The models commoditise. The corpus does not.

The first leg of this cycle paid the people who sold the shovels. The next leg pays the people who own the ground.

Every general-purpose technology follows the same arc. The novel capability is expensive, scarce and differentiating for a period, and then it is not. Steam, electrification, the relational database, the web server and cloud compute all made the same journey: from proprietary advantage, to competitive parity, to line item. The returns did not disappear when that happened. They moved to whichever layer of the stack stayed scarce.

Frontier models are making that journey now, and quickly. Capability gaps between the leading closed models and the best open-weight models are measured in months rather than generations. Price per token has fallen persistently. Enterprises are building routing layers precisely so that no single model is load-bearing. A capability that three vendors sell, that a fourth gives away, and that a buyer can swap with a configuration change is not a moat. It is an input cost.

So the question worth answering is not who has the best model. It is what a model cannot manufacture for itself. The answer is consistent across every serious enterprise deployment: the corpus, and the meaning attached to it. A model with no privileged access to a firm's operating reality is a very articulate consultant who has never seen the books.

That is the whole thesis. Value accrues to those who have something. Something proprietary, something legally or structurally difficult to replicate, something that compounds with use. Models are no longer that something.

structural valuation discipline required

Each layer answers one question: can a competitor re-derive this with money and time? Where the answer is yes, margin is competed away.

The useful discipline is to walk the stack and ask, at each layer, what stops the next entrant. Compute is capital-intensive but ultimately purchasable, and the current scarcity is a supply condition rather than a permanent structural one. Model weights are reproducible — that is what the open-weight releases have demonstrated repeatedly. Orchestration and tooling are genuinely valuable and genuinely copyable; the feature set of every agent framework converges within two quarters.

Two layers do not behave this way. The first is the data layer, where the asset is a licensed, exclusive or historically accumulated dataset that a competitor cannot re-derive at any price because it was created by a position, a relationship or a regulatory role the competitor does not hold. The second is the context layer — the encoded map of how one specific organisation actually operates.

Both layers share the characteristic that matters: they get more valuable as AI gets cheaper. Falling inference cost expands the number of workflows worth automating, and every one of those workflows needs grounding in a real corpus with real meaning. Cheap intelligence is demand for expensive context.

Twelve businesses, sorted by what they own that a well-funded competitor could not build in three years.

The first kind is the agreed reference. S&P Global, Moody's, MSCI and the financial data aggregators do not sell information that is scarce because it is hidden. They sell information that is load-bearing because the market, and in several cases regulation itself, has agreed to treat it as the reference. An index is only useful if everyone else uses the same one. A rating only matters if the mandate is written against it. No model can synthesise that, because the asset is not the data — it is the agreement to use that data. AI raises the value here: two decades of converting documents into structured, licensed, machine-readable feeds meet a new class of consumer that queries at machine speed and is billed per feed and per use case.

The second kind is consumer behavioural data at scale, and it is the least appreciated of the four. Netflix and Spotify hold something no studio or label has ever held: a complete, individual-level record of what was actually watched or listened to, where attention was abandoned, and what was substituted instead. Applied to commissioning, that is a direct attack on the cost base of content production — fewer greenlights against less uncertainty, better calibration of what a title is worth before the money is committed, and the ability to size a production to a demand curve that is measured rather than assumed. Uber holds the equivalent for physical movement: the rider and driver record across pricing, supply response, routing and reliability, which is the substrate for matching and pricing decisions no amount of model capability can substitute for. In every case the model is the tool; the behavioural record is the asset.

The third kind is the enterprise data estate and the infrastructure sitting on it. Microsoft is the current home of a major cohort of enterprise data and the productivity exhaust that goes with it, which is a positional asset before it is a software one. SAP is where the transactional spine of a large part of the industrial economy is recorded. Snowflake holds the analytical estate, Datadog the longitudinal record of how systems actually behave, and MongoDB the operational application state. None of these is a legal monopoly and each is competitively contested, but the gravity is real: the data is already there, the pipelines are already built, and the cost of moving is paid in engineering quarters rather than dollars.

The fourth kind is the context layer itself — Palantir, ServiceNow, Atlassian, Salesforce — which is the subject of the next section.

A model cannot act inside a business it cannot read. The ontology layer is the reading — and it is now contested by four different kinds of incumbent.

Systems of record — the CRM, the ERP, the ticketing system, the issue tracker — spent two decades being the place where a business wrote down what happened. They were, in effect, expensive filing cabinets with workflow attached, and the market priced them as mature seat-based software.

Agents change the job description. An autonomous system asked to reprice a contract, reroute a shipment, resolve an incident or escalate a claim must know what those objects mean in that specific firm: which fields are authoritative, which relationships are real, which actions are permitted, and who is accountable when one is taken. That map has to exist somewhere. It cannot be inferred from a schema dump, and it certainly cannot be inferred from the public internet.

This is what the ontology layer is. Palantir's Ontology sits at the core of its AI Platform and constructs a digital replica of an organisation, organising data and logic into interconnected objects, links and actions that represent real-world concepts and their relationships. The strategic point is that the model becomes the interchangeable part and the ontology becomes the durable one. Swap the model next year and the map still holds; swap the map and the entire deployment restarts.

The contest is now genuine, and each contender is approaching the same layer from a different starting position. Palantir comes from the operational deployment and works down into the data. ServiceNow comes from the workflow graph and works out into adjacent operations, which is the closest thing to a native competitive threat because its map is already the map of how work moves. Snowflake comes from underneath, owning where the analytical corpus physically sits and building semantics upward. Microsoft comes from everywhere at once, holding both the enterprise data estate and the distribution to put an agent in front of every user of it.

Salesforce, SAP and Atlassian are the same argument arriving from the record side. Salesforce's asset was never the interface — it is the accumulated record of customer interactions, pipeline states and approved processes, which is precisely the substrate an agent needs before it is allowed to touch a customer. SAP holds the transactional truth of the physical economy, which is what any supply-chain or finance agent must reconcile against. Atlassian holds the record of how software is actually built and shipped, which is the natural grounding layer for coding agents that now write a rising share of it.

The risk to be explicit about: being the system of record is necessary, not sufficient. Several of these incumbents will expose an ontology layer and still fail to charge for it, because seat-based pricing does not capture the value of work done without a seat. That is an execution and pricing question, and it is where the differentiation inside this group will show up.

ontology layer contested pricing model unresolved

The partnership matters less as a revenue event than as a proof that the ontology layer works at the hardest possible scale.

In October 2025, NVIDIA and Palantir announced an integrated stack for operational AI: CUDA-X data science libraries, GPU-accelerated processing and open Nemotron models integrated into the Ontology at the core of AIP, with Lowe's among the first adopters, building a digital replica of its global supply chain for continuous optimisation rather than weekly node-level runs.

The more recent extension is the part that deserves attention. The two companies have taken the same stack to NVIDIA's own operations, with NVIDIA as the first deployment site for a sovereign AI architecture aimed at critical supply chains. NVIDIA's supply chain is, by its own description, among the most complex in the world: millions of parts, thousands of suppliers, a global manufacturing network, and roughly 1.3 million parts in each Vera Rubin rack. The stated ambition is to compress the path from wafer to first token.

Read this as evidence, not as a press release. Three things follow. First, the hardest available test case is now a reference customer, which is the single most effective way to sell an ontology into other complex industrial operators. Second, the architecture is deployable on-premise or in cloud, which opens the set of buyers who cannot route sensitive operations through a public model endpoint — regulated industry, defence, sovereign infrastructure. Third, and most importantly for this thesis, the model in that stack is an open one supplied by the hardware vendor. The commodity layer is explicitly commodity. The scarce layer is the map.

That is the thesis demonstrated in public, by the company with the most to gain from the opposite being true.

Stated up front, so the position can be marked against evidence rather than defended. Any one of these would be disconfirming.

Models stop commoditising. If capability gaps between the frontier and open weights widen durably rather than compressing — because of a step change in architecture, a data wall that only the largest labs can cross, or an export and compute regime that entrenches two or three providers — then the model layer keeps its pricing power and the premium sits there instead. Watch the gap between frontier and best open-weight performance on tasks that enterprises actually pay for.

Models learn the context layer. The thesis assumes ontology cannot be inferred. If sufficiently capable systems can reconstruct a workable operational map directly from raw schemas, logs and documents with little human curation, the context layer thins into a feature and the premium collapses toward the data underneath it. Watch how much implementation labour a typical deployment still requires in two years.

The consumer-data claim does not convert. Viewing and listening records reduce uncertainty at the commissioning stage, but they do not underwrite a hit, and the history of data-informed commissioning includes expensive misses alongside the wins. If better demand signal simply raises the price of talent and rights rather than lowering the cost of production, the asset is real and the margin never shows up. Watch content spend per hour of engagement, not the size of the dataset.

Privacy regimes narrow the behavioural asset. The consumer-data names hold their record under consent and jurisdictional regimes that can tighten. A rule that restricts individual-level use for training or optimisation does not touch a ratings franchise, but it goes straight at Netflix, Spotify and Uber.

Hyperscalers bundle the layer at zero price. The context and observability tooling could be given away as a compute attach, which breaks the pricing umbrella for the independent names simultaneously. This is the most under-priced risk in the group.

The data owners get disintermediated by their own customers. If large buyers pool contributed data or if regulation forces reference data open, the network position that makes an index or a rating load-bearing weakens. This is slow-moving and low-probability, but it is the tail that actually kills the franchise rather than just compressing the multiple.

And the risk that is not a thesis risk at all: price. Several of these names already embed the structural case in the multiple. Being right about where value accrues over five years says nothing about the return earned buying it today. Those are two separate decisions and should be made with two separate pieces of work.

falsifiers stated hyperscaler bundling opinion, not advice