Here is a number that should stop every enterprise AI team in their tracks: inference costs dropped 100x over the last two years. Enterprise AI spend went up 300% over the same period.
And Gartner projects that 40% of agentic AI projects will be shut down by 2027 — not because the technology failed, but because the economics became unsustainable.
I heard this signal at the AI Agent Conference in New York, where I attended as a founder, investor, and LP in Firsthand VC. Across multiple sessions — from enterprise CTOs, VC investors, and practitioners running production agentic systems — the economics of agentic AI emerged as the most underappreciated risk in the space. Everyone is talking about what agents can do. Almost nobody is talking about what they cost at scale.
Why the Math Gets Dangerous at Agent Scale
Traditional AI deployments have predictable costs. A user sends a query, the model responds. Token consumption is bounded by the interaction.
Agentic AI breaks this model entirely. A single agent workflow — analyzing a document, browsing the web, calling multiple APIs, drafting a report, requesting human review — can generate tens of thousands of tokens in a single run. Multiply that by hundreds of concurrent workflows running continuously. The cost curve is not linear. It is exponential.
One VC panelist described a live incident: an agentic workflow hit a broken tool call, entered a retry loop, and generated over $10,000 in token costs before anyone noticed. This is not a theoretical risk. It is happening to production systems right now.
The Fine-Tuning ROI Warning Nobody Is Hearing
A speaker at the conference made a point that landed hard: Fortune 500 teams are wasting significant resources fine-tuning custom models for problems that the next base model update will solve for free in six months.
This is a direct challenge to the AI investment thesis of many enterprises — and many VCs. If you are building a moat on a fine-tuned model, ask yourself: is the moat in the model, or in the data and domain knowledge that trained it? Because the model will be commoditized. The domain data will not.
The enterprises winning the long game are not the ones with the best fine-tuned models. They are the ones with the best proprietary data, the clearest workflow instrumentation, and the governance infrastructure to know exactly what their agents are doing and what it costs.
Where the Real Investment Opportunity Is
1. Orchestration and Cost Governance Infrastructure
The companies building the control planes — tracking token consumption per workflow, enforcing budget limits, killing runaway agents before they generate catastrophic spend — are solving the most immediate pain point in enterprise AI. This is survival infrastructure for any company running agents at scale.
2. Observability and Audit Trail Systems
Datadog's Bits AI SRE agent is already in GA production: an agentic on-call engineer that auto-triggers on incidents and performs root cause analysis. Their core insight: observability is the verification layer for the agentic era. As AI agents automate more workflows, the bottleneck shifts from creation to verification. The companies that own the observability layer own the trust layer. In regulated industries, trust is the product.
3. Grounded, Verified Data Infrastructure
Multiple sessions converged on one architectural requirement: agents must be grounded in verified, referenceable data to produce reliable outputs. Ungrounded agents in compliance-heavy environments — financial services, healthcare, government — are not just unreliable. They are a regulatory liability. The companies building the licensed, structured data pipelines that agentic systems can trust are building infrastructure every AI enterprise will eventually need.
The Asia Angle
Asian enterprises are not insulated from these dynamics. If anything, they face them with less operational experience managing LLM cost curves at scale. The companies that emerge as the cost governance, observability, and data infrastructure layer for enterprise AI in Asia — built with deep understanding of the region's regulated industries — will find themselves in an extraordinarily defensible position.
At N+, we are specifically looking at founders building in these infrastructure categories. Not the agents themselves. The plumbing. The picks and shovels. The layer that makes every other AI investment defensible.
“Everyone has access to the same APIs. Your moat is not the model. It is the domain data, the governance infrastructure, and the proprietary workflows you have built around it.” — Conference C-Suite Panel
Building AI Infrastructure That Compounds?
N+ is looking for founders who see the infrastructure opportunity, not just the application layer. Let's build it together.
Start a conversation →