● OpenAI Cash Blitz Ignites AI Supercycle, Grid and HBM Crunch

OpenAI, the True Intent Behind ‘Money Sprinkling’ and the Potential to Trigger the AI Supercycle + Q4 Investment Playbook

All the content now consists solely of English sentences.

1) Why OpenAI’s ‘Money Sprinkling’ Is Rational: Structural Changes in the Model Economics

• Key Points
  OpenAI’s massive cash injection is not aimed at short-term performance but at achieving long-term “economies of scale” and “network effects.”
  Training is a one-time expense, while inference is a recurring fee structure; when the model becomes the standard interface, its lifetime value (LTV) expands explosively.
  Even if prices keep falling, as long as the cost per token decreases at a faster pace, margins remain protected.

• Detailed Explanation
  1) Cost Curve: Cost per token ≈ (GPU cost per hour × GPU hours per 1M tokens) / 1M. Enhancements in HBM bandwidth, MoE (Mixture-of-Experts), kernel optimizations (vLLM, Triton), dynamic batching, and KV cache reuse structurally drive down cost per token.
  2) Revenue Curve: As multimodality, agents, and tool usage increase, token consumption rises, and once integrated deeply into enterprise workflows, customer churn becomes difficult.
  3) Data Lock-in: Licenses for content and corporate domain data, user interaction logs, and agent execution data reinforce the “data moat.”
  4) Platform Play: GPT stores, partner ecosystems, and basic integrations with OS/devices (e.g., mobile and PC NPU connectivity) are mechanisms that secure sustained demand.

2) Exactly Where the Money Goes: The Four Pillars of Compute, Power, Data, and Talent

• Compute
  The full value chain—from GPUs/HBM packaging (CoWoS), OSAT (back-end processing), optical modules (800G→1.6T), switching (IB/Ethernet), cooling (immersion and liquid cooling), to EDA/equipment (from design to photolithography to etching)—is being expanded in a cascade.
  Networking is no longer a “performance bottleneck” but an “economic bottleneck.” Demand for switches, NICs, and cables increases non-linearly with the training cluster size.

• Power/Infrastructure
  The true bottleneck is power. PPA (Power Purchase Agreements), extensions of transmission lines and substations, HV cable lead times, water resources/cooling zones, and site permits are critical path items for projects.
  Delays in N+1 power backup, power distribution panels, circuit breakers, and switchgear supply slow down data center ramp-ups.
  Upward revisions in utilities’ CapEx guidance translate into medium- to long-term beneficiaries.

• Data
  Licensing for broadcasting, music, publishing, and code, along with connections of enterprise documents and tool logs, synthetic data, and reinforcement learning feedback loops, raise the performance ceiling of models.
  Compared to the legal risks of web scraping, formal licensing offers a distinctly superior risk/return profile.

• Talent
  There is a concentration of talent in compilers, distributed systems, and simulation, with frequent acquisitions and team lifts.
  Standardization of toolchains (from prompts to workflows to evaluation to deployment) is accelerating, and “ML-Ops for Agents” is emerging as the next-generation demand.

3) The Overlooked Five Bottlenecks and Investment Points

• Power Grids and Water Resources
  Long interconnection queues for power grids are critical. Key items include substation extensions, high-voltage cables, GIS switchgear, UPS systems, cooling towers, and securing industrial water supplies.
  Investment perspective: utilities, power equipment, high-voltage cables, land/REITs, cooling solutions including valves and pumps, and water treatment.

• Packaging/Back-end Processing
  HBM3E, advanced packaging (CoWoS/SoIC), and the capacity of advanced substrates (ABF) are the true constraints. Expansion of OSAT and yield improvements determine GPU shipment volumes.
  Investment perspective: memory (HBM), OSAT, substrates, and test equipment.

• Networking/Optical Communications
  As clusters grow larger, the scale of switches, NICs, and optical modules influences costs. The transition to Ethernet-based AI fabrics is also accelerating.
  Investment perspective: switching, optical modules, connectors, and cables.

• Cooling and Thermal Management
  The shift toward immersion and liquid cooling is becoming widespread. Models integrating heat reuse (for heating or greenhouses) with carbon credit schemes are drawing attention.
  Investment perspective: cooling solutions, heat exchangers, and eco-friendly refrigerant chains.

• Regulatory/Antitrust Risks
  Issues regarding content copyrights, personal data, model safety, and partnerships with big tech increase costs and time.
  Demand for compliance, model evaluation, and audit tools is structural. Companies providing governance toolchains are worth watching.

4) Technology Roadmap: Coexistence of ‘Large + Small Models’ and Agentification

• Large-Scale Models
  Multimodality, real-time voice, tool usage, and enhanced memory/planning capabilities. API prices will stabilize downward while usage rates increase upward.

• Small Models/Edge
  SLM and on-device NPUs are rapidly spreading. With advantages in privacy, latency, and cost, edge inference is growing, leading to a hybrid architecture between central clusters and edge devices.

• Agents and Workflows
  Standard business automation (from research to summarization to drafting to review to purchase approval) is emerging with RAG, tool usage, and state management.
  “Evaluation, safety, and access control” are becoming essential layers for the operation of agents.

• Pressure from Open Source
  As the quality of open-source models catches up, competitive pressures push down the prices of commercial models. However, data, ecosystems, and integrated support still serve as barriers to entry.

5) The Truth of Monetization: Profitable Areas Despite Price Cuts

• Platform/Ecosystem
  Even though prices are falling, the lock-in effect on developers and high switching costs maintain durability. API revenues grow “through usage × natural expansion.”

• Enterprise
  Security, audit, access control, and SLA form the foundation of a premium B2B offering. Solutions that secure response quality and traceability in accountability differentiate themselves.

• Data Commercialization
  Legally licensed data, building domain knowledge bases, and synthetic data training services yield high margins. The focus of the “AI tax” is shifting from compute to data.

• Service/Agent Billing
  Task-based billing and fee structures linked to KPIs are rolling out. Performance-sharing contracts drive up profit margins.

6) Q4 Scenario-Based Investment Strategies (Focused on US Stocks)

• Scenario A: Soft Landing (Moderate Growth + Easing Inflation)
  If interest rate outlooks ease, long-duration growth stocks and AI infrastructure beneficiaries will continue. Prioritize data centers, networking, HBM, OSAT, and EDA.

• Scenario B: Reheating (Steady Growth + Inflation Resurgence)
  There is a risk of rising interest rates. Defensive plays in software and subscription stocks are preferred, with workflow-based SaaS using agent automation that delivers immediate ROI. Increase exposure to power, utilities, and energy efficiency stocks.

• Scenario C: Hard Landing (Growth Slows Down)
  Rebalance defensively towards sectors such as utilities and consumer staples with robust cash flows, including “pick and shovel” sectors (semiconductor equipment, parts, and power infrastructure). Consider higher cash positions, covered calls, and protective puts.

• Common Principles
  1) Diversify across the supply chain: GPU leaders, foundries, HBM, substrates/OSAT, networking, optical modules, EDA, equipment, power infrastructure, and data center REITs.
  2) Focus on applications offering “immediate financial impact” (e.g., call center automation, coding assist, sales/back-office automation, and biotech R&D).
  3) Adjust beta/duration according to macroeconomic indicators; fine-tune weightings sensitive to inflation and interest rate trajectories.

7) Sector/Stock Basket Ideas (Categorical Examples)

• Infrastructure Core
  GPU/accelerator leaders, HBM memory, foundries, OSAT/substrate, networking (switches/NICs), optical modules, and data center REITs.

• Pick-and-Shovel
  Semiconductor equipment (photolithography, etching, measurement), EDA, chemicals/gases, and test/inspection equipment.

• Power/Cooling/Grid
  Utilities, high-voltage cables, substations/switchgear, cooling solutions, and water treatment.

• Platform/Cloud
  Hyperscalers, collaboration tools/security/AI-Ops, and solutions for evaluation, audit, and access control.

• Applications
  Agent-based process automation, vertical SaaS with guaranteed demand (healthcare, legal, manufacturing, retail), and call center, sales, and development tools.

8) Checklist: Key Points to Monitor Over the Next 90 Days

• Updates on hyperscaler CapEx guidance and data center power contracts (PPA).
• Comments on HBM/CoWoS/OSAT capacity expansion/yield, and trends in optical module orders.
• API pricing policies and token usage growth rates, along with enterprise contract indicators.
• Power grid expansions, lead times for substations/high-voltage cables, and water resource issues.
• Model updates (multimodality, agent features) and on-device NPU adoption rates.
• Changes in regulations/antitrust, content licensing, and personal data guidelines.

9) Risk Management

• Prevent Overconcentration in a Single Factor: Diversify within the same theme across different parts of the value chain as well as market cap and regions.
• Stress Test for Macro Variables (Interest Rates, Inflation, Exchange Rates).
• Mitigate Volatility: Maintain cash positions, use covered calls/protective puts, and establish predefined stop-loss/rebalancing rules.
• Reduce Information Asymmetry: Monitor “power, packaging, and networking” keywords in quarterly earnings and conference calls.

10) Conclusion: OpenAI’s ‘Mega Bet’ Is the Catalyst for the Supercycle

• The GPU shortage is only the starting point; power, packaging, networking, cooling, and data contracts are the true bottlenecks and sources of alpha.
• Although price-cutting competition is inevitable, increased usage, lower per-unit costs, and a robust data moat will protect platform profits.
• In Q4, strategies driven by macroeconomic shifts (interest rate outlook and inflation) involving rebalancing and increased allocation to “hidden bottleneck” exposed assets such as power/grid and packaging/networking appear effective.

< Summary >

• OpenAI’s aggressive investment is a rational bet aimed at economies of scale, data moat, and expansion of the agent ecosystem.
• The true bottlenecks lie in power grids, HBM packaging, networking, cooling, and water resources – these are the sources of alpha.
• Even amid price competition, platform, enterprise, data licensing, and agent billing protect margins.
• Q4 strategy involves scenario-based rebalancing and supply chain diversification: infrastructure (power/packaging/networking) + pick-and-shovel + applications with immediate ROI.
• Checklist: Monitor updates on CapEx, PPAs, HBM/CoWoS, optical modules, NPUs, and regulatory changes.

[Related Articles…]

• Core Checklist for the AI Semiconductor Supercycle
• Interest Rate Outlook and a One-Stop Strategy for U.S. Stocks

*Source: [ 소수몽키 ]

– 오픈AI 무섭게 돈 뿌리고 다닌다? AI 슈퍼 사이클 만들까

---

---

● Fed Rate Cut, Historic Blunder, Stagflation Ahead

Why the Fed’s Rate Cut Is a ‘Historic Misjudgment’ and a 2025 Global Economic Checklist Covering Stagflation, the Dollar, Bonds, and AI Investments

This article contains three core elements.
First, it addresses the cost of compromising the Fed’s independence, which other YouTube channels and news outlets have overlooked, and the structural shift of reinstating the ‘long-term interest rate risk premium.’
Second, it outlines the triggering conditions and timing signals for the transition from mini-stagflation to ‘true’ stagflation.
Third, it explains how bond duration traps, the time gap between the short and long ends of the dollar, and bottlenecks in AI infrastructure power are reshaping asset allocation.
The global and macroeconomic trends are reorganized from an investment strategy perspective, allowing you to quickly grasp the key points.

1) Why This Rate Cut Is a ‘Historic Misjudgment’

The policy’s objective function is off track.
With the U.S. not currently experiencing a sharp decline in growth and core inflation still anchored above its target, lowering rates will stimulate demand and destabilize inflation expectations once again.
The divided Fed dot plot is not merely due to ambiguous short-term data, but because the inclusion of ‘political noise’ in the policy reaction function has raised market concerns.
When doubts about independence arise, the bond market protests with prices on long-term bonds rather than short-term ones.
The result is the recent bear steepening (a sharp rise in long-term rates with relatively stable short-term rates).
Paul Volcker’s lesson is simple.
When both cannot be controlled, inflation must be tackled first to restore real incomes, followed by employment recovery.
The current rate cut reverses this order.

2) Stagflation: The Path and Conditions from ‘Mini’ to ‘Severe’

This year, even though factors like falling oil prices and the normalization of supply chains were in our favor, the rigidity of downward price pressures remained notable.
Once fortune turns (with rising oil prices, geopolitical shocks, tariff hikes, or a rebound in rent), inflation will quickly reignite.
If policy loosening is hastened under the pretext of employment, price indicators could heat up with a 6–12 month lag.
The subsequent need for another rate hike—the ‘rebound hike’—will further damage real economic activity.
The pathway is as follows: mini-stagflation (modest growth + stubborn inflation) → policy easing → re-anchoring of inflation expectations → rising long-term yields and investment contraction → an increased risk of severe stagflation (low growth + high inflation).

3) Bond Market: Duration Traps and the ‘Supply + Independence Premium’

Long-term yields are influenced by three factors.
They are shaped by expectations of policy rates, inflation expectations, and the rarely discussed ‘intangible premium (compensation for independence, fiscal discipline, and liquidity).’
The expansion of issuance (fiscal deficits), constraints on primary dealers’ balance sheet capacity, and reduced central bank demand amidst the Fed’s tightening make it difficult for long-term bonds to trade at lower prices.
In this scenario, a rate cut sends a signal to long-term bonds that “inflation risk has increased,” which fuels steepening.
Therefore, buying ultra-long-term bonds at low prices remains risky.
Having cash assets, focusing on short- and medium-term bonds, and incorporating TIPS and floating rate instruments for inflation hedge serve as the foundational ‘duration defense.’

4) The Time Gap in the Dollar: Long-Term Strength, Short-Term Volatility

The dollar is the unrivaled long-term winner.
Europe’s political divisions, Japan’s persistent easing, and the vulnerabilities in some emerging markets structurally support the dollar’s reserve currency status.
However, in the short term, concerns over compromised Fed independence, fiscal uncertainties, and a sharp rise in long-term bond yields may temporarily shake confidence in the dollar.
Even within the broader scenario of a strong dollar, there are reasons to tactically increase allocations to gold and certain commodities on a quarterly or half-year basis.
In summary, the basic scenario for the dollar is ‘long-term strength with short-term volatility.’

5) Debt and Interest Costs: The Aftermath of Short-Term Issuance

Fiscal financing that relies on short-term issuance accelerates the pace at which interest costs reset.
If rollovers are concentrated during periods of high rates, interest expenses will increase exponentially.
Rising interest expenses weaken the fiscal ‘automatic stabilizer’ and encroach on long-term investments (infrastructure, R&D).
Ultimately, this erodes potential growth and forces the market to demand an additional premium.
This could lead to a ‘new regime’ where the level of long-term yields is driven upward.

6) The AI Trend: A Supercycle with Power, Grid, and Cooling Bottlenecks

AI simultaneously brings hope and risks to the global economy.
While demand for servers, GPUs, HBM memory, and optical networking is structurally robust, the bottleneck lies in power.
A rapid surge in power demand for data centers, delays in upgrading transmission and distribution networks, prolonged delivery times for substation equipment, and rising costs for cooling infrastructure extend project timelines and increase capital expenditures.
In an environment of high rates and volatile long-term yields, companies in the ‘AI value chain’ that generate fast cash flows and those in power, utilities, LNG, and essential equipment sectors benefiting from regulatory support hold relatively favorable positions.
Even among AI beneficiary stocks, those with long-duration, far-future cash flows (often characterized by expanding deficits) are prone to multiple compression.

7) Checklist for Korean Investors: Exchange Rates, Semiconductors, Commodities

The Korean won and U.S. dollar exchange rate will likely follow short-term dollar volatility while maintaining upper-end rigidity in the face of long-term dollar strength.
Semiconductors are structurally strong due to AI demand, but valuation volatility can widen during periods of sharp rises in long-term yields.
Among commodities, intersecting items such as gold, uranium, and copper—key to both AI infrastructure and energy transition—offer a middle ground between defense and growth.
For bonds, a focus on shorter and medium durations is advisable, while equities should concentrate on companies with robust cash flow generation and pricing power to serve as core positions.

8) Portfolio Strategy: A Framework to Apply Immediately

Maintain a 20–30% allocation in cash and ultra-short-term bonds as a volatility buffer.
For government and corporate bonds, allocate 30–40% in short- and medium-term bonds, and incorporate TIPS and floating rate instruments to guard against inflation surges.
Equities should form 30–40% of the portfolio, selected as baskets focused on key AI components (semiconductor memory/HBM, networking), power and grid, and high dividend/high cash flow companies.
Diversify with 5–10% in tangible assets such as gold to hedge against periods of weakened currency confidence.
Geographically, build a concentrated yet diversified portfolio centered on the U.S. core, supplemented by AI value chains in Korea and Taiwan, and a few countries with high fiscal and monetary policy credibility.

9) Risk Triggers and Signals: “When to Change Course”

If inflation expectations begin to rise simultaneously in the 5-year and 10-year BEI, further reduce duration.
If the steepening of the gap between the 10-year and 2-year U.S. Treasury yields accelerates, resist the temptation to buy ultra-long-term bonds.
If oil prices exceed $90 and remain there for three months, increase the probability of stagflation.
If the unemployment rate rises without a corresponding drop in wage growth, interpret it as a ‘bad slowdown’ and increase defensive allocations.
Once the Fed’s communications restore the ‘price priority’ phrase, that is the time to gradually increase duration.

10) One-Sentence Conclusion and Action

We are at the very beginning of a policy dilemma where the short-term sweetness of a rate cut will incur long-term costs.
Given the rising risk of stagflation, it is advantageous to maintain a defensive stance with a duration defense, cash flow-oriented equities, gold, and power infrastructure, and to delay large bets until the Fed reverts back to a ‘price priority’ stance.
Under the premise of a strong dollar in the long term with short-term fluctuations, hedge currency exposure and diversify risks.

What Makes This Content ‘Different’

First, it does not merely focus on the direction of rate cuts but identifies the structural factors that drive up long-term rates through a ‘Fed Independence Risk Premium.’
Second, it takes into account the amplification of steepening due to market microstructures such as issuance structures and constraints on dealer balance sheets.
Third, it translates the AI supercycle from a simple theme into an investment portfolio framework by linking it with power/grid bottlenecks and capital cost variables.

< Summary >The global economy can shift from mini-stagflation to severe stagflation if policy missteps occur.Concerns over compromised Fed independence heighten the long-term rate premium, expanding the duration risk of bonds.The dollar is viewed as strong in the long term with short-term volatility; hence, defense is achieved through allocations in gold, short- to intermediate-term bonds, and cash.AI should be approached as a physical investment cycle coupled with power and grid bottlenecks, with a priority on companies with robust cash flows.Positioning should center on cash/ultra-short-term assets, short- to intermediate-term bonds plus TIPS, and cash flow-oriented equities and gold, while duration is increased only when the Fed signals a reversion to ‘price priority.’

SEO Keyword Note

This article naturally incorporates key terms such as global economy, macroeconomics, interest rates, inflation, and the dollar.

[Related Articles…]
A Comprehensive Overview of U.S. Interest Rates and Stagflation Risk
Key Points on AI Infrastructure Power Bottlenecks and Investment Strategy

*Source: [ 경제 읽어주는 남자(김광석TV) ]

– 연준의 금리 인하 “이건 역사적 오판이다”, 미국 경제 무너트리고 글로벌 시장 마저 뒤흔든다 | 김광석의 콜라보 – 경제포차 홍춘욱 1편

---

---

● AI Bubble or Power Supercycle – 2025 Playbook

AI Bubble Controversy: Bubble or Infrastructure Transformation — 2025 Investment Strategy and Risk Checklist

This article contains the real core details that most YouTube videos and news outlets skip.
The entire analysis covers, in one sweep, the single variable that determines the AI bubble—the bottleneck of power and transmission infrastructure, the cost curve per token and the profitability threshold, the sustainability of hyperscaler CAPEX, the impact of the dollar cycle and interest rate scenarios on the stock market, and the composition of a practical portfolio connecting the KOSPI with the global economic outlook.
In particular, it breaks down into numerical frameworks the actual lead times for data center power intake, transformer, packaging, and HBM supply, as well as the pathway through which GPU pricing and utilization rates transfer to corporate earnings, making it immediately usable.

1) The Essence of the Issue: It’s Not about Price, It’s about Physical Constraints

What’s most underestimated in the bubble debate are the physical infrastructures like power, cooling, and optical transmission.
It isn’t simply a matter of buying more GPUs; the capacity for power intake, transformers, switchgear, optical cables, and the heat density per server rack are creating bottlenecks.
Power intake contracts and transformer manufacturing typically require 18 to 30 months of lead time, which is slower than hardware shipments.
This means there may be a period when revenue recognition lags behind demand.
The primary factor in AI investment decisions is not valuation, but determining when, where, and how much power will be connected.

2) The Realities through the Cost Curve: Revenue Only Follows Once the Cost per Token Drops

AI investment performance is ultimately determined by the spread between the cost per token and unit revenue.
The simple formula is as follows: Token Unit Price = (GPU Depreciation + Power Cost + Data Center OPEX) ÷ (Tokens per Second × Utilization Rate × Lifespan in Months).
The biggest leverage comes from the utilization rate, memory bandwidth, and power cost.
If the utilization rate increases from 40% to 70%, the cost structurally drops, and if tokens per second increase through model and kernel optimization, the cost decreases even further.
Conversely, if the power cost is high or if there is frequent RAG and memory access, costs can spike.
The investment point is to bet on the pace of the downward movement of the cost curve rather than technological optimism.
In short, companies that improve compilers, kernels, compression, pruning, low-bit quantization, and cache hit rates to cut costs are the real beneficiaries.

3) Will CAPEX Continue: The Three “Total Demand Functions” of Hyperscalers

The sustainability of hyperscaler CAPEX can be judged on three aspects.
First, are there any signals of ARPU (average revenue per user) growth in consumer products?
If AI functionalities in search, shopping, and cloud SaaS consistently show data that positively affects conversion rates, shopping cart sizes, and bounce rates, CAPEX will be maintained.
Second, the pace of redeployment in enterprise workloads.
If copilots, summarization, code generation, and chatbots move from on-premise setups to the cloud, infrastructure expansions will follow with multi-year contracts.
Third, energy acquisition.
Securing power via PPAs, small modular reactors (SMR), and early adoption of renewable energy plus storage systems (ESS) serves as tangible collateral for sustaining CAPEX.
If these three do not align simultaneously, CAPEX may become conservative when the macro environment falters.

4) Bubble Check Framework: Six Indicators to Quickly Gauge the Situation Numerically

Quickly check the revenue-to-stock price gap using the following indicators:
– Order Backlog/Sales Ratio (BB ratio): Over 1.3 indicates overheating, while around 1.0 suggests a normalization phase.
– Packaging/High Bandwidth Memory (HBM) Lead Time: A drop below 9 months signals that supply is catching up with demand.
– GPU Resale Premium: If the used market premium converges to 0, it indicates a cooling of the bubble.
– Number/Capacity of Approved Power Intakes: A surge in new approvals by region predicts revenue elasticity in 12 to 24 months.
– Utilization: If the paid compute share over total available cloud compute exceeds 60%, revenue leverage is possible without lowering unit prices.
– Model Cost vs. ARPU: For enterprises, if the ratio of revenue per token to cost is less than 1.5 times, it is not scalable.

5) 2025~2027 Supply Timeline: Where Will the Strain Occur?

Advanced packaging such as HBM and CoWoS will see capacity expansions in 2025–2026, but power, transformers, and medium-voltage switchgear will lag behind.
Networking components (optical transceivers, switching ASICs) will continue to experience strong demand, and the transition from 800G to 1.6T around 2026 is a key point to watch.
Even if supply increases in advanced foundry processes (N3/N2), if packaging, memory, and power bottlenecks do not accompany it, the actual usage increase could be limited.
In conclusion, the true key for 2025 is not semiconductors but rather power and cooling.

6) Macro Variables: The Turbulence Caused by Interest Rates, the Dollar, and Inflation

Growth stocks’ valuations are highly sensitive to discount rates.
If interest rates rise again, the present value of distant future cash flows will be reduced, causing multiples to contract.
A strong dollar negatively affects emerging market risk assets, and it also burdens the cost structure of Korean companies with a high import share of AI infrastructure.
If inflation re-heats energy and wages, rising power costs will directly pressure the margins of AI services.
Therefore, even for AI investments, one must simultaneously consider the global economic outlook, interest rates, the dollar, and inflation.

7) Korean Investment Landscape: The Triangular Formation of “Memory-Power-Equipment”

Memory (HBM) offers structural benefits, but when ASPs and mixes are near their peak, momentum volatility increases.
The key lies in the actual conversion rate of order backlogs to revenue and diversification of clientele.
Power and electrical equipment are undervalued themes.
Companies providing transformers, cables, switching devices, and data center power solutions have long-term demand pipelines and conservative contract structures, ensuring high visibility of earnings.
Equipment and parts (packaging materials, substrates, advanced testing) need to align with expansion timing and customer capacity transitions to be optimal.
Application software is more favorably approached via an index-based strategy until a killer use case is confirmed, with individual stocks analyzed based on unit prices, channels, and retention data.

8) Portfolio Strategy: A Barbell Structure Utilizing the Bubble Debate

Core: Diversify into “power, networking, and memory” within the AI infrastructure and assess valuations based on the speed of conversion from sales to cash flow.
Defensive: Reduce volatility with utilities and consumer staples that have stable cash flows and dividends.
Optional: Capture the application and agent ecosystems with either index-based or small positions initially, increasing exposure once the winners are confirmed.
Hedge: Manage risks with index put spreads at volatility lows, dollar strength hedges, and gold diversification.
Rebalancing: Reduce exposure to growth when GPU resale premiums, power intake approval speeds, and HBM lead times slow, and increase when they accelerate.

9) Key Facts Rarely Covered in the News

– The power grid is the limiting factor for AI growth.
– Compiler, runtime, and cache optimizations reduce costs significantly more than the race for model size.
– Latency in transmission lines, optical modules, and switching fabrics has a direct impact on user experience and TCO.
– The setup of depreciation periods has a large influence on the margins of cloud service providers.
– Data quality and governance are bottlenecks for enterprise adoption, which in turn drives consulting and tooling demand.

10) Practical Checklist: Before Buying/Selling

– Customer Mix: Check the revenue proportion and contract durations of the top three customers.
– Utilization: Look for disclosures or comments on the average utilization rates of data centers and GPU clusters.
– Power: Verify the schedule of new facility power intakes and PPA contracts.
– Lead Time: Monitor trends in lead times for HBM, CoWoS, and optical transceivers.
– Price: Observe the direction of ASPs for GPUs and HBM, as well as the used market premium.
– Unit Price: Review the roadmap for reducing cost per token (compression, quantization, memory architecture).
– Cash Flow: Assess the speed of conversion from CAPEX to OCF.
– Regulation: Understand the impact range of data sovereignty and privacy regulation issues.

11) Conclusion: Bet on Speed Rather than the “Bubble”

AI is not a theme but an infrastructure cycle.
Bubbles may exist, but the factors that determine speed are power, networking, memory, and software optimization.
When macro volatility (interest rates, the dollar, inflation) shakes the market, it is rational to reduce exposure, but once evidence of a declining cost curve accumulates, increasing exposure is justified.
In 2025, the direction of stock prices is likely to be influenced more by power than semiconductor expansions.
Remember that the AI investment strategy ultimately depends on the function of physical constraints and the cost curve.

< Summary >

The core elements are power and transmission bottlenecks and the cost curve per token.
The sustainability of CAPEX is determined by ARPU, enterprise transition, and power acquisition.
Quickly check the bubble by reviewing order backlog, lead times, used market premiums, utilization, and the ARPU-to-cost ratio.
The portfolio should combine an AI infrastructure core with defensive and optional barbell components, along with hedging, while always taking into account macro factors (interest rates, the dollar, inflation).
The decisive factor in 2025 will be power rather than semiconductor capacity expansion.

[Related Articles…]

AI Power Bottleneck and Data Center Investment Map 2025

Report on the Correlation between a Strong Dollar and the KOSPI

*Source: [ Jun’s economy lab ]

– AI 버블논란에 대한 저의 투자 의견

---

---