A two-axis grid of an independent AI and technology research corpus: seven coverage sub-sectors crossed against seven altitude layers of one compute stack, from the physical substrate that creates compute to the applications that spend it. It documents structural coverage breadth; it does not make recommendations.
Point-in-time rendering, July 23, 2026. Underlying framework authored June 2026; classification pass July 2026.
Rows are sub-sectors as the source research groups them, 57 company write-ups in total. Columns are altitudes: where in the flow of compute an occupant actually collects. The two axes are deliberately independent, and the tension between them is the point: a write-up filed under enterprise software can sit at the bottom of the physical substrate, and one sub-sector's governing economics refuse to map onto the axis at all, which the grid says honestly instead of forcing a fit. Cells record the mechanism at each intersection, and name the notable public occupants of each layer as any observer of the industry would name them: the roster is derived from public knowledge of who operates where in the stack, is independent of the research corpus the counts describe, and implies no mapping between named companies and the write-up counts. Every company description on this page was written from public knowledge. Click any marked cell for the occupants and the methodology that placed them.
● mechanism with named public occupants ★ the hinge (contested) badges: P perimeter · C content gate · M meter and governance click any marked cell
The compute execution layer and the routing layer run the same play: author a standard, let everyone build on it, and let the accumulated gravity do the selling. Arm runs the version one floor down, licensing processor architecture into most of the world's smartphones and roughly a third of all processor chips. But the two software chokepoints run the play in opposite directions.
CUDA, NVIDIA's programming platform, locks compute in. In practice the dominant path to the GPU for AI workloads runs through the programming model, so the switching cost points down toward the silicon and the toll is collected on every training and inference cycle. The Model Context Protocol, authored by Anthropic and donated in December 2025 to the Agentic AI Foundation (a directed fund under the Linux Foundation), lets data out: it collapses M-by-N integration work to M-plus-N, commoditizes the connectors, and pushes competition up to model quality. One toll is a wall; the other is a door.
Open compiler paths (PyTorch 2.x, Triton, AMD's ROCm) challenge the training-side lock, while deployment runtimes (TensorRT and version-pinned inference engines) deepen the serving-side lock at exactly the moment a product ships and latency starts to matter. The grid records both dynamics. Which prevails is a question it deliberately does not answer.
The content gate splits on billing form. A per-crawl meter collects in proportion to use (TollBit and Cloudflare's pay-per-crawl program are the two most visible implementations between publishers and AI crawlers); a flat license collects once regardless of use. The two forms allocate usage risk differently between buyer and seller; the grid records the split as structural and stops there.
Challenge at training, deepening lock at serving, as described in the hinge note above. Both readings are carried simultaneously; the column is carried as contested (see the Altitude 3 key) rather than silently resolved in either direction.
A chain of narrow markets, most with one, two, or three credible suppliers: chip-design software, architecture licensing, lithography and wafer-fab equipment, foundry and advanced packaging, high-bandwidth memory, and the power, cooling, and optical assembly that make rack density physically possible. Optimization does not shrink this toll: compressing the per-token cost of inference expands the workloads run on top of it (the Jevons pattern), which increases load on downstream memory and storage. A published NVIDIA research method compressing attention KV caches roughly 20x with model weights unchanged (KVTC, arXiv:2511.01815, accepted to ICLR 2026) is the current reference case for how fast that compression is moving.
Rent on the metal: the building, the power contract, the racks, sold as raw accelerator capacity or as capacity plus an inference platform. The landlord's economics are real-estate economics priced in compute, which is why the growth ceiling is grid build velocity rather than AI demand.
The seam between the hardware and software halves of the stack. Whoever owns the programming model that addresses the accelerator collects on every FLOP that crosses it, which makes this layer simultaneously the top of the silicon story and the floor of the software story: a multi-layer resident, not a filing ambiguity. Full treatment in the hinge note above.
Extraction and preparation of raw corporate documents into AI-ready form, streaming and operational stores, and retrieval optimization between the store and the model. The toll is real because bad data induces failure: enterprises must structure their pipelines before agents can be trusted with them. It is not uniformly a clean toll; pricing models built for pre-agent query volumes can be squeezed as agentic workloads change consumption patterns.
The one stretch of the stack that everyone mistakes for the destination. Substitution is one API pointer away, open-weight releases keep compressing per-token pricing toward marginal cost, and no constraint at this altitude is durable by itself. Occupants of this layer often also appear one column to the right, in routing and standards, where the structural mechanics differ; the grid records both placements.
Protocols, gateways, sandboxes, and the payment rails agents use to pay for external services. Gateways that store every model's API keys and cache upstream calls have quietly become critical infrastructure; a March 2026 supply-chain compromise at LiteLLM, a widely deployed open-source gateway (trojanized packages published through a hijacked maintainer account, disclosed by the project and documented by multiple security firms), promoted the category from utility router to security perimeter in a single incident. One recurring structural pattern at this altitude is authoring a standard rather than owning a product.
The altitude where the toll is hardest to keep. Agents erode seat pricing, so survival runs through owning the system of record beneath the agent and re-pricing from seats to hybrid metering that passes compute cost through rather than absorbing it. Most application software does not carry a durable toll; the survivors own the record the agent must write back to.
Three tolls wrap the stack rather than sitting at one altitude, so the grid shows them as per-cell badges instead of columns. Perimeter (P): security as a non-discretionary risk tax that grows with machine identities, because every agent is an endpoint to authenticate. Content gate (C): metered or licensed access to legally cleared corpus, split down the middle by billing form (fault line 1 above). Meter and governance (M): observing and governing agentic flow, because you cannot manage what you cannot trace.
The substrate taxes the creation of compute through a chain of narrow markets: chip-design software, architecture licensing, lithography and wafer-fab equipment, foundry and advanced packaging, high-bandwidth memory, and the optical assembly that ties racks together. Most links in the chain are one-, two-, or three-supplier markets, which is what makes the toll real rather than rhetorical.
Placement lens: the four-part toll test (can the ecosystem run 48 hours without you; must traffic route through you; are you pre-provisioned rather than a discretionary purchase; do you pass your costs through). Occupants listed along the chain, from design to assembly.
The execution layer is where the silicon meets its programming model, and it belongs to whoever owns that model. It is simultaneously the top of the hardware half of the stack and the floor of the software half: a multi-layer resident, not a filing ambiguity to resolve.
Placement lens: standards gravity. Full treatment in the hinge note below the grid.
Interconnect is the substrate's bandwidth sub-layer: moving data between accelerators and between racks within power and latency budgets that electrical links cannot meet at distance. Component classes here concentrate into one or two credible suppliers with long capacity lead times.
The write-ups in this row are classified onto the substrate column by function. The row label describes the market; the column records the altitude.
Placement lens: the four-part toll test, applied at the component level.
Power is part of the substrate, not an afterthought to it: the switchgear, distribution, and backup equipment between the grid connection and the rack, plus the thermal systems that keep modern rack density physically possible. This equipment is pre-provisioned and non-discretionary, specified and booked before the first accelerator arrives.
Placement lens: the four-part toll test; this cell passes the toll test on pre-provisioning and necessity.
The landlord collects rent on standing capacity: the building, the power contract, the racks. Its binding constraint is energization speed, how fast contracted power becomes connected power; the growth ceiling is grid build velocity, not AI demand.
This sub-sector deliberately spans two columns. Equipment sits in the substrate; landlord economics sit here; collapsing the two into one archetype would misstate both. The grid records the split.
Placement lens: the four-part toll test, pass-through check in particular.
Hyperscale platforms occupy the landlord column when they rent raw accelerator capacity, and they carry a second, quieter toll: identity. Enterprise agents authenticate through platform identity services, a system-of-record chokepoint that persists no matter whose model runs on the metal.
Placement lens: system-of-record necessity; perimeter wrap badge, because every agent is a machine identity.
Managed data platforms give the landlord a second altitude: gravity. Data accumulates where the compute already is, and moving it out costs more than staying.
Placement lens: data gravity as routing necessity.
Training-corpus access is where the content gate wraps the stack. Whether access to legally cleared text is metered per crawl or sold as a flat license determines who actually carries the toll; the two forms distribute the toll differently.
Placement lens: the content gate wrap; see fault line 1 below the grid.
The model layer is the thinnest toll in the building. Substitution is one API pointer away, open-weight releases keep compressing per-token pricing toward marginal cost, and no constraint at this altitude is durable by itself. Activity at this layer frequently pairs with standard-authorship activity one column to the right; the grid records both placements.
Placement lens: the four-part toll test, which this altitude mostly fails; that is the finding.
Routing taxes connectivity: protocols, gateways, sandboxes, agent payment rails. A recurring move at this altitude is authoring the standard rather than owning the product. The Model Context Protocol collapsed M-by-N integration work to M-plus-N, commoditized the connectors, and pushed competition up to model quality; its author collects gravity, not license fees.
Placement lens: standard-authorship gravity; see the hinge note.
Chip-design software is conventionally labeled enterprise software, and this row's label follows that convention. The grid places design tooling at the bottom of the substrate instead: it is a precondition for compute to exist, not an application of it. Same conventional label, different altitude.
This is the clearest case of the grid overruling a filing convention, and the kind of reclassification a flat vendor list cannot express.
Placement lens: the four-part toll test; necessity outranks the label.
The application layer taxes the end task, and it is the altitude where the toll is hardest to keep. Survival runs through owning the system of record beneath the agent and re-pricing from seats to hybrid metering that passes compute cost through rather than absorbing it.
The wraps concentrate here: security spend scales with machine identities (P), and observability meters what the agents actually did (M).
Placement lens: system-of-record survival plus the pass-through check.
Not mapped, and honestly so. The governing economics of industrial AI run through equipment fleets, dealer networks, and aftermarket attach: a structural-necessity story in which the machine is the distribution channel. That logic does not decompose onto a compute-stack altitude axis, and forcing a placement into these cells would manufacture a mapping where none exists.
Placement lens: structural necessity, which runs on a different axis than altitude; recorded as not mapped rather than forced onto the axis.