The AI value chain
Every layer of the AI stack — from agents at the apex to lithography at the base — rated on moat, growth, margin, cost to run, and security exposure. Read it as a ledger, or drill into it as a map.
The apex layer captures the user, not always the margin. Prices compress as base-model capability converges. Winners defend with distribution, memory, and workflow, not model access, which everyone rents.
Category-defining reach, but every rival ships comparable capability within months.
Strong in code and regulated settings where reliability outweighs raw novelty.
The one horizontal assistant with a captive billion-user surface to push into.
Sharp product, structurally exposed to Google and to model-vendor encroachment.
Fastest-monetizing AI category. Real moat is codebase context and developer habit, but every entrant rides the same foundation models, a thin layer over someone else’s compute.
Proof the application layer can move fastest, and that it rents its intelligence.
Doesn’t need to be best; it ships where developers already are.
A feature race where the model does the work and the wrapper takes the margin.
Distribution shifts to whoever the developer already trusts for the model.
Where durable application value concentrates: regulated data, embedded workflow, and switching pain buy pricing power the horizontal assistants lack.
The template for defensible vertical AI: regulated domain, deep integration.
Health-system integration and reimbursement alignment make it sticky.
Charging for resolved outcomes, not tokens, is the defensible business model.
The layer where fat margin hides, if the domain has data and switching pain.
Incumbents don’t need the best model; they own the seat. AI is a feature that raises contract value on software customers already locked in.
The purest distribution play in AI: monetize AI on seats you already own.
Data gravity and workflow lock-in convert to AI pricing power.
Owns the system of action, which is where agents get deployed.
Capex-heavy race with thin per-token economics. Moat exists at the capability frontier and erodes the moment open weights close the gap. Value migrates to whoever monetizes distribution on top.
Ahead on reach, structurally dependent on Azure compute and continuous capex.
Competes at the frontier while leaning into reliability-sensitive buyers.
Structurally advantaged: owns compute, data, and the surfaces to ship into.
Capital and talent can buy proximity to the frontier, not a durable moat.
The single most important substitution force in the stack. Open weights turn "the model" into infrastructure, pushing durable value up to apps and down to compute. Watch this line to price every layer above it.
A deliberate scorched-earth play that resets the entire base-model economics.
Credible open models with a regulatory tailwind, thin economics.
The clearest signal that training-cost moats are softer than assumed.
Global open-weight competition keeps the price floor falling.
Economically rational for narrow tasks. Fragile long-term: general models get cheaper faster than specialized ones can defend a niche.
Good-enough small models erode the case for paying for frontier calls.
Defensible only where the data is proprietary and the domain is narrow.
Shifts some inference off the cloud, capping a slice of datacenter demand.
Glue code. Real but hard to monetize; model vendors keep absorbing orchestration into their own APIs, hollowing the layer.
Ubiquitous and hard to charge for as agent APIs go native.
Useful primitive, thin as a standalone business.
The vendors are eating this layer; that is the whole risk to it.
Retrieval matters; the standalone database is contestable. When a checkbox in Postgres does most of the job, the specialist has to win on scale and ops, not existence.
Has to out-operate a free Postgres extension to justify itself.
Capable and free, which is exactly the pricing problem.
The commoditizer: good-enough retrieval with zero new vendor.
A margin-thin arbitrage on GPU rental. Differentiation is speed and cost, both of which erode as open engines and hardware improve.
Competing on price for a service the open stack can self-host.
Convenience layer over commoditizing compute.
Free, fast, and the reason hosted-inference margins stay thin.
Underbuilt and structurally necessary as agents ship to production. The layer most likely to consolidate into a defensible category; worth watching for entrants.
Gets sticky once it sits inside CI; early leader position is open.
Production monitoring is a real need as agents fail in subtle ways.
Necessary for regulated deployment; likely bundled over time.
The landlords. They set compute price, vertically integrate into silicon to escape NVIDIA, and capture recurring enterprise spend. Structurally advantaged, but their own custom chips are the tell on where GPU pricing power is headed.
Building its own silicon is the clearest signal on NVIDIA dependence.
Owns the highest-profile model tenant and the enterprise relationships.
The only hyperscaler that controls its own accelerator at scale.
Late but buying share with capacity commitments and concentrated deals.
Explosive revenue, brutal economics: debt-funded GPU fleets that depreciate on NVIDIA’s release cadence. The forensic risk sits in financing structure and customer concentration, not demand.
The purest neocloud test case: demand is real, the balance sheet is the risk.
Same treadmill: fleet value tracks NVIDIA’s next launch.
Differentiates on power access, still exposed to depreciation cycles.
Growth driven by policy rather than product. Durable only where law forces it; commoditizes wherever it doesn’t.
Exists because regulation requires it, not because it competes on merit.
A genuinely defensible niche gated by clearance and compliance.
Policy and capital create instant scale with uncertain economics.
High revenue, low margin: passing NVIDIA silicon through to buyers. Value is captured up (chips) and down (cloud), squeezing the integrator in the middle.
Revenue tracks GPU allocation; it earns an assembly spread, not a moat.
Distribution and service, squeezed between silicon and cloud.
Pure contract build; the definition of a replaceable middle.
The interconnect is a quiet chokepoint: training clusters are bottlenecked by the network, and NVIDIA owns much of that fabric. Under-appreciated leverage.
Positioned as the open alternative to NVIDIA’s proprietary fabric.
Extends the GPU moat into the network the cluster depends on.
The silicon under most Ethernet fabrics, sold to everyone.
A physical bottleneck as clusters scale. Optics scarcity is a genuine constraint, not a narrative; supply lead times gate buildout.
Component scarcity gives real, if cyclical, pricing leverage.
High-value silicon at the center of the interconnect scaling problem.
Next-gen bottleneck-breaker; unproven at volume, high option value.
You can’t rack GPUs you can’t power or cool. This has flipped from afterthought to gating factor, one of the most defensible growth lines in the chain.
Direct leverage on every megawatt of AI capacity built.
Air cooling fails at Blackwell densities; this becomes mandatory.
The unglamorous electrical layer now gating datacenter timelines.
The apex predator of silicon. CUDA is a decade-deep software moat, not just faster chips. Every layer above pays the NVIDIA tax, and every hyperscaler is spending billions to escape it. That escape effort is the one crack to monitor.
Near-total pricing power this cycle; the chip everyone queues for.
The workhorse; depreciation risk if Blackwell supply catches up.
Rivals can match silicon faster than they can match the software estate.
Where NVIDIA’s moat gets tested. Hyperscaler in-house silicon and merchant challengers won’t dethrone CUDA soon, but they cap pricing power and re-rate the whole compute layer over time.
The proof that a credible non-CUDA accelerator can run at scale.
A direct attempt to cut the NVIDIA tax on its own fleet.
The only merchant GPU credibly chasing NVIDIA; ROCm is the hurdle.
Arms the buyers building their way off NVIDIA; a picks-and-shovels win.
Interesting performance angles, unproven durability against incumbents.
The clearest "valuable or replaced?" verdict in the map. AI training value migrated to GPUs, where Intel is a non-factor. Arm (Graviton, Apple) is eating general compute on efficiency, and AMD took server share. The x86 franchise still prints cash, but it sits in the one vector that is flat-to-declining while every high-growth vector routes around it. Former chokepoint, increasingly substitutable.
Still cash-generative, but the AI buildout routes around it.
Executed the share shift x86 leadership failed to defend.
Perf-per-watt wins in the cloud where power is the constraint.
Owns its stack end to end; irrelevant to datacenter, potent on device.
The under-covered chokepoint. High-bandwidth memory gates GPU output; supply is tight and sold out ahead. A three-player oligopoly with genuine pricing power, arguably a cleaner scarcity story than the GPUs it feeds.
Sets the pace on the memory that gates every top GPU.
Capable challenger; qualification timing is the swing factor.
The domestic supply-chain hedge in a two-player Korean race.
Broadcom in particular sits at two chokepoints at once: switching silicon and co-designed accelerators for hyperscalers. Under-narrated leverage on the AI buildout.
Double exposure to the buildout; wins whether the fabric is Ethernet or the chip is custom.
High-value silicon across interconnect and bespoke accelerators.
The physical bottleneck under all AI silicon. NVIDIA, Apple, AMD, and the hyperscaler ASICs all fabricate here. The moat is measured in years and hundreds of billions of capex. The tail risk is geopolitical, not competitive.
If any single company is load-bearing for AI, this is it.
The real near-term bottleneck within the foundry; advanced packaging is sold out.
The bull case for Intel is here, not in CPUs: a credible Western leading-edge foundry with policy tailwinds. But it is unproven at volume yield versus TSMC and burning extreme capex to prove it. Optionality with real execution risk, the opposite profile from its legacy franchise.
The whole thesis rides on yield at volume, still unproven.
Marquee customers would prove the model; absence would indict it.
Structurally capable, persistently behind TSMC on advanced-node yield. Matters as a second source, but hasn’t converted capability into leadership.
The standing second source that never quite catches the leader.
If any node is truly irreplaceable, it is this one. No leading-edge chip exists without ASML’s EUV machines, and there is no second source. A monopoly at the base of everything, and the ultimate geopolitical control point.
A literal monopoly; export policy on these machines shapes the whole industry.
Extends the monopoly forward another node generation.
The hidden chokepoint most AI narratives skip. You cannot design a modern chip without this duopoly’s tools, a software toll booth in front of all silicon, with software-grade margins and switching costs measured in years.
Half of an unavoidable duopoly with recurring, high-margin revenue.
The other half; every chip company is effectively a customer.
The pick-and-shovel layer beneath the foundries. Concentrated, high-margin, and export-controlled, durable exposure to every fab that gets built.
Sells to every fab regardless of which chip wins.
Process-critical tools with few substitutes at the leading edge.
Concentrated supply base; export controls add leverage and risk.
Yield is everything at advanced nodes, and KLA effectively owns the inspection that protects it. A quiet, high-margin near-monopoly in its niche.
Advanced-node economics depend on the tools it dominates.
The literal raw material. Unglamorous, concentrated, and impossible to skip, the kind of supply-chain base that only gets attention when it breaks.
A quiet duopoly on the substrate every chip starts from.
Choke risk hides here; a few suppliers, export-sensitive.
As models converge, data becomes the differentiator. Rights-cleared proprietary corpora and quality labeling increasingly separate winners, but the vendor layer (labeling shops) is itself contestable as models self-generate training signal.
Real business, exposed as models increasingly self-label.
The actual durable moat as base models commoditize.
Could erode the labeling vendors it competes with.
Compute is now gated by electrons, not just chips. Power availability is re-rating from macro backdrop to primary investable theme: the buildout is capped by what the grid and generation can deliver.
The literal ceiling on how fast capacity can come online.
High-conviction long-term supply, slow to build, heavily hyped.
The pragmatic near-term answer to grid constraints.
Cuts through every layer: model theft, prompt injection, supply-chain integrity, data leakage. Under-provisioned relative to how fast AI is shipping to production, a whitespace category, and a systemic risk when it fails.
Nascent tooling for a threat surface expanding faster than the defenses.
Hard, unsexy, and increasingly mandatory for regulated buyers.
Where breaches will actually happen as agents touch live systems.