this round of power semiconductors is, instead of only focusing on the price hike range, it is better to look at four indicators at the same time.

The first one is delivery lead time.

If the product price rises but the delivery lead time is not significantly extended, it may just be that manufacturers are adjusting their pricing strategies.

If both the price and the delivery lead time rise, the signal of tight supply and demand is usually stronger.

The second one is capacity utilization rate.

Whether the wafer fab is really busy is more important than the price hike notice.

The third one is inventory.

One of the most worrying situations in the semiconductor industry is that downstream parties hoard goods in advance for fear of price hikes.

The sudden increase in short-term orders seems to indicate strong demand, but in fact it may just be that inventory is transferred from the upstream to the downstream.

The fourth one is demand duration.

The growth of AI servers is a real trend, but how many years the demand for power devices can last and how many products it can penetrate still need to be observed.

Therefore, we cannot simply draw the conclusion of "full reversal of power semiconductors" just because some products have raised prices.

What is really important is that the demand structure is changing.

Why does the domestic supply chain pay special attention to this round of changes?

Power semiconductors are also a field that the Chinese semiconductor industry has long focused on layout.

Compared with the most advanced logic chips, the domestic industrial chain foundation of mature process-related products is relatively more complete.

From wafer manufacturing, to MOSFETs, IGBTs, to power modules and related packaging, China has formed a relatively large industrial system.

In the past few years, this market has also experienced a round of fierce competition.

After a large number of enterprises expanded production, the prices of some products kept falling.

Therefore, for domestic enterprises, what is really worth paying attention to in this round of changes is not "overseas manufacturers raise prices, so domestic products will definitely benefit".

This logic is too simple.

The more critical are three questions.

First, who has actually entered the AI data center supply chain?

Power devices used in ordinary consumer electronics do not have exactly the same requirements for efficiency, stability and service life as power supplies for data centers.

Having production capacity does not mean being able to enter the high-value customer system.

Second, who can move from single devices to solutions?

Data centers are paying more and more attention to the overall power supply efficiency.

The future competition may not only be how much a single chip sells for, but the power supply architecture, module design and system efficiency.

Third, who has the ability to continue investing?

Although many power semiconductors use mature processes, it does not mean that there is no technological upgrading.

Materials, packaging, thermal management and device structures are still making continuous progress.

Especially in high power density applications, every small improvement in efficiency may become an important cost advantage for customers.

Therefore, the enterprises that can really share the dividends of AI data centers may not be the enterprises with the largest production capacity, but more likely the enterprises that can enter high-end applications, pass long-term certification and continuously improve efficiency.

AI is rewriting the value of "mature chips"

One of the most noteworthy things in the AI industry is that it is constantly breaking the original industrial classification.

In the past, people used to divide semiconductors into advanced and mature categories.

Advanced processes represent high growth and high value; mature processes mean full competition and falling prices.

But the reality is becoming more complex.

In an AI data center, there are not only the most advanced GPUs, but also a large number of analog chips, power chips and control devices produced with mature processes.

Without the latter, the former cannot operate normally.

This is why the larger the investment in AI infrastructure, the longer the industrial chain that will eventually be affected.

GPU is just the most noticeable link.

Server power supplies, power devices, heat dissipation, connectors, PCBs, energy storage and even power grid infrastructure may continue to expand outward along the computing power demand.

From this perspective, what is really worth paying attention to in this round of power semiconductor price adjustments is not that a certain chip has risen by 10% or 15%.

Rather, it has released a more important signal:

AI investment is moving from "buying computing chips" to the stage of "building a complete computing power infrastructure".

When an industry enters the stage of infrastructure construction, what determines the cost is no longer only the most expensive core component.

Those components that were once hidden inside machines, with unremarkable unit prices but determine efficiency, stability and energy consumption, will become more and more important.

The 8-inch mature process being re-focused by the market follows the same logic.

In the future, when judging the boom of the AI industrial chain, we may no longer only ask:

Are GPUs still in short supply?

We should also ask one more question:

To make more and more GPUs actually run, how much electricity, how many power supplies, and how many unremarkable power chips are still needed?

This may be the change that is more worth observing in the next stage after the demand for AI hardware continues to spread.

This article is from the WeChat Official Account "BT Finance Data Platform", the author is Jiang Xu, and published with authorization from 36Kr.