French observability startup TSUGA raises €30 million as AI agents push monitoring costs to breaking point

AI is often presented as a cost-cutting technology

Task automation, streamlined operations, productivity gains — the dominant narrative is one of companies able to produce more with fewer resources.

Yet as AI agents move out of labs and into operational processes, a different reality is emerging. Every decision made by an agent, every interaction between models, every call to an external tool generates a growing volume of data that must be monitored, stored, analyzed, and governed. AI isn’t just creating new streams of value — it’s also creating new operational debt.

It’s in this context that Tsuga is announcing a $35 million raise, close to €30 million, led by General Catalyst and Singular, with participation from DST Global Partners and Quantumlight. Founded in Paris in 2024, the company argues a simple thesis: the architecture underlying modern observability is no longer suited to the era of autonomous agents.

An industry built for the cloud era

Observability is one of the most understated yet critical layers of modern software. Its job is to let technical teams understand what’s happening in their infrastructure by collecting logs, traces, and metrics.

For fifteen years, the market has been structured around players like Datadog, Splunk, Dynatrace, New Relic, and Elastic. Their model is fairly simple: customers send their data to the provider’s infrastructure, which stores, indexes, and analyzes it. The more the volume grows, the higher the bill.

This logic tracked perfectly with the rise of cloud computing. As companies adopted distributed architectures, microservices, and multi-cloud environments, the need for visibility grew — and observability platform revenues grew right along with it.

For a long time, the interests of providers and their customers seemed aligned: growing infrastructure mechanically created more data, and therefore more value. The arrival of artificial intelligence, however, is fundamentally changing that equation.

When every agent becomes a telemetry factory

An AI agent doesn’t behave like a traditional application. When a user queries a classic system, only a handful of events are typically generated: a request, a response, a few service calls. When an agent is involved, the chain becomes far more complex.

The system might call on several models, invoke external tools, query different databases, generate chains of reasoning, trigger other specialized agents, and then produce a final response.

Each step produces its own telemetry — prompts, tokens, API calls, execution graphs, confidence metrics, intermediate decisions. Observability is no longer just an infrastructure question; it’s becoming a question of understanding decision-making mechanisms itself.

This shift creates a paradox: AI is supposed to cut operational costs, yet it simultaneously increases the need for oversight.

In some organizations, monitoring costs are now growing almost as fast as spending on the models themselves.

AI may not be the real culprit

Blaming this cost inflation solely on artificial intelligence would be an oversimplification. Companies have already been grappling with an explosion in infrastructure spending for several years. Storage is getting more expensive, architectures are growing more complex, data flows are multiplying, and distributed systems are generating more signals to monitor.

AI acts more as an accelerant than as a sole cause — observability may simply be the visible symptom of a broader phenomenon: the ongoing growth of digital complexity.

That distinction matters, because it shapes what kind of solution is needed. Is this an AI-specific problem, or a structural problem of the cloud economy itself?

The emergence of a new market

One thing is certain: traditional metrics are no longer enough. Companies no longer just want to know whether an application is working properly. They want to understand why an agent made a given decision, which tools it used, which models were involved, and how much confidence can be placed in the result.

This is giving rise to a new software category. Concepts like AI Observability, Agent Observability, AI Governance, and AI Traceability are starting to converge. Behind the sometimes-different terminology lies the same underlying need: making AI systems auditable.

As companies deploy agents across finance, HR, legal, and industrial functions, the question of accountability becomes central. Understanding how a decision was made is no longer optional — it’s becoming an operational requirement, and soon a regulatory one.

An already crowded market

Tsuga isn’t entering an empty field, however — established players quickly spotted the opportunity. Datadog is already developing advanced model- and agent-monitoring features. Dynatrace is pushing its Davis AI offering. New Relic, Splunk, and Elastic are progressively enriching their platforms with capabilities tailored to AI workloads.

At the same time, a new generation of specialists has emerged. Arize AI, Langfuse, Helicone, and WhyLabs focus on model observability, prompt analysis, hallucination detection, and performance tracking for generative systems.

So the question is no longer whether a market exists — it’s how to differentiate within it.

The real bet: architecture

This is precisely where Tsuga is trying to stand out. Rather than presenting its innovation as an extra feature, it’s directly challenging the industry’s dominant architecture.

Instead of centralizing data on its own infrastructure, the platform is deployed directly inside the customer’s cloud environment. Data stays in AWS, Azure, Google Cloud, or a sovereign cloud — it never passes through Tsuga’s own systems.

The argument is twofold: on one hand, this approach cuts the costs tied to data duplication and transfer; on the other, it addresses growing concerns around sovereignty and governance.

This strategy reveals a strong underlying intuition: observability data is itself becoming a strategic asset. Traces now contain prompts, decisions made by agents, business information, and sometimes sensitive data. For some companies, outsourcing that data is becoming just as fraught as handing over customer data.

The question is therefore no longer purely technical — it’s becoming regulatory and economic as well.

The unexpected return of software bundled with services

Another notable point: Tsuga isn’t just selling a platform — it’s also highlighting teams of engineers dedicated to helping customers continuously optimize their observability environment.

This approach echoes recent shifts in the AI market. After two decades of standardized SaaS, several software categories are swinging back toward hybrid models blending product and expertise.

Tsuga’s real competitors

The most dangerous competition may not come from Datadog or Splunk, but from Microsoft, AWS, and Google. The hyperscalers already control the infrastructure, the data, the observability tools, and, increasingly, the AI models themselves.

That gives them everything they need to natively bake these capabilities into their own platforms — the main strategic risk facing any startup in this category.

If AI observability becomes a standard feature of cloud environments, differentiation will need to rest on something beyond technical capability alone.

A battle that’s only just beginning

The tech industry has spent the last three years building models, copilots, and agents. It’s now discovering that value doesn’t lie solely in the ability to automate — but also in the ability to control that automation.

The cloud era produced its monitoring champions. The AI era could produce its governance champions.

By raising close to €30 million, Tsuga is betting that this new infrastructure layer will become just as indispensable tomorrow as observability became yesterday. The success of that thesis will depend less on the company’s ability to monitor agents than on its ability to answer a question every organization will soon have to face: who watches the systems that make decisions on our behalf?