NVIDIA has just made an argument that should matter far beyond the semiconductor industry.
In a recent post, the company described AI-factory compute as an emerging investable infrastructure asset, supported by financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR that are designed to mobilize more than $500 billion in third-party capital over time. The underwriting logic is familiar: customer quality, demand, utilization, cash flow and residual value.
That matters because it signals a broader transition. Artificial intelligence is no longer being treated only as software expense. It is becoming productive infrastructure. NVIDIA puts the point plainly: an AI factory turns energy and data into valuable intelligence.
But that raises the next economic question.
If capital markets are learning how to finance the infrastructure that produces intelligence, how should insurance markets measure the autonomous systems that deploy that intelligence into economic activity?
That is the missing layer.
From Compute to Agency
The emerging AI economy can be understood as a chain: Capital → AI Factory → Model → Agent → Business Process → Economic Output
NVIDIA and its financing partners are primarily addressing the upper half of that chain. They are asking whether AI infrastructure is productive, durable, redeployable and capable of generating sufficient economic return to support financing.
Insurance must answer a different set of questions about the lower half. What happens when an AI system is no longer merely generating an answer, but is authorized to act?
What happens when it can:
- allocate cloud resources,
- approve invoices,
- negotiate contracts,
- purchase inventory,
- transfer funds,
- alter production systems,
- manage customer accounts,
- or execute workflows without continuous human approval?
At that point, the economic significance of the system cannot be captured by the cost of the model or the compute underneath it.
The relevant question becomes: How much economic value has been placed within the sphere of this system’s agency?
That is the problem Agent Insurable Value is intended to measure.
Agent Insurable Value
Agent Insurable Value, or AIV, is a framework for estimating the economic exposure created when an organization delegates meaningful authority to an AI system.
AIV is not the market value of the software. It is not the replacement cost of the model. And it is not simply the cost of the compute consumed by the agent.
A relatively inexpensive agent could influence hundreds of millions of dollars in procurement, revenue, payments or operational activity.
Conversely, an AI system consuming enormous amounts of compute may have very little authority over economically consequential decisions.
The value of an agent therefore has to be separated from the value of its infrastructure.
AIV focuses instead on factors such as:
- economic dependency,
- delegated authority,
- transaction authority,
- operational criticality,
- resource-consumption authority,
- data dependency,
- intellectual-property exposure,
- concentration risk,
- governance maturity,
- and human oversight.
The more economic activity entrusted to the agent, and the greater its ability to act autonomously, the greater the exposure.
NVIDIA’s Argument Sharpens the Insurance Question
NVIDIA’s financing model is particularly useful because it reinforces a distinction insurance markets need to make. The economic value of AI infrastructure is increasingly measurable through variables such as utilization, cash flow and residual value.
AI agents require a parallel but different measurement system.
Consider two systems.
The first is an AI writing assistant used by a marketing department. The second is an autonomous procurement agent that influences $100 million in annual purchasing and can independently approve transactions up to $500,000.
Both may use the same foundation model. Both may consume similar compute. But their insurance exposures are radically different. The difference is agency.
One produces content. The other controls economically significant actions. That difference should appear in underwriting.
Compute Is Revenue — and Compute Can Also Be Loss
NVIDIA’s formulation that “compute is revenue” is useful because it highlights the productive nature of AI infrastructure. But autonomous systems introduce the opposite possibility as well.
For an AI agent, compute can also become loss.
An improperly constrained agent could:
- provision unnecessary GPU capacity,
- generate runaway API usage,
- create infinite execution loops,
- purchase excessive cloud resources,
- acquire unnecessary licenses,
- or repeatedly execute costly automated workflows.
These are not traditional cyberattacks. There may be no hacker, malware or unauthorized network intrusion. The system may simply be exercising authority exactly as it was permitted to do — badly. That is why autonomous resource allocation deserves its own risk category.
In the InclusionScore framework, this is Autonomous Resource Allocation Risk. And it is only one of several emerging AI-native exposures.
Others include Autonomous Economic Decision Risk, Autonomous Operational Failure Risk, Intellectual Property Exposure Risk, Personal Data Exploitation Risk, AI Concentration Risk and Informational Labor Risk.
The common feature is that the loss originates from the economic behavior of AI, not merely from the compromise of an information system.
The Underwriting Model Has to Follow the Economic Model
Traditional insurance is ultimately concerned with three variables: Frequency × Severity × Correlation
How often might something go wrong? How large could the loss become? And could the same underlying dependency cause many insureds to experience losses simultaneously?
Agent Insurable Value provides the exposure layer beneath those questions.
A simplified underwriting framework might look like:
Economic Exposure × Delegated Authority × Business Dependency × Control Adjustment = Agent Insurable Value
That exposure can then be evaluated through expected-loss assumptions:
AIV × Frequency × Severity × Correlation Adjustment = Expected Loss
This is not intended to replace actuarial judgment. It is intended to give underwriters a more structured object to evaluate.
The assumptions should remain visible, challengeable and adjustable. An underwriter may disagree with a modeled event frequency or severity estimate. That is healthy. The value of the framework lies in making the economic assumptions explicit.
Governance Is Necessary but Not Sufficient
Standards such as ISO/IEC 42001 and frameworks such as the NIST AI Risk Management Framework provide increasingly important structures for governance, accountability, monitoring, documentation and risk management. But governance alone does not answer the insurance question.
Two companies may have comparable governance programs while deploying agents with radically different economic authority. An agent that drafts procurement recommendations and an agent authorized to execute purchases may both satisfy a governance checklist. Their potential severity is still different. That is why AI assurance must connect governance to economic exposure.
The relevant chain is:
Governance Evidence → Economic Exposure → Delegated Authority → Loss Modeling → Insurance Readiness
The insurance market needs all five.
The Next Capital Market for AI
NVIDIA argues that its financing partnerships represent the beginning of an open capital market for AI infrastructure. That is likely correct. But infrastructure financing is only one financial architecture the AI economy will require.
The next one will be risk transfer.
As autonomous systems become responsible for larger portions of enterprise operations, organizations will increasingly ask whether those systems can be insured. Carriers will then face a problem. They cannot price what they cannot measure. And they cannot reliably measure agentic risk by asking only whether a company “uses AI.”
They need to know what the AI can do, what economic activity depends on it, what resources it can control, what data it consumes, what systems it can modify, what controls constrain it and what shared dependencies could create correlated loss. That requires a new underwriting language. Agent Insurable Value is an attempt to build one.
Infrastructure Produces Intelligence. Agency Produces Exposure.
NVIDIA is helping establish a framework for answering one of the defining economic questions of the AI era: What is productive AI infrastructure worth?
Insurance markets now face the complementary question: How much economic value have organizations entrusted to autonomous intelligence, and what happens when that agency fails?
An AI factory turns energy and data into intelligence. An AI agent turns intelligence into action. And once intelligence is authorized to act, it creates a new category of economic exposure.
Capital markets are beginning to price the infrastructure.
Insurance markets now need to price the agency.
Was this article valuable?
Here are more articles you may enjoy.
Jury Finds No Gun Defect, Sides With Sig Sauer in Unintended Discharge Case 
UnitedHealth Investor Suit Alleges Security, Governance Lapses 
Allstate Introduces Large Language Model, ALLIE 
AI Data Center Boom Is ‘Maxing Out’ P/C Insurers, AIG CEO Says 
