Author: Hu Xuanfeng (Director of Digital Asset Business at Fosun Wealth, CMO of FinChain, Executive Dean of Hong Kong Blockchain Application and Investment Research Institute, Deputy Director of Yangtze River Delta Blockchain Industry Promotion Center)

Risk Warning: This article does not constitute any securities trading advice and does not make definitive judgments on stock prices of any industry or company. The terms "opportunity," "revaluation," "wealth map" mentioned in the text refer to the directions that industry trends and capital markets may focus on. Real investment still requires independent judgment based on valuation, performance, orders, competitive landscape, financial quality, and risk tolerance.

Introduction: Who is the next optical module?

After the optical modules increased tenfold, many felt they had missed the best opportunity in the AI industry chain. Companies like Zhongji Xuchuang, New Ease, Tianfu Communication, and Yuanjie Technology have become the brightest stars in the capital market over the past year. Particularly Yuanjie Technology, whose stock price once exceeded Kweichow Moutai during intraday trading in April 2026, becoming a new representative of high-priced stocks in A-shares. This event itself indicates that AI computing hardware has shifted from a technical theme to a real capital market pricing event.

But what I really want to discuss is what everyone is most concerned about: "Who is the next optical module?"

To find the answer, we cannot look superficially; we must understand a fundamental logic. In every round of industrial revolution, the capital market may assign high valuations to well-narrated stories, but this does not last long. Those companies that manage to break through new bottlenecks often receive long-term capital pricing.

Whoever breaks through the new bottleneck will have new pricing power; whoever owns new pricing power may be re-evaluated. The recent surge in optical module prices is essentially not due to the market suddenly liking optical communication but rather the AI data centers pushing the bottleneck of "high-speed interconnection" to the forefront.

1. AI is a new information infrastructure revolution

Today, many people view AI with a thematic mindset: ChatGPT became popular, so they hype large model valuations; Nvidia rose, so they hype GPUs; optical modules rose, and everyone chased after optical modules; applications haven't yet made large-scale profits, and they claim AI is a bubble.

This view is too short-sighted, making it more likely to pursue what leads to a loss. What should truly be done is an in-depth judgment of why optical modules gained capital recognition and what patterns can be explored.

AI is a new information infrastructure revolution. Like the telegraph, telephone, and mobile internet, it is redefining the production, transmission, processing, and monetization of information.

Recently, I have been writing a new book titled "Token Economy: New Development Paths in the Intelligent Era," and after in-depth research, I found that every generation of information revolution first produces a new commercial pricing unit.

In the telegraph era, the most important unit was "word"; in the telephone era, the most important unit became "minute"; in the mobile internet era, the most crucial unit was "traffic"; and in the AI era, the most essential unit is Token, which is---word unit.

Token is superficially the smallest unit of information processed by AI, but behind it lies the comprehensive costs of computing power, electricity, video memory, storage, network, model architecture, and inference efficiency. When you ask AI a question, it consumes tokens; when a company runs a process using an Agent, it also consumes tokens. In the future, as AI enters customer service, investment research, design, programming, education, healthcare, finance, and manufacturing, there will be real token costs behind every task execution.

Therefore, the long-term wealth distribution of the AI industry chain cannot just consider who can sell GPUs. The truly important aspects are four things: who can produce tokens, who can reduce token costs, who can schedule tokens, and who can turn tokens into results that users are willing to pay for.

2. History is not background, but the rules of industry distribution

To understand the future of AI, one must first understand the history of the information industry over the past century.

The telegraph, telephone, and internet may seem like three different industries on the surface, but underneath, they repeatedly enact the same script: when a new unit first appears, it is expensive, and infrastructure profits first; as unit costs decrease, efficiency takes over; eventually, the access layer redistributes commercial value.

The first act is the telegraph era.

In 1866, the transatlantic submarine cable was put into real use, changing information transmission between Europe and North America from weekly to minute calculations. But at first, telegrams were very expensive—$10 per word, with a minimum of ten words to send. Sending the shortest transoceanic telegram was equivalent to the wages of a skilled worker for ten weeks.

At that time, the ones who first made profits were the people laying submarine cables and controlling the transnational telegram network, as they held the information highways of global finance, trade, shipping, and diplomacy.

However, because one word was too expensive, it would inevitably drive the emergence of compression techniques. Merchants began using telegraphic codes and commercial cipher books. One word could now represent a whole sentence, a transaction directive, or a specification for a type of goods. Today, when we talk about AI model compression, quantization, and speculative decoding, it sounds cutting-edge, but the underlying logic is not new. Humanity has been doing the same thing since the telegraph era: can the same information be sent out using fewer pricing units?

Later on, access emerged. Reuters is a typical example. In 1850, Paul Julius Reuter used 45 pigeons to transmit stock prices and news between Brussels and Aachen, speeding up delivery by about six hours compared to the railroad; after the English Channel cable was laid, he quickly connected to the telegraph network to sell financial information, business news, and market trends to banks, newspapers, and merchants.

The cleverness of Reuters was not in laying cables but in knowing what information was important and who was willing to pay for faster information. Telegraph companies earned revenue from character transmission, while Reuters made money from information distribution rights. One profited from access, while the other profited from the value of information.

The second act is the telephone era.

In the telephone era, the pricing unit became "minute." In 1915, during the early commercialization of transcontinental telephone service in the United States, a three-minute call from New York to San Francisco cost approximately $20.70, equivalent to several hundred dollars today. The first wave winner was AT&T. The telephone network had strong physical monopoly attributes; the connections, switches, relay stations, and end-users formed a massive network.

Later, automatic switches, signal amplifiers, and upgrades to communication equipment caused the cost of telephone minutes to continually decrease. The capital market began to reevaluate companies that improved system efficiency. Afterward, the Yellow Pages emerged on the telephone network. The Yellow Pages did not charge for calls; instead, it charged businesses for visibility. When users wanted to find businesses, and businesses needed to be found by users, commercial access was formed.

The third act is the mobile phone and internet era.

In early wireless communication, infrastructure builders were the most valuable, but later on, telecom operators gained prominence by controlling cell numbers, networks, packages, and pricing, earning high profits as SMS, voice, and data traffic all charged by usage. The internet, based on wired and wireless network infrastructure, greatly lowered communication costs and improved execution efficiency, ushering in the era of traffic. With the rapid decline in unit traffic costs, infrastructure was no longer overvalued, while companies managing user access became increasingly valuable, resulting in the rise of internet platform giants. WeChat, Taobao, Meituan, Douyin, Xiaohongshu, and Pinduoduo seized control of users' time, transactions, and consumption decisions.

Operators control bytes, while internet platforms control the commercial intent behind those bytes. This is the rule observed across three generations of the information industry: infrastructure rises first, the efficiency layer takes over, and the access layer ultimately redistributes higher value. AI is currently at a critical transition from the first stage to the second and third stages.

3. Why the first wave has fallen on GPUs, HBM, and optical modules

In the past two years, the first wave of AI surges started with Nvidia, storage, and optical modules, which is not surprising. This is because the first phase of AI is focused on large model training and the construction of computing power clusters.

Training large models requires a large amount of GPUs; GPUs require high bandwidth memory, which is HBM; and many GPUs need to work in coordination, which necessitates high-speed interconnection, which includes optical modules, switching chips, PCBs, connectors, and network devices. Traditional data centers resemble a group of servers handling many ordinary tasks, while AI data centers more closely resemble a gigantic supercomputer. Thousands or even hundreds of thousands of GPUs must operate as a unified whole, and any slow link in the chain can drag down the entire system.

GPUs are expensive. If the network isn’t fast enough, the GPUs will be idle, waiting for data. If GPUs wait for data, they become costly assets that are unproductive. Therefore, the rise of optical modules has industrial foundations; the rise of HBM represents the genuine bottleneck within the industry chain sought after by capital markets.

However, the market won’t always focus solely on the first batch of bottlenecks. Once these visible links of GPUs, HBM, and optical modules are fully discussed, the questions will continue to push further: once computing power is established, how can it run stably? How can it run cheaply? How can it integrate into corporate processes? How can it become results that users are willing to pay for?

4. The next bottleneck in AI development: electricity, liquid cooling, and computing power industrial real estate

The next most certain line, I believe, is electricity and liquid cooling. The reason is simple: AI data centers are transitioning from a "machine room business" to an "energy business."

In the past, people understood data centers as buildings filled with many servers. AI data centers are not like that. The core constraints of AI data centers are changing to power access, cabinet power density, cooling capacity, energy scheduling, and infrastructure delivery. When NVIDIA officially introduced the GB200 NVL72, it emphasized that it connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design.

This means that the competition in AI is no longer just about single GPUs but about the entire rack, room, and data center systems. In the future, cabinet power density will continue to push toward tens of kilowatts or even hundreds of kilowatts, and liquid cooling and power supply and distribution will no longer be backend support but prerequisites for computing power deployment.

More importantly, it's about electricity. The International Energy Agency predicts in “Energy and AI” that global electricity consumption by data centers will nearly double by 2030, reaching about 945 TWh, accounting for close to but less than 3% of total global electricity consumption; AI is one of the most important drivers of this growth.

GPUs can be ordered, optical modules can be produced in bulk, and servers can be assembled, but electricity grids, substations, transmission lines, backup power sources, and cooling systems cannot be produced out of thin air in just a few months. The stronger AI becomes, the higher the power consumption; the denser the computing power, the greater the heat generated; the more centralized the data centers, the more extreme the demands on electricity and heat dissipation will be.

Therefore, transformers, UPS, distribution cabinets, switching power supplies, busbars, data center power systems, liquid cooling plates, CDU, pumps, heat exchangers, rack liquid cooling solutions, and comprehensive data center infrastructure packages will all be re-evaluated. They have historically been categorized as traditional manufacturing industries, but with AI's arrival, they have become prerequisites for computing power delivery.

Going further, AI data centers will shift from traditional IDCs to a new type of industrial real estate. Traditional IDCs focus on cabinet numbers, rack rates, PUE, rents, and customers; AI data centers will focus on electrical performance, substations, long-term energy contracts, liquid cooling capacity, high-speed network access, long-term contracts with major clients, GPU cluster operational capabilities, and land expansion space.

This is no longer just a simple "build a building and put servers in it" business. It resembles stations in the railroad era, docks in the port era, airports in the aviation era, and hubs in the highway era. The best AI data center companies will not only lease out their rooms but will also be able to organize land, electricity, cooling, networks, chips, and customer long-term contracts into a set of infrastructure assets with cash flow, barriers to entry, and scarcity.

This line also has a subsequent change: the financialization of data center assets. Once AI data centers form stable cash flows, they could be packaged into REITs, RWAs, infrastructure funds, revenue rights products, and long-term leasing assets. In the past, during the cloud computing era, data centers were backend assets for cloud vendors; in the AI era, data centers will be repriced as "computing power industrial real estate."

5. After training, it’s the war over inference costs

Now many people believe Nvidia's strength means the opportunities for AI chips have already been exhausted by Nvidia. This assessment is only half correct.

In the large model training phase, Nvidia's advantages are very strong. It's not just that their GPUs are powerful; CUDA, the developer ecosystem, network systems, complete machine solutions, and software toolchains are also robust. But once AI enters the large-scale inference stage, the logic will change. In the training phase, the most important aspect is creating the model; in the inference phase, the most critical factor is allowing the model to serve vast numbers of users daily. Training resembles capital expenditures, while inference is more akin to operating costs.

As AI enters customer service, business operations, programming, finance, education, healthcare, and manufacturing, a massive number of calls will be generated daily. At this point, it becomes evident that the logic of the token economy differs from that of the traffic economy; in the traffic economy, the marginal cost decreases, so user acquisition can happen on a large scale before considering revenue because the network cost of adding a user diminishes. But the logic of the token economy makes companies that operate large models and cloud vendors experience entirely different economic costs: their marginal cost remains unchanged or even rises. Training involves a one-time investment for long-term benefits, whereas inference does not. If servicing each user using AI incurs losses, then if a task is called several million or even hundreds of millions of times a day, no company can withstand this. That’s why even ByteDance’s large model Doubao had to initiate charging.

This is when new opportunities arise as everyone begins to consider the issue of cost reduction. Why must all tasks use the most expensive general-purpose GPUs? Can dedicated chips be used instead? Can we utilize lower power consumption, higher throughput, and ASICs more suitable for fixed scenarios?

This is why cases like Broadcom, AMD, and Google TPU are worth paying attention to.

Reuters reported that Broadcom expects that by 2027, revenue opportunities from custom AI chips could exceed $100 billion, driven by the rapid rise in demand for custom AI chips from major tech companies. AMD disclosed in its 2024 annual report that its revenue from data center AI has surpassed $5 billion annually, with clients like Meta, Microsoft, and Oracle deploying the AMD Instinct MI300 accelerators. Google Cloud also emphasizes that TPU v5e focuses on cost efficiency, providing greater query volumes at the same cost.

Thus, in the future, AI chips won’t be limited to a single form. Nvidia will continue to be strong, but in addition, cloud vendors will develop their own chips, custom ASICs, inference acceleration chips, and edge AI chips, each finding their place. This isn’t just a simple replacement of Nvidia but rather sharing a portion of the profit pool during the inference era. As AI moves from the training to the inference era, cost optimization will become a new pricing power.

6. After optical modules, the entire AI network

Many people believe that the optical modules have risen as far as they can, and thus, the AI trend is over. I don’t see it that way. Optical modules are merely the first layer of visible components in the AI network. Behind them lie switching chips, switches, DPUs, SmartNICs, CPOs, silicon photonics, cluster scheduling, and network operating systems.

The essence of AI data centers is to connect numerous GPUs into a supercomputer. The most expensive asset involved is the GPUs, and the most unacceptable situation is the idle GPUs. If network latency is high, the GPUs wait for data; if switching efficiency is low, the GPUs wait for data; if the communication architecture is poor, the GPUs still wait for data.

Thus, the value of the AI network it constitutes is not merely in transmitting data but in improving the overall utilization rate of the GPU cluster. In typical internet data centers, a slight delay in network speed merely results in slow loading for users; however, in AI data centers, a network that is a bit slow can lead to a decline in equipment utilization worth several hundred million or even billions of dollars.

NVIDIA's Quantum-X800 InfiniBand platform focuses on an end-to-end 800 Gb/s network aimed at serving trillion-parameter-level AI models; Spectrum-X Ethernet emphasizes improving AI network performance and supporting the expansion of large-scale GPU clusters. TrendForce also points out that the demand for optical transceiver modules of 800G and higher in AI server clusters is rapidly increasing, and the market for AI optical transceiver modules is expected to continue to expand.

So the future of the AI network will continue to upgrade: from 400G to 800G, and then to 1.6T; from traditional optical modules to CPOs; from electronic switching to optoelectronic integration; from regular networks to AI fabrics; from single-point devices to overall cluster scheduling. The capital market will not only look at the optical module business but will also consider who can enhance the connection efficiency of AI clusters, who can reduce GPU wait times, and who can stabilize clusters of thousands or hundreds of thousands of cards.

7. After tokens become less expensive, the entry point will change hands

The true large-scale application in the AI era depends on whether the cost of tokens can continue to decrease. The more expensive the tokens, the harder it is for AI to penetrate; the cheaper the tokens become, the easier it is for AI to enter corporate processes and daily life.

The Stanford 2025 AI Index report shows that the querying cost of models reaching GPT-3.5 level has dropped from about $20 per million tokens in November 2022 to about $0.07 by October 2024, a reduction of over 280 times in approximately 18 months; under different tasks, the price drop rates of LLM inference vary significantly, ranging from nine times to 900 times annually.

This data indicates that the true long-term deflationary force within the AI industry has started to appear. Whoever can ensure that the same task consumes fewer tokens, less video memory, less electricity, and less inference time holds value.

I refer to these types of companies as the token compression faction.

They can be model companies, inference platforms, chip companies, cloud vendors, or enterprise AI infrastructure companies. The key is not what they are called but whether they can execute the same task at a lower cost, shorten the inference chain, reduce ineffective calls, and produce more stable results.

A few important technologies in this include MoE, quantization, distillation, caching, speculative decoding, and model routing. Especially model routing, not every task needs to utilize the most expensive model. A mature AI system will automatically select the most suitable model and path according to task difficulty, cost budget, speed requirements, privacy requirements, and accuracy demands. Of course, model routing is also susceptible to disruption by major manufacturers, and its moat is not firm.

Once costs come down, the question of entry becomes more significant. Many believe that the entry point in the AI era will be a model scheduling platform, akin to a Meituan for the AI era. This analogy carries weight but lacks depth. The true AI entry point may not be a platform for model selection but rather a system embedded within workflow.

Ordinary users are unlikely to open a model scheduling platform every day. Corporate users won’t call models just for the sake of calling models. Users want tasks completed, companies seek process efficiency, and employees want working results. Ultimately, AI will be embedded in Office, Feishu, DingTalk, WeChat for Work, ERP, CRM, code editors, browsers, email, searches, finance systems, customer service systems, and trading systems. Whoever controls the workflow controls the power to invoke AI.

Microsoft’s 2025 annual report reveals that the Copilot product family has over 100 million monthly active users combined from business and consumer sides, further integrating Microsoft 365 Copilot into office processes. This indicates that the entry point for AI is not necessarily a standalone app but might be an intelligent layer within an existing workflow.

Programmer entry may be through code editors and code hosting platforms; office entry may be through Microsoft 365, Google Workspace, Feishu, or DingTalk; enterprise management entry may come from ERP, CRM, or finance systems; personal entry may be through mobile operating systems, browsers, search boxes, or smart glasses. The true entry into the AI era lies not in a model list but in workflow entry.

8. The real difficulty for enterprise AI is entering processes

For AI to become the entry point to workflows, another premise is that it must enter the real processes of enterprises. The hardest part of enterprise AI is not just connecting a chatbot; it’s about whether the model can securely read enterprise data, understand business processes, invoke systems, leave logs, accept audits, and integrate with human approval mechanisms.

Many enterprises today are using AI but still remain in the stage where employees ask a question, write something, or summarize. This can improve individual efficiency but does not genuinely alter organizational structure. True enterprise AI is when Agents enter the processes.

A customer service Agent doesn’t merely answer questions; it must check orders, track logistics, assess refund conditions, and invoke after-sales systems; a finance Agent doesn’t just write reports; it must read vouchers, reconcile accounts, identify anomalies, and generate approval opinions; an investment research Agent doesn’t only summarize news; it must extract data, construct models, compare companies, and track risks; a legal Agent doesn’t merely draft contracts; it must search clauses, identify risks, relate cases, and maintain modification records.

This requires an entire set of infrastructures: databases, vector retrieval, permission management, data governance, system integration, workflow engines, audit logs, security compliance, enterprise knowledge bases, and Agent orchestration platforms. These aspects may not sound as glamorous as large models, but they are the foundations for AI's real implementation in enterprises. The first investment that a company makes in truly deploying AI should be buying security, buying data, buying permissions, buying processes, buying integration, and buying compliance, not deploying a lobster and purchasing some tokens to execute flashy tasks.

Here lies an even greater change: the real big money from AI applications might not come from software budgets but from labor budgets. SaaS sells tools, while AI Agents deliver results. Tools require people to operate, whereas Agents directly complete tasks.

An AI customer service system, if it only sells software, has a ceiling in the customer service software market; if it effectively replaces a large number of human customer service representatives, the ceiling becomes the customer service outsourcing and corporate labor cost for customer service. An AI legal system, if it only sells document tools, has limited ceiling; but if it can replace junior lawyers for contract review and adjustment, the ceiling becomes the legal services cost pool.

Harvey is an interesting case to observe in legal AI. TIME 2025 reported that Harvey is valued at around $5 billion, has over 300 clients, operates in 53 countries, and reaches seven of the top ten law firms in the U.S. by revenue. It indicates that AI applications in high-value knowledge work scenarios are not merely simple tool replacements, but they are penetrating the labor cost pools of professional services.

The most successful AI application companies in the future will not just define themselves as software companies; they will quantify how much work they can complete for clients, how much labor they can save, how many mistakes they can reduce, how much conversion they can increase, and how much they can shorten delivery timelines. In the past, capital markets looked at ARR, but in the future, they will also assess how much of the labor cost pool they can capture.

9. Don’t overlook local AI and the financialization of computing power

There are also two lines which may not be the hottest right now but cannot be neglected in the medium to long term. One is local AI. Currently, most tokens are still produced in cloud data centers; when you ask a model, fundamentally, it is the distant data center that helps calculate for you. However, in the future, not all AI inference will necessarily take place in the cloud.

The reason is straightforward: cloud-based inference is too costly, and many scenarios require low latency, while a lot of data cannot be uploaded to the cloud, and terminal devices will become increasingly intelligent. Therefore, in the future, a portion of tokens will migrate from the cloud to local, or what we can refer to as edge computing. Smartphones will run AI, PCs will run AI, cars will run AI, robots will run AI, and smart glasses will run AI; local workstations will also run AI.

Once edge AI takes off, it will bring about a new cycle of hardware. Edge AI chips, NPUs, low-power storage, power management, heat dissipation, sensors, camera modules, microphone arrays, AI PCs, AI smartphones, AI glasses, robots, and in-car intelligent computing platforms will all enter a new phase of supply chain revaluation.

However, this line needs to be viewed objectively. The edge AI direction is correct, but currently, there isn’t a truly killer application. AI PCs and AI smartphones are still primarily driven by hardware manufacturers, and the user side hasn't formed an indispensable demand yet. Thus, edge AI will not become the first main line to explode, but it will be an important main line in the medium to long term.

The other line is the financialization of computing power. AI infrastructure is heavy. GPUs are expensive, data centers are costly, electricity contracts are high, construction cycles are long, and funds are heavily tied up. Relying solely on tech companies' balance sheets to bear this weight may not be the best solution.

In the future, several types of new financial assets might emerge: GPU leasing contracts, computing power revenue rights, data center REITs, AI infrastructure funds, long-term purchase agreements for electricity, GPU collateral financing, structured financing based on inference revenues, as well as RWA-oriented computing power assets.

The digital asset business of Fosun Wealth, where the author works, is one of the most professional RWA issuance teams in Hong Kong. From frontline business analysis, computing power assets RWA has extremely high financial asset value and potential for future compliant global transactions. FinChain’s Star Link and Star Path are helping large traditional computing power manufacturers achieve tokenization through compliant financial new pathways.

There are already successful cases of the financialization of computing power overseas, with CoreWeave being one of the most typical. In March 2026, CoreWeave announced it had completed an $8.5 billion delayed draw term loan facility, calling it the first investment-grade GPU-backed financing. This indicates that GPUs, cabinets, and computing contracts are being re-evaluated by the financial market as collateralizable and financeable infrastructure assets.

This is quite similar to the railroad, telecommunications, and cloud computing sectors. During the railway era, railway companies financed their construction through bonds; during the telecommunications era, operators used long-term capital to lay down networks; in the cloud computing era, cloud vendors constructed data centers with massive capital expenditures. In the AI era, GPU, cabinets, electricity contracts, and future inference revenues will also be repackaged, priced, and circulated by financial markets.

10. The highest level opportunity: AI-native companies will rewrite profit statements

What has been discussed previously pertains to the industry chain. However, the most significant long-term impact of AI is not just on the industry chain, but rather on how organizational structures will be rewritten.

In the past, companies were structured around human-managed departments. Sales department, customer service department, finance department, legal department, investment research department, operations department—each had positions, workflows, approvals, and performance reviews. After AI Agents enter, the organization will change: one person can manage multiple Agents, one department can be compressed through Agent workflows, back-office jobs will be automated, management radius will expand, and companies will shift from labor-intensive organizations to human-machine collaborative organizations.

This means that in the future, capital markets will re-evaluate a new type of company: AI-native companies. These are not simply companies buying a few AI tools nor just having employees use ChatGPT to draft copy, but rather companies designed from the ground up with AI in mind, leading to fewer employees, higher revenues, increased output per person, lower marginal costs, and faster delivery times.

Therefore, the most significant capital market impact from AI is not just about "who in the AI industry chain rises," but also "who in all industries can use AI to rewrite their profit statements." In the future, the market will reward two types of companies: one type sells AI infrastructure and AI capabilities; the other type restructures its cost and revenue structures using AI. The latter may not outwardly appear to be an AI company, but their organizational efficiency, profit margins, and output per person will undergo fundamental changes.

Conclusion: AI is redefining scarcity

At this point, if we only see GPUs, optical modules, electricity, liquid cooling, ASICs, data centers, and edge devices, we are still viewing AI as merely a technological industry chain. The more profound change is that AI will redefine what is deemed scarce.

In the past, GPUs were scarce, hence Nvidia’s rise; later HBM and optical modules became scarce, prompting rises in storage and optical modules. Next, what will become scarce are electricity, liquid cooling, AI networks, inference chips, data pipelines, workflow entry, enterprise data, and organizational execution capabilities.

If we break down this round of AI boom, the first phase bought computing power construction, the second phase focused on whether computing power could run stably and cheaply, while the third phase will emphasize whether computing power can enter corporate processes and translate into real income and profit.

The tenfold increase in the price of optical modules does not signal the end of the story; it marks the first clear visibility of the physical bottlenecks in AI infrastructure to the capital market. Greater revaluation is likely to happen in the next batch of new bottlenecks, which have yet to be fully priced.

Electricity, liquid cooling, AI data centers, custom ASICs, AI networks, token compression, model routing, enterprise data pipelines, workflow entry, edge AI, financialization of computing power, and AI-native companies—all these directions will together form the next wealth map of the AI industry chain.

Of course, this does not mean that every company will rise, nor that every concept is worth buying. Wealth in every round of industrial revolution is not evenly distributed. Those companies that will be rewarded by the capital market in the long term will definitely be those that break through bottlenecks, have customers, have orders, possess technological barriers, cost advantages, and ecological positions.

To sum it up: the first wave of opportunities in AI lies in who can build computing power; the next wave of opportunities is in who can support computing power, optimize computing power, schedule computing power, and ultimately turn computing power into real commercial outcomes.

Notes and References

The following materials support the historical facts, public data, and industrial cases involved in the article. To facilitate financial media editing and review, priority is given to official agencies, company announcements, authoritative media, or primary data.

[1] Regarding Yuanjie Technology's intraday stock price exceeding Kweichow Moutai and becoming a new high-priced stock representative in A-shares: Sina Finance, 2026-04-17, "Beyond Moutai, a new king of A-shares is born." https://finance.sina.com.cn/wm/2026-04-17/doc-inhuupte2305062.shtml

[2] About the transatlantic telegraph fees in 1866: PBS American Experience, "How the Early Cable Was Used," explaining that the initial charge for the transatlantic telegram in 1866 was $10 per word, with a minimum of 10 words, roughly equating to the wages of a skilled worker for ten weeks. https://www.pbs.org/wgbh/americanexperience/features/cable-how-early-cable-was-used/

[3] About the Reuters pigeon case: Reuters, "The long history of speed at Reuters," mentioning Reuter's early use of pigeons to relay financial information. https://www.reuters.com/article/business/the-long-history-of-speed-at-reuters-idUSKBN2761WD/

[4] Regarding the pricing of transcontinental telephone calls in the U.S. in 1915: JSTOR Daily, "AT&T: Birth of the First Social Network," mentioning that a three-minute coast-to-coast call cost $20.70 in 1915. https://daily.jstor.org/birth-first-social-network/

[5] About NVIDIA GB200 NVL72: NVIDIA official page, explaining that GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design. https://www.nvidia.com/en-us/data-center/gb200-nvl72/

[6] About global data center electricity consumption: International Energy Agency, "Energy demand from AI," predicting global data center electricity consumption will be about 945 TWh by 2030, accounting for approximately less than 3% of total global electricity consumption. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai

[7] About Broadcom's custom AI chips: Reuters, 2026-03-04, "Broadcom forecasts second-quarter revenue above estimates," reporting Broadcom expects AI chip revenues to exceed $100 billion by 2027. https://www.reuters.com/technology/broadcom-forecasts-second-quarter-revenue-above-estimates-2026-03-04/

[8] About AMD’s data center AI business: AMD 2024 Annual Report, disclosing that its data center AI business annual revenue has exceeded $5 billion and mentioning clients like Meta, Microsoft, and Oracle deploying AMD Instinct MI300 accelerators. https://ir.amd.com/financial-information/sec-filings/content/0001193125-25-067185/0001193125-25-067185.pdf

[9] About the cost efficiency of Google TPU v5e: Google Cloud Blog, "Performance per dollar of GPUs and TPUs for AI inference," mentioning TPU v5e improving query volume at the same cost. https://cloud.google.com/blog/products/compute/performance-per-dollar-of-gpus-and-tpus-for-ai-inference

[10] About NVIDIA AI network platform: NVIDIA Quantum-X800 official page states it is an end-to-end 800 Gb/s InfiniBand network. https://www.nvidia.com/en-us/networking/products/infiniband/quantum-x800/

[11] About the AI optical transceiver module market: TrendForce, 2026-04-20, "Global AI Optical Transceiver Market to Reach US$26 Billion," indicating the rapid rising demand for optical transceiver modules of 800G and above in AI server clusters. https://www.trendforce.com/presscenter/news/20260420-13017.html

[12] About the drop in AI inference costs: Stanford HAI, "AI Index 2025: State of AI in 10 Charts," stating the querying cost for models reaching GPT-3.5 level has dropped by over 280 times in about 18 months. https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts

[13] About Microsoft Copilot user scale: Microsoft Annual Report 2025, disclosing that the Copilot product family has over 100 million monthly active users across business and consumer segments. https://www.microsoft.com/investor/reports/ar25/index.html

[14] About the Harvey legal AI case: TIME 2025 global influential companies report states Harvey is valued at about $5 billion, serving over 300 customers, covering 53 countries. https://time.com/collections/time100-companies-2025/7289586/harvey/

[15] About CoreWeave GPU-backed financing: CoreWeave investor relations announcement, in March 2026 announcing it completed an $8.5 billion delayed draw term loan facility, calling it the first investment-grade GPU-backed financing. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx

Note: This article is a long-form industrial viewpoint, and the notes are used to explain the source of facts and do not constitute any investment advice.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。