HomeArticle

A bearish report on Innolight has hit the biggest soft underbelly of the AI hardware industry.

硅基观察Pro2026-11-18 21:39
AI hardware cannot escape cycles after all.

A valuation calculation that bears a bearish view on Innolight has almost laid the valuation dilemma of the entire AI hardware assets on the table.

On November 17, a bearish calculation document about Innolight spread rapidly in the investment circle. According to the forecast of this material, by 2029, Innolight's annual net profit will only be about 300 billion yuan, corresponding to a reasonable market value of 3 trillion to 4.5 trillion yuan.

Considering the discount factor, even in the most optimistic scenario, there is still a pullback space of about 25% in the company's current stock price.

As soon as this calculation was released, it quickly sparked huge controversy.

It should be noted that Innolight's performance this year can be described as quite amazing. In the first three quarters of 2026, the company achieved an operating revenue of 695.47 billion yuan, a year-on-year increase of 178.13%; the net profit attributable to shareholders reached 222.38 billion yuan, a year-on-year increase of 211.81%. Calculated on this basis, the net profit margin attributable to shareholders is close to 32%.

Not only the operating data remains quite strong, but there is no obvious deterioration in orders, product prices and gross profit margins.

Sell-side institutions are even continuously raising their profit expectations. At the end of October, Huatai Securities predicted that the net profit attributable to shareholders of Innolight from 2026 to 2028 would reach 335 billion yuan, 807 billion yuan and 1132 billion yuan respectively.

On the one hand, there is the optimistic expectation that the net profit will exceed one trillion yuan in 2028, and on the other hand, there is the bearish calculation that the net profit will only be 300 billion yuan in 2029. For a leading AI hardware company like Innolight that has attracted much market attention, it is very rare that the future profit expectation can differ by several times.

In fact, the divergence faced by Innolight is also the epitome of the entire AI hardware sector.

Just today, NVIDIA released its latest third-quarter financial report, and its revenue still maintains an amazing growth. But on the other hand, with the fall of its stock price, its expected P/E ratio has dropped to about 12 times, hitting the lowest level since 2012.

Obviously, the performance is still growing at a high speed, but the market has generated huge concerns about AI hardware assets. This is also why a calculation based on the 2029 hypothesis has caused such a big reaction at this time.

300 Billion Yuan of Profit, How Is It Calculated?

Let's first talk about how this 300 billion yuan profit is calculated. The whole calculation process is actually very simple, divided into four steps.

Step 1, Assume that global AI capital expenditure will reach 800 billion US dollars in 2029.

This number does not come out of thin air. On November 13, semiconductor research institution SemiAnalysis updated its forecast for AI accelerator capital expenditure, lowering the 2029 forecast from 1.352 trillion US dollars to 796 billion US dollars, a drop of as high as 41%.

There are three main concerns behind this: AI application revenue falls short of expectations, data center financing is blocked, and the high premium of AI chips is difficult to sustain.

Step 2, How much of the budget can optical interconnection take?

This calculation assumes that by 2029, pluggable optical modules and near-package optical interconnection such as NPO will account for 7% of the total AI hardware capital expenditure. Among them, Scale-out interconnection across servers and cabinets accounts for 5%, and Scale-up interconnection between accelerators in the same cabinet contributes the other 2%.

Calculated at 800 billion US dollars, it corresponds to an optical interconnection market of 56 billion US dollars.

Step 3, Assume that Innolight takes 50% of the market share.

This assumption is actually quite optimistic. At present, Innolight's share in the global high-speed digital communication optical interconnection market is about 28.1%. If it reaches 50% by 2029, it means an increase of 21.9 percentage points.

Of course, this is not completely impossible. Innolight's share in the 800G market has exceeded 40%, and for new generation products such as 1.6T, the market estimates that its share even reaches 50%-70%.

With the continuous upgrading of AI data center network rate and the increase in the proportion of high-end products, Innolight does have the opportunity to further expand its market share.

But the problem is that the future optical interconnection market is not a simple extension of existing products.

The technical requirements for Scale-out and Scale-up are different, and there is also competition and substitution between technical routes such as NPO, CPO and pluggable optical modules. The fact that an enterprise occupies an advantage in 800G and 1.6T today does not mean that it will definitely take 50% share of the entire optical interconnection market in the future.

However, even according to this optimistic assumption, 800 billion US dollars of capital expenditure corresponds to a 56 billion US dollar optical interconnection market, and Innolight takes half of it, which means 28 billion US dollars of revenue, equivalent to about 1.89 trillion yuan.

Step 4, the most critical step, assume that the net profit margin drops to 15%.

This figure is significantly lower than the current level. It should be noted that in the first three quarters of this year, Innolight's net profit margin attributable to shareholders is about 32%.

Calculated at 15%, the revenue of 1.89 trillion yuan can only finally generate a net profit of about 283.5 billion yuan, which is rounded to 300 billion yuan in the calculation.

Next, give a P/E ratio of 10 to 15 times, corresponding to a market value of 3 trillion to 4.5 trillion yuan in 2029. Taking a three-year cycle, and discounting back to the present at a discount rate of 8% to 12%, the valuation range falls between 2135 billion yuan and 3572 billion yuan.

If the market value of 4850 billion yuan is taken as a reference, the neutral scenario corresponds to a drop space of about 44%, and even in the most optimistic scenario, there is a pullback space of about 26%.

Seeing this, everyone should find a very interesting point.

This bearish calculation, on the one hand, assumes that Innolight can take half of the global optical interconnection market, and on the other hand, believes that its net profit margin will drop from 32% to 15%, almost cut in half.

This seems somewhat contradictory. After all, according to the usual business logic, the higher the market share, the more obvious the scale effect, and the stronger the profitability of the enterprise tends to be.

Then why is this calculation so optimistic about market share but so pessimistic about profitability? Let's move on to the following part.

Demand Is Still Growing, But the Good Days for Optical Module Companies May Be Over

To understand why Innolight's net profit margin is only 15%, we must first figure out a question: Why was Innolight able to improve its profit margin so much in the past two years?

In the past two years, Innolight's gross profit margin has increased by more than 12 percentage points. For a hardware manufacturing enterprise, such an improvement is quite amazing.

There are two main reasons behind this.

First, the proportion of high-end products has increased rapidly. According to the 2025 annual report, the sales volume of Innolight's optical communication modules increased from 14.59 million units in 2024 to 21.09 million units, an increase of 44.6%, while the revenue of related businesses increased by 63.7% in the same period.

Further calculation shows that the average sales revenue of each optical module in 2025 increased from 1569 yuan to 1776 yuan, an increase of 13.2%, while the average sales cost decreased slightly from 1025 yuan to 1019 yuan.

Between the one increase and one decrease, the average gross profit of a single optical module increased from 544 yuan to 757 yuan, an increase of about 39%.

Second, high-end products are in short supply, and manufacturers have stronger bargaining power.

It should be noted that the optical module industry was originally an industry where prices fell very fast. According to the public disclosure of Cambridge Industries in 2023, optical modules usually reduce their prices by 10%-25% every year, and some mature products even drop by more than 30%.

But this round of AI wave has changed this situation. Innolight mentioned in its communication on November 3 that the annual price reduction of 800G and 1.6T products in 2027 is limited, and some products have not even been reduced in price.

This means that Innolight's profit explosion in the past two years has benefited from both product upgrading and the price dividend brought by the shortage of supply.

The problem is, how long can such a market environment last?

On the demand side, some signals worthy of vigilance have already appeared.

As mentioned earlier, SemiAnalysis has lowered its 2029 AI capital expenditure forecast by 41%. Christopher Wood, a strategist at Jefferies, also raised concerns that the AI investment cycle may peak in 2026 or enter a mid-cycle adjustment.

The reason is simple: higher and higher financing costs will completely end the expansion model of the AI industry that is highly dependent on external financing.

From early June to late October this year, Oracle's 5-year CDS spread expanded from 153 basis points to 329 basis points at one point, and CoreWeave's spread expanded from 452 basis points to 1163 basis points. The expansion of CDS spreads reflects rising concerns in the credit market about the risks of related enterprises.

On the other hand, the supply is still accelerating its release.

At present, leading manufacturers such as Innolight, Eoptolink, and HGTECH are all expanding production, and key component suppliers in the upper reaches such as DSP and lasers are also increasing their production capacity.

For example, Coherent, the US giant in optical devices, disclosed that its indium phosphide output capacity doubled year-on-year in 2026, and plans to more than double it again by the end of 2027.

In September this year, Jefferies predicted that the shipment of 1.6T optical modules in 2027 will increase by about 206% compared with 2026.

On the one hand, the growth rate of capital expenditure may slow down, and on the other hand, the industrial chain continues to expand production. Once the boom cycle reverses, the reflexivity of the industrial chain will amplify the impact of demand changes on corporate profits.

There may even be such a situation: data centers are still expanding, and optical module shipments are still growing, but the revenue and profits of enterprises have begun to decline.

This kind of story once happened in 2018.

At that time, North American cloud manufacturers built data centers on a large scale, driving the rapid growth of demand for 100G optical modules, and leading manufacturers made a lot of profits.

But the good times did not last long. As more and more enterprises had the mass production capacity of 100G, the originally tight supply-demand relationship began to ease, and price competition intensified accordingly.

According to a special report of China International Capital Corporation in 2021, the price reduction of some 100G optical module products in 2018 even reached 50%. Data cited by Guosen Securities from LightCounting also shows that the average selling price of 100G optical modules fell by about 39% year-on-year that year.

The result is that although the shipment of 100G optical modules is still growing, the price falls faster than the sales volume rises, and the global optical transceiver module market sales have dropped by about 3% instead.

What's more troublesome is that once the optical module industry has an obvious inventory cycle.

When supply is in short supply, cloud manufacturers often purchase in advance to ensure the construction progress of data centers, and even place duplicate orders to multiple suppliers.

In July this year, LightCounting pointed out that almost all key components of high-speed optical modules are in short supply, and the phenomenon of duplicate orders has appeared.

What problems will this bring? In the upward stage of the boom, the orders seen by suppliers may be higher than the actual consumption quantity of customers. But once the supply catches up, customers don't need to continue hoarding goods, and even may start to digest inventory.

Therefore, the weakening of optical module demand may come earlier than the actual decline in cloud manufacturers' capital expenditure.

LightCounting's review of industry data from 2012 to 2025 found that in history, the growth rate of optical module sales turned negative about half a year earlier than the growth rate of cloud manufacturers' capital expenditure.

It also proposed a scenario in July this year: even if the growth rate of cloud manufacturers' capital expenditure in 2027 only slows down from 70% to 50%, the growth rate of optical module sales may fall back to 30%-40%.

Of course, this does not mean that today's optical modules will inevitably repeat the mistakes of 2018. After all, AI is changing the application scenarios and value of optical interconnection, and the new generation of products also has higher technical thresholds.

But history has at least illustrated one thing: for cyclical industries, what is most likely to be overestimated is often not demand, but the profitability at the peak.

This is the reflexivity of the cyclical industry. As the saying goes, what you start with is what you will end up with.

Summary

There is a saying circulating in the investor circle: performance reflects the past, and stock price reflects expectations.

This is undoubtedly the best interpretation of the current AI hardware assets. On the other side of the ocean, NVIDIA's dynamic P/E ratio has dropped to about 12 times, the lowest level since 2012.

Early this morning, NVIDIA released its third-quarter financial report as scheduled. Although the revenue growth is still explosive, the gross profit guidance has begun to reflect the pressure and uncertainty of the supply chain.

According to the management's disclosure, the GAAP gross profit margin in the next quarter is expected to be 71%-72%, a month-on-month decrease of about 3 percentage points.

NVIDIA almost has the strongest technical barrier and pricing power in the entire AI hardware industrial chain, but it still cannot get rid of the constraints of the cycle. Not to mention those hardware manufacturers that mainly earn excess profits by relying on the shortage of supply.

At least for now, the bull market of AI hardware is most likely over.

This article is from the WeChat official account "Silicon-Based Observation Pro", written by A Qi, and published with authorization from 36Kr.