Semiconductor on Uncle Xiang's Notebook

The big model development has indeed drawn the internet giants into the same competitive arena.

Tue, 12 May 2026 00:59:12 +0800

My previous article covered the semiconductor cycle, and I feel like there’s a piece of background/context missing.

Your judgment/conclusion regarding this point—the overall direction is correct. Furthermore, I believe it is a prerequisite that is easiest to overlook when trying to understand this current semiconductor boom.

A more accurate way to put it is not that “all internet giants are fighting in the same field,” but rather: Large Models have, for the first time, brought together major players previously scattered across different domains—such as search, advertising, social media, e-commerce, office productivity, cloud computing, and content distribution—into direct competition within the same technical stack.

This technology stack includes models, computational power, inference, cloud, Agents, distribution gateways, and commercialization closed loops. Everyone’s original “moat” is different, but now we must all fill the same gap. Those who fail to do so will see their future search entry points, ad pricing, office suites, e-commerce conversion, and social traffic distribution rewritten by others.

Why is your viewpoint basically valid

Previous internet competition was more like everyone staying within their own established territories (or: was more siloed).

Google / Baidu primarily focus on search.
Meta / Tencent primarily focus on social networking and traffic distribution.
Amazon / Alibaba primarily focus on e-commerce and merchant ecosystems.
Microsoft focuses on office software and enterprise software.
Although AWS, Azure, and Google Cloud are involved in the fight, that is more of a battle over cloud infrastructure and enterprise IT.

It’s different now.

Large models are not a single function; they are more like an “master switch” that will reverse-engineer/supersede all entry points. Search can be rewritten, ad placements can be rewritten, customer service can be rewritten, code generation can be rewritten, e-commerce recommendation can be rewritten, and enterprise knowledge bases can also be rewritten.

Therefore, although these giants are still making money in their respective established domains on the surface, fundamentally they are all competing for the same thing:

\[ \text{AI Competitiveness}=\text{Model Capability}+\text{Computational Resource Supply}+\text{Distribution Channels}+\text{Commercialization Closed Loop} \]

The difference merely lies in who is superior in which aspect.

Microsoft is strong in enterprise distribution and Azure.

Alphabet is strong in search, advertising, and proprietary model stacks.

Amazon is strong in AWS, chips, and enterprise cloud customers.

Meta is strong in traffic entry points and advertising scenarios.

Tencent is strong in super apps, gaming, and advertising implementation.

Ali is strong in e-commerce, cloud, and industrial clients.

Baidu is strong in search, AI Cloud, and ERNIE.

So your judgment is correct, but we need to add one caveat: Everyone is fighting on the same large battlefield, not using the same set of weapons or competing for the same group of users.

This is also why semiconductors are linked/grouped together

The previous text discussed storage and South Korean semiconductors; at that time, I focused on specific product categories like HBM, DDR5, and eSSD.

However, looking at a higher level, what truly fueled all these developments was the way the capital expenditure of internet giants began to concentrate in one direction.

It’s not that Google needs to buy a little more server capacity this year, but Meta might follow suit next year.

Instead, during the 2025 to 2026 phase, Microsoft, Google, Amazon, Meta, Alibaba, Tencent, and Baidu are almost all focusing on AI infrastructure, model training, inference services, Agent platforms, and AI distribution entry points. While their specific terminology varies, money is pouring into areas like GPUs, HBM, networking, SSDs, data centers, and power.

This is uncommon in internet history.

In the mobile internet era, many companies are involved, but not every single one needs to build its own operating system.

The short video era is booming, but not every company needs to build and train its own foundational recommendation model infrastructure.

The era of cloud computing has been long, but traffic platforms and gaming companies like Meta and Tencent have not all shifted their focus to the cloud.

Large models are something that almost every platform giant believes cannot be missed this time. This is what makes this semiconductor cycle different from previous ones.

What Is the Core Area of Competition Among All Players Right Now?

Let’s first clarify the definitions/scope. (Alternative options depending on context:)

We need to establish the parameters first. (If discussing data or technical metrics)
Let’s get aligned on the basis first. (If referring to an agreement or overall strategy)

For the table below, revenue and net profit should use the official disclosures based on each company’s latest completed fiscal year as of May 12, 2026. However, regarding the “AI Investment” column, disclosure methods vary greatly among companies; some report Capex, others provide a three-year investment plan, some report R&D figures, and some only provide quarterly expenditures.

So this column can only be used to view the strength and weakness direction, not for mechanical ranking.

Company	Existing Cash Cow / Core Business	Current AI Leverage Points	Latest Full Fiscal Year Revenue	Latest Full Fiscal Year Net Income	Most Recently Disclosed AI Investment Scope
Microsoft	Office, Windows, Enterprise Software, Azure	Azure + OpenAI Ecosystem + Copilot + Enterprise Agent	FY2025 Revenue $281.7 Billion	FY2025 Net Income $101.8 Billion	Approx. $80.1 Billion in CapEx for the first 9 months of FY2026; AI business annualized revenue reached $37 Billion in Q3 2026
Alphabet	Search, Ads, YouTube, Android, Cloud	Gemini + Search AI + TPU + Google Cloud	FY2025 Revenue $403.0 Billion	Total GAAP Net Income for four quarters of 2025 was approx. $132.2 Billion	FY2025 CapEx $91.4 Billion; FY2026 guidance $175B–$185B
Amazon	E-commerce, AWS, Advertising	Bedrock + Trainium/Inferentia + AWS AI infra + Nova	FY2025 Revenue $716.9 Billion	FY2025 Net Income $77.7 Billion	Approx. $147.3 Billion in CapEx over the last 12 months as of March 2026; Over 2.1 million AI chips delivered in the past 12 months
Meta	Facebook / Instagram / WhatsApp Advertising	Llama + Ad Recommendation + Meta AI + Smart Glasses / Agent	FY2025 Revenue $200.97 Billion	FY2025 Net Income $60.46 Billion	FY2025 CapEx $72.2 Billion; FY2026 guidance $125B–$145B
Tencent	Gaming, Social, Advertising, Fintech	Hunyuan + WeChat /

The most notable thing about this table is not who makes the highest profit, but rather who possesses enough robust “cash cow” resources to withstand the upfront investment period required for AI.

Microsoft, Alphabet, Amazon, and Meta—these four companies are all essentially printing money while expanding production.

Tencent and Alibaba are similar; their old businesses are still providing blood, so they can ramp up their AI investments.

Baidu’s problem is more apparent: they bet early, but their scale and cash flow depth are weaker than the previous few companies, making them naturally less resilient under pressure.

Although all are involved in the AI craze, each company’s approach is actually different

Upon closer examination, it can be seen that although these companies are all engaged in intense competition, their strategic positioning is different.

Category 1: Selling the Shovels While Also Entering the Field

Microsoft, Alphabet, and Amazon belong to this category.

They have cloud, chips or accelerators, enterprise customers, and models or model ecosystems.

What is most frightening about these companies is that, for them, AI isn’t a standalone product, but rather an upgrade tax on the entire platform.

You want to train a model, so you need to buy cloud services.

You need to infer, you need to rent a GPU.

You will need to become/develop an Agent, and you will need to purchase platform services.

You need to provide employees with AI tools, which requires purchasing related capabilities for Copilot, Gemini, and Bedrock.

So, they don’t just want to create a blockbuster AI App; they aim to integrate AI into every part of the entire IT budget (or: IT expenditure).

Type Two: The Traffic King – Optimize Your Distribution Channels First

Meta, Tencent, and Baidu are closer to this category.

What they value most are not enterprise contracts, but rather traffic funnels/entry points, advertising systems, content ecosystems, and high-frequency user time spent.

Meta is the most typical example. It doesn’t need to sell AI to all enterprise clients; merely by applying AI in ad targeting, content recommendation, creative generation, and conversational entry points, it is enough to enhance advertising efficiency and commercialization capabilities.

Tencent is similarly positioned. Areas like WeChat, advertising, gaming, and cloud are inherently natural use cases for embedding AI. While it doesn’t necessarily have to compete with Microsoft by tackling the entire Office suite, it will certainly focus on dominating AI assistants within WeChat, improving ad placement efficiency, enhancing game content production, and deepening collaboration in enterprise WeChat.

Baidu, on the other hand, is more like a traditional search company striving to make a turnaround. It has models, cloud capabilities, and search, but it also faces the greatest pressure from its old advertising business. Therefore, it needs to both maintain its traditional traffic base while simultaneously boosting AI commercialization.

Category Three: Ecological King, Aiming to Transform AI into Trading and Industrial Infrastructure

Alibaba is the most typical example among this category.

Its ambition is not merely to create Qwen, nor is it simply about divesting Alibaba Cloud. What it truly seeks is to integrate AI throughout the entire ecosystem of e-commerce, merchant tools, customer service, marketing, supply chain, and industrial cloud.

So Alibaba has repeatedly emphasized “AI + Cloud” in recent years. These four words are not just a slogan because it realized earlier than many companies that: the model itself might not be the most profitable, but the model will redefine cloud and transaction platforms.

Why Has This Conflict Concentrated Semiconductor Demand So Heavily

Seeing this, we can actually get back to semiconductors.

If only a single company is optimistic about AI, what it generates will primarily be themed investing (or thematic investments).

If there are seven or eight platform companies with the strongest cash flows, and simultaneously converge capex, R&D, models, inference, Agents, and distribution channels onto the same layer/level, then it will no longer be a theme, but actual orders (or: real demand).

But in reality, the final orders will turn into these things:

GPU and customized AI chips.
HBM and high-end DRAM.
Enterprise-grade SSD and higher bandwidth storage.
High-speed networking, switching chips, optical modules.
Data centers, power, cooling.

This is why when analyzing the semiconductor sector lately, you shouldn’t only focus on a single memory manufacturer or a single South Korean company.

The true background is: Internet giants have unusually ramped up their efforts simultaneously in the same arms race.

Conclusion

As of May 12, 2026, my assessment of your point is:

Basically correct.
A more accurate statement should be: internet giants are not completely engaged in “the same product track,” but rather in an arms race over the same foundational AI capabilities.
The importance of this matter is not just because AI will produce blockbuster applications, but because it is drawing originally scattered profit pools—such as search, advertising, social media, e-commerce, office work, and cloud services—back into one common underlying competitive framework.
This is also one of the core backgrounds explaining why the semiconductor industry, especially the chains involving GPU, HBM, storage, and data centers, has been so intense recently.

Well, how do I say it? Previously, everyone could maintain their own separate domains. That’s not feasible anymore. The matter of large models has evolved from merely a new trend into an infrastructure war that platform companies cannot afford to exit.

References

Notes on Writing

Original Prompt

The previous text mentioned high AI demand and surging semiconductor prices, but there is a background that is easy for everyone to overlook: large AI models. This is rare—it’s a field where all major internet giants have started heavily competing. Previously, they developed in their own respective domains; it is uncommon now for them all to be intensely vying within a single sector. First, can I confirm if my viewpoint is correct? Then, let’s outline the sectors each company is developing, along with their respective revenues, net profits, and investments in AI.

Writing Idea Summary

First, judge whether the point of view is valid, and then clarify “to what degree it is valid,” avoiding making sweeping statements with one sentence.
In the main body, rewrite “same industry track” as “same layer of technical stack,” as this better aligns with the true positioning of various companies.
In the middle, use a summary table to place cash cows, AI entry points, revenue, net profit, and AI investment metrics side-by-side, allowing readers to quickly survey the whole picture.
The article intentionally writes out that “AI investment cannot be directly compared,” to avoid forcing rankings using disparate disclosure metrics.
This piece deliberately does not include OpenAI, Anthropic, xAI, and ByteDance together in the table, as they are non-public or have varying disclosure methods, which could easily bias the comparison.

Expanded Brainstorming

Area	Whether to include in the main text	Reason
Including Apple in the comparison?	No	It is certainly affected by AI, but it is not a typical sample within this

The endpoint of this semiconductor cycle is unlikely to be in 2026.

Tue, 12 May 2026 00:19:22 +0800

Regarding this round of semiconductor trends, I temporarily do not see a peak in 2026.

If forced to give an initial judgment, as of May 12, 2026, I am more inclined to place the truly critical period between the second half of 2027 and the first half of 2028, rather than now. The core driver of this current uptrend—particularly in US listed storage and Korean semiconductors—is not a general recovery, but rather AI pulling HBM, DDR5, and enterprise SSD up simultaneously. If supply expansion fails, both prices and profits will rise together.

This also explains why companies like Micron, SK hynix, and Samsung seem to be “printing money” lately. The semiconductor cycle hasn’t vanished, but this time it is unlikely to collapse when demand first kicks in; rather, it is more likely to crash when capacity expansion finally catches up, and the market has already front-loaded two or three years’ worth of profit.

Executive Summary: Why I Doubt the ‘End by 2026’ Prediction

Place a few pieces of primary and semi-primary source materials on the table first.

Connecting these signals makes the answer clearer.

This cycle is neither the traditional PC cycle, nor is it driven by smartphone replacement cycles, and certainly not a simple rehash of the “semiconductor shortage due to the pandemic.” Its engine is the capital expenditure from major hyperscale AI data centers, and the massive throughput requirements for memory and storage needed by AI servers are simultaneously stressing HBM, DDR5, high-capacity DIMMs, and enterprise SSDs. Micron was quite direct in its investor materials last December: “Overall industry supply is significantly below demand in the foreseeable future, and this tightness will persist well beyond 2026.”

In other words, the market isn’t betting on “whether things will be better next year,” but rather on “whether the period of tension before 2027 will end.” If this premise is not broken, the cycle will struggle to wind down on its own by 2026.

Why Did This Round Ignite Both US Storage Stocks and Korean Semiconductors

Many people tend to view the US memory/storage sector and South Korean semiconductors as two separate issues. In reality, these two trends are expected to largely converge by 2026.

Micron represents the most direct memory sector beta on US stock markets. Reuters, in a report on March 19, 2026, noted that Micron’s stock price has already risen by over 61% this year. Furthermore, the company increased its capital expenditure plan for fiscal year 2026 by another $5 billion, bringing the total amount to over $25 billion. This move itself indicates one thing: management acknowledges that existing capacity is insufficient to meet the demand generated by the current wave of AI development.

Korea was even more dramatic. SK Hynix hit an all-time high on May 4, 2026, closing up 12.52% that day. The same Reuters report also mentioned that high-ranking officials from the Bank of Korea judged that this chip cycle upturn might last longer than previous cycles. This statement is not groundless, as the two most core memory manufacturers in Korea have repeatedly released similar information during their earnings reports and conference calls.

Samsung even stated directly during its conference call at the end of April 2026 that, “based solely on the already received demand for 2027, the supply-demand gap in 2027 will be even larger than in 2026.” This statement is very significant. It means that the market’s current surge isn’t due to people competing for 2026 earnings, but rather they are preemptively trading the 2027 shortage.

Therefore, the most crucial characteristic of this market cycle is not the “overall semiconductor boom,” but rather:

AI has turned the most lucrative memory sub-sector into a bottleneck for the entire industrial chain.
HBM will consume DRAM capacity and advanced packaging resources, simultaneously constraining commodity DRAM as well.
The demand for SSDs in servers and the inference side no longer simply follows PCs or mobile phones; rather, it is directly tied to large model infrastructure.

When these three factors are combined, Micron, SK hynix, and Samsung’s profit elasticity will be enormous. The memory industry is inherently a high operating leverage sector, so when ASP rises, profits often do not increase linearly; instead, they jump significantly.

How Have Past Semiconductor Cycles Typically Concluded?

I don’t want to turn this section into a semiconductor chronicle (or history). What is truly powerful/disruptive in terms of investment, and therefore worth remembering, are actually the following cycles.

Cycle Phase	Approximate Duration	Termination Details	Data Observed at End of Cycle
1998-2000 Uptrend	Approx. 2 years	Dot-com bubble burst, decline in PC and mobile demand, overstocking pressure	Global semiconductor sales dropped from $2.04 trillion in 2000 to $1.39 trillion in 2001, a year-on-year drop of 32%
2002-2007 Uptrend	Approx. 6 years	Global Financial Crisis first depressed valuations, then suppressed end-user demand	Sales dropped from $255.6 billion in 2008 to $248.6 billion, and again by 9% to $226.3 billion in 2009
2016-2018 Uptrend	Approx. 3 years	Slowing growth rate in the second half of 2018, price cycle reversal combined with trade friction	Global sales dropped to $412.1 billion in 2019 (a year-on-year drop of 12.1%); memory sales dropped 32.6%, and DRAM dropped 37.1

Looking at this table, the pattern is actually quite straightforward.

The semiconductor cycle truly ends not usually because of “overvaluation, so it must fall.” It often requires two out of three conditions (or even all three) to appear simultaneously:

\[ \text{Cycle Peak} \approx \text{Supply growth rate catches up to demand growth rate} + \text{Inventory reversal from low levels} + \text{Marginal weakening of end-user demand} \]

To put it more simply:

The release of new capacity begins.
Clients are shifting from frantic purchasing to cautious observation.
Leading companies no longer talk about shortages but start discussing inventory, cost, and pricing pressures.

Died in a demand collapse and inventory in 2001.

Died in the macro crisis of 2008–2009.

Died in the price cycle and trade disruption of 2019.

De-stocking following the depletion caused by the pandemic in 2023.

So, one theory I am most skeptical of right now is that “the semiconductor has risen too much, so it must end soon.” Cycle peaks do not form this way. They require the underlying supply and demand relationship in reality to ease first.

What exactly is the difference between this round and the 2021-2022 round?

During the 2021-2022 cycle, many chips saw price increases, but the underlying logic was rather scattered. Automotive, home appliances, mobile phones, PCs, and servers were almost all engaged in restocking inventory, while the pandemic severely disrupted the supply chain. The conclusion of that cycle was also quite typical: end-user demand declined, channel inventories remained high, and the industry began a collective correction.

The cycle in 2026 will be more concentrated and more dangerous.

The focus is on money being poured into AI data centers. Although there are fewer sources of demand, the intensity at specific points is much greater.

The danger is that these types of needs are not everyday consumer goods in a completely market-driven way, but rather driven by the capital expenditure of several super large companies. Once cloud vendors and platform providers discover:

GPU utilization was not as high as expected;
Inference revenue realization is slower than capital expenditure;
Agentic AI has failed to achieve commercialization;
Or macroeconomic environment, tariffs, energy, and exchange rates are extending the return cycle.

The demand growth rate will drop very quickly.

The industry has not entered this stage yet. On the contrary, what can be seen is that multiple tech giants are continuing to heavily invest in AI infrastructure. Under SIA figures, global semiconductor sales already reached $791.7 billion in 2025 and are projected to approach one trillion US dollars by 2026. Samsung also clearly stated that server memory demand will remain strong in the second half of 2026 as hyperscalers accommodate enterprise AI and LLM services.

This is why I judge that: 2026 looks more like a year of concurrent price increases and expanded production, rather than a peak year.

When will this round most likely end?

I’ll give you a baseline assessment; nothing too mystical/over-the-top. (Alternative translations depending on context: I’ll provide a basic evaluation, keep it simple.)

My Benchmark Scenarios

If there is no sudden global recession, nor a dramatic collapse in AI capital expenditure, this semiconductor uptrend is more likely to enter a danger zone starting in the second half of 2027, and genuine cyclical peak characteristics will be easier to observe in the first half of 2028.

There are four reasons.

First, supply has a physical lag.

Whether it’s Micron scaling up its CAPEX, or Samsung and SK hynix expanding their HBM and associated production lines, establishing cleanrooms, acquiring equipment, optimizing packaging processes, and achieving stable yields all require significant time. While every company knows that money is lucrative right now, announcing expansion today does not mean they can deliver products tomorrow.

Secondly, the demand/requirements for 2027 have already been pre-booked/committed.

Samsung already mentioned multi-year binding contracts, but the conference call content reported by Reuters was even more direct: the potential supply-demand gap in 2027 might be larger than in 2026. This signal is very critical. It means that leading customers are not buying on a quarterly basis, but rather locking resources on an annual or even multi-year basis.

Third, this round is not just about HBM.

If it were only about HBM, one could still explain the situation as “advanced products are strong, but everything else is average.” However, Samsung, SK hynix, and Micron’s current consensus indicates broader tightness across DRAM and NAND, particularly for high-capacity server DRAM modules, AI-oriented eSSDs, and storage related to KV cache. As long as this diffusion continues, the cyclical trend will keep propagating outward.

Fourth, the stock market typically peaks ahead of earnings reports, but it does not peak before the overarching narrative (or storyline).

The current main narrative is “shortages, price hikes, locked orders, and insufficient expansion.” Only when the main narrative shifts to “expansion realized, peak prices, and customers no longer rushing to buy” will the stock price truly seem to have peaked.

Under what circumstances might it end early

There is also a possibility that it could reach its peak sooner.

If two or three of the following events occur simultaneously between Q4 2026 and H1 2027, I would significantly become more cautious:

Cloud vendors are starting to revise down their AI capital expenditure growth rate.
Leading memory manufacturers no longer emphasize shortages, but instead focus on CAPEX, depreciation, and yield ramp-up.
Standard DRAM / NAND prices flatten or even reverse before HBM.
Budgets for mobile phones, PCs, and enterprise IT cannot keep pace with high-priced memory, thereby cooling demand.
Macro-level factors such as recession, tariff escalation, energy shock, or local production disruptions in Korea.

The fifth point, in particular, should not be underestimated. The 2008-2009 cycle already proved that even if the semiconductor industry is strong, it cannot withstand a simultaneous decline in macroeconomic credit and demand.

What to Focus On: Not ‘How Much It Has Risen,’ But These Turning Points

If you care about when the cycle ends, I suggest watching these signals rather than looking at daily stock price colors.

Indicators to Monitor	Suggests if It Remains Strong	Suggests if It Starts Weakening
The description of supply/demand for 2027 in leading company earnings reports	The cycle is still extending	Leading companies start preparing for a slowdown
Capital expenditure and progress of new cleanrooms / packaging facilities	Supply is still catching up to demand	Could plant seeds of oversupply in the next 12-18 months
Pricing and delivery times for DRAM / NAND / eSSD	Tightness is spilling from HBM to broader sub-markets	Supply starts moderating/softening
Scope of Hyperscaler AI CAPEX	The end-market engine is still pressing the accelerator	The most important demand source of the cycle has eased up
Inventory and booking behavior	Customers are still scrambling for capacity	Customers shift from rushing to buy to waiting for goods

My own criterion for judgment is simple: As long as the industry continues to emphasize supply shortages rather than inventory cleanup/recovery, this cycle has not truly ended.

Conclusion

As of May 12, 2026, my conclusion is:

This semiconductor cycle is highly unlikely to end in 2026.
The massive surge in US storage and Korean semiconductors is due to one factor: AI has completely unlocked the profit elasticity of the memory complex.
The end mechanism of past cycles is not mysterious; the core elements are simply falling demand, rising inventory, and supply catching up.
The most probable critical period this time is neither now, but from the second half of 2027 to the first half of 2028.

Of course, this doesn’t mean we should rush into investing blindly now. How can I put it? For an industry like semiconductors, what they truly fear is never the economic “boom” cycle itself, but rather everyone believing that the boom will continue indefinitely, leading them to desperately overexpand capacity during high-profit years.

The cycle usually doesn’t die when it’s worst, but when it’s best, hottest, and most out of stock.

References

Writing Notes

Original Prompt

Recently, semiconductor-related stocks have been soaring, covering U.S. storage/memory and Korean semiconductors. I need to gather relevant information to predict when this current semiconductor cycle will end. Also, how did past semiconductor cycles conclude, and for how long did they last?

Summary of Writing Approaches

First, clarify why this rally is so strong before discussing when it will end; otherwise, the prediction lacks foundation.
The body article intentionally separates firsthand facts from my judgment, avoiding presenting time-sensitive information as permanent conclusions.
The historical section does not try to cover every small fluctuation but only focuses on the few major downturns most significant for investment.
The predictive section avoids making “fortune teller”-style single-point guesses on the peak; instead, it provides a baseline timeframe and trigger conditions for an early top.
This article intentionally did not elaborate on the independent cycles of wafer

Extended Brainstorming

Topic	Include in Main Text?	Reason
Analyzing NVIDIA and TSMC together	Partially Downplayed	They are important, but they will pull the theme from the memory cycle toward the entire AI industry chain, which is easy to lose focus on.
Expansion of domestic Chinese storage and mature processes	No	It affects supply, but this piece’s main thread is US memory/storage and South Korean semiconductors

After AI stocks skyrocketed

Fri, 08 May 2026 19:40:15 +0800

The most unusual aspect of this current AI market cycle is not that Nvidia has risen sharply, but that the increase in value has been transmitted throughout the entire industrial chain: first GPUs, then servers, switches, ASICs, HBM, and finally to NAND, hard drives, power, and data centers.

If it were just a concept, the market trend shouldn’t last this long. But saying that it has already formed a complete profit cycle might be premature.

I prefer to view it as a “bull market driven by certain expenditures”: cloud vendors and model companies are genuinely spending money, and upstream companies are indeed collecting revenue, which is why stocks rose first; however, terminal applications have not yet proven that these investments can reliably generate enough profit, meaning the risk of a bubble also exists.

First, Clarify the Definitions/Scope

This is not investment advice. The stock price data uses an approximate retrospective based on public market quotes and historical closing prices, focusing on stages and logic rather than aiming for decimal precision for every trading day.

I set the starting point to 2022-11-30, which is near the date ChatGPT was released. The endpoint is based on an understanding of public market conditions around the writing time, 2026-05-08.

The gain can be roughly understood as:

\[ 
\text{Increase Rate}=\frac{\text{Period-end Price}-\text{Period-start Price}}{\text{Period-start Price}}
\]

There are two sources of potential error here: First, different websites do not handle adjusted pricing, stock splits, and intraday prices consistently; second, the US stock market on May 8, 2026, has not yet closed, so real-time prices will continue to change. Therefore, the main body focuses more on “multiple levels” and “relative strength,” rather than presenting it as a trading system.

Timeline: Where AI Goes, Stocks Rise

Phase	AI Development Status	Top Gainers	What the Market is Buying
Nov 2022 to Mar 2023	ChatGPT goes mainstream

May 2023 was the most critical turning point in the first phase.

When ChatGPT gained mainstream attention, the market could still question: Is this a chatbot bubble? However, Nvidia’s earnings guidance in May 2023 directly shattered that doubt. The data center revenue and next quarter’s revenue guidance were clearly higher than market expectations, marking the first time the market saw “model capability” translating into actual “GPU orders.”

This is why Nvidia did not begin its rise in 2024, but rather entered the major uptrend phase as early as 2023.

As of 2024, the market trend has shifted from merely “buying GPUs” to requiring the acquisition of entire “AI factories.” Training large models is not just about buying several graphics cards; the true expense lies in the complete cluster: GPU, HBM, networking equipment, servers, liquid cooling, power supply, data center space, and software stack. If any single component is missing, the system cannot function.

Therefore, Super Micro Computer will surge, Broadcom will surge, TSMC is expected to rise, and Oracle will also increase. They are not the same companies, but they all stand on the value chain of AI infrastructure.

After 2025, the market began searching for a second level of certainty: whether models could actually be integrated into enterprise workflows. Palantir’s AIP is representative of this phase; the market isn’t buying merely a software company, but rather the imaginative potential of “AI entering enterprise decision-making and operational systems.”

The speculative potential here is significantly higher. GPU companies are already generating revenue, while enterprise AI software is still proving its capacity for sustained revenue generation.

Approximate Gains of Major Companies

Judging by the trend since the release of ChatGPT, the stocks experiencing the most dramatic gains are not large-cap tech giants like Microsoft, Google, and Amazon, but rather companies with relatively small market capitalizations whose earnings potential has been amplified by AI.

What is most noteworthy in this table is that price appreciation is determined not only by the “degree of AI relevance,” but also jointly by the original market capitalization, profit elasticity, stock holding pattern, and the industry cycle.

Microsoft is certainly important, but it is simply too large. For Microsoft to rise 50%, the required capital and resulting increase in market capitalization are quite exaggerated. A small or medium-cap company, if suddenly perceived by the market as being on the main AI track, will find its stock price much easier to multiply several times.

This is also one of the core rationales behind the massive surge in flash memory.

Why Did Silicon Industry Surge So Much?

Solidigm is not a stock that has risen along with the AI trend since 2022. Its specialty is that, as of 2025, it spun off from Western Data to become a purer NAND and flash memory stock.

It skyrocketed, not just because “AI requires storage.” More accurately, it is due to several overlapping factors:

Factor	Impact
AI data centers require more high-performance storage	Training data, inference cache, vector search, data lakes, logs, and checkpoints all increase storage requirements
The NAND industry itself is undergoing a cyclical reversal	The storage industry has gone through a trough; after supply contraction, price recovery will be significant, leading to high profit elasticity
The listed target company becomes purer after going public independently	After being spun out from Western Digital, the market is easier to price based on the NAND/SSD cycle
Original market capitalization is not high	Compared to giants like Nvidia or Microsoft, less absolute capital is needed to boost the stock price
Short selling or low expectations are easily counterattacked	Once the earnings and guidance of a cyclical stock exceed expectations, valuation recovery can be very aggressive

I agree with the user’s assessment that the company has a low market capitalization and low capital expenditure, but I must add one point: Low market cap only provides potential elasticity, it is not the inherent catalyst for growth itself.

Without fundamental catalysts such as NAND price recovery, AI data center SSD demand, or improvement in financial performance following the company’s independence, a low market cap can only make it easier to speculate on, and also easier to fall back from. A truly significant market rally typically occurs when “low market cap + low expectations + marginal improvement in fundamentals” appear simultaneously.

Micron’s current movement appears to be driven by the powerful AI tailwind hitting the bottom of the storage cycle. The wind itself is enormous, and the market ground is perfectly dry.

Does this market cycle resemble the internet bubble?

Like, but not like.

The key point is that valuation precedes profit realization. The rise in many companies’ stock prices reflects expectations for the next 5 or even 10 years, as the market preemptively prices in companies with the potential to become infrastructure.

The difference is that: upstream companies in this round are already earning real money. Nvidia, TSMC, Broadcom, memory manufacturers, and server manufacturers are not selling PPTs; they are delivering hardware and services.

So, I do not quite agree with summarizing it in one sentence as “all bubbles.” It is more like:

Layer	Current Status	Risk
Compute Hardware	The profit loop is clearest, orders are real	If capex slows down, valuation and inventory will recoil/be hit by it
Cloud Infrastructure	Revenue is real, but depreciation and electricity cost pressure are high	Whether customers are willing to pay long-term for AI computing power
Enterprise Software	Has cases and growth, but ROI has not been widely proven	Many pilots, little scaling; easily shifts from enthusiasm to budget scrutiny
Consumer Applications	Many users, clear monetization divergence	The balance between customer acquisition, retention, inference costs, and subscription pricing may not be achieved

The biggest contradiction currently is that the AI upstream segment has formed a profitable closed loop, but the downstream applications have not.

Nvidia profits from cloud vendors, which spend capital expenditures (CapEx). This CapEx ultimately gets paid for by enterprise customers and consumers. If the end-users do not pay enough, or if AI fails to deliver sufficient cost reduction and efficiency gains to enterprises, someone in the chain will ultimately bear the depreciation and valuation pressure.

The stock price might not crash immediately, but the market will begin to ask a harder question: Where is the return on investment (ROI) for these GPUs?

How Do Research Reports and Institutions View the Risk of a Collapse?

The institutional views I found are not consistent, but they can be grouped into three categories.

The first camp is the cautious one. The title of the Goldman Sachs 2024 report on generative AI was very direct: it suggests that investment is too high and returns are too low. Its core argument is not that AI is useless, but rather questioning whether massive short-term capital expenditure can generate sufficient returns.

The second type is the moderate view. Sequoia has raised the issue of an “AI revenue gap”: to support GPU investments, the entire ecosystem needs to generate very large end-user revenues, but current application layer revenue has not yet caught up with infrastructure investment. This is not a bearish take on AI; it is a reminder that the commercial loop has not been closed yet.

The third category represents the optimists. They believe that AI will follow a path similar to cloud computing: first requiring years of infrastructure investment, followed by a gradual release of software and service revenues. This assessment also has merit; after all, cloud computing itself was questioned for its high costs in its early days.

The problem is that the stock market won’t wait 10 years to price it. It will buy early, and it will also kill early.

My judgment is:

The market may not face a sharp downturn in the short term because capital expenditure remains high, orders are still flowing, and the AI race is ongoing. As long as major companies like Microsoft, Google, Amazon, Meta, and Oracle continue to expand their data centers, the revenue of upstream hardware companies will remain supported.

But it will definitely undergo a rigorous ROI review in the mid-term. The triggers might be:

Cloud vendors slowing down AI Capex;
The large model price war results in insufficient inference revenue coverage for costs;
The failure rate of enterprise AI projects transitioning from pilot phase to production is too high;
High inventory levels in certain upstream links;
Interest rates or the macroeconomic environment make the market reluctant to assign high valuations to long-term narratives.

This does not mean AI technology has failed. After the dot-com bubble burst, the internet did not disappear. What truly disappeared was the portion that had been prematurely overvalued in the valuations.

Which metrics should I focus on

If I continue observing this round of market trends, I won’t just look at model announcement conferences/releases.

Several more useful indicators are:

Metric	Why It Is Important
CAPEX growth of four major cloud providers	Determines the sustainability of upstream hardware orders.
NVIDIA data center revenue and gross margin	Indicates whether compute power demand is genuinely strong or if prices are starting to ease.
HBM / NAND / SSD pricing	Shows whether storage market recovery is occurring, or if it

Especially the last one. AI cannot only look at revenue, but it must also account for depreciation.

Buying GPUs is not free, nor is building a data center. If AI service revenue growth is very impressive, but free cash flow becomes increasingly concerning, the market will eventually reprice it.

Conclusion

The surge in AI stocks this round started with the technological shock brought by ChatGPT, then moved to Nvidia’s orders, then to the entire AI factory, and finally spread to storage, memory, power, and enterprise software.

Cymbet’s surge is not an isolated event. It stands at the intersection of AI storage demand, NAND cycle reversal, independent listing, and low market cap elasticity; therefore, its gains will be more exaggerated than those of many mega-cap companies.

However, the more such market conditions arise, the more it is unwise to only look at “infinite AI demand.” Capital markets prefer to incorporate long-term trends into stock prices all at once, and they are also adept at correcting in reverse when the rate of realization is insufficient.

AI is probably not fake. The problem is that current stock prices have already assumed that AI will soon become a very profitable, very stable, and very large-scale business.

This default value is where problems are most likely to occur later.

References

OpenAI: Introducing ChatGPT, November 30, 2022.
OpenAI: GPT-4 research, March 14, 2023.
NVIDIA: FY2024 Q1 Earnings Report, May 24, 2023.
NVIDIA: FY2025 Q1 Earnings Report, May 22, 2024.
Western Digital: Completes SanDisk Spin-Off, February 24, 2025.
SanDisk: FY2026 Q3 Earnings Report, April 30, 2026.
Goldman Sachs: Gen AI: Too Much Spend, Too Little Benefit?, 2024.
Sequoia Capital: AI’s $600B Question, 2024.
Gartner: 30% of GenAI Projects Will Be Abandoned After Proof of Concept, July 29, 2024.
MIT NANDA: The GenAI Divide: State of AI in Business 2025.

Author’s Notes

Original Prompts

The AI boom has caused many companies' stock prices to soar. We need to organize the stock gains of related companies since ChatGPT was released according to a timeline, marking periods of steepest increases, what state of AI corresponded to those times, and why the corresponding stocks surged. Why did [Shandili/Company Name] surge so much? Besides the influence of AI, there are other factors. [Shandili]'s original market capitalization was not high; lifting it requires capital, and if the market cap is low, less capital needs to be expended. Search through research reports: Will this wave of AI eventually collapse? Currently, everything is burning cash, and there is no complete profit cycle established yet.

Writing Outline Summary

Instead of presenting all AI stocks as a market data list in the main body, it writes according to categories such as “model capability, order fulfillment, infrastructure diffusion, enterprise implementation, and storage cycle.”
The section on Flash Memory deliberately did not only focus on AI demand, but also included NAND cycles, independent listing status, and low market cap elasticity.
Regarding the bubble issue, it was not presented directly as collapsing or stable; instead, it was broken down into upstream profit cycles and downstream ROI cycles.
The article minimized details of many individual companies (e.g., AMD, TSMC, and power stocks), otherwise it would become a mere stack of data/information dumping.
Stock price metrics are limited to the multiples level, used for explanatory purposes, not for trading judgment.