The AI Compute
Value Chain
AI infrastructure is the dominant marginal capital cycle in US equities — but returns accrue where scarcity and pricing power persist. This study maps where, then builds a hypothetical portfolio around the answer.
One capital cycle now drives the market’s growth
Not the market’s only engine — its dominant marginal engine. A disproportionate share of S&P 500 earnings and capital-spending growth now traces to this one buildout. The study’s question: where along the chain do the economics ultimately accrue?
- The buildout is creating bottlenecks across several layers of the chain.
- Not every bottleneck yields durable economic rent.
- The task is to identify who captures that rent — and what could erode it.
One more number frames the concentration. Since ChatGPT’s launch, 42 AI-related companies have contributed ~78% of the S&P 500’s price return, ~66% of its earnings growth and ~71% of its capex-and-R&D growth since November 2022 (JPMorgan Asset Management, 2026 Outlook) [10] — without them, the index would have trailed Europe, Japan and China. Concentration is usually a risk to diversify away. Here, the concentration is the market’s growth — and the index itself carries the same fragility: if the cycle turns, the benchmark falls with it, so a portfolio cannot diversify this risk by owning the market. The construction in Section 05 answers it with sizing, drift ceilings and tripwires, not with diversification it cannot have. This study’s response is to understand where in the chain the rent is collected, and how durably.
- Capex: Goldman raised 2026 to ~$800B and 2027 to $1.2T; debt is now expected to fund roughly a third of the spend [1].
- Semiconductors: July sales +135% y/y; WSTS’s Q2-adjusted calculation ~$1.66T for 2026 [5].
- Memory: Micron’s fiscal Q4 at $54.2B revenue and 87% gross margin; most calendar-2027 high-bandwidth memory (HBM) bits contracted under take-or-pay agreements with price floors [16].
- Custom silicon: Broadcom AI semiconductor revenue $16.7B, +221% y/y, guided to $21.7B [19].
- Platforms: Azure past $100B of annual revenue, growing 43%; Microsoft’s commercial remaining performance obligation (RPO) $678B, +84% [14].
- Tripwires: none triggered at the cutoff; 03 on watch as debt-financed capex rises — readings in Section 08.
Thirty-six names, three distinct roles in the construction
The figures in Section 01 describe one buildout. Who keeps its economics depends on where, along the supply chain, scarcity and pricing power coincide — so the map that follows is organised by bottleneck, not by sector. Scarcity alone earns nothing, and not every name here is a tollbooth.
- The rent test: something is scarce; a company controls access to it; customers have few alternatives; the advantage persists.
- Rings are portfolio roles; layers are positions in the supply chain.
- A bottleneck is a constraint; a tollbooth is a company able to charge for it.
Where scarcity is strongest — and how durable it may be
Sixteen companies spanning the most constrained parts of the chain — some tollbooths, some challengers, some second-source optionality. Durability is the variable to keep testing.
- Tollbooth: passes all four stages of the rent test today.
- Challenger: carried by customers’ incentive to qualify an alternative, not by equivalent economics of its own.
- Optionality: passes none of the stages yet; sized for the chance that it will.
- Each card states the mechanism, the class, the rent, what erodes it, and the tripwire that tests it.
Leading-edge fabs depend on this concentrated equipment cohort. ASML is the world’s only maker of extreme-ultraviolet (EUV) lithography systems [18]. Etch, deposition and metrology are two- to three-player markets. Decades of process lock-in let these suppliers collect a toll across advanced-silicon capital expenditure.
TSMC prints an estimated ~95% of the 3nm-class leading edge (Yole) [11] — including most leading merchant and custom AI accelerators; its share of “advanced” nodes overall is definition-dependent. Intel is second-source optionality, not a present tollbooth: a policy-backed option on buyers securing the alternative they want.
The layer pairs the merchant-silicon leader with its challenger. CUDA’s two decades of developer lock-in remain the deepest software moat in semiconductors. AMD is the credible second source that large buyers actively qualify — a position that rests on the buyers’ incentive to qualify an alternative, not on equivalent economics today.
Hyperscalers design their own accelerators. Broadcom and Marvell supply the IP, physical implementation and connectivity that make custom silicon shippable [19–20] — plus the Ethernet and optical links wiring clusters together. Corning supplies the fibre and glass those links run over. Substitution is not binary: it redistributes economics among chip designers, IP providers, networking vendors, foundries and the hyperscalers themselves, and a gain in one part of the chain need not offset a loss elsewhere. Broadcom’s AI semiconductor revenue reached $16.7B in its fiscal Q3, +221% y/y, with $21.7B guided for Q4 [19] — the custom route now carries reported scale.
HBM created a structurally richer, technically differentiated segment inside an otherwise cyclical industry. Three global manufacturers; calendar-2026 supply was committed by late 2025, and by September Micron had contracted the vast majority of its calendar-2027 HBM bits under multi-year take-or-pay agreements with price floors [16]. WSTS now calculates 2026 memory revenue growth near 300% [5] — against which stands Gartner’s warning: “memflation is profound, but it is not perennial.” [6] Take-or-pay floors are the industry’s answer; whether they hold through a downturn is the open test. Contracted supply improves visibility, but it does not remove exposure to future capacity additions, demand shifts or margin normalisation.
The second act is inference on-device. Qualcomm is a major beneficiary if that expands — across Android handsets, Windows PCs, automotive and other edge devices. It is deliberately a different cycle from the data centre.
Compute cannot be deployed faster than it can be energised, and “power” is four constraints with different beneficiaries. Generation equipment: GE Vernova supplies the gas turbines and grid equipment behind the build — a $176B backlog, 116 GW of gas equipment under contract or reservation heading past 125 GW by year-end, and data-centre orders above $5B year-to-date [24]. Grid and interconnection: transmission and connection queues, which no holding owns but GE Vernova’s electrification business serves. On-site generation: Bloom Energy supplies fuel-cell power where those queues run years — Q2 revenue +166% y/y, 2026 guidance raised to $3.9–4.2B, all major US hyperscalers among approved customers [25]. Time to energise: the common denominator, measured by tripwire 08. Both sell into the facility slice of the capex dollar. Neither is a tollbooth: the turbine backlog is the scarcity, and fuel cells compete with grid power, reciprocating engines and batteries on cost.
Illustrative economics of a stylised GB200-era 1 GW build Author synthesis · two models [12–13]
39¢ accelerators + HBM · 17¢ implied other server · 13¢ network · 30¢ facility · ~1¢ land & utility
Who keeps the economics — the seller of scarcity, or the converter of capacity?
Follow a capex dollar through the system: revenue to a supplier, then gross profit; depreciation on the buyer’s books; re-earned — or not — as AI revenue and free cash flow. Where the rent finally rests is the central open question of this cycle.
The demand side: ~$800B of expected 2026 capex, revised up again in September, with Goldman expecting debt to fund roughly a third of it [1–4]. Holding both suppliers and spenders is not double-counting a dollar — it is owning both sides of a bargaining relationship whose outcome is uncertain. Nor is it diversification: both sides share the risk that end-customer monetisation disappoints, and the sleeve bands assume that correlation. Today the suppliers capture disproportionate economics; the platforms are underwriting the conversion of capacity into recurring revenue. Cloud growth of 82% at Google, 43% at Azure and 37% at AWS confirms demand [2–3, 14] — not yet the return on incremental AI capital, which is the rent question. Apple is different by design: minimal capex, maximal distribution — more than 2.5 billion active devices, though only a subset supports advanced on-device inference [15].
Proprietary data and distribution are among the few durable moats at the application layer — alongside deployment, reliability and ecosystem. Palantir operationalises AI inside institutions. Snowflake is embedded in important enterprise data estates and workflows. Cloudflare runs edge delivery and defence. CrowdStrike and Palo Alto sell one of the most resilient budget lines in enterprise IT — resilient, though not immune to cuts. This ring is contested; positions here lean on switching costs, not scarcity.
Not one theme, and not part of the compute chain — a basket of separate long-duration exposures, each underwritten on its own rationale and its own failure mode, to the proposition that compute reshapes the physical world. Tesla and SpaceX: embodied AI and orbit. Rocket Lab: launch supply. Oklo: a speculative answer to part of the power constraint, complementing the Ring 1 power names and held at option size. Axon: public safety. Intuitive: surgical robotics. IBM: the quantum option. Indirect exposure, high dispersion — which is why Ring 3 is budgeted, not accumulated.
Outside the compute chain, and labelled as such. IBIT is monetary-regime optionality — against persistent fiscal deficits, financial repression, and an eventual easing response to capital-cycle stress. Two honest caveats: the cycle so far is substantially internally funded [10], and evidence on Bitcoin as an inflation hedge is mixed and period-dependent. A view, not an established property. Robinhood operates a major retail distribution platform — equities, options, crypto and, increasingly, tokenised assets — its Robinhood Chain testnet has processed 100M+ transactions [17] — an activity measure, not evidence of commercial monetisation.
The rules that make it a portfolio, not a list
Exact weights are withheld by design. The rules are not.
- The industry conclusions map to four portfolio roles: Ring 1 bottlenecks, Ring 2 platforms, substitution beneficiaries within Ring 1, and Ring 3 optionality.
- Sleeve bands, sizing limits, drift ceilings and rebalancing triggers turn those roles into a construction.
| Sleeve | Illustrative band |
|---|---|
| Ring 1 · Direct infrastructure (16 names) | 35 – 50% |
| Ring 2 · Platforms & distribution (6) | 30 – 40% |
| Ring 2 · Software, data & security (5) | 5 – 15% |
| Ring 3 · Frontier themes (7) | 5 – 12% |
| Ring 3 · Monetary & market rails (2) | 3 – 8% |
Initiate, build, drift — three limits
Initiation: 0.5–3%; core-eligible names (criteria below) up to 5%. Built position, after formal re-underwriting: 6% for core-eligible names, 4% otherwise. Drift ceiling: 10% — a breach forces a reduction. Every initiation and build must clear a prospective-return test under conservative assumptions. A strong company is not automatically a strong investment.
On triggers, not on the calendar
Rebalance when a sleeve breaches its band, a position breaches its ceiling, or a thesis breaks. Not on anniversaries. The construction refreshes regularly; it is a point-in-time snapshot, not a fixed roster.
Core eligibility — five tests before core sizing, and the trade-off they create
Hard to route around, or owns the customer at scale.
Reported segment revenue, disclosed AI run-rates [14] or backlog. AI-specific return on incremental capital is not yet observable — tripwire 03 tracks it.
Finances its own investment cycle from cash flow.
A durable non-AI business supports value if the cycle slows.
Prospective return must clear a hurdle under conservative assumptions — at initiation and at every re-underwriting. The gate is time-varying and intentionally not published.
Supply-side monopolies pass centrality easily — and concentrate single-point cycle and geopolitical risk that the resilience test penalises. Reasonable constructions weigh this differently.
A historical replay of the current basket
A hypothetical, hindsight-built replay of the current basket. Past performance, actual or simulated, is not indicative of future results.
- Sample holdings with share counts frozen, replayed across five years.
- Newly listed names enter on their first trading day.
- No capital was at risk.
Sample portfolio (simulated) vs indices — periods to 4 Aug 2026, market open Hypothetical · author calculation · provisional [21]
| Return | 1M | 6M | YTD | 1Y | 3Y ann. | 5Y ann. |
|---|---|---|---|---|---|---|
| Sample portfolio (simulated) | +0.4% | +36.6% | +36.5% | +94.8% | +45.4% | +26.2% |
| Nasdaq-100 Index | −1.1% | +18.0% | +16.3% | +29.0% | +24.1% | +14.1% |
| S&P 500 Index | +1.6% | +11.3% | +11.9% | +22.8% | +19.3% | +11.6% |
| Excess vs S&P 500 (pp) | −1.2 | +25.3 | +24.6 | +72.0 | +26.1 | +14.6 |
The same figures, charted — portfolio · Nasdaq-100 · S&P 500 · panels independently scaled: compare within a period, not across
How the thesis fails
“Volatility is the price” is acceptance, not management. A construction this concentrated needs named failure modes — and, on the next page, tripwires that force a response.
Hyperscaler guidance stops rising, then falls. Ring 1 revenue follows with a lag.
Depreciation outruns AI revenue. Platforms cut spending to defend free cash flow.
In-house chips take share from merchant GPUs faster than networking gains offset.
“Memflation” reverses as supply lands. HBM margins compress toward history.
Geopolitics interrupts the leading edge, or policy fragments the customer base.
Energisation and interconnection queues stall deployments; capex timelines slip.
The thesis holds but the price paid does not. Rates or risk appetite reset valuations.
Vendor financing and concentrated AI-lab customers can overstate underlying demand; that support can reverse.
Nine indicators, amber and red
Each indicator carries two thresholds. Amber: investigate and document. Red: mandatory re-underwriting — retain only with a documented exception, otherwise reduce.
| Indicator | Amber — investigate (window stated) | Red — re-underwrite | Reading · 30 Sep 2026 |
|---|---|---|---|
| 01 · Hyperscaler capex — primary demand signal | Forward-12-month consensus capex for the five largest hyperscalers falls by any amount vs the prior quarter | Forward-12-month aggregate falls >10% vs prior quarter | No trigger — estimates revised up: ~$800B 2026, $1.2T 2027 [1] |
| 02 · Cloud AI revenue & backlog — demand conversion | Cloud segment revenue growth decelerates for two consecutive quarters while capex growth does not | RPO or backlog falls both sequentially and year-on-year with no offsetting revenue acceleration | No trigger — MSFT RPO +84% to $678B; AWS +37%, Google Cloud +82% [2–3, 14] |
| 03 · Depreciation vs AI revenue — the rent test | Year-on-year increase in cloud operating profit plus disclosed AI revenue is smaller than the year-on-year increase in depreciation and lease expense, one quarter | The same shortfall for two consecutive quarters while cloud gross margin or free-cash-flow conversion deteriorates | Watch — debt to fund about a third of 2026 capex (GS); no disclosed shortfall yet [1] |
| 04 · Compute pricing — the thesis’ clearing price | Published rental indices [23] fall more than 10% over 30 days on a performance-adjusted basis | Performance-adjusted prices fall more than 20%, alongside declining utilisation | No trigger — H100 ~$3.8/hr, flat over 30 days; B200 +79% in three months [23] |
| 05 · HBM — memory normalisation watch | HBM lead times shorten in two consecutive quarterly disclosures | Contract pricing and gross-margin guidance are cut in the same quarterly revision | No trigger — CY27 HBM bits largely contracted at higher ASPs under take-or-pay floors [16] |
| 06 · Foundry, packaging & equipment — the scarcity gauge | Equipment order intake falls sequentially (direct for equipment names); leading-edge utilisation declines alongside softer lead times (direct for foundry, proxy for equipment) | Equipment order cancellations or push-outs (direct); price concessions or wafer-order reductions at AI-relevant nodes (direct for foundry, proxy for equipment) | No new reading — TSMC Q3 results due mid-October |
| 07 · Merchant vs custom balance — substitution watch | Custom-accelerator (ASIC) share rises while AVGO and MRVL custom-silicon revenue growth decelerates, one quarter | Merchant-GPU economics deteriorate faster than custom-silicon and networking positions benefit | No trigger — Broadcom AI semiconductor revenue +221% y/y; merchant orders intact [19] |
| 08 · Power & energisation — the physical constraint | Reported median energisation delay lengthens in two consecutive quarterly readings | Delays trigger announced capex cancellations | No new reading — the constraint now shows in financing more than energisation [1] |
| 09 · Valuation — the price paid | Base-case expected return falls below the stated hurdle | Downside exceeds remaining upside, thesis intact | n/a — hurdle not published |
Important disclosures and limitations
Purpose, circulation & status
An educational paper for the exchange of ideas and the study of an analytical framework, intended for general circulation: it does not take into account the specific investment objectives, financial situation or particular needs of any person, and readers should take their own advice on suitability. Nothing herein is an offer to sell or a solicitation of an offer to buy any security in any jurisdiction, nor investment, legal, accounting or tax advice, nor a research report. This document has not been reviewed or approved by the Monetary Authority of Singapore or any other regulatory authority.
Hypothetical portfolio
The portfolio presented is hypothetical: a teaching construct illustrating how a value-chain framework could be expressed in listed securities. It does not represent any actual fund, account, client assets or investment vehicle, and no such vehicle is offered, described or implied. Sleeve bands, sizing rules and core-eligibility criteria are illustrative; the construction is refreshed at the author’s discretion and will differ at other dates.
Author’s interest
The author may personally hold, or transact in, securities mentioned in this document, and may do so at any time without notice. Nothing herein is a recommendation to buy, sell, or hold any security.
Hypothetical performance
Section 06 presents hypothetical, simulated performance; its methodology and limitations are stated with the exhibit. Such results are constructed with hindsight, involve no capital at risk, and reflect no actual account, trading, costs, taxes or financing; no representation is made that any portfolio achieved, or would achieve, results resembling those shown. Indices are unmanaged; one cannot invest directly in an index. Past performance, actual or simulated, is not indicative of future results.
Forward-looking statements
Market and industry statistics describe a period of unusually strong AI-related demand that may not persist. This document contains opinions and forward-looking statements involving significant uncertainty; actual outcomes may differ materially. Narrative figures reflect disclosures through the 30 September 2026 data cutoff; the reference table in Section 11 is stated at the 16 October 2026 close and the replay in Section 06 to 4 August 2026, as each exhibit notes. Opinions are the author’s as of the 30 September 2026 data cutoff and may change without notice.
Risk factors of the concept discussed
A concentrated, technology-focused approach of the kind studied may experience substantial and prolonged losses if implemented; individual securities may suffer a total loss of value. Semiconductor and AI-infrastructure markets are cyclical and have suffered severe historical drawdowns. Digital-asset exposure is highly volatile and speculative. ADRs and recently listed securities carry additional liquidity, currency, and disclosure risks. Leverage, where used in any real implementation, magnifies losses. Sections 07–08 list thesis-specific failure modes and tripwires; neither is exhaustive.
Evidence labels
Figures are tagged in context: GUIDANCE (company-stated), FORECAST (third-party projection), ESTIMATE (third-party model or analyst estimate), AUTHOR SYNTHESIS (author’s combination of sources), HYPOTHETICAL (simulated result; author’s calculation). Untagged qualitative claims are the author’s views. Third-party figures are believed reliable but not independently verified, are approximate, and speak only as of their publication dates.
Distribution
Not directed at, or for distribution to or use by, any person in any jurisdiction where such distribution or use would be contrary to law or regulation. Do not reproduce or redistribute without written permission. © 2026 Compute Edge Capital — Educational Series.
Compute Edge Capital · Educational Series · Q4 2026 Issue · Section 09 of 11
Every figure, its origin, its status
All sources accessed 1–8 August, 1–5 October and 18–19 October 2026 · Status: REPORTED = company-reported actual · GUIDANCE = company-stated outlook · FORECAST = third-party projection · ESTIMATE = analyst or model figure · HYPOTHETICAL = simulated result; author’s calculation, not reproducible from the information presented
Compute Edge Capital · Educational Series · The AI Compute Value Chain · Q4 2026 Issue · v2.8.2 · Data cutoff 30 Sep 2026 · Section 10 of 11
The names, in numbers
Reference data for every holding in the sample construction — one line each, no commentary. As of 16 October 2026, market close [22].
Reference table — 36 names, alphabetical by ticker 16 Oct 2026 close
| Ticker | Company | Role | Mkt cap | Beta | YTD | 1Y | 3Y ann. |
|---|---|---|---|---|---|---|---|
| AAPL | Apple | R2 · Platforms & distribution | $5.01T | 1.07 | +26.7% | +39.3% | +24.9% |
| AMAT | Applied Materials | R1 · Equipment | $324.81B | 1.59 | +59.8% | +80.7% | +43.8% |
| AMD | Advanced Micro Devices | R1 · Logic · challenger | $863.17B | 2.45 | +146.9% | +125.4% | +70.6% |
| AMZN | Amazon | R2 · Platforms & distribution | $2.70T | 1.46 | +8.3% | +16.6% | +23.6% |
| ASML | ASML Holding ADR | R1 · Equipment · EUV | $638.23B | 1.30 | +53.8% | +63.6% | +41.0% |
| AVGO | Broadcom | R1 · Custom silicon & networking | $1.57T | 1.48 | −4.5% | −6.5% | +55.5% |
| AXON | Axon Enterprise | R3 · Frontier themes | $31.71B | 1.44 | −31.3% | −40.2% | +21.0% |
| BE | Bloom Energy | R1 · Power · on-site generation | $65.18B | 3.78 | +154.7% | +92.3% | +161.4% |
| CRWD | CrowdStrike | R2 · Software, data & security | $233.33B | 1.20 | +94.0% | +88.6% | +69.2% |
| GEV | GE Vernova listed Apr ’24 | R1 · Power · turbines & grid | $245.35B | 0.97 | +41.2% | +53.4% | n/a |
| GLW | Corning | R1 · Optics & interconnect | $113.70B | 1.12 | +51.5% | +53.5% | +70.4% |
| GOOGL | Alphabet | R2 · Platforms & distribution | $4.17T | 1.21 | +9.0% | +35.8% | +35.2% |
| HOOD | Robinhood Markets | R3 · Rails · outside chain | $88.89B | 2.35 | −12.6% | −24.8% | +119.7% |
| IBIT | iShares Bitcoin Trust ETP · 2024 | R3 · Rails · outside chain | $66.21B | 2.14 | −7.0% | −24.9% | n/a |
| IBM | IBM | R3 · Frontier themes | $218.71B | 0.72 | −20.0% | −13.7% | +22.3% |
| INTC | Intel | R1 · Foundry · optionality | $472.12B | 2.23 | +143.5% | +143.9% | +35.7% |
| ISRG | Intuitive Surgical | R3 · Frontier themes | $150.74B | 1.46 | −24.7% | −1.8% | +15.9% |
| KLAC | KLA | R1 · Equipment | $216.29B | 1.45 | +36.9% | +51.6% | +51.2% |
| LRCX | Lam Research | R1 · Equipment | $330.30B | 1.85 | +54.6% | +86.2% | +60.8% |
| META | Meta Platforms | R2 · Platforms & distribution | $1.81T | 1.19 | +7.7% | 0.0% | +30.6% |
| MRVL | Marvell Technology | R1 · Custom silicon & networking | $199.79B | 2.26 | +168.6% | +158.7% | +62.4% |
| MSFT | Microsoft | R2 · Platforms & distribution | $3.93T | 1.10 | +10.1% | +4.3% | +17.6% |
| MU | Micron Technology | R1 · Memory · HBM | $994.34B | 2.23 | +208.2% | +334.4% | +134.2% |
| NET | Cloudflare | R2 · Software, data & security | $105.51B | 1.65 | +50.3% | +40.2% | +67.1% |
| NVDA | NVIDIA | R1 · Logic · accelerators | $5.17T | 2.22 | +15.1% | +18.1% | +67.0% |
| OKLO | Oklo listed 2024 | R3 · Frontier themes · power | $5.20B | 1.22 | −61.0% | −82.8% | +39.5% |
| ORCL | Oracle | R2 · Platforms & distribution | $388.73B | 1.77 | −33.3% | −58.5% | +6.9% |
| PANW | Palo Alto Networks | R2 · Software, data & security | $279.76B | 0.93 | +85.7% | +66.4% | +37.8% |
| PLTR | Palantir Technologies | R2 · Software, data & security | $447.04B | 1.60 | +4.7% | +4.4% | +120.5% |
| QCOM | Qualcomm | R1 · Edge silicon | $183.79B | 1.65 | +2.2% | +7.1% | +18.4% |
| RKLB | Rocket Lab | R3 · Frontier themes | $30.43B | 2.94 | −27.1% | −24.1% | +125.1% |
| SKHY | SK hynix ADR · listed Jul ’26 | R1 · Memory · HBM | $695.40B | n/m | n/a | n/a | n/a |
| SNOW | Snowflake | R2 · Software, data & security | $108.15B | 1.35 | +39.9% | +27.8% | +23.7% |
| SPCX | SpaceX listed Jun ’26 | R3 · Frontier themes | $1.98T | n/m | n/a | n/a | n/a |
| TSLA | Tesla | R3 · Frontier themes | $1.40T | 1.92 | −21.1% | −17.3% | +11.8% |
| TSM | Taiwan Semiconductor ADR | R1 · Foundry | $1.93T | 1.24 | +38.8% | +36.8% | +68.1% |
Compute Edge Capital · Educational Series · Q4 2026 Issue · Section 11 of 11