The AI Value Chain

The AI product chain is a connected economic system: electricity and manufacturing enable computing; computing enables models; models enable workflow tools; distribution turns those tools into customer outcomes. Its investment opportunity is not one homogeneous market. Profit can migrate between layers even while aggregate use increases. The central question is which business keeps a meaningful share of customer value after paying for computing, talent, reinvestment and competition.

Our strongest ten-year business-economics thesis is the combination of qualified electrical equipment, advanced manufacturing bottlenecks and software-rich compute systems. They sell scarce complements during a physical expansion whose construction cannot happen at software speed. Our strongest twenty-year thesis is distribution-owned workflow automation: businesses already inside customer identity, data, transactions and decision processes can turn improving models into recurring cash flow. These are business-quality conclusions, not assertions that every associated stock is inexpensive.

For a public shareholder, entry valuation can overturn the industry ranking. A highly profitable lithography supplier bought at an extreme cash-flow multiple can deliver a lower return than a less celebrated platform bought cheaply. The ten-stock ranking below therefore uses modeled cash returns at the dated market snapshot, normalized reinvestment and a downside case. It is an investable research shortlist, not a guarantee, an exhaustive global optimization or a recommendation to buy all ten immediately.

No credible analysis can identify the single highest-return investment over twenty years with certainty. We distinguish corporate return on invested capital, customer payback and stock return. Our forecasts are conditional illustrations designed to expose what must happen for an investment to work. If the price required for a 12% model return is below the market price, that is a valuation warning, not an instruction to raise the terminal multiple until the stock looks attractive.

The upstream chain starts with power generation, transmission, substations, transformers, switchgear, backup generation, cooling, water and permitted land. Semiconductor inputs add electronic design software, intellectual property, lithography, deposition, etching, metrology, wafers, chemicals, advanced packaging and test. A finished AI system combines accelerators, CPUs, high-bandwidth memory, ordinary DRAM, storage, optical interconnects, switches and management software. Every important bottleneck is a possible profit pool, but none has an unlimited market.

Above hardware sit cloud infrastructure and orchestration, foundation-model training, inference serving, evaluation, security, observability, data governance and enterprise integration. Applications include coding, search, advertising, customer support, document processing, design, healthcare administration and industrial optimization. Physical AI extends the chain into sensors, actuators, robots, autonomous systems and maintenance. Finance and insurance sit across all layers: they fund construction, cover liabilities and absorb utilization or technology risk.

Do not add revenues across these layers and call the result an AI total addressable market. A cloud operator's purchase becomes a chip designer's revenue, part of which becomes a foundry's revenue, part of which funds equipment purchases. The same end-customer dollar appears repeatedly. Conversely, substantial AI value may appear as an existing advertiser's improved conversion rather than a separately disclosed AI revenue line. Our sector comparison consequently uses economics and operating measures, not a fabricated summed market size.

A useful investment map classifies the constraint, buyer, payment mechanism, replacement cycle and substitute at each layer. An accelerator is not interchangeable with an installed switchboard: its technological life is shorter and its value is more sensitive to software performance. A workflow product is not interchangeable with a model API: it may own the customer relationship, consented data and regulatory accountability. Those differences explain why one growth rate or one valuation multiple cannot describe the entire chain.

Stanford's 2026 AI Index reports organizational adoption of 88% and emphasizes a jagged capability frontier. Adoption answers whether people use a tool; it does not answer how often they pay, whether it replaces another expense or whether the provider earns an economic profit. Strong benchmark performance is an input to a business case, not a substitute for measuring actual task reliability, completion time and required supervision.

The customer-support field research summarized by NBER found meaningful average productivity improvements, especially for less experienced workers. METR's early-2025 experiment found experienced open-source developers took 19% longer with AI tools. Its February 2026 update explained why changing tools and participant selection complicated follow-up estimates. These are different tasks and populations; neither proves that AI is universally productive or universally unproductive. We use the contrast to require task-specific measurement instead of extrapolating one headline to the economy.

A deployment should be judged on completed, accepted work. Start with a controlled baseline, record human review and error remediation, and include integration, training, security and model costs. A twenty-percent improvement in drafting speed can coexist with no saving if approval takes longer or demand is fixed. An apparent productivity gain can also become higher output rather than lower payroll. For investors, what matters is the portion of this value a supplier can charge for without destroying the customer's incentive to use it.

IFR's September 2026 release places the operating industrial robot stock at roughly five million units in 2025. That demonstrates a substantial existing automation base, not proof that general-purpose humanoid robots already have attractive economics. Physical deployments need durable uptime, safe interaction, spare parts and acceptable liability. We treat general robotics and AI-driven drug discovery as valuable twenty-year options, but do not insert speculative breakthrough revenue into the base valuation of an unrelated listed company.

The IEA's April 2026 central outlook moves global data-centre electricity consumption from 485 TWh in 2025 to around 950 TWh in 2030, roughly 3% of global consumption at that date. It is a forecast and includes non-AI data centres. The implication is substantial local construction pressure, not unlimited demand for every energy stock. Transmission access, reliability, permitting and time-to-power can be more valuable than undifferentiated generation capacity.

Electrical equipment and thermal management are attractive complements because they solve a real constraint and can serve markets beyond AI. Qualification, maintenance networks and installed equipment create switching friction. Nevertheless, rising order books can invite capacity additions, customer double ordering and eventual margin normalization. Compare backlog conversion, cancellation terms, cash advances, factory utilization and warranty claims rather than treating every reservation as guaranteed revenue.

Generation investments need contract-by-contract underwriting. A long-term power purchase agreement is valuable only if the price, escalator, collateral, delivery obligations and customer credit compensate for capital and fuel risk. Merchant power prices may rise during shortages and fall after new capacity arrives. Nuclear projects add licensing, construction duration and cost-overrun risk. Gas facilities add fuel and emissions exposure. Storage improves dispatch flexibility but is not itself a source of net energy.

Cooling demand should be measured by delivered heat-removal capacity, service revenue and customer architecture, not simply rack count. Higher density may favor liquid systems, but equipment standards and design choices can change. Eaton and Vertiv represent different exposures: diversified power management versus more concentrated data-centre infrastructure. We prefer qualified, cash-generative suppliers as business models, while acknowledging that a high stock valuation can already discount years of successful expansion.

Accelerator economics depend on system-level usefulness: how much accepted output a cluster produces per dollar and per watt. Software compatibility, developer familiarity, networking, utilization and support can sustain an advantage beyond raw arithmetic throughput. NVIDIA therefore has an integrated-system thesis rather than merely a GPU-volume thesis. But large customers have a strong incentive to optimize their own workloads, making custom accelerators and inference specialization a real competitive constraint.

Broadcom's opportunity is to participate in that optimization through custom silicon and connectivity. The attractive model is capital-light design combined with scarce expertise; the risk is concentration in a few customers and design programs. Foundry economics are different. TSMC must continually invest in process transitions, yield learning and packaging. It can serve multiple competing designers, yet it bears significant physical capital requirements, geographic concentration and the lower initial productivity of new fabs.

ASML's lithography position is a hard-to-replicate manufacturing complement. Installed-base service can soften, but not eliminate, equipment cycles. A dominant position does not mean customers can indefinitely accelerate investment or accept any price. Analyze customer concentration, shipment acceptance, service attachment and process adoption rather than projecting an order spike forever. A normalized margin and explicit maintenance investment are essential even for a near-monopoly.

Memory is a particularly important warning against extrapolation. Micron's September 30 release reports FY2026 revenue of $133.188B and a very strong earnings and cash-flow year, much newer than its available companyfacts quarter. Those reported results should be recognized, not silently replaced with older data. They also should not become a permanent twenty-year margin assumption. HBM qualification can improve differentiation while ordinary memory still responds to bit supply, inventory and capacity additions. We leave MU outside this ten-stock long-duration model because a reliable cycle-specific cash model deserves separate treatment, not because memory is unimportant.

A cluster with powerful accelerators but a congested network wastes expensive capacity. Networking can therefore participate in AI growth without owning the winning foundation model. The relevant measures are job completion time, fabric reliability, congestion control, bandwidth per accelerator and total cost of ownership. Nominal port speed is not sufficient if optics, cabling, software or failure recovery become the limiting factor.

Arista's software-rich Ethernet proposition benefits when customers value a manageable open fabric across different compute architectures. It is not immune to alternatives: integrated accelerator platforms, customer-designed switching, changes in optical interconnect and aggressive bundling can alter its share of system spending. A large AI fabric order also does not automatically imply high recurring revenue; investors should track service attachment and the proportion of growth supplied by a few hyperscalers.

Optics and storage contain both differentiated intellectual property and commoditized manufacturing. Count qualification cycles, yield, component redundancy, inventory commitments and customer bargaining power. Faster traffic can increase bandwidth demand while integration moves value into fewer packages. For storage, retrieval and memory efficiency can reduce cost per task even as total data expands. Growth in bits or ports is not identical to growth in gross profit.

Our base thesis favors integrated operations software and qualified components over unqualified capacity announcements. We nevertheless give Arista a material valuation and concentration penalty in the score. Its strong current operating margin is evidence of a good business; it is not evidence that paying any multiple will produce a good twenty-year stock return.

Cloud platforms connect infrastructure spending with paying users, enterprise budgets and software deployment. Their advantage is distribution, integration, identity, data management and the ability to optimize several workloads. Their risk is that compute becomes a fiercely competitive commodity while the balance sheet absorbs buildings, chips and contractual commitments. Revenue growth must be tested against cash return on the capital installed to produce it.

Microsoft's FY2026 cash flow illustrates the tension: $182.935B of operating cash flow and $115.948B of cash PP&E investment, before adjusting for compensation and financing obligations. Alphabet's first-half 2026 CFO of $84.859B was close to its $80.598B PP&E expenditure. Meta's 2026 capex guidance of $130-145B includes lease principal. These are different fiscal periods and definitions; they should not be added into an apparently precise comparable total without reconciliation.

Foundation-model revenue can grow quickly while research and inference costs also rise. Competition from open models, falling prices, distillation and customer-owned infrastructure can transfer value from the model producer to the application or end customer. Proprietary models may retain an advantage in difficult tasks, evaluation, safety and enterprise support, but that advantage needs repeated investment. We do not assume that the frontier model layer will automatically have the highest sustainable margin.

The model's cloud capex declines as a share of revenue only gradually and still exceeds depreciation at maturity. That is a consequential assumption, not a discovered fact. If expanding usage continually requires equally intensive chip replacement, free-cash-flow recovery will be weaker than our base case. Useful leading indicators are net incremental cloud revenue per dollar of cumulative capex, availability versus paid utilization, backlog conversion and depreciation growth relative to revenue.

Applications can capture the value between cheap intelligence and an expensive business problem. A model may be replaceable while a product embedded in permissions, audit trails, customer history and business processes is harder to remove. This is the most attractive long-duration opportunity in our framework, provided the vendor genuinely improves an outcome and does not simply repackage a model behind an interface.

There are two competing revenue effects. AI can improve a product enough to support higher pricing and broader adoption; it can also automate tasks previously sold as individual software seats. A seat-priced vendor may lose revenue as customers need fewer workers, even while customers become more productive. Outcome pricing, transaction pricing and consumption pricing may defend revenue, but they require accurate measurement and can shift delivery risk onto the vendor.

Data and security matter because enterprise automation must respect access controls, trace decisions and prevent unauthorized actions. Retrieval quality, data freshness, evaluation coverage and incident response can be more commercially important than adding another model option. But 'proprietary data' is not inherently a moat: a company must have lawful usage rights, differentiated information and a workflow in which that information changes decisions. Compliance expense belongs in the margin forecast, not just the risk section.

Advertising offers a different route: Meta and Alphabet can monetize better ranking and conversion inside an existing marketplace without convincing every user to buy an AI subscription. The investment test is incremental advertiser value after extra compute, content integrity and privacy expense. Longer-term success requires maintaining distribution and trust. We favor measurable business outcomes over generic claims about agents replacing entire professions.

Robotics can address labor scarcity, dangerous work and repetitive physical tasks. The relevant market is not a count of all human labor hours multiplied by a robot wage. Tasks differ in physical variability, required dexterity, failure consequences and integration cost. An automated warehouse station with a defined environment can have a far clearer payback than a general robot expected to work safely anywhere.

Underwrite an installation using accepted units per hour, uptime, interventions, maintenance, energy, insurance, installation and replacement. A demonstration of a successful task is not evidence of fleet reliability. Customer financing, lease guarantees and service obligations can make a hardware company look recurring while leaving it exposed to asset residual values. Public-company exposure may be diluted by legacy businesses or priced as if broad adoption were already certain.

Healthcare administration, imaging support and scientific tools may reach cash-generative workflows sooner than fully autonomous diagnosis or de novo drug discovery. Drug economics still depend on experimental validation, clinical trials, manufacturing and reimbursement. An AI discovery platform can improve search efficiency without removing biological failure risk. Forecast probability-adjusted milestone receipts separately from hypothetical blockbuster sales.

Over twenty years these fields could create exceptionally large winners. They also present wide outcome dispersion and often require expertise beyond a general public-equity model. We keep them in the industry map and monitoring plan rather than choosing a speculative stock solely to make the ranking appear comprehensive. A large addressable problem is not the same as an investable company with a defensible share of future profits.

For incremental corporate capital, asset-light workflow ownership and differentiated design software can be more attractive than owning fleets of rapidly depreciating compute. Their marginal delivery cost can fall as models improve, while existing customer relationships reduce acquisition costs. However, research spending and stock compensation are real capital costs, and competition can pass efficiency gains to the customer. High accounting ROIC may also reflect expensed investment rather than a truly small economic capital base.

For the next ten years, qualified power equipment and thermal systems have a comparatively visible physical demand driver; foundry, packaging, lithography and integrated compute have powerful bottleneck positions but higher cycle or concentration risk. We regard raw GPU rental, generic model reselling and undifferentiated hosting as less attractive unless secured by strong utilization, financing and customer contracts. The best private-project ROI could still occur in a niche unavailable through a listed stock.

For the next twenty years, we prefer customer-owned workflows, identity, transactions and trusted data to a static bet on today's model leaderboard. This does not require the same model vendor to win forever. The owner's product can swap models and benefit from falling unit costs. The opposing risk is that an incumbent's old pricing or distribution becomes obsolete. Long duration should never mean ignoring technological substitution.

The sector table below contains our qualitative research assessments, not measured sector-wide return forecasts. It purposely avoids invented precise ROIC ranges. Our numerical stock returns instead come from explicit company-level assumptions and dated equity values. This separation matters: a sector with the highest business return can still have the lowest shareholder return if the price already assumes perfection.

Our project example is deliberately an analyst illustration, not a quoted market price or a forecast of any company's facility. It assumes 100 MW of IT load, PUE 1.2, $5B of installed capital, $3.5B of compute replaced every four years and $1.5B of facility capital replaced over twenty years. Power is conservatively charged at full-load consumption for the whole year at $70/MWh, regardless of billable utilization. Other annual operating cost is $200M.

With these assumptions electricity costs about $73.6M a year, while steady-state replacement costs $950M a year. At a hypothetical $25,000 of annual revenue per billed kW and 65% billed utilization, revenue is $1.625B. Pre-tax owner cash is approximately $401M after operating expense and replacement; after a simplified 20% tax on that economic cash amount it is about $321M, or 6.4% of initial capital. This is a steady-state cash yield, not a full construction-period IRR or accounting net profit.

To obtain a twelve-percent after-tax cash yield under the same simplified assumptions, billable utilization must be around 79%. At 50% utilization the economics deteriorate sharply. Faster chip replacement, lower rental prices or a higher construction bill weaken results; slower renewal or better software efficiency improve them. We exclude land residual value, debt, ramp-up, price decay and customer defaults, which are all necessary in a real project valuation.

The point is not that all data centres have these economics. It is that the dominant cost can be compute renewal rather than power. A provider reporting strong revenue but spending equally fast to maintain competitive capacity may not generate attractive owner returns. Compare customer willingness to pay per accepted task, not only megawatts or installed accelerators. Power scarcity creates an opportunity for suppliers without proving the renter of expensive compute earns a high return.

The comparison model begins with a disclosed annual revenue or explicitly labeled annual research estimate; it does not annualize one exceptional quarter and call it audited history. Years 1-5, 6-10 and 11-20 have separate growth assumptions. Operating margin, depreciation and capital intensity interpolate to a normalized year-ten level and then remain at that level. Forecast years are relative model years because issuers have different fiscal calendars, not precise calendar-year guidance.

Owner cash equals operating income less modeled net interest and tax, plus depreciation, less capital expenditure, incremental working capital and a preferred-holder reserve. Operating margin already includes recurring stock compensation; the model does not add that compensation back. It values common-equity distribution capacity, not enterprise FCFF. Therefore it does not add a net-cash bridge or subtract debt a second time. Existing non-operating investment holdings and acquisitions are excluded, which can materially understate some companies' equity value.

Positive owner cash is assumed distributed annually. A deficit is assumed funded at 8% and deducted, with accumulated financing cost, from maturity proceeds; external funding availability is assumed rather than guaranteed. Terminal value is next year's positive owner cash multiplied by an explicit exit multiple. The bear case cuts growth, margin and terminal franchise retention while raising capex. It is an adverse scenario, not a calibrated probability. It does not capture outright insolvency or a geopolitical shutdown; those can produce much larger losses.

The displayed IRR is a hypothetical distribution-capacity return, not a forecast of actual dividends or share prices. Fixed quote-equivalent shares avoid counting buybacks twice, but omit future issuance and dilution changes. Alphabet's annual $1.203B preferred reserve approximates the current coupon; it is retained throughout the model as a conservative simplification, not a correct conversion schedule. TSM uses five common shares per ADS; ASML uses a fixed 1.17 USD/EUR modeling assumption. Currency changes, preferred conversion, lease renewal and repurchases require issuer-specific work before investing.

A 12% hurdle price discounts the same cash stream at twelve percent; it is not a sell-side target. The ranking score equals 100 times the average of the base ten- and twenty-year IRRs, plus 20 times the bear ten-year IRR, plus 1.5 times subjective durability (0-10), less subjective risk (0-10). There are no estimated probabilities or promises in this formula. The score is designed to be inspectable, not to make a subjective shortlist look statistically objective.

The candidate set intentionally covers distribution, cloud, advertising, integrated compute, custom silicon, foundry, lithography, networking and electrical equipment. It is limited to ten large, liquid listed companies where we can articulate a long-duration economic case. The financial appendix additionally covers Vertiv, Micron, AMD, Constellation and Vistra. Those five are not numerically ranked here; excluding them is a scope choice, not a claim that their returns must be inferior.

Rankings are generated from the model, not manually arranged to follow a popular AI narrative. Each row shows the dated price, base ten-year and twenty-year returns, bear ten-year return and the price consistent with the stated hurdle. A high score can still have an unattractive absolute base return. A lower-ranked stock can become attractive after a price decline or after evidence improves. The dashboard lets readers change a candidate's entry price without pretending the publication-date ranking has changed.

These businesses are economically correlated. A hyperscaler pause can reduce accelerator, networking, foundry, equipment and electrical orders together. Owning one stock from each layer is not automatically diversification. Customer budgets, Taiwan production, interest rates and capital-market sentiment are shared factors. A real portfolio also needs position limits, liquidity planning and exposure to assets outside the AI thesis; this article does not prescribe allocations.

A falsifiable thesis is more useful than a precise-looking distant target. We provide a failure condition for each candidate: sustained capex intensity, weakened software economics, lost custom design wins, customer concentration, manufacturing disruption or backlog normalization. Revisit the ranking when those variables change, rather than defending it because a company remains important to AI. Great technology and disappointing shareholder returns can coexist.

The first risk is economic demand: users may like AI but resist prices sufficient to cover infrastructure and model renewal. The second is efficiency: cheaper inference can expand usage, yet the demand response may be insufficient to preserve supplier revenue. The third is capital overbuild: project announcements, customer reservations and financed purchases can create the appearance of stronger independent demand than actually exists. The fourth is substitution: value can move from frontier models to open models, from GPUs to specialized chips, or from legacy seats to automated outcomes.

The fifth risk is financing. Leases, preferred shares, customer prepayments and special-purpose structures can move liabilities away from a simple cash-capex headline without eliminating them. Strategic investors that are also major customers can support a supplier while complicating the interpretation of its revenue quality. Demand funded by a partner's financing deserves a different confidence level from an unrelated customer's recurring cash payment. Track commitments, collateral and credit exposure, not just recognized revenue.

The sixth risk is geography and policy: advanced-node production and essential inputs are concentrated; export permissions, tariffs, cyber incidents and military conflict can interrupt supply. The seventh is social permission: power prices, water, land use, privacy and copyright can slow deployments or increase cost. The eighth is valuation duration: a large part of a twenty-year estimate may be terminal value. Small changes in sustained margin or the exit multiple can outweigh several good near-term quarters.

Our downside case is not a worst case. A stock can lose most of its value, and a compounder can cease to be a compounder. The right response is not to label every uncertain variable bullish; it is to price uncertainty, demand a margin of safety and monitor evidence. Longer time horizons increase both the potential for reinvestment and the chance that the competitive landscape changes fundamentally.

Review the chain quarterly, but separate operating evidence from the share price. For platforms, track revenue quality, net retention, paid usage, accepted task outcomes and cash after infrastructure renewal. For chips, track end-customer concentration, inventory, purchase commitments, software adoption and returns on development spending. For foundries and equipment, track utilization, yield, geographic mix, capex per incremental revenue and customer acceptance. For power suppliers, track firm orders, cancellations, capacity additions and warranty costs.

Refresh the market snapshot before using the valuation outputs. Reconcile fiscal calendars, stock splits, ordinary shares and ADS ratios, currencies, cash versus lease capex and GAAP versus adjusted profit. Remove paper investment gains from operating profitability. Retain the original filing date and source, and mark management guidance and analyst estimates differently from historical facts. A convenient data feed is a starting point for validation, not an authority that makes unit errors disappear.

Use a thesis checklist with measurable invalidation triggers. Examples include several quarters of cash conversion deterioration despite revenue growth, utilization below break-even, backlog cancellations above new bookings, or persistent maintenance capex exceeding the model. A trigger should lead to investigation and a revised valuation, not an automatic trade based on one quarter. Timing effects, working capital, acquisitions and accounting changes can explain a temporary mismatch.

The practical conclusion is selective ownership rather than indiscriminate exposure. Search for businesses that solve scarce problems, retain pricing power, reinvest at sensible returns and distribute economic cash after compensation and renewal. Then insist on an entry price that can withstand weaker growth and a less generous exit valuation. That discipline is more likely to survive twenty years of technological change than a portfolio built around today's most impressive demonstration.