qinbaFrank: A Review and Outlook on the AI Computing Power Wave - From Nvidia's Three Major Debates to Optical Interconnection, How is Capital Rotating?

Source: Cynthia, Hong Kong Ethereum Community Hub

Guest: qinbaFrank— US stock and crypto secondary market investor; he has long analyzed macro, industry, and individual stock logic using first principles.

On June 8, 2026, during a VIP event co-hosted by Futu, SNZ, ETH HK Hub and Sharplink, senior investor qinbaFrank delivered a presentation titled "Review and Outlook of the AI Computing Power Wave," systematically sorting out the complete path taken by the AI market from 2023 to present: from three major debates on "Is computing power necessary?" to how penetration rate dividends determine commercialization efficiency, and then to the current critical stage shifting from hardware shortage to commercialization validation.

He also provided a framework for assessing the level of this adjustment—three scenarios of killing valuations, killing performance, and killing logic, and explained why this round of AI market resembles the internet bubble of 2000 "in form but not in essence."

Disclaimer: The content of this article is a true presentation of the guest's shared viewpoints and does not constitute any investment advice, product sales invitation, or profit promise.

1. Why I pointed out risks and reduced positions on June 3

Since 2023, I have been writing a series of thoughts on macroeconomics and this round of AI/computing power market. In June 2024, I recommended Palantir on X, believing that as a representative of defense and military AI, it still had 3 to 5 times upside potential. There was considerable controversy in the market regarding this judgment at the time, but looking back, it indeed led to a very considerable market wave.

This was my first time sharing in person here. Taking this opportunity, I want to systematically outline my overall framework regarding this round of AI market: how it has progressed, its current position, and possible future directions.

Last Wednesday (June 3), I participated in an interview with a US stock community on X called 168X, lasting over two hours. The core viewpoint was: the market has become somewhat "too hot" recently, necessitating appropriate cooling and adjustments. The specific reasons are as follows:

First, the emotional aspect is overly crowded, and FOMO is excessive. The concentration of funds in popular directions has reached a relatively extreme position; parabolic increases are challenging to sustain, and orders along with financial reports have not yet fully materialized.
Second, SpaceX's IPO roadshow triggered institutional portfolio adjustments. During the SpaceX roadshow, many institutions started to reduce relevant holdings and free up funds in advance, rather than waiting until the official listing moment to take action—this fund rotation and extraction effect tends to show itself early.
Third, geopolitical situations have raised risk aversion sentiment. There are still fluctuations in US-Iran negotiations, combined with last Friday's non-farm payroll data and this week's CPI data, leading to an overall decline in market risk appetite.
Fourth, non-farm payroll data impacts interest rate cut expectations. If the non-farm payroll data in May significantly exceeds expectations, it would prompt the market to readjust to a higher interest rate path.
Fifth, this week's CPI data is the real policy variable. Strong non-farm payroll data alone is insufficient to determine whether interest rates should be raised; the real key is the core CPI—particularly whether the rise in energy prices will spread to service sector prices, which is a core variable that needs close attention over the next week or two.

The core boundary for judging this adjustment level is: purely market liquidity/crowding digestion usually only leads to small adjustments; if inflation data overshoots expectations, it may upgrade to small to medium adjustments; and only when AI commercialization or cloud revenue shows a significant slowdown does it mean the entire narrative is reset. Overall, I believe that the market needs some time to digest and wait in the short term; the excessively crowded popular directions may enter a moderate or medium amplitude correction phase until the next "macro signal" appears to alleviate conditions.

2. Review: The "three major debates" of the AI market over the past three years

To understand the current position, it is necessary to review the complete path taken by this round of AI market from 2023 to now. I believe this is not a simple linear rise but rather a wave market driven by multiple "market debates—verifications—re-debates."

First Debate (Second Half of 2023): Is capital expenditure really necessary?

In the first half of 2023, the main theme was primarily valuation-driven—performance had not significantly improved, yet stock prices had already risen considerably (roughly several times). At that time, the global semiconductor industry was in a downturn, leading to significant divergence in the market concerning "How much computing power does AI actually need?" Therefore, the second half of 2023 exhibited a high-level oscillation.

Second Debate (Early 2024 to Early 2025): Will major companies' capital expenditures continue to accelerate?

In the first quarter of 2024, Nvidia's performance began to improve quarter-on-quarter, and large tech companies also started to accelerate capital expenditures, which gradually confirmed to the market that "the demand for computing power is a real trend." A landmark event was at the Davos Forum in early 2024, where OpenAI's Sam Altman suggested that future investments of several trillion dollars would be necessary for chip manufacturing capacity. This statement prompted considerable debate in the industry, including public disagreement from the management of Nvidia and TSMC, who stated that such large-scale investments were unnecessary. However, as subsequent capital expenditures by large cloud providers continued to exceed expectations, the market gradually accepted this judgment—that the electricity and computing power required for newly built data centers in the US were indeed on the order of trillions of dollars.

During this stage, funding flowed from major tech companies' capital expenditures to Nvidia and upstream supply chains, driving the main upward wave of 2024.

Third Debate (Early 2025): Is computing power overestimated?

In the first quarter of 2025, the release of a large model with significantly enhanced training efficiency sparked market doubt about "Is this much computing power really necessary?" resulting in a notable stock price correction. Shortly thereafter, in February, shifts in U.S. tariff policies led to another significant decline, causing key related stocks to drop significantly from their peaks—this marked the second major adjustment of this round of market activity.

Third Stage (Second Half of 2025): Formation of consensus

By the second and third quarters of 2025, the market could widely sense a significant improvement in the capabilities and practicality of large models; application scenarios shifted from "training-driven" to "inference-driven," and the enhancement of model parameter scales and multi-modal capabilities further boosted the demand for computing power. During this phase, capital expenditures of major tech companies entered a new period of acceleration, and the market subsequently entered a new round of growth.

3. Core framework: Penetration rate determines commercialization efficiency

In my personal judgment of how far a technological wave can go, the core aspect is penetration rate, not merely whether "a trend exists."

Many people compare this round of AI market to the internet bubble of 2000. I believe the two are "similar in form but different in essence": both experienced parabolic rises with valuations advancing ahead of performance, but the industrial environment is vastly different.

Around 2000, US internet penetration was just over 30%, and business models (advertising, e-commerce, gaming, value-added services) were still in exploratory stages; thus, after the bubble burst, the Nasdaq took quite a long time to recover from its low.
In contrast, around 2010 for mobile internet: The iPhone was released in 2007, and following the opening of the Android system, the penetration of mobile internet in the US and China completed its transition from early to mainstream within about ten years (2010-2018)—far quicker than the internet's two to three-decade progression. This was facilitated by the previous generation's infrastructure (internet accessibility, information dissemination efficiency) laying an excellent foundation for the next generation.

Today we face an environment where billions around the globe have become accustomed to using WeChat, social media, and various apps—speed of information dissemination and overall acceptance of new technology among the public is on a completely different level than in 2000. This is precisely what differentiates this round of AI industrial environment from the internet of 2000.

Specifically regarding judgment methods, I resonate with a key point in the "technology adoption lifecycle" (crossing the chasm theory): a 10% penetration rate is the critical point. Below 10%, the technology is still in the "early validation" stage; whether it’s sufficiently revolutionary will determine whether scaling occurs; once it crosses 10%, it indicates it has entered the mainstream market, and the growth rate typically becomes steeper; the 10%-50% range is the core observation window and the "golden period" for industry investment—user scale expansion and increased willingness to pay occur simultaneously, leading to rising token consumption; beyond 50%, incremental space will then diminish marginally.

According to a survey data from a major investment bank on corporate AI procurement intentions, this ratio has increased from about 10% in September last year to about 18% at the end of March this year—indicating that corporate AI penetration has crossed the critical point and has officially entered a rapid growth period.

If we compare this round of AI wave to three generations of technological waves: The PC internet took about 20 years from 1990 to 2010 to reach saturation; the mobile internet took less than 10 years from 2010 to 2019; whereas AI, starting in 2023, may spread even more quickly. The core reason is that the more complete the infrastructure, the shorter the commercialization cycle—during the mobile internet era, smartphones, 4G, app stores, and mobile payments drove mainstream adoption; today’s AI stands atop cloud computing power, model APIs, social dissemination, and agents’ infrastructure, making information dissemination and commercialization methods more mature than in any previous generation.

4. AI and the Internet: The fundamental differences in commercialization logic

The core issue that the internet addressed is "the efficiency of connectivity and information dissemination"—it lowered the intermediate costs of information flow, logistics, and capital flow, but it did not directly replace "humans."

AI, on the other hand, is different: it directly replaces human cognition and labor. When an AI's abilities reach or even exceed those of a "social average" human employee, it brings not only efficiency improvements but a true sense of substitution—this means that companies paying for AI essentially pay what they previously would have paid to hire that labor. This is also why many (including myself) upgrade the amount they pay for AI tools quickly from free versions to tens, hundreds of dollars monthly, or even paying simultaneously for multiple large models—once experiencing "it really does it better and faster than I can," the willingness to pay increases very decisively. Therefore, once AI crosses the average human intelligence level, its commercial value will rise rapidly and exponentially.

This also echoes a question the guest mentioned previously: under the rapidly replacing cognitive labor trend of AI, how will the value of personal professional knowledge and experience "moat" change? This is one of the fundamental reasons why AI commercialization is more complex than the internet.

5. The investment logic of the computing power industry chain: From "GPU single-point narrative" to systematic revaluation

The logic of this round of computing power investment is shifting from merely betting on GPUs to a full-chain systematic revaluation that includes storage, CPUs, interconnects, power supply, packaging, and edge hardware. Overall, this can be summarized with a three-part framework: Short-term view "resource scarcity," mid-term look "system upgrade," long-term view "Physical AI penetration rate."

1. Scarcity pricing: GPU demand spills over to storage and CPUs

The logical chain is: long context, multi-modal applications and agents increase storage demand—HBM is first under pressure, and then affects DRAM/GDDR, NAND/SSD/HDD layer by layer, eventually reaching CPU scheduling and then power supply.

First is the GPU shortage. From 2022 to 2023, the global storage industry was in a downturn, and much capacity was eliminated. Entering 2024, as major cloud providers accelerated capital expenditures, the effects of this capacity elimination began to show.

Then it is storage/HBM scarcity. HBM production processes are complex, and yield improvements are slow; after experiencing a previous round of severe overcapacity, major storage manufacturers have been very cautious about expanding production, with new capacity gradually released only after the second half of 2027. This has led storage manufacturers to significantly enhance their bargaining power when signing long-term supply agreements—a long-term contract lasts five years and requires 10% to 30% upfront payment, and even demands that downstream customers provide financial guarantees. This is also why these companies demonstrate "performance preceding valuation increases": they have consistently exceeded expectations in earnings over the past few quarters, but valuations have been suppressed due to market concerns about "repeating semiconductor cyclical pitfalls," until adherence to long-term agreements gradually convinces the market that cyclical fluctuations will be "smoothed out," allowing valuations to begin their recovery.

Next is CPU scheduling tightness, followed by power supply tightness. The core reason is that many orchestration and scheduling tasks in data centers are not suitable for GPU processing and must rely on CPUs. Taking Nvidia's NVL72 cabinet as an example, the current configuration is roughly 72 GPUs paired with 36 Vera CPUs, indicating a CPU:GPU ratio of about 1:2 (early versions used to be about 1:8); the market expects this ratio could further approach 1:1 in the future, implying that the importance of CPUs (whether Intel, AMD, or self-developed ARM chips) in the computing power infrastructure is being repriced. Then, this touches upon the power and grid capacity issues within data centers.

2. Upgrade pricing: Optical interconnects, power supply, and advanced packaging upgrading in sync

The second main line is the "upgrade logic"—the core issue is not simply "Is this module available?" but whether conversion efficiency, power consumption, power density, and packaging yield can continue to improve.

Optical interconnects: Optical modules evolving to LPO/NPO/CPO. Co-packaged optics (CPO) aims to integrate optical chips and electronic chips more closely, theoretically reducing power consumption, but it has not been mass-produced yet. Some visits and surveys indicate that large cloud providers are unlikely to adopt CPO on a large scale before 2027—core concerns revolve around reliability: a traditional optical module can be directly replaced if it fails, whereas a CPO issue entails the replacement costs and validation periods at the entire board level, hence major manufacturers need time to thoroughly test yield and failure rates.

Power supply networks: Evolving from 48/54V to 800V HVDC. This path is quite similar to the high-voltage trend in the electric vehicle industry—early electric vehicles commonly utilized lower voltage power supply architectures, which were less efficient; subsequently, companies like BYD and Huawei began transitioning to higher voltage direct current structures, resulting in higher voltage, lower current, and smaller losses. The power supply systems in data centers are undergoing similar upgrades, fostering demand in related industries like power semiconductors (such as silicon carbide) and power management.

Advanced packaging: 3D stacking + glass/ceramic substrates. This resembles the evolution of smartphone chips over recent years—when reliance on scaling via process node reductions yields diminishing returns, the industry shifts towards utilizing more advanced packaging methods (like 3D stacking, glass or ceramic substrates) to break through physical limitations and enhance overall performance using better materials and packaging processes.

3. Long-term pricing: Edge computing and Physical AI

The long-term rationale is that edge computing and Physical AI are entering the application validation stage—from small model edge inference to robots, autonomous driving, and finally mass production and cost reduction, ultimately forming new penetration rate curves. The short to mid-term tracking focus lies on storage, CPU/ARM, optical interconnects, power equipment, and advanced packaging; long-term, it's essential to look at the production curves for robots and autonomous driving.

6. Evolution of investment mainlines: From physical constraints to vertical AI OS

Once the tight supply of computing power eases, the market's focus will experience a migration path: Physical constraints (insufficient computing power/capacity) → Enterprise deployment layer (can companies turn AI into production systems?) → Vertical AI OS (grasping industry workflow entry points) → Physical AI (entering the real physical world).

The essence of the enterprise deployment layer is not merely to integrate a chatbox but to re-engineer enterprises' workflows: first, identify high-frequency, high labor-cost, verifiable workflows, and then integrate them with the enterprise's private data (involving RAG, permission management, data lineage, knowledge graphs), enabling agents to genuinely take action (calling APIs, SaaS, completing approval and rollback procedures), continually measuring task completion rates, takeover rates, costs, and ROI.

The so-called "vertical AI OS" can be understood as the intelligent control layer of the industry—distinct from traditional SaaS "human-operated software," AI OS signifies "AI invocation tools advancing processes, with humans overseeing, approving, and making decisions," essentially combining System of Intelligence + Action + Governance. Core indicators for assessing progress in this phase include: whether commercialization continues to accelerate (model ARR, cloud revenue, corporate customer count), whether deployment quality genuinely surpasses production lines (task completion rates, human takeover rates, accuracy), whether economic viability is closed-loop (unit inference cost, ROI, gross margins), and whether moats have been established (private data, process depth, compliance audits).

7. Underlying anchors of wave-like upturns: Model ARR and cloud revenue

Whether the market narrative can continue, the essence is not "whether valuations are high," but whether model vendors' ARR (annual recurring revenue) and cloud business revenues continue to maintain high growth—this determines whether large tech companies' capital expenditures are rational and whether the overall healthiness of the entire computing power chain can persist. The conveyance chain is: Real demand (real payment from B/C ends) → Model vendors' ARR growth → Cloud business surpassing expectations → Computing power chain continues to benefit.

Around this conveyance chain, three scenarios can be discussed:

Scenario 1: Growth velocity isn't slowing; logic is not reversing. If model vendors' ARR continues to grow and cloud business keeps exceeding expectations, it indicates that the rationality of capital expenditures remains intact, and the ordering logic of the computing power chain continues to be effective. In such circumstances, even if there is a short-term surge and valuations are "perceived as high" leading to minor to medium pullbacks, the fundamentals have not deteriorated—often declines occur quickly, but repairs do as well; when earnings reports or new applications emerge, they may quickly reverse the trend.

Scenario 2: Growth velocity is below expectations; narrative resets. If model vendors' performance clearly stalls, or if there is a clear slowdown in the cloud business demand chain, it suggests a situation closer to the "commercialization origin"—because a lot of the computational resources in the cloud actually derive from these model vendors. In this case, at least a medium-level adjustment is required, and new evidence must be awaited to prove that scale and growth can exceed expectations once more for confidence to return.

Scenario 3: Macro/funding aspects are "magnifiers" but not the fundamental reason. Macroeconomic elements and funding conditions can affect market sentiment and discount rates; however, only when it truly impacts the commercialization level does it escalate to core risk. Specifically, this can be broken down into three tiers: simple withdrawal of funds or a singular CPI overshoot typically results in a minor adjustment; if compounded by ongoing inflation, no interest rate cuts, and geopolitical risks, it could elevate to a small to medium adjustment; only when model ARR or cloud revenue experiences genuine slowdown does it indicate entry into a medium-level narrative reset.

Simply put: As long as large model ARR and cloud revenues are not slowing, this adjustment seems more akin to a repricing of valuations and funding aspects rather than a 2000-style collapse; only when the fundamentals genuinely stall is new reversal evidence needed.

8. Current phase: From hardware scarcity to commercialization verification

From April to June this year, the core assumption in the market is that the capital expenditure guidance of major cloud vendors will continue to exceed expectations, backed by actual payment demand from corporations and consumers for cloud services (i.e., the growth rate of cloud business revenues). If this assumption holds, it suggests that capital expenditures are "rational and sustainable," so the entire supply chain—storage, optical, CPUs, chips, and even power and grid—will benefit.

Looking ahead, I believe the market's focus will gradually shift from "hardware scarcity" to "commercialization realization." A report released in May this year mentioned that in the corporate services market, the best-selling product category is actually AI implementation/consulting services—specifically, the capability to help enterprises effectively integrate AI into specific business processes. The underlying logic is that many industries' core production processes and knowledge aren't publicly documented but are ingrained in the experiences of senior employees; this implicit knowledge is typically not part of the large model's training data. Those who can help enterprises marry this industry know-how with AI will capture opportunities in the next phase.

My personal judgment is that as long as this growth rate itself does not significantly worsen, subsequent corrections due to macro factors (like interest rates, tariffs, etc.) are more likely to be minor to medium periodic adjustments rather than a reversal of the trend. What truly needs to be guarded against is a large-scale underperformance of the overall growth rate of AI commercialization—only then will there be a true need to reassess the valuation logic of the entire sector.

9. Historical reference: The three-tier framework of US stock adjustments

Judging the level of adjustments in US stocks based simply on the magnitude of declines has little meaning; the key is whether the triggering sources have overturned the long-term logic—whether it is merely a "killing the valuation" impulse, a macro event shock, or if the entire industry narrative has been reset. Taking the Nasdaq as a benchmark (due to its purer tech attribute), the adjustments over the last 20 years can generally be categorized into three levels:

L1 small level (single-digit decline): Triggering sources are typically the "killing valuations" impulse after a rapid rise, compounded by liquidity shocks or disturbances in inflation/interest rate cut expectations. This kind of adjustment is not a crisis; the fundamentals have not changed, and once disturbances are confirmed to have eased, reversions usually occur quickly. A recent example is last November's approximately 7%-8% correction, primarily due to liquidity shocks compounded with emerging doubts about AI capital expenditures.

L2 medium level (approximately 15% decline): Typically accompanies certain macro major events or market mechanics shocks, requiring risks to be re-evaluated, but does not signify a collapse of underlying order; the market must wait for new data to confirm that risks have not further expanded. For instance, the approximately 15% correction from August to October 2023 was against a backdrop of the 10-year US Treasury yield nearing 5%; the corrections in July and August 2024 related to unwinding carry trades and market concerns regarding recession.

L3 large level (decline of 25% or more): Indicates that the previously familiar macro logic has been reset, or the industry's long-term narrative has been overturned; risk preference will undergo systematic reevaluation, needing completely new evidence to rebuild confidence. Historical examples include the 2008 financial crisis (halving), the fourth quarter of 2018 (approximately 25%-30%), the pandemic shock in March 2020 (approximately 30%-40%), the interest rate hike cycle in 2022 (approximately 33%-35%), along with declines of about 28% due to tariff or global trade order shocks.

Applying this to the current round of AI market, the core boundary remains whether the growth rate of AI commercialization is declining: if model ARR, corporate user count, token revenue, and cloud business revenues remain above expectations, it indicates that the business logic has not been reversed, and pullbacks are more likely small to medium adjustments caused by liquidity or macro disturbances; conversely, if model vendors' performances fall below expectations, it indicates that they are approaching the commercialization origin, necessitating medium-level repricing and waiting for new evidence; only when the growth rate of AI declines, alongside surging inflation, geopolitical conflict, or fractures in the global order could it potentially escalate into a large-level adjustment.

Simply put: as long as AI commercialization does not slow down, this round of adjustments is more akin to "repricing"; it will only signify that the entire framework needs resetting when evidence of commercialization verging on a halt emerges.

10. Conclusion: AI is a foundational leap in civilization's underlying capabilities

Finally, I'd like to share my personal understanding of the nature of this wave. Historically, gunpowder, steam engines, electricity, and the internet have fundamentally been "single-point industrial revolutions"—they upgrade a certain tool, energy source, or information channel, solving a critical bottleneck, and then diffusing through the产业链, presenting a singular S-curve of technological cycles. These revolutions alter a "specific dimensional ability," rather than directly enhancing intelligence itself.

I believe AI is different—it enhances "intelligence" as the most fundamental underlying capability. It can be likened to humanity’s use of fire: shifting from not knowing to knowing how to use fire brings not just "one more tool," but cooked food alters bodily structure, influencing brain capacity, ultimately leading to an expansion of overall civilizational abilities. AI is similarly transforming foundational capabilities—perception, reasoning, generation, decision-making, and action—all of which are collectively elevating; this represents a bottom-up upgrade on a "civilizational production function" level, instead of merely making one specific tool more usable.

Because it represents a leap in underlying capabilities, the upper levels will continuously and progressively yield new industrial revolutions: the agent revolution, robotics revolution, drone revolution, followed by defense and military applications, space technology, as well as the restructuring of processes across various industries. This progression will not manifest as a single event but will come as wave after wave. Therefore, I think the truly worthwhile line to follow is not betting on a specific application explosion but instead continuously observing "how intelligent capabilities overflow into the physical world and various industry processes"—this is the core clue for judging how far this round of AI wave can go.

Looking ahead to the next year or two, I believe everyone will continue to feel this "accelerating acceleration"—the interplay and mutual verification of technological capabilities and commercialization processes. However, the market itself will not follow a straight line, and there will be wave-like features reflecting shifts between "scarcity—upgrades—long-term realizations."