The Board Room

The loudest graph vendors hold the least data

The pricing signal arrived before the product matured, and the reason is structural. Atlassian and ServiceNow hold thin slices of customer data, development and collaboration metadata, IT service records, next to what Databricks, Snowflake, Amazon, Salesforce and Microsoft already store. Their graphs are a claim on relevance in an agent-mediated stack, not evidence of ownership, which is why they are the ones signalling a meter. Free access buys the dependency; the meter turns on after switching costs are sunk. ServiceNow already bills customers who point outside agents at its knowledge graph.

A skeptic would call the demand a marketing cycle. Neo4j, valued at $2.4B in 2021, posted a quarter with more revenue than its entire prior fiscal year, and ClickHouse's recurring revenue passed $350M on OpenAI and agent workloads. Beside strong quarters at Databricks and Snowflake, data infrastructure is capturing the spend while application vendors work to prove they still sit in the value chain. A stale 2021 mark plus that growth rate, in a category incumbents want to bundle, makes Neo4j an obvious acquisition target inside four quarters. That conversation is cheaper while still being the buyer.

The second chokepoint is authorization, and it has moved

Anthropic's enterprise-managed authentication for MCP connectors, the standard way agents reach outside tools, is generally available. Authorization runs through the identity provider, and per-user consent disappears: one approval instead of hundreds, plus a permission record that turns an audit into a query rather than a reconstruction project. The stickiness is the point. Model quality churns every two quarters; connector registries and permission histories do not. Scoping to Team and Enterprise tiers is the SSO-tax playbook applied to a layer with far higher switching costs.

Two consequences belong in the architecture review. A model vendor's control plane now decides which tools and which data reach the model. And one compromised administrator equals org-wide tool and data access, so hardware-key authentication on AI admin roles stops being optional hygiene.

Where the sources diverge

All the reporting reviewed agrees value is migrating away from the model: Gemma passed 150 million downloads under Apache 2.0, setting a zero-cost floor, and Kiro credited its spec scaffolding, not the model, for a claimed ~82% lower cost per task. The disagreement is which layer above the model wins. Databricks says its platform beats standalone graphs on completeness and freshness; the application vendors say relationships matter more than volume. Both cases arrive with vendor-sourced numbers and no disclosed methodology. The ~50% token reduction, the 82% figure, and a sponsored report concluding that agent winners have "strong, trusted data foundations" are direction, not planning inputs.

One reason to slow the wiring before widening it: a researcher-built malicious "skill", a configurable instruction set distributed much like a browser extension, coaxed Copilot into pushing Outlook, SharePoint and Teams data to an attacker-reachable proxy, and Microsoft's own skills scanner missed it. The same architecture exists on Claude, ChatGPT and Perplexity. One queryable substrate holding every relationship in a business multiplies blast radius where that detection has already failed once.

Access to your context graph is priced at zero exactly once: while your agents can still do their work without it.

Concurrency, not intelligence, was the binding constraint

Ramp liked Claude Code on day one and hit its ceiling the same day, because a local machine caps a developer at one or two agent sessions. The fix was not a better model. It was a centrally configured remote environment carrying a full development stack and provisioning in under five seconds. Central configuration also killed per-machine setup, which is the reason designers and non-developers got in the door at all. The second fix was closed-loop verification: the agent runs tests, reads telemetry, queries feature flags and screenshots its own frontend work. Ramp shipped that screenshot check roughly a year before third-party vendors did.

Once the substrate existed, the marginal cost of an internal agent collapsed. A per-team-customizable code review system that outperformed the third-party tools was built by one engineer in one week. Over 200 internal agents now run on the platform, past a million cumulative sessions. That number is the procurement consequence: every AI code review, QA, incident-assistant and BI-copilot renewal becomes a renew-or-rebuild decision priced at roughly one engineer-week of internal capacity.

The perishable part and the durable part

The temptation is to copy the plumbing and call it a moat. Vendors closed the screenshot-verification gap in about a year, which is what integration leads are worth. What compounds is measurement. Microsoft shipped Agent Lightning as a portable skill that installs inside Claude Code, Codex and Cursor, and it demands an eval suite as the price of entry. Microsoft is not fighting for the harness. It is claiming the optimization layer and running it inside competitors' products. Prompts, model choice and harnesses are all disposable, and the eval suite is the asset that survives all three.

That asset is also the gate on the price cuts. Nothing moves from a frontier model to a cheap tier without a scored regression check.

The margin sitting on the table

Marketplace data shows the top-end GPT-5.6 tier taking more than 50% of U.S. business spend inside OpenAI's family, while the cheap tier is roughly ten times less expensive. A skeptic would call that a reasonable premium for reliability. It is not a premium for reliability. It is a subsidy paid to avoid making a routing decision, and the best public enterprise reference point runs the other way: AT&T has open models at 25% of workflows and 40% of employee queries, routed through an open-source router rather than a paid one. The harvestable margin is in routing and evaluation, not in more seats and not in more capex.

Where this breaks

A 5.5-person team is a single point of failure for three quarters of a company's code output and 200-plus dependent agents. Agents hold access to sanitized production replicas and observability inside a regulated fintech, with every session visible and no opt-outs. Compute spend across a million sessions is never quantified anywhere in the account. And once review itself is delegated to an agent, model regressions propagate under thin human oversight.

The emerging fix — score each agent run for risk and confidence, then route low-risk work straight to a pull request — comes from a single case study whose vendor is a disclosed paid partner and whose delivery metrics sit behind a paywall. The mechanism looks sound. The returns are unproven. Two details survive that caveat: those agents are triggered by customer support chats and monitoring telemetry, which makes a support inbox a code-change entry point, and no confidence score should ever unlock authentication, payments, personal data paths, migrations or infrastructure permissions.

Seats buy a commodity. The verification loop between your agents and your own telemetry is the thing no vendor can invoice you for.

The deal tape is the thesis

No large platform buyout appears anywhere on the current transaction list. What appears instead is bolt-ons and sponsor-to-sponsor trades: Rotunda folding Revv into AirPro Diagnostics, Thompson Street's ATIS adding AuditMate, Linden buying ArtesRx out of Flexpoint Ford. Descartes paid roughly $100 million for Tai, and because nearly every private-equity transaction on the tape carried undisclosed terms, that print is one of the only real price points available in logistics software. Price discovery has collapsed. That is a governance problem before it is a data problem, because acquisition committees are now benchmarking against vendor-database estimates and seller assertions.

The mechanism behind the jam is unglamorous and irreversible. Assets bought with capital that cost next to nothing are marked against rates that hit 40-year highs in 2023. Financial engineering carried value creation for fifteen years and no longer functions, so general partners have to manufacture operational improvement at the moment it is hardest to produce. PitchBook's Kyle Walters puts the trap plainly: bought at 12x, maybe worth 10x, with "no real way to get out." That describes a seller with no leverage and a deadline that never formally arrives. The same tape holds the counter-example. Flexpoint Ford got out of ArtesRx while its peers sit stranded.

Three exposures, pointing in different directions

The vendor exposure is the one most teams underprice, and the cleanest proof is not in the private-equity data at all. Minimus raised $51 million, failed to find customers, and gave clients 60 days to migrate. Vendor viability is a security control now, not a procurement footnote. A single-source dependency owned by a fund past year seven carries the same migration risk with a longer fuse and no announcement. The roadmap simply stalls while support quality degrades.

What would break this call

Walters is explicit that the condition is not yet systemic, and that a real breakdown requires stacking another risk factor on top of the zombie layer: a credit-spread move, a rate shock, a first visible sponsor default. General partners hold the timing advantage unless their hand is forced, and private markets are, in Fortune's framing, quite adept at kicking the can down the road. A strategy premised on imminent distressed pricing will look foolish for several more quarters. The tradeoff worth naming is optionality bought cheaply now against capital committed on a timing call, with the trip-wire defined in advance. When forced selling starts, the winner is whoever already has board approval rather than the better thesis.

Capital, meanwhile, is not closed. It is surgical. Quintessent raised $40 million for optical interconnects in AI data centers, and Airbound raised $37 million with DoorDash participating strategically. Money is reaching the AI infrastructure bottleneck and the tooling built to monitor distressed portfolios. Everything between those two poles is being rationed, which is the market operators and their PE-backed competitors are both working in.

These sellers need a transaction more than they need a price, and that only helps the buyer who built the target list before the credit event.