Source: Cynthia, Hong Kong Ethereum Community Hub
Guest: qinbaFrank — Investor in US and crypto secondary markets, who deconstructs macro, industry, and individual stock logic based on first principles.
On June 8, 2026, at a VIP event jointly hosted by Futu, SNZ, ETH HK Hub, and Sharplink, senior investor qinbaFrank gave a presentation titled "AI Computing Power Wave: Review and Outlook." He systematically reviewed the complete trajectory of the AI market from 2023 to the present: from the three major debates on whether 'computing power is necessary' to how penetration rates determine commercialization efficiency, and to the current critical phase shifting from hardware shortages to commercialization verification.
He also provided a framework for judging the level of the current adjustment—three scenarios: valuation correction, earnings correction, and logic correction—and explained why this AI wave is 'superficially similar but fundamentally different' from the 2000 internet bubble.
Disclaimer: The content of this article faithfully presents the views shared by the guest and does not constitute any investment advice, product sales solicitation, or profit guarantee.
I. Why Risk Was Flagged and Positions Reduced on June 3rd
Since 2023, I have written some thoughts on macroeconomics and this wave of the AI/computing power market. In June 2024, I recommended Palantir on X, believing it had 3~5x upside potential as a representative of defense/military AI. This view was highly controversial at the time, but in retrospect, it indeed experienced a significant rally.
This is my first time giving such a presentation offline. I'd like to take this opportunity to systematically outline my overall framework for this AI wave: how it has evolved, where it stands now, and its potential future direction.
Last Wednesday evening (June 3rd), I gave an interview lasting over two hours for the US stock community 168X on X. The core viewpoint was: the market recently has been 'too hot' and needs appropriate cooling and adjustment. Specific reasons include:
- First, excessive sentiment and overheated FOMO. Capital concentration in hot sectors has reached an extreme level; parabolic rises are difficult to sustain, while orders and earnings have not fully materialized.
- Second, SpaceX's IPO roadshow triggered institutional portfolio adjustments. During the SpaceX roadshow, many institutions began trimming related holdings and shifting capital in advance, rather than waiting until the official listing—such capital rotation and extraction effects often manifest early.
- Third, geopolitical tensions bring risk aversion. US-Iran negotiations remain volatile, compounded by the non-farm payroll data released last Friday and this week's CPI data, leading to a decline in overall market risk appetite.
- Fourth, non-farm payroll data impacted rate-cut expectations. If May's non-farm payroll additions significantly exceeded expectations, the market would reprice a higher interest rate path.
- Fifth, this week's CPI data is the real policy variable. Strong non-farm data alone isn't enough to decide on rate hikes; the key is core CPI—especially whether rising energy prices will transmit and spread to service prices. This is the core variable to watch closely in the next week or two.
The core dividing line for judging the level of this adjustment is: Pure digestion of liquidity/crowding typically leads to minor corrections; CPI data exceeding expectations may escalate it to a small-medium level; only a clear slowdown in AI commercialization or cloud revenue would imply a reset of the entire narrative. Overall, I believe the market needs time to digest and wait in the short term. Previously overheated sectors may enter a period of mild to moderate correction until the next 'macro signal' provides relief.
II. Review: The 'Three Debates' of the AI Market Over the Past Three Years
To understand the current position, it's necessary to review the complete path of this AI wave from 2023 to now. I believe this wasn't a simple linear rise but a wave-like progression driven by cycles of 'market debate—verification—re-debate.'
First Debate (Late 2023): Is the Capital Expenditure Really Necessary?
In the first half of 2023, this theme was primarily valuation-driven—earnings hadn't significantly improved, yet stock prices had already surged (roughly several-fold). This coincided with a global semiconductor downturn, and the market had significant分歧 on 'how much computing power AI actually needs.' Consequently, the latter half of 2023 saw high volatility and consolidation.
Second Debate (Early 2024 to Early 2025): Will Big Tech CapEx Continue to Accelerate?
In Q1 2024, NVIDIA's earnings began to improve sequentially, and major tech companies' capital expenditure also started accelerating, gradually convincing the market that 'computing power demand is a real trend.' A landmark event was at the 2024 Davos Forum, where OpenAI's Sam Altman suggested future needs for trillions of dollars in chip manufacturing capacity. This statement was highly controversial within the industry, with executives from NVIDIA and TSMC publicly expressing skepticism, arguing such massive investment wasn't needed. However, subsequent continued outperformance in cloud providers' capital expenditure led the market to gradually accept this view—the scale of electricity and computing power needed for new US data centers is indeed a trillion-dollar level.
During this phase, capital flowed from big tech's capital expenditure to NVIDIA and its upstream supply chain, driving the main upward wave of 2024.
Third Debate (Early 2025): Is Computing Power Overestimated?
In Q1 2025, the release of a large model with significantly improved training efficiency sparked market质疑 about 'whether this much computing power is truly needed,' leading to a noticeable stock price correction. Following that in February, changes in US tariff policy caused another sharp decline, with core stocks correcting significantly from their highs—the second major adjustment in this wave.
Third Phase (Late 2025): Consensus Formation
By Q2 and Q3 2025, the market widely perceived clear improvements in large model capabilities and practicality. Use cases shifted from 'training-focused' to 'inference-focused,' with increases in model parameter scale and multimodal capabilities further driving computing power demand. In this phase, big tech capital expenditure entered a new round of acceleration, and the market rally entered a new upswing.
III. Core Framework: Penetration Rate Determines Commercialization Efficiency
Personally, I judge how far a technological wave can go primarily by its penetration rate, not just whether 'the trend exists.'
Many compare this AI wave to the 2000 internet bubble. I believe they are 'superficially similar but fundamentally different': both experienced parabolic rises where valuation preceded earnings, but the industrial environments are worlds apart.
-
Around 2000, US internet penetration was only over 30%, and business models (advertising, e-commerce, gaming, value-added services) were still being explored. Thus, after the bubble burst, the Nasdaq took considerable time to recover.
-
The mobile internet around 2010 was different: After the iPhone's 2007 launch and Android's openness, mobile internet penetration in China and the US crossed from early to mainstream in about a decade (2010-2018)—much faster than the internet's two-to-three-decade process. This was because previous infrastructure (internet普及, information dissemination efficiency) laid a solid foundation for the next generation.
Today, we face an environment where billions globally are accustomed to using WeChat, social media, and various apps—the speed of information dissemination and public acceptance of new technologies are incomparable to 2000. This is the biggest difference between the current AI industrial environment and the 2000 internet era.
Specifically regarding judgment methods, I subscribe to a key node 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' phase, and whether it's revolutionary enough determines if it can scale; once it crosses 10%, it意味着 crossing into the mass market, and the growth slope typically steepens; the 10%~50% range is the core observation window and the 'golden period' for related industry investment—user base expansion and willingness to pay increase simultaneously, driving token consumption upwards; beyond 50%, incremental space diminishes marginally.
Referencing survey data: A major investment bank's survey on corporate AI procurement intentions showed this比例 increased from about 10% last September to about 18% by the end of March this year—indicating corporate AI penetration has crossed the critical point and entered a rapid growth phase.
Comparing this AI wave to three generations of tech waves: PC internet took about 20 years from 1990 to 2010 to完成渗透; mobile internet took less than 10 years from 2010 to 2019; AI, starting from 2023, may扩散 even faster. The core reason is that more complete infrastructure leads to shorter commercialization cycles—the mobile internet era was propelled by smartphones, 4G, app stores, and mobile payments; today's AI stands on the shoulders of cloud computing, model APIs, social dissemination, and Agent infrastructure, making information diffusion and commercialization手段 more mature than any previous generation.
IV. AI vs. Internet: The Fundamental Difference in Commercialization Logic
The internet's core solved the problem of 'connection and information dissemination efficiency'—reducing intermediary costs in information, logistics, and capital flows—but it didn't directly replace 'humans.'
AI is different: it directly substitutes for human cognition and labor. When an AI's capabilities reach or exceed the 'societal average level' of human employees, it brings not just efficiency gains but genuine substitution—meaning companies paying for AI is essentially equivalent to their past costs of hiring that labor. This is why many people (including myself) quickly upgrade AI tool payments from free to tens, hundreds of dollars per month, or even pay for multiple large models simultaneously—once experiencing 'it indeed does things better and faster than me,' willingness to pay rises decisively. Therefore, once AI surpasses the average societal智力水平, its commercial value rises exponentially.
This also echoes a question raised earlier by a guest: Under the trend of AI rapidly replacing cognitive labor, how will the value of an individual's professional knowledge and experience 'moat' change? This is one of the fundamental reasons why AI commercialization is more complex than the internet's.
V. Investment Logic in the Computing Power Industry Chain: From 'Single GPU Narrative' to Systemic Re-rating
The logic of computing power investment is shifting from单纯押注 GPUs to a systemic re-rating across the entire chain: storage, CPUs, interconnect, power supply, packaging, and edge hardware. Overall, it can be概括 by a three-stage framework: Short-term看 'resource scarcity,' medium-term看 'system upgrade,' long-term看 'Physical AI penetration.'
1. Scarcity Pricing: GPU Demand Spills Over to Storage and CPUs
The logical chain is: long context, multimodal, and Agent applications drive storage demand—HBM tightens first, then progressively transmits to DRAM/GDDR, NAND/SSD/HDD, then to CPU scheduling, and finally to power supply.
First, GPU scarcity. 2022-2023 coincided with a global memory industry downturn, with大量产能 cleared. Entering 2024, as large cloud providers' capital expenditure accelerated, the impact of this capacity clearance began to show.
Then, storage/HBM scarcity. HBM itself has complex production processes and slow yield improvement. After the previous cycle's severe产能过剩, major memory makers are very cautious about expansion, with new capacity only gradually释放 by late 2027. This significantly increased memory makers' bargaining power when signing long-term supply agreements—contracts are for 5 years, requiring 10%~30% prepayments and even financial担保工具 from downstream customers. This is why these companies show 'earnings rising before valuation' characteristics: earnings consistently beat expectations in recent quarters, but valuation was suppressed due to market fears of 'repeating the semiconductor cycle.' Only after the existence of long-term agreements gradually convinced the market that cyclical波动 would be 'smoothed' did valuation begin to修复.
Next, CPU scheduling scarcity, and finally, power scarcity. The core reason is that many orchestration and scheduling tasks in data centers aren't suitable for GPU processing and must rely on CPUs. Taking NVIDIA's NVL72 rack as an example, the current configuration is roughly 72 GPUs配 36 Vera CPUs, i.e., a CPU:GPU ratio of about 1:2 (early schemes were about 1:8). The market expects this may further trend towards接近 1:1, meaning CPUs (whether Intel, AMD, or custom ARM chips) are being re-priced for their importance in computing infrastructure. Further down the chain lies data center power and grid capacity issues.
2. Upgrade Pricing: Simultaneous Upgrades in Optical Interconnect, Power Supply, Advanced Packaging
The second主线 is the 'upgrade logic'—the core isn't 'whether this module exists,' but whether conversion efficiency, power consumption, power density, and packaging yields can continue to improve.
Optical Interconnect: Optical modules evolving towards LPO/NPO/CPO. Co-Packaged Optics (CPO) integrates optical and electrical chips more closely, theoretically reducing power consumption, but it's not yet mass-produced. Some research indicates large cloud providers are unlikely to大规模 adopt CPO before 2027—core concerns are reliability: traditional optical modules can be replaced directly if faulty, while CPO issues involve更换 entire boards, with higher costs and validation cycles. Big players need time to充分验证 yields and failure rates.
Power Supply Network: Evolving from 48/54V to 800V HVDC. This is very similar to the high-voltage evolution in the electric vehicle industry—early EVs普遍 used lower-voltage architectures with lower efficiency; later,包括比亚迪, Huawei转向 higher-voltage直流 architectures, offering higher voltage, lower current, and smaller losses. Data center power systems are undergoing a similar upgrade path, driving demand for power semiconductors (like SiC) and power management相关产业链.
Advanced Packaging: 3D Stacking + Glass/Ceramic Substrates. This resembles the evolution path of smartphone chips in recent years—when性能提升 from单纯工艺节点缩小 reaches diminishing marginal returns, the industry turns to more advanced packaging methods (like 3D stacking, glass/ceramic substrates) to突破 physical limits, using better materials and packaging工艺 to continue提升 overall性能.
3. Long-term Pricing: Edge Computing and Physical AI
The long-term logic is edge computing and Physical AI entering application verification—from small-model edge inference to robotics, autonomous driving, then大规模量产 and cost reduction, ultimately forming new penetration curves. Short-to-medium-term tracking focuses on storage, CPUs/ARM, optical interconnect, power equipment, and advanced packaging; long-term要看 robotics and autonomous driving量产 curves.
VI. Evolution of Investment Themes: From Physical Constraints to Vertical AI OS
After computing power supply tightness eases, market focus will migrate along a path: Physical constraints (computing power/capacity shortage) → Enterprise deployment layer (can companies turn AI into production systems) → Vertical AI OS (controlling industry workflow entry points) → Physical AI (entering the real physical world).
The essence of the enterprise deployment layer is not simply adding a chatbox but rewriting enterprise workflows: first identify high-frequency, high-human-cost, verifiable-result workflows, then integrate private enterprise data (involving RAG, access control, data lineage, knowledge graphs), enabling Agents to truly execute actions (calling APIs, SaaS, completing approval and rollback processes), and continuously measure task completion rates, takeover rates, costs, and ROI.
'Vertical AI OS' can be understood as the industry's intelligent control layer—unlike traditional SaaS where 'humans operate software,' AI OS is 'AI调用 tools,推进 processes, with humans负责 supervision, approval, and decision-making.' It本质上 combines System of Intelligence + Action + Governance. Core metrics for judging progress in this phase include: whether commercialization continues accelerating (model ARR, cloud revenue, enterprise customer count), whether deployment quality truly passes production lines (task completion rate, manual takeover rate, accuracy), whether economics close the loop (unit inference cost, ROI, gross margin), and whether moats form (private data, process depth, compliance/audit).
VII. The Underlying Anchor for Wave-like Upswings: Model ARR and Cloud Revenue
Whether the market narrative continues depends not on 'whether valuation is expensive,' but on whether model vendors' ARR (Annualized Recurring Revenue) and cloud business revenue maintain high growth—this determines whether big tech capital expenditure is justified and whether the entire computing power chain's景气度 can persist. The transmission chain is: Real demand (B/C端真实付费) → High model vendor ARR growth → Cloud business beats expectations → Computing power chain continues benefiting.
Around this chain, three scenarios can be discussed:
Scenario 1: Growth hasn't slowed, logic not reversed. If model vendors' ARR is still growing and cloud business continues beating expectations, it means capital expenditure justification holds, and the computing power chain's order logic remains valid. In this case, even short-term overbought conditions and valuation 'being deemed expensive' cause minor-to-moderate pullbacks, the fundamentals aren't impaired—often falling fast and recovering fast.一旦财报季或新应用出现, it may quickly drive a reversal.
Scenario 2: Growth falls short, narrative reset. If model vendors' performance clearly slows, or cloud business demand链 shows clear deceleration, it indicates the problem is closer to the 'commercialization origin'—as much cloud computing采购本身 comes from these model vendors. In this case, it's at least a moderate adjustment, requiring new evidence that scale and growth can重新 beat expectations before信心 returns.
Scenario 3: Macro/liquidity factors are 'amplifiers,' not root causes. Macro and liquidity affect market sentiment and discount rates, but only when they truly impact the commercialization level do they升级为核心风险. Specifically, three tiers:单纯 liquidity withdrawal or a single CPI beat usually leads to minor adjustments; if叠加持续通胀, no rate cuts, and geopolitical risks, it may升级 to minor-moderate; only when model ARR or cloud revenue shows real deceleration does it enter moderate-level logic重置.
Simply put: As long as large model ARR and cloud revenue don't slow, this adjustment resembles a valuation and liquidity re-pricing, not a 2000-style crash; once fundamentals truly失速, then new reversal evidence is needed.
VIII. Current Phase: Shifting from Hardware Scarcity to Commercialization Verification
From April to June this year, the market's core assumption was: large cloud providers' capital expenditure guidance would continue beating expectations, supported by real付费需求 for cloud services from enterprises and consumers (i.e., cloud business revenue growth). If this holds, it means capital expenditure is 'reasonable and sustainable,' and the entire supply chain—storage, optics, CPUs, chips, all the way to power and grids—would benefit.
Looking ahead, I believe market focus will gradually shift from 'hardware scarcity' to 'commercialization realization.' A May report mentioned that in the enterprise services market, the best-selling product category is actually AI implementation/consulting services—i.e., capabilities helping enterprises truly deploy AI into specific business processes. The underlying logic is: core production工艺 and experience in many industries aren't公开的文档资料 but reside in资深员工的经验, which大模型 training data doesn't include. Those who can help enterprises combine this industry know-how with AI will capture下一阶段 opportunities.
My personal judgment is: as long as this growth rate itself doesn't明显恶化, subsequent pullbacks due to macro factors (e.g., rates, tariffs) are more likely to be minor-to-moderate阶段性调整, not trend reversals. What truly warrants caution is a scenario where overall AI commercialization growth大幅低于预期—that's when re-evaluating the sector's valuation logic becomes necessary.
IX. Historical Reference: A Three-Tier Framework for US Stock Adjustments
Judging the severity of US stock adjustments, the magnitude of decline itself isn't very meaningful; the key is whether the trigger推翻 long-term logic—whether it's单纯 valuation correction impulse, macro event shock, or a reset of the entire industry narrative. Using the Nasdaq as a benchmark (purer tech属性), corrections over the past ~20 years can be大致分成 three levels:
L1 Minor (single-digit % decline): Triggers are usually 'valuation correction' impulses after rapid rises,叠加 liquidity shocks or inflation/rate-cut expectation扰动. This isn't a crisis; fundamentals unchanged.一旦扰动缓解确认, reversals are typically swift. A recent example is last November's ~7%~8% correction, primarily a liquidity shock叠加萌芽 market质疑 about AI capital expenditure.
L2 Moderate (~15% decline): Usually伴随 certain macro events or market mechanism shocks; risks need repricing but don't mean underlying秩序崩塌; the market needs new data to confirm risks haven't further扩散. Examples: the ~15% correction from August to October 2023, against the backdrop of 10-year Treasury yields approaching 5%; the July-August 2024 correction was related to carry trade unwinding and recession fears.
L3 Major (25%+ decline): Means past familiar macro logic is重置, or the industry's long-term narrative is推翻; risk偏好 undergoes systemic重估, requiring entirely new evidence to重建信心. Historical examples include the 2008 financial crisis (~50%), Q4 2018 (~25%~30%), March 2020 pandemic (~30%~40%), 2022 rate-hike cycle (~33%~35%), and corrections around ~28% from tariffs or global trade order shocks.
Applying this to the current AI wave, the core dividing line remains whether AI commercialization growth slows: If model ARR, enterprise user count, token revenue, and cloud business revenue still beat expectations, business logic isn't reversed, and pullbacks are more liquidity- or macro-driven minor-to-moderate adjustments; if model vendor performance disappoints, it's closer to the commercialization origin, requiring at least moderate re-pricing and等待新证据; only when AI growth decelerates, simultaneously叠加 inflation爆表, geopolitical冲突, or systemic global order破裂, could it升级 to a major adjustment.
Simply put: As long as AI commercialization doesn't slow, this adjustment resembles 're-pricing'; only when commercialization evidence断档 does the entire framework need resetting.
X. Conclusion: AI is a Foundational Leap in Civilizational Capability
Finally, sharing my personal understanding of the nature of this wave. Historically, gunpowder, steam engines, electricity, and the internet were本质上 'single-point industrial revolutions'—they upgraded specific tools, energy sources, or information channels, solving a key bottleneck before diffusing along industry chains,呈现单一技术周期 S-curves. These revolutions changed 'one-dimensional capabilities,' not directly提升 intelligence itself.
I believe AI is different—it enhances the most foundational capability: 'intelligence.' An analogy is humanity 'using fire': transitioning from not using fire to using it didn't just add 'one more tool'; cooked food altered physical structure, impacting brain capacity, ultimately leading to civilizational capability expansion. AI similarly changes foundational capabilities—perception, reasoning, generation, decision-making, action—this整套能力 is整体上移. It's a底层 upgrade at the 'civilizational production function' level, not just making a specific tool more useful.
Precisely because it's a foundational capability leap,上层 will持续,分批地 grow new industrial revolutions: Agent revolution, robotics revolution, drone revolution, then defense/military, space tech, and更多行业流程重构. This process won't materialize一次性 but will emerge in successive waves. Therefore, I believe the truly值得跟踪的主线 isn't betting on a specific application爆 but continuously observing 'how智能能力外溢 to the physical world and各行业流程'—this is the核心线索 for judging how far this AI wave can go.
Looking ahead one to two years, I believe people will持续感受到 this 'acceleration within acceleration'—technological capabilities and commercialization进程 mutually验证,推动. But the market行情 itself definitely won't be a straight line; it will呈现波浪式的 characteristics amid逻辑切换 of 'scarcity—upgrade—long-term realization.'
Disclaimer: The content of this article faithfully presents the views shared by the guest and does not constitute any investment advice, product sales solicitation, or profit guarantee.
