Ask an AI security vendor what next quarter will cost and watch how the answer gets qualified. Depends on alert volume. Depends on case complexity. Depends on how much reasoning each investigation requires.
That’s not evasiveness. It’s what happens when the cost of your product is a metered API call to someone else.
Per-Token Inference Makes Security Spend Unforecastable
Every AI security platform built on a public model API pays per token. That cost flows to you, directly as usage-based billing or indirectly as margin the vendor has to protect.
The variables are all outside anyone’s control:
Alert volume. How many events need analysis this month. Nobody knows in advance, and a bad month is exactly when you need the platform working hardest.
Case complexity. A straightforward triage consumes a fraction of what a multi-step investigation does. Complexity is determined by what attackers do, not by your budget.
Reasoning depth. Modern models are more conversational and do more internal reasoning before answering. Every one of those tokens is billed. Model quality improvements make inference more expensive, not less.
The subsidy unwinding. Public model pricing has been underwritten by investor capital. As those providers move toward profitability, inference pricing moves in one direction.
“It’s very difficult when you don’t use your own infrastructure to price it accordingly to customers, because you have this variable of how many tokens is it going to use, how many alerts they’re going to get this month,” says Greg Scasny, CTO and Founder of IntelliThreat. “All that leads to cost variability, which means it just ends up costing your end customers more.”
That last point is the one buyers underestimate. A vendor facing unpredictable input costs prices in a risk premium. You pay for the volatility whether or not you ever hit the ceiling.
IntelliThreat Owns the Inference Layer
No third-party model API sits anywhere in the path.
Models at the edge. For the XDR product, small models run directly on the machine performing log analysis — cyber threat analysis nodes that triage inside the customer environment. Local compute, no metered call.
Owned GPU infrastructure. Heavier analysis runs on IntelliThreat’s own GPU servers, inside infrastructure IntelliThreat owns and operates.
Fixed cost structure. The GPU infrastructure was a capital investment. Once it’s running, an additional investigation costs electricity, not an invoice line.
“We can use tokens as long as we want. It’s our own hardware, it just runs,” says Scasny.
What That Buys Beyond Predictable Pricing
Pricing that holds. IntelliThreat can quote a price and stand behind it, because the underlying cost doesn’t move with your alert volume.
No incentive to under-analyze. This is the subtle one. When each reasoning step costs money, there’s pressure to cut the investigation short. On owned hardware, running one more correlation pass is free — so the harness runs as many as the case warrants.
Lower latency. IntelliThreat controls the network path between the analysis and the model. No public API round trip, no queueing behind other tenants, no rate limits during an incident.
No third-party dependency during an incident. A provider outage or rate limit doesn’t degrade your detection and response.
Data never leaves. The same architecture that fixes the cost structure is what makes Private AI genuinely private. Security logs are not sent to any external provider for inference, which is a compliance and data-residency answer as much as an economic one.
Why Smaller Models Are Enough
The obvious objection: owned infrastructure means running models smaller than the frontier, and smaller models mean worse results.
That assumes accuracy scales with parameter count. In practice, accuracy comes from the harness — the programmatic scaffold that determines what context the model receives, how its output is re-queried, and what it’s permitted to conclude. The same model in two different harnesses produces substantially different accuracy. The scaffold is doing the work.
Which means a well-scaffolded mid-size model on owned hardware delivers the outcome without the per-token exposure. Frontier models remain available as a configuration option for customers willing to pay the token costs. None have chosen to.
The Question for Your Next Renewal
Ask where inference runs, what happens to your bill in a high-alert month, and whether the vendor’s cost of goods is fixed or metered. If it’s metered, you’re carrying that volatility — in your invoice, in the risk premium, or in how thoroughly each case actually gets investigated.
Get a price that doesn’t move with your alert volume — request a demo or start a 30-day free trial.