In brief

In December 2025, venture capital firm Foundation Capital published a thesis: context graphs will constitute the next trillion-dollar infrastructure for enterprise, by capturing the “decision traces” that current systems do not preserve. This proposition triggered an immediate technical debate. Todd Blaschka identifies a structural “Logic Gap” — these graphs record the what, not the why. Gartner considers the term redundant, Verdantix the concept immature. Between real opportunity and premature hype, the state of discussions in 2026.

In short: a system like Salesforce records that a 12% discount was granted on a deal. It records neither why, nor which rules were weighed, nor who decided at 11 PM on Slack. The context graph is the archive of reasoning — not just the result. The Foundation Capital thesis says: whoever captures this memory of “whys” captures the next enterprise value. The critics reply: capturing the “whats” without a solid ontological model does not yield meaning, just a giant log.

The Foundation Capital thesis

On December 22, 2025, Jaya Gupta and Ashu Garg (Foundation Capital) published “AI’s trillion-dollar opportunity: Context graphs”. The central thesis: the next trillion-dollar platforms will not be built by adding AI to existing systems of record — Salesforce, Workday, SAP — but by capturing what these systems never see.

What the thesis calls decision traces: the traces showing how a rule was applied, where an exception was granted, why an action was authorized. This decision context is currently dissolved — it lives in Slack threads, deal desk conversations, escalation calls, and in the heads of operators.

The proposed definition of the context graph: “a living record of decision traces stitched across entities and time, where precedent becomes searchable”. It does not merely document what happened, but why it was authorized. Each new automated decision adds a trace to the graph, precedents become queryable, and the advantage compounds.

The thesis articulates a value shift: from systems of record (state recording) to systems of agents (decision recording). The previous generation of enterprise software created a trillion-dollar ecosystem by becoming systems of record. Foundation Capital projects that context graphs open an equivalent opportunity for the next generation.

Aaron Levie (CEO, Box) responded to the thesis by declaring “the era of context”: in a world where AI models become commodities, the differentiator becomes the organizational knowledge you provide them — and the critical decision traces that agents would need are found in no existing software today. The thesis reportedly reached Fortune 500 CIOs and unicorn AI founders according to Foundation Capital, becoming one of the most discussed AI ideas of the year in the weeks following its publication.

The structural advantage of startups

Foundation Capital’s argument rests on an architectural distinction: the execution path versus the read path.

When an AI agent handles a customer escalation, responds to an incident, or decides on a discount, it is in the execution path — it extracts context from multiple systems, evaluates rules, and because it executes the workflow, it can capture this context at the exact moment of decision. The structure of a decision trace is precise: what inputs were collected, what policies were evaluated, what exceptions were granted, who approved, what state was written.

Data warehouses — Snowflake, Databricks — are in the read path. They receive data via ETL after decisions have been made. By the time data arrives via ETL, the decision context is lost. These systems can say what happened, not why. SaaS incumbents can add AI to their data, but they cannot capture what they never see: cross-system decisions that occur outside their application perimeter.

Vertical agent startups, by contrast, are natively in the execution path. The sales agent startup captures the context graph of renewals. The support agent startup captures that of escalations. Foundation Capital projects that these startups will become the owners of the context graph for their domain — a structural advantage that incumbents cannot easily replicate. The incumbents will likely respond through acquisitions and by locking their APIs with egress fees to make data extraction costly.

In short: the metaphor is a court clerk versus an accountant. The clerk (execution path) hears the plea live, sees the exhibits filed, notes the argument retained — they are in the room. The accountant (read path) receives the final verdict by mail, without the context. Snowflake is the accountant. The agent startups orchestrating the workflow are the clerks. And an accountant, however competent, cannot reconstitute what was said orally between two breaks.

The Logic Gap

Todd Blaschka published a direct critique of this narrative titled “Context Graph Hype: Why the Holy Grail is Leaking”. His central argument is the Logic Gap: the gap between recording a decision and understanding its meaning. Context graphs as presented capture the what but not the why.

Without a formal knowledge graph foundation, Blaschka identifies a failure triad:

  1. Identity crisis — the entity cannot be stably identified from one system to another, from one decision to the next.
  2. Hallucinated judgment — the system generates inferences without solid semantic grounding, in the absence of a structured ontological model.
  3. Context rot — context degrades over time without an explicit maintenance mechanism.

The critique joins a broader position in the practitioner community. Gartner (Afraz Jaffri) argues that the term “context graph” is redundant: a graph implicitly implies context. Using “context” as a modifier adjective would be a marketing phenomenon, not a conceptual innovation. Subramanya N raises a different architectural tension: most enterprise decisions extract context from 6 to 10 systems simultaneously — context graphs could be fundamentally a platform problem rather than a vertical application problem.

Context graph vs knowledge graph

The terminological debate is structurally important. Positions are clear and divided.

Andreas Blumauer (Graphwise) defends the community consensus: “a Context Graph is an operationalized Knowledge Graph”. A context graph is a kind of knowledge graph enriched with a meta-layer, using standards like RDF-Star or Named Graphs to attach timestamps, provenance, and confidence scores to links between nodes. It is not a replacement — it is an evolution that adds the temporal dimension and decision lineage.

Kurt Cagle refines this position: the central idea of context graphs is to be able to determine exactly when a decision was made and why — something a knowledge graph alone can only provide indirectly, even with RDF-Star annotations. The distinction concerns operational temporality, not the underlying architecture.

Jessica Talisman is more critical. She notes numerous similarities between the language of her work “Process Knowledge Management” (November 2025) and Gupta & Garg’s article (December 2025), and considers “context graph” a rebranding. Her question remains open: “Are we discussing context relative to tokens or context designed for AI reliability?” — two meanings of the term that do not necessarily overlap.

HackerNoon summarizes the underlying tension: the absence of an accepted standard, coherent implementation model, and stable definition of the term “context graph” creates conditions for maximum vendor noise without commensurate progress. The community identifies the genuinely unresolved problems: temporal validity of facts, resolution of contradictory facts, and decision traces — gaps that existing ontologies (Schema.org, CDM, WAND) do not fill.

A convergence point emerged in early 2026: the W3C Context Graphs Community Group was proposed on February 23, 2026 by Ron Itelman (with Kurt Cagle among its supporters). Its declared mission — developing specifications, vocabularies, and best practices to resolve “contextual misalignment” between global knowledge representations and local interpretive contexts in decision systems — signals that the debate has reached sufficient maturity to justify a standardization attempt.

Limitations and maturity

Verdantix, an independent analyst firm, assesses context graphs in 2026 as “immature in enterprise, with most usage confined to pilots and research projects.” The distinction between knowledge graph and context graph is characterized as incremental rather than transformational. Challenges of data integration, ontological alignment, and governance remain entirely unresolved.

The market context of knowledge graphs is illuminating: only 27% of enterprise knowledge graphs are reportedly currently in production, most remaining static and devoid of operational context (lineage, policies, usage patterns). Context graphs are supposed to bridge this gap — but bridging a gap in a technology that is itself underdeployed is a cascading dependency.

Glean constitutes one of the few documented implementations at scale: its Enterprise Graph relies on a 6-year-old knowledge graph (mapping people and company knowledge) and adds a layer for projects, processes, and data. Its Agentic Engine 2 claims 94% task completion with this context. This is an architecture that corresponds to the Foundation Capital vision — but built on 6 years of knowledge graph investment, not on a standalone context graph deployment.

Statistics on enterprise generative AI adoption are sobering: 95% of organizations that invested in enterprise generative AI reportedly have not observed measurable ROI, and only 5% of custom AI solutions have reached production with sustained value. 40% of Chief Data Officers cite recurring data quality and integration problems as the main obstacle. These figures contextualize the thesis: the opportunity is real, but the blockers are upstream of the context graph itself.

In short: the gap between vision and field is massive. Foundation Capital projects a trillion-dollar market. Verdantix replies “immature, pilots only”. 95% of enterprises that invested in enterprise GenAI have not seen measurable ROI. Glean, one of the rare cases that resembles the promise, took 6 years to build its knowledge graph before layering on its context layer. The thesis points to a real target; the path to reach it goes through a decade of data integration, not through a self-service product.

Key takeaways

  • The thesis has solid technical grounding: current systems of record do indeed lose decision context. AI agents in the execution path have structural access to this context that data warehouses do not capture. This is not hype — it is a real architectural gap.

  • The Logic Gap is a serious critique: recording a decision trace and understanding its meaning are two distinct problems. Without an ontological foundation and maintenance mechanism, a context graph produces unstable identities and unfounded inferences. Blaschka’s triad — identity crisis, hallucinated judgment, context rot — points to concrete engineering gaps.

  • The terminological debate masks underlying agreement: Blumauer, Cagle, and the majority of practitioners converge on “context graph = knowledge graph operationalized with temporal dimension and decision lineage.” This is not an architectural break — it is an extension. The novelty lies in the application to agent workflows, not in the graph structure.

  • Maturity is low, preconditions demanding: Verdantix says “immature.” Glean took 6 years to lay the foundations. 27% of enterprise knowledge graphs are in production. Building context graphs on fragmented enterprise data, of uneven quality, without a stable standard, is an integration problem as much as an architecture problem.

  • The value capture question is open: Subramanya N asks the right question — who actually captures the value? If enterprise decisions cross 6 to 10 systems, the context graph may be a platform problem, not a vertical application problem. The answer will determine whether vertical startups, orchestration platforms, or incumbents (who are fighting back) capture the bulk of the value.