Overview
AI investing has moved beyond the first-stage question of whether a company has an AI narrative. The more important question is now which companies can convert exponential growth in token consumption into sustainable profits. Over the past three years, markets first priced the scarcity of GPUs, high-bandwidth memory, networking equipment, servers and data-center capacity. They then shifted toward the capability gains delivered by frontier models, reasoning systems and coding agents. In 2026, the variables driving equity returns are shifting away from model parameter counts and the scale of capital expenditure toward token production cost, task-completion reliability, usage intensity, enterprise-workflow penetration and the ability of enormous AI investments to generate durable free cash flow.
The AI industry and AI equities are not currently at the same point in their respective cycles. The technological diffusion of large models, agents and multimodal systems remains early, while enterprise adoption is in an early-to-middle phase. Hyperscaler capital expenditure, however, has entered a late-stage acceleration phase, and equity valuations, concentration and investor sentiment are closer to the second half of a bull market. J.P. Morgan Asset Management estimates that five U.S. hyperscalers will spend approximately $697 billion in 2026, with capital expenditure rising from roughly 33% of their operating cash flow in 2023 to an estimated 93% in 2026. At the same time, the S&P 500 had risen approximately 14% in 2026 and was again trading near record levels in mid-August.
The correct question is therefore not whether AI is a bubble in its entirety, but which layers of the AI value chain have become overcapitalized or overpriced. Cloud revenue, coding-agent adoption, semiconductor sales and enterprise demand are real. However, capital expenditure, external financing, private-model valuations, data-center projects and several high-multiple second-tier equities are displaying increasingly speculative characteristics.
The appropriate strategy is not to abandon AI. It is to move away from undifferentiated AI beta and toward businesses with real cash flows, resilient balance sheets, proprietary context, distribution and control over the cost of intelligence.
At current prices and at this stage of the cycle, we believe Alphabet, Meta, TSMC, NVIDIA, Microsoft and Amazon offer the most attractive combinations of valuation, business quality and upside-to-downside asymmetry. Alphabet offers the strongest overall risk-reward. Microsoft offers the highest fundamental probability of success, although its potential for multiple expansion is more limited. Meta has an attractive relationship between headline valuation and revenue growth but faces significant capital-spending risk. TSMC is the highest-quality picks-and-shovels exposure, subject to geopolitical tail risk. NVIDIA remains the most attractive pure AI semiconductor stock on a growth-adjusted basis. Amazon provides considerable AWS and Trainium optionality, although its reported earnings materially overstate its normalized operating valuation.
This report was prepared by HTX Research, the dedicated research arm of HTX. HTX Research tracks structural changes across global capital markets, with particular attention to how technology cycles, capital expenditure cycles and asset pricing interact. Drawing on public financial disclosures, industry data and observable operating indicators, this report assesses the valuation state and risk-reward structure of each layer of the AI value chain, framing its conclusions around cycle position and asymmetry rather than directional price forecasts.
1. The Macro Structure of the Model Industry and Token Economics
1.1 AI Investing Is Moving from Compute Scarcity to Token Value Capture
1.1.1 Three Stages Require Three Different Investment Frameworks
The U.S. AI trade can be divided into three stages. The first was the compute-scarcity trade, centered on GPUs, HBM, switching, servers, power and data centers. The second was the model-capability trade, driven by scaling laws, reasoning systems, test-time compute, multimodal models and coding agents.
The third stage is the token value-capture trade now beginning to dominate investor attention. Markets must determine who can produce high-quality tokens at lower cost, who owns valuable context, who can embed models into real workflows and who can convert rapidly growing AI revenue into free cash flow.
The AI value chain can be expressed as:
Power and Data Centers → Chips and Networking → Cloud → Foundation Models → Context and Agents → Products and Workflows → Users → Additional Token Consumption.
The central investment question is no longer simply whether token volumes will grow. It is about which layer will retain pricing power, scale economies, customer lock-in and attractive returns on capital.
1.2 The First Principle of AI Is Computation, Not Connectivity
The internet created value primarily through connectivity. Search connected users with information, social networks connected people with one another, and e-commerce connected products with consumers.
AI has a different foundation. Models use chips, electricity, data and algorithms to manufacture intelligence and distribute it through tokens. The economic process can be expressed as:
Compute → Intelligence → Token → Action → Economic Value.
Tokens can be evaluated based on quality, cost and speed. Quality refers to complex reasoning, coding, mathematics, multimodal understanding, tool use and long-horizon reliability. Cost should include not only nominal token prices, but also context processing, retries, failures and human review. Speed includes latency, throughput and useful work completed per unit of time.
When capability differences are large, quality dominates because small differences in step-level reliability compound across long agent workflows. As models become more interchangeable, cost and speed gain importance. Enterprise customers ultimately buy successful outcomes per dollar, not benchmark scores.
1.3 Token Demand Is Shifting from User Growth to Task Intensity
There is no unified database covering token usage across all model providers, clouds and private deployments. Claims that global token demand is growing by nearly tenfold year over year should therefore not be treated as audited market data. OpenRouter nevertheless provides a useful cross-model window. Its 2026 research sample covered more than 450 trillion tokens through mid-June, while its earlier State of AI study analyzed more than 100 trillion tokens of real-world usage.
A conventional chatbot interaction often consists of one prompt and one response. An agent may plan a task, retrieve information, read files, call tools, execute code, inspect results, repair errors and validate the output. A single agentic assignment can therefore generate dozens or hundreds of model calls.
Coding agents have already validated this pattern. Anthropic reported in February 2026 that Claude Code had exceeded a $2.5 billion revenue run rate, having doubled from the beginning of the year and generating more than half of its revenue from enterprise customers. OpenAI reported in June that Codex had more than five million weekly active users, with non-developers representing approximately 20% of the user base.
The most informative demand metrics will consequently shift away from monthly active users alone and toward tokens per user per day, concurrent agents per user, average task duration, cost per successful task and human work hours replaced by machines.
1.4 Four Different AI Cycles
1.4.1 The Technology Cycle Remains Early
From a technology and product-diffusion perspective, AI is far from maturity. Chatbots were the first major product form, while coding agents are only beginning to demonstrate the transition from answering questions to completing work. Full-duplex voice, multimodal understanding and generation, browser operation, enterprise agents and long-horizon systems remain underdeveloped.
A large correction in AI equities would not necessarily invalidate the industry thesis. Technological revolutions and equity returns often move on different timelines. The internet ultimately transformed the economy, but the Nasdaq still experienced a long valuation reset after 2000.
1.4.2 The Application Cycle Is in Its Early-to-Middle Phase
Enterprises are moving from purchasing AI seats to embedding agents into workflows. OpenAI’s enterprise data show that Codex usage in legal, sales, recruiting and marketing has recently grown faster than engineering usage, indicating that agent adoption is expanding beyond software development.
Most enterprises, however, have not yet rebuilt their data access, permissions, workflows and organizational structures around AI. Large-scale application revenue and productivity gains may therefore lag current infrastructure investment by several quarters or years.
1.4.3 The Capital-Expenditure Cycle Is in a Later Acceleration Phase
Five major U.S. hyperscalers are expected to spend approximately $697 billion in 2026, which is $173 billion above estimates made at the beginning of the year. Capital expenditure has risen to an estimated 93% of operating cash flow, and markets increasingly reward higher spending only when it is accompanied by higher revenue expectations.
This is no longer the beginning of the AI infrastructure cycle. Total spending may continue to rise, but the rate of spending growth, financing quality, cloud utilization and return on invested capital have become decisive valuation variables.
1.4.4 The Equity and Sentiment Cycle Is in the Second Half of a Bull Market
By mid-August 2026, the S&P 500 had gained approximately 14% for the year and was again near record highs. J.P. Morgan raised its year-end target to 8,000 but maintained a roughly 20-times forward multiple assumption because of high interest rates, geopolitical risks and rising debt and equity issuance.
The current environment is therefore best described as a later-stage pricing phase within a long-term structural bull market, rather than a final top that can already be dated with confidence. Elevated volatility, widening dispersion and repeated valuation compression are more likely than the sudden disappearance of AI demand.
2. Token Business Models and Foundation-Model Competition
2.1 The Market Is Splitting into Premium Tokens and Value Tokens
Premium tokens derive their value from higher success rates on difficult tasks, reduced human-review requirements and superior coding, tool-use and long-context capabilities. OpenAI, Anthropic and Google remain the most important frontier providers.
Value tokens target standardized, price-sensitive workloads. DeepSeek, Zhipu, Qwen and other open systems can gain share through lower prices, higher throughput, private deployment and local ecosystems. OpenRouter’s data show that low-cost open models have become real production substitutes rather than merely benchmark demonstrations.
While scaling remains rapid and capability gaps remain wide, frontier tokens retain meaningful pricing power. As models converge, performance per dollar, latency and deployment cost become more important.
2.2 Raw Tokens Resemble Electricity
Value-token markets naturally trend toward intense competition. Architectures converge, capabilities spread through distillation, synthetic data, fine-tuning and open weights, and customers can route workloads among several providers.
Frontier tokens are economically stronger but do not possess permanent pricing power. Leading labs face direct competition from one another, substitution from lower-cost systems and enormous expenses related to training, inference, infrastructure and talent.
Tokens increasingly resemble electricity. Electrification transformed the economy, but value did not remain confined to electricity producers. It spread into equipment, manufacturing, applications and new business models. Models may similarly become general-purpose production inputs, with economic value flowing into chips, clouds, proprietary data, workflows, products and distribution.
The long-term winner is more likely to control:
Model × Compute × Context × Product × Distribution.
2.3 Anthropic and OpenAI: Exceptional Growth Still Requires Valuation Discipline
Anthropic has become a defining company in coding agents. Claude Code had exceeded a $2.5 billion revenue run rate by early 2026, with weekly users doubling and enterprise subscriptions expanding rapidly.
The user-supplied estimate that Anthropic exceeded $69 billion of ARR by the end of June 2026 is not supported by a sufficiently authoritative company disclosure and should be treated as a third-party estimate rather than an audited fact. A scenario in which Anthropic reaches $300 billion of ARR by the end of 2027 and receives a ten-times revenue multiple should be classified as a strong bull case rather than a base case.
OpenAI has a stronger consumer gateway and broader product distribution, but it must also demonstrate that rapid revenue growth can translate into sustainable inference margins and free cash flow. Private model companies should be valued not only on ARR, but on customer concentration, cloud-provider bargaining power, infrastructure obligations, dilution and retention.
3. Coding Agents, Context and the Next Product Form
3.1 Coding Agents Are the Second Major AI Application After Chatbots
Coding combines high-value labor, strong feedback and dense context. Code can be compiled, tested and executed, giving models clearer feedback than open-ended content generation. Repositories also provide files, dependencies, historical changes, tests and development standards.
A coding agent can therefore read a project, plan work, modify files, run tests, diagnose failures and iterate. Its token intensity is substantially higher than that of a conventional chatbot, and the same operating model is spreading into research, finance, legal work, sales, recruiting and operations.
3.2 Context Is More Defensible Than a Standalone Harness
A harness connects models with tools, terminals, files and interfaces. It is important, but relatively replicable in isolation. As model providers add native file access, browser control, code execution, parallel agents and long-running tasks, standalone harness products risk being compressed.
Context is more difficult to reproduce. It includes enterprise code, documents, email, meetings, customer records, permissions, user history and business processes. Models can be switched; deeply integrated context is costly to migrate.
The most durable application moat is therefore more likely to be:
Proprietary Context + Deep Workflow Integration + Distribution + Continuous Feedback Data.
3.3 The Next Product Form Will Move from Answers to Persistent Execution
The next generation will center on full-duplex, long-horizon and multimodal agents. The defining characteristic of a long-horizon agent is not a longer answer, but a longer period of autonomous operation.
Moving from seconds to minutes, hours and days creates more model calls, tool actions and feedback loops. This is why slower chatbot-user growth does not imply that token demand has peaked. The ultimate market may extend beyond software spending into global knowledge labor itself.
4. Core U.S. Platform and Cloud Assets
The latest reported year-on-year growth rates were approximately 93% for Oracle IaaS, 82% for Google Cloud, 43% for Azure and 37% for AWS. Fiscal periods and segment definitions differ, but the common conclusion is that AI-led cloud demand remains strong.
4.1 Alphabet: The Best Overall Risk-Reward
Alphabet controls TPU, Google Cloud, Gemini, DeepMind, Search, Workspace, Android, Chrome and YouTube, creating the most complete public full-stack AI architecture.
In the second quarter of 2026, Alphabet generated $119.8 billion of revenue, up 24%. Google Cloud revenue rose 82% to $24.8 billion, Cloud operating margin expanded from 20.7% to 35.6%, and Cloud backlog reached $514 billion.
Alphabet’s headline P/E should not be used without adjustment. At the August 14 close, GOOGL traded at approximately $345.90 with a headline trailing P/E of 17.4 times. However, second-quarter other income was $98 billion, primarily from unrealized gains on equity securities, materially inflating net income.
A rough normalization based on operating profit and a standard tax rate suggests that Alphabet’s core operating valuation is closer to the low-to-mid 30-time range than 17 times. It is not an absolute deep-value stock, but its combination of Cloud growth, Search cash flow, TPU commercialization, Gemini optionality and balance-sheet resources provides the strongest overall asymmetry among the mega-cap platforms.
The principal risk is capital intensity. Alphabet raised its 2026 capital-expenditure guidance to $195 billion–$205 billion, generated negative $5.9 billion of quarterly free cash flow and has materially increased debt.
4.2 Microsoft: The Highest Probability of Success, but Not the Highest Upside
Microsoft controls Office, GitHub, Teams, Windows, Dynamics, Azure and enterprise identity, making it the strongest business-to-business distribution and context platform.
In the fiscal fourth-quarter of 2026, Azure annual revenue exceeded $100 billion and Microsoft 365 Copilot passed thirty million paid seats. Quarterly capital expenditure was approximately $41 billion, but free cash flow remained $19.6 billion.
At the August 14 close, Microsoft traded at approximately $495.40 and 29.5 times trailing earnings.
Microsoft’s advantages are recurring cash flow, enterprise lock-in and business-model stability. It probably offers the highest fundamental probability of success among the core AI platforms. However, the market already awards it a substantial quality premium, and Copilot must still prove deep usage and usage-based monetization. Microsoft is therefore a defensive core holding rather than the highest-upside opportunity.
4.3 Amazon: High Optionality, but Distorted Headline Valuation
AWS generated $42.2 billion of second-quarter revenue, up 37%, and $16.6 billion of operating income.
Amazon traded at approximately $262.65 and a headline trailing P/E of 21.1 times at the August 14 close. However, second-quarter net income included approximately $53.4 billion of non-operating income, primarily related to Anthropic, making the headline multiple unsuitable for normalized valuation.
The investment thesis is not a low headline P/E. It is renewed AWS acceleration, Trainium’s potential to reduce inference costs, and the strategic value of a model-neutral cloud platform. Amazon offers more upside optionality than Microsoft, but less near-term free-cash-flow visibility.
4.4 Meta: The Most Attractive Conventional Earnings Valuation
Meta traded at approximately $589.85 and 22.2 times trailing earnings at the August 14 close. Second-quarter revenue rose 28% to $60.8 billion, but operating income declined 8%, operating margin fell from 43% to 31%, capital expenditure reached approximately $31.1 billion and free cash flow was only $784 million.
Meta’s advantages are advertising efficiency, the global social graph and consumer distribution. Unlike Alphabet and Amazon, its headline P/E is not distorted to the same extent by enormous unrealized investment gains.
If AI continues to improve ad conversion and engagement while WhatsApp, Threads, smart glasses and personal agents become monetizable, Meta offers significant earnings upside. Its principal risk is that capital expenditure continues to rise without a clearly visible external AI-revenue stream.
Meta has one of the most attractive relationships between conventional earnings valuation and growth, but also one of the largest capital-expenditure tail risks.
4.5 Oracle: High Payout, Lower Probability
Oracle’s fiscal fourth-quarter IaaS revenue grew 93%, while remaining performance obligations reached $638 billion, up 363%. Fiscal-year free cash flow, however, was negative $23.7 billion.
Oracle traded at approximately $150.52 and 27 times trailing earnings at the August 14 close.
If backlog converts smoothly and customer prepayments reduce capital needs, Oracle could generate substantial operating leverage. In a credit-tightening or AI-capex downturn, however, its financing, free-cash-flow and customer-concentration risks are materially greater than those of Alphabet, Microsoft and Amazon.
Oracle is a high-payout, lower-probability asset rather than a defensive late-cycle core holding.
5. Compute, Custom Silicon and AI Memory
5.1 NVIDIA: The Preferred Semiconductor on a Growth-Adjusted Basis
NVIDIA generated $81.6 billion of fiscal first-quarter 2027 revenue, up 85%, while Data Center revenue rose 92% to $75.2 billion and GAAP gross margin reached 74.9%.
At the August 14 close, NVIDIA traded at approximately $225.16 and 34.3 times trailing earnings.
A 34-times multiple is not low in absolute terms, but remains attractive relative to 85% revenue growth, 75% gross margin and the combined moat of CUDA, GPUs, networking and rack-scale systems.
NVIDIA is more sensitive than Alphabet or Microsoft to a peak in hyperscaler spending. If token demand continues to grow faster than unit compute cost declines, earnings can expand further. If capex decelerates, both earnings expectations and valuation could reset simultaneously.
5.2 Broadcom: Exceptional Quality, but Less Attractive Entry Asymmetry
Broadcom generated $10.8 billion of AI semiconductor revenue in fiscal second-quarter 2026, up 143%, and $10.26 billion of quarterly free cash flow. It guided to approximately $16 billion of AI semiconductor revenue in the following quarter, representing growth of more than 200%.
Broadcom is one of the most direct beneficiaries of custom accelerators and AI networking. Its headline GAAP P/E of nearly 98 times is distorted by VMware-related amortization and is not directly comparable to other companies. Even after adjustment, however, expectations are demanding.
Broadcom is a high-quality company, but it is more attractive after a valuation or price reset than as the leading value opportunity today.
5.3 TSMC: The Highest-Quality Picks-and-Shovels Exposure
TSMC generated $40.2 billion of second-quarter revenue, a 67.7% gross margin and a 60.3% operating margin. Third-quarter revenue guidance was $44.6 billion–$45.8 billion.
TSMC does not need to identify whether OpenAI, Anthropic, Google, NVIDIA or AMD ultimately wins. As long as demand for advanced nodes, packaging and AI/HPC compute expands, it participates in industry-wide value creation.
TSM’s ADR traded at approximately $426.35 at the August 14 close.
Its advantages are manufacturing leadership, customer trust, profitability and balance-sheet quality. Its principal risk is geopolitical concentration in Taiwan, together with overseas-fab costs, customer concentration and capital intensity.
For investors willing to accept geopolitical tail risk, TSMC offers one of the most balanced combinations of operating probability and upside across the AI supply chain.
5.4 AMD, Micron, Arista and Vertiv: Strong Businesses Do Not Necessarily Offer Strong Entry Odds
AMD generated $11.5 billion of second-quarter revenue and more than 100% growth in Data Center, but traded at approximately 132 times trailing earnings at the August 14 close.
AMD must continue gaining substantial accelerator share and improving its software ecosystem to justify the current valuation. The business trend is strong, but the price requires near-perfect execution.
Micron generated $41.46 billion of fiscal third-quarter revenue, $18.3 billion of free cash flow and more than $25 billion of data-center revenue.
Its headline trailing P/E was approximately 22 times.
Memory remains highly cyclical. A low P/E may indicate peak earnings rather than genuine undervaluation. Micron is best treated as a tactical HBM-cycle exposure rather than a perpetual compounder.
Arista grew second-quarter revenue by 37.7% and generated a 45.4% GAAP operating margin, but traded at approximately 61.9 times trailing earnings. Vertiv grew revenue 24% and adjusted free cash flow 234%, but traded at approximately 66.5 times trailing earnings.
6. AI Capital Expenditure, Bubble Risk and Equity Asymmetry
6.1 The Market Is Moving from Revenue Growth to Return on Capital
Alphabet, Microsoft and Meta spent approximately $44.9 billion, $41 billion and $31.1 billion respectively in their latest quarters. Microsoft still generated $19.6 billion of free cash flow, while Alphabet produced negative $5.9 billion and Meta approximately $784 million.
Markets will not reward capex growth indefinitely. AI investment must eventually be monetized through cloud revenue, agent subscriptions, advertising efficiency, custom-silicon savings and usage-based workflows.
The key metrics are becoming AI revenue relative to AI capex, cloud gross margin, inference utilization, successful tasks per dollar and free-cash-flow conversion.
6.2 AI Is Not a False Industry, but Its Financial Architecture Is Becoming Speculative
The principal difference from the 2000 bubble is that today’s leaders generate real revenue, profits and cash flow. Google Cloud, Azure, AWS, NVIDIA and Broadcom are producing genuine growth.
A bubble, however, does not require the underlying technology to be false. It occurs when capital supply grows faster than the cash flows that projects can ultimately generate. At least seventy-five U.S. data-center projects worth approximately $130 billion faced local opposition in the first quarter of 2026, while hyperscaler spending and external financing continued to expand.
The most speculative areas are therefore more likely to be private-model valuations, leveraged data centers, projects without long-term customers, second-tier AI accelerators and agent harnesses without proprietary context or distribution.
6.3 Headline P/E Ratios Do Not Represent True Relative Value
Alphabet and Amazon earnings are inflated by unrealized investment gains. Broadcom’s GAAP earnings are reduced by acquisition-related amortization. Micron may be near a memory-cycle earnings peak. AMD, Arista and Vertiv embed substantial future-growth expectations.
AI equities must therefore be valued through normalized operating profit, free cash flow, capital expenditure and cycle position—not by a single headline P/E table.
6.4 Which Stocks Offer the Best Value and Risk-Reward Today?
6.4.1 First Choice: Alphabet
Alphabet is not the deep-value stock implied by a 17-time headline P/E, but its normalized valuation, Cloud growth, Search cash flow, TPU, Gemini, distribution and balance-sheet resources form the strongest overall package.
Upside would come from sustained Cloud growth, external TPU commercialization, improved Gemini product positioning and a reclassification from “Search disruption victim” to “full-stack AI platform.” Search cash flow and diversified businesses offer partial downside protection.
Alphabet currently provides the best balance between the probability of success and upside potential.
6.4.2 Second Choice: Meta
Meta’s roughly 22-time earnings multiple against 28% revenue growth provides an attractive conventional valuation framework. Upside comes from advertising efficiency, WhatsApp, Threads, smart glasses and personal agents. Downside comes primarily from capital expenditure and margin pressure.
Meta offers more upside than Microsoft, but with greater fundamental volatility.
6.4.3 Third Choice: TSMC
TSMC benefits from nearly every winning AI architecture through advanced manufacturing and packaging. It offers a higher operating probability than most chip designers because it does not depend on one model or accelerator architecture.
Its principal risk is geopolitical rather than competitive. For investors willing to bear that tail risk, TSMC offers one of the best supply-chain asymmetries.
6.4.4 Fourth Choice: NVIDIA
NVIDIA’s valuation is high in absolute terms but attractive relative to current growth and profitability. It offers greater upside than Microsoft or TSMC, but greater sensitivity to capital-expenditure deceleration and expectations.
NVIDIA is more attractive on market or earnings-driven pullbacks than after periods of accelerated price momentum.
6.4.5 Fifth Choice: Microsoft
Microsoft offers the highest fundamental probability of success and the strongest enterprise distribution. Its returns are more likely to come from earnings compounding than from major multiple expansion.
It is the best defensive AI core holding, but not the highest-payout opportunity.
6.4.6 Sixth Choice: Amazon
Amazon offers substantial upside through AWS acceleration, Trainium and model neutrality. Its headline valuation is distorted, however, and capital-expenditure and free-cash-flow visibility remain weaker than at Microsoft.
It is a high-optionality name with somewhat lower valuation transparency.
6.4.7 Broadcom: Wait for a Better Price
Broadcom’s custom-silicon and networking position is scarce and valuable, but current expectations are demanding. It is more attractive as a buy-on-reset asset than as the leading value opportunity at current levels.
6.4.8 Micron and Oracle: High Payout, Lower Probability
Micron can generate enormous cash flow if HBM shortages persist, but a memory-cycle reversal would pressure both earnings and valuation. Oracle can produce substantial upside if backlog converts, but negative free cash flow and financing risk create significant downside in a credit or capex slowdown.
Both are better suited to tactical exposure than to late-cycle core allocation.
6.4.9 AMD, Arista and Vertiv: Excellent Companies, Insufficient Entry Asymmetry
AMD requires sustained share gains to justify approximately 132 times earnings. Arista and Vertiv are executing exceptionally well, but multiples above sixty times already require a long period of strong performance.
The businesses may continue to grow, but the balance between upside and valuation-compression risk is currently less attractive than at Alphabet, Meta, TSMC, NVIDIA or Microsoft.
Our qualitative ranking is therefore:
Best overall risk-reward: Alphabet, Meta and TSMC.
High-quality core holdings: NVIDIA, Microsoft and Amazon.
High quality, but wait for a better entry: Broadcom.
High-payout, lower-probability tactical positions: Micron and Oracle.
Strong businesses, but insufficient current asymmetry: AMD, Arista and Vertiv.
These probability and payout assessments are qualitative research judgments based on fundamentals, valuation and cycle position. They are not statistical forecasts or guaranteed returns.
7. A Time-Cycle Investment Framework
7.1 The Next Zero to Six Months: Shift from Beta Maximization to Drawdown Control
In the second half of a bull market, the objective should no longer be to maximize AI exposure. Portfolio quality should rise.
The most defensible core assets are Alphabet, Microsoft, Meta, NVIDIA and TSMC because they possess existing cash engines or structural bottlenecks. Exposure to companies dependent on one customer, external financing or distant market-share assumptions should be reduced.
Entry timing should be diversified. Gradual accumulation around earnings verification and market corrections is more appropriate than chasing one-week price acceleration.
7.2 The Next Six to Eighteen Months: The Critical AI Return-on-Investment Window
The next six to eighteen months will determine whether the current infrastructure cycle can generate sustainable returns. Investors should monitor Google Cloud, Azure, AWS and Oracle IaaS growth; cloud margins; hyperscaler capex relative to operating cash flow; GPU utilization; model API pricing; and enterprise-agent renewal and usage depth.
If two consecutive quarters show rising capital expenditure, slowing cloud growth, weaker backlog conversion and deteriorating free cash flow, the AI capex cycle may be approaching a genuine inflection point.
Oracle, Micron, AMD, Arista and Vertiv would likely be more sensitive to such a transition than the major platform companies.
7.3 The Next Eighteen to Thirty-Six Months: A Potential First Systemic Shakeout
If enterprise agents and workflow revenue validate current investment, 2027–2028 could produce a second major AI upcycle, with value migrating from GPUs and data centers toward models, proprietary context, enterprise workflows and consumer interfaces.
If revenue fails to validate spending, the industry may experience its first systemic shakeout. Data centers without stable customers could face financing stress, second-tier models could enter severe price competition, standalone harness products could be absorbed by model providers, and semiconductor and memory capital expenditure could decline.
In either scenario, the long-term strategy is not to abandon AI. It is to rotate from highly valued compute beta toward compounders that control models, context, distribution and free cash flow.
7.4 The Most Important Monitoring Variables
The key variables are the quality of token growth, cost per successful task, autonomous-agent duration, AI revenue relative to capex, cloud gross margin, ASIC substitution of GPUs and the migration of consumer interfaces among ChatGPT, Gemini, Claude, Search, browsers and operating systems.
Markets will not permanently reward the number of GPUs purchased. They will ultimately reprice companies according to the revenue, profit and free cash flow those GPUs generate.
8. HTX Case Study: AI Is Accelerating the Convergence of Crypto and U.S. Equities
8.1 The Strategic Importance of HTX’s U.S. Equity Business
For many years, competition among crypto exchanges centered on native digital assets such as Bitcoin, Ethereum, perpetual futures and meme tokens. Since 2026, however, a structural shift has become increasingly visible. Crypto-native investors are no longer satisfied with trading only digital assets. Instead, they want to allocate capital across cryptocurrencies, U.S. equities, ETFs, commodities and even pre-IPO opportunities within a single account. AI semiconductor companies, HBM suppliers, cloud leaders, precious metals and major U.S. equity indices are increasingly being viewed as part of the same global risk portfolio alongside Bitcoin and Ethereum.
HTX has been one of the earliest crypto exchanges to systematically execute this strategy. According to the company’s H1 2026 report and July operational update, TradFi trading volume has exceeded US$2.5 billion, while the platform now supports more than 170 TradFi-related assets, including U.S. equities, ETFs, gold, silver, crude oil, AI semiconductor companies, memory suppliers, aerospace companies and pre-IPO themes such as OpenAI and Anthropic. As product coverage expands and user experience continues to improve, TradFi is becoming one of HTX’s most important new growth engines, transforming the platform from a pure crypto exchange into a broader global digital asset and traditional finance gateway. (globenewswire.com)
8.2 AI Is Reshaping How Crypto Investors Allocate Capital
Historically, crypto investors focused primarily on digital assets, while stocks, commodities and other traditional financial instruments required separate brokerage accounts and fiat settlement systems. As AI has become one of the most important investment themes in global capital markets, that allocation framework is changing rapidly. Companies such as NVIDIA, Micron, TSMC, Broadcom, Meta and Alphabet, together with gold, oil, ETFs and pre-IPO opportunities, are increasingly becoming part of the everyday investment universe of crypto-native investors.
This shift suggests that users no longer see crypto and U.S. equities as two independent markets. Instead, they increasingly manage them as components of a unified global portfolio, reallocating capital dynamically between Bitcoin, Ethereum, AI leaders, commodities and equity indices according to macro conditions, industry trends and risk appetite. The rapid expansion of HTX’s TradFi business therefore reflects a broader transformation in investor behavior—from single-asset crypto trading toward multi-asset global portfolio management.
8.3 HTX’s Distribution Advantage
HTX’s competitive advantage is not simply that it offers access to U.S. equities. More importantly, it already possesses a large installed base of crypto users who have completed KYC, funded their accounts and hold stablecoins such as USDT. These users do not need to open traditional brokerage accounts, convert funds into fiat currencies or move capital into another financial ecosystem before investing in AI stocks, ETFs, commodities or pre-IPO assets.
This significantly reduces customer acquisition costs for HTX’s TradFi business while simultaneously lowering friction for users. Instead of acquiring an entirely new audience of traditional equity investors, HTX can leverage its existing crypto community and extend stablecoin liquidity, trading habits and user relationships into a much broader range of financial products.
TradFi therefore represents far more than a new product category. It increases trading frequency, improves customer retention and expands lifetime value (LTV). Users can rotate between crypto assets, AI leaders, ETFs and defensive assets such as gold without leaving the platform, allowing HTX to participate continuously in users’ global asset-allocation decisions throughout different market environments.
8.4 What HTX’s Growth Suggests About the Future
The recent increase in HTX’s U.S. equity and TradFi-related revenue should not be interpreted merely as a temporary consequence of strong AI stock performance. More importantly, it validates a new business model: crypto-native users are willing to trade traditional financial assets inside a crypto exchange. Once this behavior becomes established, the competitive landscape for digital asset platforms changes fundamentally.
Future competition will no longer revolve solely around spot trading, derivatives, liquidity or token listings. Instead, exchanges will increasingly compete to become comprehensive financial platforms capable of offering crypto assets, U.S. equities, ETFs, commodities, pre-IPO investments, wealth management services and AI-powered investment tools within a unified user experience.
Within the Token Economy framework presented throughout this report, HTX provides an important real-world case study. AI is not only transforming model capabilities and computational demand; it is also reshaping capital allocation, investor behavior and the architecture of financial platforms. As AI becomes a common investment theme across both traditional finance and crypto, the distinction between these two ecosystems is likely to continue narrowing. In the long run, the key competitive advantage for exchanges will evolve from pure trading execution toward Global Asset Distribution, positioning platforms such as HTX to benefit from the ongoing convergence of digital assets and traditional capital markets.
Conclusion
The AI industry and AI equities are not at the same point in their cycles. Large-model and agent diffusion remains early, and the long-term opportunity to automate global knowledge work remains enormous. Capital expenditure, financing and equity valuations, however, have moved materially ahead of application penetration.
The correct 2026 investment framework is therefore not the linear proposition that all AI stocks must continue rising because AI will continue developing. Genuine technological revolutions can coexist with severe asset bubbles. The technology cycle remains early, the application cycle is early-to-middle, the capital-expenditure cycle is later-stage, the earnings cycle is entering verification and the equity cycle is in the second half of a bull market.
Summary
The most attractive AI equities are not necessarily those with the lowest headline P/E ratios. The relevant question is which companies offer the best combination of normalized valuation, competitive advantage, free cash flow and AI optionality.
Alphabet’s headline multiple is distorted by investment gains, but its full-stack architecture and multiple growth engines produce the best overall asymmetry. Meta offers an attractive conventional earnings valuation but must control capital spending. TSMC possesses the most durable supply-chain moat but carries geopolitical risk. NVIDIA is the most attractive pure AI asset on a growth-adjusted basis. Microsoft provides the highest probability of fundamental success. Amazon offers substantial cloud and silicon optionality.
Oracle and Micron offer high payouts but lower probabilities, while AMD, Arista and Vertiv are high-quality businesses whose current prices require near-perfect execution.
The most important future question is no longer how many people have not yet used a chatbot. It is how many hours of high-value human work machines can complete each day. Models determine whether a company has a ticket to compete. Compute cost determines whether tokens can be produced at scale. Context and workflow determine whether models can complete real work. Distribution determines whether revenue can grow efficiently. The valuation and capital-cycle position at which investors enter ultimately determine how much of that long-term value they capture.
The cycle divergence within AI will continue to evolve. Whether capital expenditure converts into free cash flow, whether token costs decline with scale, and whether enterprise-workflow penetration materializes are questions that will be answered over the coming quarters rather than settled today. HTX Research will continue tracking this process and how the relationship between technological diffusion, capital returns and asset pricing develops.
Risk Disclosure: This report is intended for industry research and investment-framework discussion only. It does not constitute a recommendation to buy or sell any security, private-company interest or financial product. Private-company ARR, revenue run rates and valuations are not equivalent to audited public-company GAAP results. Public-market prices and valuation multiples are point-in-time observations around August 14, 2026, and AI technology, capital expenditure and competitive conditions may change rapidly.
Reference
The post first appeared on HTX Square.
