Related context: this article follows our work on late July big tech earnings and the memory thesis, Micron’s FY3Q26 AI memory earnings, and the 1H26 AI infrastructure bottleneck review. Related hubs are the Exclusive Analysis Hub and the AI HBM Hub. The question is not simply whether AI demand is strong. The question is who keeps the economic value created by AI tokens.
TL;DR
If we decompose one current dollar of AI token revenue, the largest share of value added still sits at the model and application layer. The equity market, however, is being driven by data center investment that is being pulled forward before final application revenue has fully arrived. That is why today’s AI value chain looks less like a normal pyramid and more like an inverted one.
The current value-added split of one dollar of AI token revenue is roughly as follows.
| Layer | Current value-added estimate |
|---|---|
| AI model company | 45 to 55 cents |
| Cloud and data center | 10 to 16 cents |
| GPU supplier | About 13 cents |
| HBM, server DRAM, SSD | About 2.3 cents |
| Other infrastructure software and storage | 3 to 5 cents |
| Foundry | About 1 cent |
| Power | About 1.3 cents |
The market capitalization and capex cycle cannot be explained by this table alone. Current AI service revenue is estimated around 60 to 100 billion dollars per year, while the leading GPU supplier’s data center revenue run rate has reached around 300 billion dollars, and the big four cloud companies’ 2026 capex is estimated near 725 billion dollars. The industry is building future demand before current revenue catches up.
Memory is a real bottleneck in this structure. But if memory captures only about 2 to 5 cents of value added per token dollar, the central question for memory stocks is not “is the P/E low?” The real question is whether the industry can sustain mid-cycle net income of roughly 100 to 140 trillion won.
| Question | View |
|---|---|
| Where does AI token value accrue? | Near term: models and cloud. Long term: workflow applications and physical bottlenecks |
| Is memory a structural beneficiary? | Yes, but its value-added share is small and the stocks already price in a strong mid-cycle profit base |
| What should investors monitor? | HBM4 contract pricing, customer allocation, blended memory ASP, cloud backlog, and application revenue conversion |
1. The Real Question: Who Keeps the Dollar?
The most common mistake in AI investing is mixing up revenue attribution with value-added attribution. If a customer pays one dollar for an AI service, that dollar may first appear as revenue at a model company. But inside that dollar are cloud costs, GPU depreciation, memory, foundry capacity, power, networking, servers, cooling, labor, software, taxes, and financing costs.
So “who books the revenue?” is not the same question as “who keeps the surplus?”
This article uses five rules.
| Rule | Meaning |
|---|---|
| Revenue attribution | The first seller that invoices the customer |
| Value-added attribution | The economic value left after removing double counting |
| Token dollar | A dollar of API or model usage revenue |
| Application dollar | A dollar of final workflow or agent software spend |
| Depreciation basis | Data center capex is annualized over a five-year asset life |
The capex treatment matters. AI data center spending does not disappear as an immediate cost. It is capitalized into GPUs, servers, power gear, buildings, cooling, and networks, then depreciated over several years. Therefore, the right cost in a token-dollar model is annual depreciation, not total capex.
Time also matters. Inference cost in 2025 is not the same as inference cost in 2028. Cost per unit of performance is falling quickly.
| Period | Inference cost as share of revenue | Model gross margin |
|---|---|---|
| 2025 | 50 to 67% | 33 to 50% |
| 2026 | 40 to 55% | 45 to 60% |
| 2027 to 2028 | 30 to 40% | 60 to 70% |
This leads to an important point. Token prices can fall while model gross margins rise, if inference costs fall faster than prices. Much of that margin is then reinvested into training, compute reservations, and customer acquisition. The value does not simply stay on the income statement. It flows back into GPUs, memory, power, and foundry capacity.
2. The Current Value Waterfall
The table below decomposes one current AI token dollar. Revenue attribution is the amount a layer effectively invoices or absorbs. Value added is the economic share after removing double counting.
| Layer | Revenue attribution | Value-added estimate | Interpretation |
|---|---|---|---|
| AI model company | 100 cents | 45 to 55 cents | Direct customer layer. Falling inference cost is the largest lever |
| Cloud and data center | About 50 cents | 10 to 16 cents | Operating layer that bundles GPUs, networks, power, and buildings |
| GPU supplier | About 18 cents | About 13 cents | Main physical bottleneck today |
| HBM, server DRAM, SSD | About 3 cents | About 2.3 cents | Critical to system performance, but small as a token-dollar share |
| Infrastructure software, storage, security, observability | About 5 cents | About 3 cents | More important as operations become more complex |
| Foundry | About 1.5 cents | About 1 cent | Common bottleneck across AI chip designs |
| Server OEM and ODM | About 3 cents | About 1.4 cents | High throughput, limited margins |
| Third-party networking | About 1.3 cents | About 0.7 cents | Important at cluster scale, but supplier power varies |
| Power | About 2.5 cents | About 1.3 cents | Small cost line, large site bottleneck |
| CPU and IP | About 0.5 cents | About 0.3 cents | Control and support compute |
| Construction, components, materials, residual | About 9 cents | Distributed | Lower layers of the data center stack |
Two observations matter.
First, power is not the largest cost item inside a token dollar. But it is one of the largest physical bottlenecks. Power matters through grid access, substations, transmission, permits, and long-term power contracts.
Second, memory’s value-added share per token dollar is still small. This does not mean HBM is unimportant. HBM is indispensable. But the economic share retained by memory companies is far smaller than the share retained by GPUs or model/application layers. If the equity market ignores that difference, valuation errors appear.
3. The Inverted Pyramid
The current AI stack looks like this.
| Item | Estimated scale |
|---|---|
| AI model and application revenue | 60 to 100 billion dollars per year |
| GPU data center revenue run rate | About 300 billion dollars |
| 2026 big four cloud capex | About 725 billion dollars |
| AI revenue needed to normalize 700 billion dollars of capex | About 2.8 to 3.0 trillion dollars |
If AI service revenue is currently 60 to 100 billion dollars, and memory revenue is 3 to 7 cents per token dollar, AI memory flow revenue should be only about 2 to 7 billion dollars per year. Yet the leading memory companies are producing revenue and profit numbers far above what current token revenue alone can explain.
For example, SK Hynix reported 97.1 trillion won of revenue in 2025 and 52.6 trillion won of revenue in the first quarter of 2026. That scale is hard to explain only from current AI token revenue. Much of today’s memory revenue is therefore tied to capex being pulled forward for future AI demand.
This is not automatically wrong. Data centers cannot be built instantly after demand appears. Power, land, substations, GPUs, servers, cooling, and networking must be secured years in advance. Pre-investment is necessary.
The question is speed. If application and cloud revenue catch up, today’s capex is normal growth investment. If final demand lags, the correction begins upstream: memory and server components first, then GPUs, then cloud capex plans.
4. Three Convergence Paths
AI revenue is not determined by token price alone.
AI revenue = token volume × token price
Token price per unit of performance is falling quickly. It may decline by roughly 10x per year through the early phase, then slow to 1.5 to 2x per year after 2027. But volume can explode at the same time. A simple question may use a few hundred tokens. A workflow, document process, coding task, or agent job can use tens of thousands, hundreds of thousands, or even more than a million tokens.
That means price declines do not automatically mean revenue declines. If price falls by 90% but volume rises 100x, revenue still rises 10x.
| Scenario | Probability sense | Description | Market implication |
|---|---|---|---|
| Demand convergence | 45% | Application revenue scales quickly and 1.0 to 1.6 trillion dollars of capex normalizes | Limited upstream drawdown, bottleneck premium remains |
| Digestion then reacceleration | 40% | 2027 to 2028 capex pauses by 20 to 30%, then workflow AI demand returns | Memory ASP could correct 40 to 60%, then recover |
| Structural oversupply | 15% | Usage elasticity disappoints and data center buildout proves excessive | Capex cuts of 50% or more and upstream recession |
On-device AI changes the distribution, not the whole demand picture. More inference on phones, PCs, cars, robots, and industrial devices can reduce some cloud token revenue. But silicon demand does not disappear. LPDDR, mobile SoCs, edge NPUs, embedded storage, and local servers may all benefit.
5. Long-Run Equilibrium: Workflows Capture the Dollar
Over the long run, tokens are unlikely to remain scarce. As token prices fall and model performance gaps narrow, value moves from token production toward workflow ownership.
On a pure API dollar, the long-run value split may look like this.
| Layer | Long-run value-added estimate |
|---|---|
| Model and API company | 45 to 55 cents |
| Cloud and data center | 12 to 15 cents |
| GPU supplier | 8 to 10 cents |
| Custom ASIC | 4 to 5 cents |
| Memory | 4 to 5 cents |
| Foundry | 3 to 4 cents |
| Power and data center infrastructure | 4 to 5 cents |
| CPU and IP | About 1.5 cents |
| Server OEM | About 1.5 cents |
| Other infrastructure software | 4 to 6 cents |
| Residual | About 5 cents |
But customers do not really buy tokens. They buy outcomes: customer support, coding productivity, claims processing, contract review, factory quality control, and logistics automation. On a final application dollar, the distribution changes.
| Layer | Long-run revenue attribution | Long-run value added |
|---|---|---|
| Workflow applications, agents, distribution | 100 cents | 45 to 50 cents |
| Model and API | 35 to 40 cents | 18 to 24 cents |
| Cloud and data center | 13 to 15 cents | 4 to 6 cents |
| Total silicon | 9 to 12 cents | 6 to 8 cents |
| Power, land, cooling | 1.5 to 2 cents | About 1 cent |
| Server OEM and generic infrastructure | About 1 cent | Low |
| Other software | 3 to 5 cents | Medium |
The long-run conclusion is simple: tokens commoditize, while workflow ownership and physical bottlenecks retain value.
6. Capture Tiers
The best AI businesses are not simply the ones with the most revenue. The best businesses have three traits.
- Customers cannot switch easily.
- Cost declines become margin expansion.
- Supply cannot expand quickly when demand rises.
| Tier | Layer | Why it matters |
|---|---|---|
| Tier 1 | Foundry, workflow applications, integrated cloud platforms | Bottleneck supply, customer lock-in, full-stack control |
| Tier 2 | GPU, platform AI suppliers, HBM leaders, custom ASIC | Current bottlenecks and high growth, but with price and competition risk |
| Tier 3 | Server OEM, neocloud, commodity memory, regulated utilities | Volume grows, but rent capture is weaker |
Foundry belongs in Tier 1 because AI wafer demand exists whether the winning chip is a GPU or a custom ASIC. Workflow applications belong there because model costs can fall while end-user value remains tied to business outcomes.
Integrated cloud platforms need case-by-case judgment. Pure GPU rental can become a capital-heavy utility. But a company that owns search, ads, productivity software, cloud, chips, models, and data centers can turn capex into margin.
7. Multiples: Cheap Looking Is Not the Same as Cheap
AI value-chain multiples should differ by layer.
| Layer | Appropriate valuation lens |
|---|---|
| Foundry | EV/S 8 to 11x, P/E 20 to 26x if growth persists |
| GPU supplier | Normalized EV/S 9 to 13x, scarcity 15 to 18x, P/E 18 to 24x |
| Integrated cloud | EV/S 5 to 8x, full-stack 8 to 10x, P/E 20 to 28x |
| Platform AI supplier | Hypergrowth EV/ARR 10 to 18x, deceleration 5 to 8x |
| Workflow AI software | Simple tools 2 to 4x, workflow owners 8 to 12x, regulated data owners 12 to 20x |
| HBM and memory | EV/S 3 to 5x, leaders 5 to 7x, mid-cycle P/E 8 to 14x |
| Storage | EV/S 2 to 3.5x, platform 4 to 7x, normalized P/E 8 to 14x |
| CPU and IP | CPU 12 to 18x, IP conditionally 30 to 50x |
| Server OEM | EV/S 0.5 to 1.5x, P/E 8 to 15x |
| Power and data center infrastructure | Contracted assets at EV/EBITDA 10 to 18x or project IRR |
Memory is the layer where investors should be most careful. A 5 to 7x P/E can signal cheapness. At a cycle peak, it can also simply mean that the earnings denominator is temporarily inflated.
Useful valuation formulas include:
| Formula | Use |
|---|---|
| EV/NOPAT = (1 - g/ROIC) / (WACC - g) | Normalized multiple with growth and capital efficiency |
| Justified PBR = (sustainable ROE - g) / (COE - g) | Long-run capital-intensive industry PBR |
| Mid-cycle EPS = (peak + trough + 2×normal) / 4 | Cycle-aware earnings |
| Pre-profit value = terminal NOPAT × terminal multiple / (1 + discount rate)^t | Early-stage or loss-making firms |
Capital costs also differ.
| Layer | WACC or discount rate |
|---|---|
| Hyperscalers | 8.5 to 9.5% |
| GPU and foundry | 9 to 10.5% |
| Memory | 10.5 to 12% |
| Vertical AI software | 12 to 18% |
| Loss-making model layer | 14 to 20% |
8. The Memory Question: Can Mid-Cycle Net Income Reach 100 to 140 Trillion Won?
For memory companies, the key question is no longer whether HBM is good. The market already knows that. The key question is whether the current market cap requires a realistic or unrealistic mid-cycle profit base.
Suppose a leading memory company is valued near 1,400 trillion won. To sustain that value after one or two peak years, mid-cycle net income must be very large.
| Item | Assumption |
|---|---|
| Peak net income | 240 trillion won |
| Trough net income | 50 trillion won |
| Normalized net income | 60 trillion won |
| Mid-cycle net income | (240 + 50 + 2×60) / 4 = about 102 trillion won |
Applying an 8 to 14x mid-cycle multiple gives a fair-value range of about 820 to 1,430 trillion won. If the current market cap is already near the top of that range, the stock is not automatically expensive, but the margin of safety is thin.
Further upside requires one or more of the following.
| Requirement | Meaning |
|---|---|
| Evidence that mid-cycle net income can exceed 140 trillion won | HBM premium and high-margin mix remain after the peak |
| HBM4 custom premium is defended | Product customization protects pricing |
| ASP holds after Samsung’s supply entry | More supply does not crush pricing |
| General DRAM and NAND pricing also improves | The cycle broadens beyond HBM |
| Long-term customer agreements and volume commitments | Downcycle earnings risk falls |
9. Samsung, SK Hynix, and Micron Are Not the Same Trade
| Company | Core logic | Risk |
|---|---|---|
| Samsung Electronics | HBM share recovery and memory mix improvement. Expectations are lower than for the leader | HBM qualification, yield, foundry and system LSI losses |
| SK Hynix | Highest HBM profit quality and customer reference | Much of the good news is already in the price |
| Micron | Leading indicator for AI memory contracts, pricing, FCF, and capex | Volatility, short interest, and supply-cycle sensitivity |
Samsung has more catch-up optionality. SK Hynix has the highest-quality memory profit pool, but expectations are embedded. Micron is the canary because its pricing, long-term agreements, data center revenue, and capex guide influence Korean memory estimates directly.
10. Cloud, Power, and Vertical AI
Cloud companies split into two groups: integrated platforms and capital-heavy rental platforms. The most important numbers are backlog, revenue conversion, operating margin, internal silicon mix, free cash flow, and bond issuance.
Power is a small token-dollar cost line but a large site bottleneck. Grid access, substations, transmission, PPAs, gas turbines, nuclear, SMRs, water, land, cooling, and permits matter more than the simple power bill.
Vertical AI companies can capture the largest long-run share if they own the workflow. The winners will have regulated or physical domains, proprietary data, action authority, measurable ROI, model replaceability, a path to 70%+ gross margin, and compliance moats. Generic chat wrappers, basic retrieval integration, API resale, data-less automation, GPU brokerage, and server assembly should be treated with caution.
11. What to Monitor
| Metric | Why it matters |
|---|---|
| Combined cloud backlog | Whether capex is tied to future revenue |
| Self-built capex ratio | How much demand is moving outside public cloud |
| On-device inference share | Cloud token revenue substitution risk |
| HBM4 contract pricing | Durability of memory premium |
| Customer allocation by HBM supplier | Supplier bargaining power |
| Blended memory ASP | Whether the cycle is broadening |
| GPU lead times | Persistence of scarcity |
| Cloud operating margin | Whether capex turns into profit |
| AI application revenue growth | Whether final demand catches infrastructure spending |
The thesis weakens if open-weight systems compress model pricing faster than expected, on-device AI reduces cloud demand more quickly, 2027 cloud capex slows sharply, HBM4 premiums compress, or AI application revenue fails to scale.
Final View
The AI token economy can be summarized in one sentence:
Tokens get cheaper, while value moves to workflow ownership and hard-to-expand physical bottlenecks.
Memory is one of those bottlenecks. HBM, high-performance DRAM, and SSDs are essential to AI servers. But memory’s value-added share per token dollar remains limited. That is why memory stocks require more than the statement that HBM is strong. Current prices already demand a strong mid-cycle profit base.
The four key questions are:
- Does HBM4 preserve custom pricing?
- Do leading suppliers hold ASP and margins after Samsung’s entry?
- Do cloud backlog and AI application revenue catch up to capex?
- Can memory companies sustain mid-cycle net income of 100 to 140 trillion won?
If those four conditions are met, today’s large memory market caps can be justified. If not, low P/E ratios may prove to be a cycle-peak illusion rather than a margin of safety.