Author: Hu Xuanfeng, Director of Digital Assets at Fosun Wealth, CMO of FinChain, Executive Director of the Hong Kong Institute for Blockchain Applications and Investment, and Deputy Director of the Yangtze River Delta Blockchain Industry Promotion Center

Risk Disclaimer: This article does not constitute any recommendation to buy or sell securities, nor does it make definitive judgments regarding the stock prices of any industry or company. The terms “opportunities,” “revaluation,” and “wealth map” used herein refer to potential directions that industry trends and capital markets may focus on. Actual investment decisions should be made independently, taking into account valuation, performance, orders, competitive landscape, financial quality, and risk tolerance.

Introduction: Who will be the next optical module?

After optical modules surged tenfold, many felt they had missed the best opportunity in the AI industry chain. Companies such as Zhongji Xuchuang, Xinyisheng, Tianfu Communications, and Yuanjie Technology became the most prominent theme in capital markets over the past year. In particular, Yuanjie Technology’s stock price briefly surpassed that of Kweichow Moutai in intraday trading in April 2026, becoming a new representative of high-priced stocks on the A-share market. This alone demonstrates that AI computing hardware has evolved from a technological theme into a tangible event driving market valuation.

But what I really want to discuss is the question everyone is most concerned about: “Who will be the next optical module?”

To understand this answer, we must look beyond the surface and grasp the underlying logic: during each industrial revolution, capital markets may assign high valuations to companies with compelling narratives, but such valuations are not sustainable. What truly earns long-term capital valuation are companies that overcome new bottlenecks.

Whoever controls the new bottleneck gains new pricing power; whoever has new pricing power is the one who can be revalued. The current surge in the optical module market isn’t due to the market suddenly favoring optical communications—it’s because AI data centers have brought the “high-speed interconnection” bottleneck to the forefront.

I. AI is a new revolution in information infrastructure

Today, many people view AI through the lens of speculative themes: because ChatGPT became popular, they inflate valuations of large models; because NVIDIA rose, they chase GPU stocks; because optical modules surged, everyone rushes into optical modules; and since applications haven’t yet generated substantial profits, they claim AI is a bubble.

This perspective is too short-sighted and makes it easier to chase trends and end up trapped. What truly matters is to conduct a deeper analysis: why are optical modules gaining capital recognition, and what patterns can be identified?

AI is a new revolution in information infrastructure. Like the telegraph, telephone, and mobile internet, it is redefining how information is produced, transmitted, processed, and monetized.

I recently wrote a new book titled "Token Economy: A New Development Path in the Intelligent Age," and after in-depth research, I discovered that each generation of information revolution has first produced a new commercial unit of account.

In the telegraph era, the most important unit was the "word"; in the telephone era, it was the "minute"; in the mobile internet era, it was "data"; in the AI era, the most important unit is the Token—also known as a token.

Tokens are表面上 the smallest units of information processed by AI, but behind them lie the combined costs of computing power, electricity, GPU memory, storage, network resources, model architecture, and inference efficiency. Asking AI a question consumes tokens; having an agent execute a business process also consumes tokens. As AI enters customer service, investment research, design, programming, education, healthcare, finance, and manufacturing, every task execution will carry a real token cost.

Therefore, the long-term wealth distribution in the AI industry cannot be determined solely by who sells GPUs. What truly matters are four things: who can produce tokens, who can reduce token costs, who can orchestrate tokens, and who can turn tokens into results users are willing to pay for.

Two: History is not background, but the rule of industry allocation

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

Telegram, telephones, and the internet appear to be three distinct industries, but beneath the surface, they repeatedly play out the same script: when a new unit first emerges, it’s expensive, and infrastructure profits first; as unit costs decline, the efficiency layer takes over; finally, the access layer redistributes commercial value.

Act one: the telegraph era.

In 1866, the transatlantic submarine cable was officially put into service, reducing the time it took to transmit information between Europe and North America from weeks to minutes. However, telegrams were initially very expensive—ten dollars per word, with a minimum of ten words. Sending the shortest possible transatlantic telegram equated to about ten weeks' wages for a skilled worker at the time. [2]

At that time, the first to make money were those who laid undersea cables and controlled the international telegraph network, because they held the global information superhighway for finance, trade, shipping, and diplomacy.

But when a word is too expensive, it will inevitably drive the development of compression technologies. Merchants began using telegraph codes and business cipher books. A single word no longer meant just a word—it could represent an entire sentence, a trading instruction, or a product specification. Today, we talk about AI model compression, quantization, and speculative decoding, which sound highly advanced, but the underlying logic is not new. Humans have been doing the same thing since the telegraph era: Can the same information be transmitted using fewer units of cost?

Later, the entrance appeared. Reuters is a classic example. In 1850, Paul Julius Reuter used 45 pigeons to transmit stock prices and news between Brussels and Aachen, arriving about six hours faster than the railway; after the English Channel cable was laid, he quickly connected to the telegraph network, selling financial information, business news, and market data to banks, newspapers, and merchants. [3]

What made Reuters remarkable wasn't that it laid cables, but that it knew which information mattered and who was willing to pay for faster access. Telegraph companies earned money by charging per character transmitted; Reuters earned money by controlling the distribution of information. One profited from the channel, the other from the value of the information.

Act two is the era of telephones.

In the era of telephones, the unit of pricing became the minute. In the early commercialization of transcontinental telephone service in the United States in 1915, the first three minutes of a call from New York to San Francisco cost approximately $20.70, equivalent to hundreds of dollars today. [4] The first major winner was naturally AT&T. Telephone networks possess strong physical monopolistic characteristics, with lines, switches, relay stations, and end users together forming a vast network.

However, later on, automatic switches, signal amplifiers, and telecommunications equipment upgrades continuously reduced the cost per phone minute. Capital markets began to revalue companies that improved system efficiency. Eventually, the Yellow Pages emerged on the telephone network. The Yellow Pages did not charge for calls; instead, it charged businesses for advertising space. When users sought businesses and businesses sought to be found by users, a commercial gateway was formed.

Act three is the era of mobile phones and the internet.

In early wireless communications, infrastructure providers were the most valuable. Later, telecom operators gained prominence by controlling phone numbers, networks, plans, and billing, charging for SMS, voice calls, and data based on usage. The internet, built on wired and wireless infrastructure, dramatically reduced communication costs and improved efficiency, ushering in the era of data traffic. As the unit cost of data plummeted, infrastructure no longer commanded high valuations; instead, companies controlling user access points became increasingly valuable, giving rise to internet platform giants. WeChat, Taobao, Meituan, Douyin, Xiaohongshu, and Pinduoduo now control users' time, transactions, and purchasing decisions.

Operators control bytes; internet platforms control the commercial intent within those bytes. This is the pattern consistently demonstrated across three generations of the information industry: infrastructure rises first, efficiency layers follow, and access layers ultimately capture the highest value. AI is now at a critical juncture, transitioning from the first stage to the second and third stages.

III. Why Did the First Wave Hit GPU, HBM, and Optical Modules?

It’s no surprise that over the past two years, the first wave of AI-driven market movement lifted NVIDIA, storage, and optical modules first, as the initial phase of AI involves large model training and compute cluster construction.

Training large models requires a large number of GPUs; GPUs need high-bandwidth memory, known as HBM; and to work together efficiently, numerous GPUs require high-speed interconnects, including optical modules, switching chips, PCBs, connectors, and networking equipment. Traditional data centers are like clusters of servers handling many general-purpose tasks, whereas AI data centers function more like a single massive supercomputer. Tens of thousands or even hundreds of thousands of GPUs must operate as a unified whole—any bottleneck in the system will slow down the entire operation.

GPUs are expensive; if the network isn't fast enough, the GPU sits idle waiting for data. An idle GPU means expensive assets are underutilized. That’s why optical modules are rising—they have a solid industrial foundation; HBM prices are rising because capital markets are buying into the real bottlenecks within the supply chain.

But the market won’t focus solely on the first bottlenecks forever. Once visible components like GPUs, HBM, and optical modules have been thoroughly discussed, the questions will shift further downstream: Once computing power is built, how do we ensure stable operation? How do we make it cost-effective? How do we integrate it into enterprise workflows? And how do we turn it into results users are willing to pay for?

IV. The Next Bottleneck in AI Development: Electricity, Liquid Cooling, and Compute Infrastructure Real Estate

The most certain trend ahead, in my view, is power and liquid cooling. The reason is simple: AI data centers are shifting from a “data center business” to an “energy business.”

Previously, people understood data centers as buildings filled with many servers. AI data centers are different. The core constraints for AI data centers are now power access, rack power density, cooling capacity, energy scheduling, and infrastructure delivery. When introducing the GB200 NVL72, NVIDIA emphasized that it connects 36 Grace CPUs and 72 Blackwell GPUs within a rack-scale, liquid-cooled design—that is, an entire rack-level liquid cooling system. [5]

This means that AI competition is no longer just about individual GPUs, but about entire racks, server rooms, and data centers as integrated systems. In the future, rack power density will continue to advance to tens of kilowatts and even hundreds of kilowatts, making liquid cooling and power distribution not just backend support, but prerequisites for compute deployment.

More importantly, it's electricity. According to the International Energy Agency's report "Energy and AI," global data center electricity consumption is projected to nearly double by 2030, reaching approximately 945 TWh, accounting for nearly but less than 3% of global electricity demand; AI is one of the most significant drivers of this growth. [6]

GPUs can be ordered, optical modules can be scaled up, and servers can be assembled—but the power grid, substations, transmission lines, backup power systems, and cooling infrastructure cannot be created out of thin air in a few months. The stronger the AI, the higher the power consumption; the denser the computing power, the greater the heat generation; the more centralized the data centers, the more extreme the demands on electricity and cooling.

Therefore, transformers, UPS systems, distribution cabinets, switch-mode power supplies, busways, data center power systems, liquid-cooled plates, CDUs, pumps and valves, heat exchangers, integrated cabinet liquid cooling solutions, and comprehensive data center infrastructure EPC services will all be revalued. Previously, they were all categorized as part of traditional manufacturing industries, but with the advent of AI, they have become prerequisites for computing power delivery.

Taking this one step further, AI data centers will evolve from traditional IDCs into a new form of industrial real estate. Traditional IDCs are evaluated by the number of racks, utilization rate, PUE, rent, and customers; AI data centers are assessed by power metrics, substations, long-term energy contracts, liquid cooling capabilities, high-speed network connectivity, long-term agreements with major clients, GPU cluster operational expertise, and land expansion potential.

This is no longer simply a business of “building towers and placing servers.” It is more like the railway stations of the railway era, the ports of the maritime era, the airports of the aviation era, or the hubs of the highway era. The best AI data center companies of the future will not just rent out server space—they will organize land, power, cooling, networking, chips, and long-term customer contracts into a cohesive infrastructure asset with cash flow, barriers to entry, and scarcity.

There is also a subsequent development for this line: the financialization of data center assets. Once an AI data center generates stable cash flow, it may be structured as REITs, RWA, infrastructure funds, revenue rights products, or long-term lease assets. In the era of cloud computing, data centers were backend assets for cloud providers; in the AI era, they will be revalued as “computing industrial real estate.”

Five: After training comes the battle of inference costs.

Many people now believe that NVIDIA is too dominant, so the opportunities in AI chips have already been fully taken by NVIDIA. This judgment is only partially correct.

During the large model training phase, NVIDIA holds a significant advantage—not only because of its powerful GPUs, but also due to its superior CUDA ecosystem, developer community, networking systems, integrated solutions, and software toolchain. However, once AI enters the large-scale inference phase, the logic shifts. In training, the priority is to build the model; in inference, the priority is to enable the model to serve millions of users daily. Training is more like capital expenditure, while inference is more like operational cost.

When AI enters customer service, office work, programming, finance, education, healthcare, and manufacturing, it generates massive volumes of requests daily. At this point, people realize that the token economy logic differs from the traffic economy logic: in traffic economics, marginal costs decrease, allowing companies to acquire users at scale before focusing on revenue, since the network cost per additional user keeps declining. But token economy logic is different—it imposes entirely different economic costs on large model companies and cloud providers, as marginal costs remain constant or even rise. Training involves a one-time investment with long-term returns, but inference does not. If each AI service request results in a loss, and a task is invoked tens of millions or even hundreds of millions of times per day, no company can sustain this. This is why even ByteDance’s large model, Doubao, has begun implementing paid services.

At this point, new opportunities emerge, and people begin to consider how to reduce costs. Why must all tasks rely on the most expensive general-purpose GPUs? Could specialized chips be used instead? Could ASICs with lower power consumption, higher throughput, and better suitability for fixed scenarios be employed?

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

Reuters reported that Broadcom expects revenue opportunities from custom AI chips to exceed $100 billion by 2027, driven by rapidly growing demand from major tech companies for custom AI chips. [7] In its 2024 annual report, AMD disclosed that its data center AI business has surpassed $5 billion in annual revenue, with customers such as Meta, Microsoft, and Oracle deploying AMD Instinct MI300 accelerators at scale. [8] Google Cloud emphasized that the TPU v5e is designed for cost efficiency, delivering higher query volumes at the same cost. [9]

Therefore, AI chips will not take on just one form in the future. NVIDIA will remain strong, but cloud providers will carve out their own space with self-developed chips, customized ASICs, inference acceleration chips, and edge AI chips. This is not a simple replacement of NVIDIA, but rather a sharing of part of the profit pool in the inference era. As AI transitions from the training era to the inference era, cost optimization will become the new source of pricing power.

Sixth, after the optical modules comes the entire AI network.

Many believe that the optical module sector has already peaked, signaling the end of the AI rally. I don’t see it that way. Optical modules are merely the first visible layer of AI networking. Behind them lie switching chips, switches, DPUs, SmartNICs, CPO, silicon photonics, cluster scheduling, and network operating systems.

The essence of an AI data center is connecting a large number of GPUs to form a supercomputer. The most valuable asset here is the GPU, and the least acceptable scenario is GPU idle time. If network latency is high, GPUs wait for data; if switching efficiency is low, GPUs wait for data; if the communication architecture is poor, GPUs still wait for data.

Therefore, the value of an AI network lies not just in transmitting data, but in improving the overall utilization of GPU clusters. In a conventional internet data center, a slower network might only result in slower loading times for users; in an AI data center, a slightly slower network could lead to a drop in utilization for equipment worth hundreds of millions or even billions of dollars.

NVIDIA's Quantum-X800 InfiniBand platform is designed for end-to-end 800 Gb/s networking to serve AI models with trillions of parameters; Spectrum-X Ethernet focuses on enhancing AI network performance and supporting the scaling of large-scale GPU clusters. [10] TrendForce also notes that demand for 800G and higher optical transceivers in AI server cluster interconnects is rising rapidly, and the market size for AI optical transceivers is expected to continue expanding. [11]

In the future, AI networks will continue to evolve: from 400G to 800G, then to 1.6T; from traditional optical modules to CPO; from electronic switching to photonic-electronic integration; from standard networks to AI fabric; from individual devices to full-cluster orchestration. The capital market will no longer focus solely on optical module businesses, but will instead prioritize those who can enhance AI cluster connectivity efficiency, reduce GPU wait times, and ensure greater stability in clusters of ten thousand or even hundreds of thousands of GPUs.

Seven: After the token becomes cheaper, the entry point will change hands.

The large-scale adoption of the AI era depends on whether token costs can continue to decline. The more expensive tokens are, the harder it is for AI to become widespread; the cheaper tokens become, the easier it is for AI to integrate into business processes and daily life.

The Stanford 2025 AI Index Report shows that the cost of querying models reaching GPT-3.5 level dropped from approximately $20 per million tokens in November 2022 to about $0.07 in October 2024—a reduction of over 280 times in roughly 18 months. The rate of decline in LLM inference costs varies significantly across different tasks, ranging from 9-fold to 900-fold annually. [12]

This data indicates that the true long-term deflationary forces in the AI industry have begun to emerge. Those who can accomplish the same task using fewer tokens, less VRAM, less electricity, and less inference time hold the most value.

Companies of this type, I call token compressionists.

They could be model companies, inference platforms, chip manufacturers, cloud providers, or enterprise AI infrastructure companies. What matters isn’t what they’re called, but whether they can make the same task cheaper, shorten the inference chain, reduce unnecessary calls, and deliver more stable results.

Several technologies are crucial here: MoE, quantization, distillation, caching, speculative decoding, and model routing. In particular, not every task requires invoking the most expensive model. Mature AI systems will automatically select the most appropriate model and path based on task complexity, cost budget, speed requirements, privacy needs, and accuracy demands. However, model routing is also vulnerable to competition from large tech companies, as its competitive advantage is not strong.

Once costs come down, the issue of entry points will become even more important. Many believe that the entry point in the AI era will be a model orchestration platform, like Meituan for AI. This analogy makes sense, but it’s not deep enough. The true AI entry point may not be a platform where you choose models; it’s more likely to be a system embedded within your workflow.

Ordinary users won't actively open a model orchestration platform every day. Enterprise users won't invoke models just to invoke them. Users want to complete tasks; enterprises want to streamline processes; employees want tangible work outcomes. AI will ultimately be embedded into Office, Feishu, DingTalk, WeCom, ERP, CRM, code editors, browsers, email, search, financial systems, customer service systems, and trading systems. Whoever controls the workflow controls the authority to invoke AI.

In its 2025 annual report, Microsoft disclosed that the combined monthly active users of the Copilot product family across commercial and consumer segments exceeded 100 million, and further integrated Microsoft 365 Copilot into workflow processes. [13] This demonstrates that the AI entry point does not necessarily have to be a standalone app, but can instead be an intelligent layer embedded within existing workflows.

For programmers, the entry point may be a code editor or code hosting platform; for office work, it may be Microsoft 365, Google Workspace, Feishu, or DingTalk; for enterprise operations, it may be ERP, CRM, or financial systems; for individuals, it may be mobile operating systems, browsers, search bars, or smart glasses. In the AI era, the true entry point is not a list of models, but a workflow entry point.

Eight: The real challenge for enterprise AI is integrating into workflows.

For AI to become an entry point for workflows, one prerequisite is that it must integrate into actual enterprise processes. The greatest challenge for enterprise AI is not simply deploying a chatbot, but whether the model can securely access corporate data, understand business processes, interact with systems, generate audit logs, comply with audits, and seamlessly integrate with human approval mechanisms.

Many companies today use AI only at the level of employees asking questions, writing, and summarizing on their own. This can improve individual efficiency, but it doesn’t truly transform organizational structures. True enterprise AI involves agents integrated into workflows.

The customer service agent does more than answer questions—they must check orders, track shipments, assess refund eligibility, and invoke after-sales systems; the finance agent does more than generate reports—they must review vouchers, reconcile accounts, detect anomalies, and produce approval recommendations; the investment research agent does more than summarize news—they must pull data, build models, compare companies, and monitor risks; the legal agent does more than draft contracts—they must search clauses, identify risks, link relevant cases, and preserve revision histories.

Behind this lies an entire infrastructure: databases, vector retrieval, access control, data governance, system integration, workflow engines, audit logs, security and compliance, enterprise knowledge bases, and agent orchestration platforms. These may not sound as glamorous as large models, but they are the foundation upon which AI truly takes root in enterprises. The first money a company should spend on deploying AI should be on security, data, access control, processes, integration, and compliance—not on deploying a flashy model to execute superficial tasks with some tokens.

There’s also a bigger shift here: the real money in AI applications may not come from software budgets, but from human resource budgets. SaaS sells tools; AI agents sell outcomes. Tools require human operation; agents complete tasks directly.

An AI customer service system, if merely selling software, has a ceiling limited to the customer service software market; but if it truly replaces a large number of human customer service agents, its ceiling becomes the entire market for outsourced customer service and corporate labor costs. An AI legal system, if only selling document tools, has limited potential; but if it can replace junior lawyers, contract reviews, and due diligence organization, its ceiling becomes the entire pool of legal service costs.

Harvey is a noteworthy case study in legal AI. According to TIME in 2025, Harvey is valued at approximately $5 billion, serves over 300 clients across 53 countries, and has been adopted by seven of the top ten law firms in the U.S. by revenue. [14] This demonstrates that AI applications in high-value knowledge work are not merely replacing tools, but are penetrating the labor cost pool of professional services.

In the future, truly outstanding AI application companies won't just describe themselves as software companies—they will highlight how much work they can accomplish for their clients, how much labor they can save, how many errors they can reduce, how much conversion they can improve, and how much they can shorten delivery cycles. Capital markets once focused on ARR; in the future, they will also assess how large a labor cost pool they can eliminate.

Nine: Don't Ignore Local AI and the Financialization of Computing Power

There are two other trends that are not currently the hottest but cannot be ignored in the medium to long term. One is local AI. Today, most tokens are still generated in cloud data centers—when you ask a model a question, it’s essentially a distant data center doing the computation for you. But in the future, not all AI inference can remain in the cloud.

The reason is simple: cloud-based inference is too expensive, many scenarios require low latency, and much data cannot be uploaded to the cloud. Terminal devices will also become increasingly intelligent. In the future, some tokens will migrate from the cloud to the local device—also known as the edge. Smartphones will run AI, PCs will run AI, cars will run AI, robots will run AI, smart glasses will run AI, and local workstations will run AI.

Once edge-side AI takes off, it will trigger a new hardware cycle. Edge AI chips, NPUs, low-power memory, power management, thermal solutions, sensors, camera modules, microphone arrays, AI PCs, AI smartphones, AI glasses, robots, and in-vehicle intelligent computing platforms will all undergo a new supply chain reassessment.

But this trend should be viewed objectively. The direction of edge-side AI is correct, but it still lacks a truly killer application. Currently, AI PCs and AI phones are primarily driven by hardware manufacturers, and users have not yet developed a compelling need to upgrade. Therefore, edge-side AI will not be the first major trend to explode, but it will remain an important long-term theme.

Another line is the financialization of computing power. AI infrastructure is too capital-intensive: GPUs are expensive, data centers are expensive, power contracts are expensive, construction cycles are long, and capital is heavily tied up. Relying solely on tech companies’ balance sheets to bear these costs may not be the optimal solution.

Future types of financial assets may include: GPU leasing contracts, computing power revenue rights, data center REITs, AI infrastructure funds, long-term power purchase agreements, GPU-backed financing, structured financing based on inference income, and tokenized computing power assets.

The digital assets business at Fosun Wealth, where I am based, is among Hong Kong’s most professional RWA issuance teams. Based on my frontline business analysis, compute power assets as RWA possess significant financial asset value and a promising future for global compliant trading. FinChain StarChain and StarRoad are helping large traditional compute power providers forge compliant, tokenized financial pathways—from Bitcoin mining power to AI compute power.

Successful cases of compute power securitization already exist overseas, with CoreWeave being the most prominent example. In March 2026, CoreWeave announced the completion of an $8.5 billion delayed draw term loan facility, labeling it the first investment-grade, GPU-backed financing. [15] This indicates that GPUs, server racks, and compute contracts are being revalued by financial markets as collateralizable and financiable infrastructure assets.

This is very similar to railways, telecommunications, and cloud computing. During the railway era, railway companies financed track construction through bonds; in the telecommunications era, carriers used long-term capital to build networks; in the cloud computing era, cloud providers made massive capital expenditures to construct data centers. In the AI era, GPUs, server racks, power contracts, and future inference revenues will also be repackaged, priced, and traded by financial markets.

Ten: The Highest-Level Opportunity—AI-Native Companies Will Rewrite the Income Statement

The previous discussion focused on the industrial chain. However, AI's greatest long-term impact goes beyond the industrial chain—it will rewrite organizational structures.

In the past, companies were structured around people forming departments—sales, customer service, finance, legal, investment research, and operations—each with defined roles, processes, approvals, and performance metrics. With the introduction of AI Agents, organizational structures will change: one person can manage multiple Agents, entire departments can be streamlined into automated Agent workflows, mid- and back-office roles will be automated, managerial span will expand, and companies will transition from labor-intensive organizations to human-AI collaborative ones.

This means that capital markets will revalue a new type of company in the future: AI-native companies. It’s not about simply purchasing a few AI tools or having employees use ChatGPT to write copy—it’s about redesigning the entire organizational structure from the ground up around AI. Fewer people, higher revenues, greater output per person, lower marginal costs, and faster delivery times.

Therefore, AI’s greatest impact on capital markets isn’t just about “who in the AI supply chain will rise,” but also about “which companies across all industries can use AI to rewrite their profit statements.” The future market will reward two types of companies: one that sells AI infrastructure and AI capabilities; the other that uses AI to reconstruct their cost structure and revenue model. The latter may not appear to be AI companies on the surface, but their organizational efficiency, profit margins, and output per employee will undergo fundamental changes.

Conclusion: AI is redefining scarcity

So far, if you only see GPUs, optical modules, power, liquid cooling, ASICs, data centers, and edge devices, you're still viewing AI as merely a technology supply chain. The deeper change is that AI will redefine what is scarce.

Previously, GPUs were scarce, so NVIDIA rose; later, HBM and optical modules became scarce, so memory and optical modules rose; next, electricity, liquid cooling, AI networks, inference chips, data pipelines, workflow entry points, enterprise data, and organizational execution will be scarce.

If we break down this AI market cycle, in the first stage, investors bought into compute infrastructure; in the second stage, they bet on whether compute could operate stably and cost-effectively; and in the third stage, they invested in whether compute could be integrated into enterprise workflows to generate real revenue and profits.

The optical module surged tenfold—not because the story is over, but because the capital market first clearly saw the physical bottlenecks in AI infrastructure. Even larger revaluations will occur on the next set of bottlenecks that have not yet been fully priced in.

Electricity, liquid cooling, AI data centers, custom ASICs, AI networks, token compression, model routing, enterprise data pipelines, workflow entry points, edge AI, compute financialization, and AI-native companies will collectively form the next wealth map of the AI industry.

Of course, this doesn't mean every company will rise, or that every concept is worth investing in. In every industrial revolution, wealth is never distributed evenly. The companies that are truly rewarded over the long term by capital markets are those that control critical bottlenecks, have customers, orders, technological barriers, cost advantages, and a strong position within their ecosystem.

The first wave of AI opportunities belongs to those who can build computing power; the next wave belongs to those who can support, optimize, schedule, and ultimately turn computing power into tangible business outcomes.

Notes and Sources

The following information supports the historical facts, public data, and industry cases mentioned in the text. For the convenience of financial media editors in verifying the content, primary sources such as official institutions, company announcements, authoritative media, or original materials are prioritized.

[1] Regarding Yuanjie Technology's intraday stock price surpassing Kweichow Moutai, becoming the new highest-priced stock on the A-share market: Sina Finance, April 17, 2026, “Surpassing Moutai: A New King of A-Share Stocks Emerges.” https://finance.sina.com.cn/wm/2026-04-17/doc-inhuupte2305062.shtml

[2] Regarding the 1866 transatlantic telegraph rates: PBS American Experience, “How the Early Cable Was Used,” states that the initial rate for the 1866 transatlantic telegraph was $10 per word, with a minimum of 10 words—equivalent to about ten weeks’ wages for a skilled worker. https://www.pbs.org/wgbh/americanexperience/features/cable-how-early-cable-was-used/

[3] Regarding the Reuters pigeon case: Reuters, “The long history of speed at Reuters,” mentions that Reuters used pigeons in its early days to transmit financial information. https://www.reuters.com/article/business/the-long-history-of-speed-at-reuters-idUSKBN2761WD/

[4] Regarding the 1915 U.S. transcontinental phone call charges: JSTOR Daily, “AT&T: Birth of the First Social Network,” mentions that a 3-minute coast-to-coast call in 1915 cost $20.70. https://daily.jstor.org/birth-first-social-network/

[5] Regarding the NVIDIA GB200 NVL72: The official NVIDIA page states that the GB200 NVL72 connects 36 Grace CPUs with 72 Blackwell GPUs using a rack-scale, liquid-cooled design. https://www.nvidia.com/en-us/data-center/gb200-nvl72/

[6] On global data center electricity consumption: International Energy Agency, “Energy demand from AI,” estimates that global data center electricity consumption will reach approximately 945 TWh by 2030, accounting for less than 3% of total global electricity use. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai

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

[8] Regarding AMD’s data center AI business: AMD’s 2024 Annual Report discloses that annual revenue from data center AI exceeded $5 billion and mentions deployments of AMD Instinct MI300 accelerators by Meta, Microsoft, Oracle, and others. https://ir.amd.com/financial-information/sec-filings/content/0001193125-25-067185/0001193125-25-067185.pdf

[9] On the cost efficiency of Google TPU v5e: Google Cloud Blog, “Performance per dollar of GPUs and TPUs for AI inference,” mentions that TPU v5e delivers higher query volumes at the same cost. https://cloud.google.com/blog/products/compute/performance-per-dollar-of-gpus-and-tpus-for-ai-inference

[10] Regarding the NVIDIA AI Network Platform: The official NVIDIA Quantum-X800 page describes it as an end-to-end 800 Gb/s InfiniBand network. https://www.nvidia.com/en-us/networking/products/infiniband/quantum-x800/

[11] Regarding the AI optical transceiver market: TrendForce, 2026-04-20, “Global AI Optical Transceiver Market to Reach US$26 Billion,” notes rapidly growing demand for 800G and higher optical transceivers. https://www.trendforce.com/presscenter/news/20260420-13017.html

[12] On the decline in AI inference costs: Stanford HAI, “AI Index 2025: State of AI in 10 Charts,” shows that the cost of querying models at the GPT-3.5 level has decreased by more than 280 times over approximately 18 months. https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts

[13] Regarding Microsoft Copilot user base: Microsoft Annual Report 2025 discloses that the Copilot product family has over 100 million monthly active users combined across commercial and consumer segments. https://www.microsoft.com/investor/reports/ar25/index.html

[14] Regarding the Harvey legal AI case: TIME’s 2025 list of the World’s Most Influential Companies reports that Harvey is valued at approximately $5 billion, serves over 300 clients, and operates in 53 countries. https://time.com/collections/time100-companies-2025/7289586/harvey/

[15] Regarding CoreWeave’s GPU-backed financing: CoreWeave’s investor relations announcement, dated March 2026, confirmed the closing of an $8.5 billion delayed draw term loan facility, described as 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 an industry perspective piece; annotations are provided to cite sources and do not constitute any investment advice.