The same AI wave that has made software developers faster is quietly making infrastructure engineers’ jobs harder — and a new startup that spent three years being built inside Sequoia Capital’s own server room thinks it has a solution that no one else has tackled head-on. Sequoia incubated the company from 2023 and spun it out as an independent entity today.

Empirik, an AI infrastructure company that officially spun out from Sequoia Capital today, raised $21 million in seed funding from Sequoia, Canapi Ventures, and Alumni Ventures. The company’s core bet is simple and specific: rather than helping engineering teams understand what went wrong after a system fails, Empirik intervenes upstream, modeling the likely effects of a change before it is deployed. That upstream focus puts it in a different lane from every other AI site reliability engineering (SRE) tool on the market — and from the five companies highlighted in TechTimes’ own roundup of the AI SRE space published last October.

The timing is not accidental. Gartner’s 2026 Market Guide for AI Site Reliability Engineering Tooling projects 85% enterprise adoption of AI SRE tooling by 2029 — up from under 5% in 2025. In the same report, Gartner warns that 90% of those same organizations will experience an AI-caused outage in that period, even as they continue using AI for speed and scale gains. Empirik is betting it can be the tool that closes that gap.

Why the Problem Is Getting Worse, Not Better

Enterprise infrastructure has always been complex, but the AI coding revolution is adding pressure at a new rate. When tools like Cursor and Claude Code allow developers to ship code significantly faster than before, every additional deployment is another opportunity for a configuration change, a new service dependency, or a version mismatch to silently cascade through interconnected systems.

The financial stakes are rising in parallel. According to Splunk and Cisco’s 2026 “Hidden Costs of Downtime” report, conducted with Oxford Economics and surveying 2,000 executives across the Global 2000, unplanned outages cost the world’s largest companies a combined $600 billion annually — up 50% in just two years. Each organization in that group loses an average of $300 million per year to outages, or more than $900,000 per hour. The median enterprise downtime event now costs $9,000 per minute, according to ITIC’s 2025 Hourly Cost of Downtime Survey.

Ninety-three percent of organizations surveyed by Spacelift in its 2026 State of Infrastructure Automation report have already experienced an AI-caused infrastructure incident. The culprit, according to the research, is AI-generated infrastructure code being adopted faster than governance frameworks were designed to handle.

From Sequoia’s Server Room to Stealth Startup

Empirik’s origin story is unusual in an industry that usually draws a bright line between investors and operators. Avon Puri joined Sequoia Capital in 2020 as chief digital and information officer, bringing more than a decade of infrastructure experience from Rubrik and VMware. Around 2023, as large language models began demonstrating their potential for reasoning across complex systems, Puri and fellow Sequoia IT leader Sudheer Dhurjati recognized that AI could be applied to infrastructure failure prevention — not just incident response after the fact.

The two built their initial product inside Sequoia’s own enterprise IT environment. That choice meant Empirik was stress-tested in a real, complex production environment from day one — not a synthetic lab. When Sequoia decided to spin the company out, it recruited Kartik Chandrayana as CEO. Chandrayana previously served as chief product officer at Quantum Metric and as Salesforce’s vice president of observability — two roles that gave him a direct view of the operational challenges Empirik is trying to solve.

How an AI Traffic Cop for Infrastructure Works

Existing observability platforms — Datadog, Splunk, Dynatrace, New Relic — are built to answer the question “what broke?” They collect telemetry (logs, metrics, and traces) from running systems and alert engineers when something goes wrong. Root-cause-analysis tools like Causely go one step further, automatically identifying which upstream component triggered a failure. Incident-response tools like Resolve.ai, which raised $125 million at a $1 billion valuation in February 2026, automate the investigation and remediation workflow after an incident has begun.

Empirik operates at none of those stages. It intercepts the change itself.

As software deployments, configuration updates, and network changes flow through an organization’s infrastructure pipeline, Empirik monitors each one and uses AI to model how that change is likely to ripple across the full dependency graph — the web of interconnected services, databases, cloud APIs, and internal tools that make up a modern enterprise stack. Based on that analysis, it acts as what Sequoia partner Bogomil Balkansky called an autonomous “traffic cop”: automatically permitting low-risk changes to proceed, adding guardrails and conditions to medium-risk ones, and escalating the highest-risk updates to human engineers for manual review before deployment.

The critical distinction is sequence. Traditional observability tells engineers what happened. Empirik tells them what a change is likely to cause before it happens — and acts on that prediction before the first alert fires.

AI Acceleration Creates the Problem, Then Supplies the Solution

Chandrayana’s analogy for Empirik’s market position is precise: what tools like Cursor and Claude Code did for software developers, Empirik wants to do for infrastructure engineers. Both categories of tool automate enough routine cognitive load that skilled practitioners can operate at a fundamentally different pace.

But the analogy conceals an important asymmetry. AI coding tools make developers faster, which means more code is shipped, which means more changes enter infrastructure — more deployments per day, more configuration updates, more version bumps — each one a potential cascade if something is misconfigured. The same AI wave that is the product’s analogy is also the direct cause of the market need Empirik addresses.

The market for the solution is growing rapidly. The global AI observability market was valued at approximately $2.94 billion in 2025 and is forecast to reach $44.2 billion by 2035, expanding at a compound annual growth rate of 31.1%, according to Next Move Strategy Consulting. Gartner’s forecast of 85% enterprise AI SRE adoption by 2029 represents a near-vertical adoption curve from today’s single-digit baseline.

Enterprise Validation Before Public Launch

Despite operating in stealth until today, Empirik has already signed enterprise customers. Named accounts include S&P Global, Guardant Health, and at least one major consumer packaged goods company, according to the company. The customer list gives the launch announcement more credibility than a typical seed-stage startup’s claims would carry.

Balkansky acknowledged the competitive landscape directly, describing Empirik as operating in a complementary layer to existing AI SRE platforms rather than as a direct replacement. Specifically, he named Resolve.ai and Sequoia-backed Traversal — which focuses on dynamic resource provisioning and load balancing for AI workloads — as peers that Empirik works alongside. That positioning is strategically smart at this stage: the AI SRE market is crowded (a mid-2026 landscape guide mapped 64 tools across seven categories), and a startup that can argue it fills a gap nobody else addresses is easier to sell than one that must displace an incumbent.

Does Sequoia’s Server Room Give Empirik a Real Edge?

The claim that Empirik was “built inside Sequoia” is marketing language, but it points to something real. The practical advantage is not the Sequoia brand; it is that the product was deployed in production at a real enterprise — one with complex, high-stakes infrastructure — and survived that environment for approximately three years before the company went public. Most seed-stage infrastructure tools have far shorter production track records when they first seek enterprise customers.

The risk, which no early-stage product escapes, is that a single VC firm’s IT environment — however sophisticated — is not the same as deploying across the heterogeneous stacks of hundreds of enterprise customers. Empirik’s dependency graph must be reconstructed for each new environment, and the accuracy of its change-impact predictions is only as good as the telemetry it can access and the dependency relationships it can infer. The company has not publicly disclosed how well the system performs on change-impact prediction accuracy, false positive rates, or coverage gaps in partially instrumented environments.

What an AI-Caused Infrastructure Crisis Looks Like

The Gartner warning that 90% of organizations will experience an AI-caused outage by 2029 deserves context. Gartner’s meaning is specific: as autonomous AI systems — AI agents, AI SRE tools, AI-assisted code deployment — are granted more authority over production infrastructure, errors in those systems’ reasoning or in the training data underlying them will produce failures that no human reviewer caught. The irony of Empirik’s market position is that an AI-caused outage scenario is exactly what its product is designed to intercept, yet the product’s own AI reasoning is subject to the same limitations.

Balkansky was candid about this in the Sequoia-backed Traversal’s own published materials on AI SRE, which note that “humans must remain in the loop to govern complex architectural failures and unpredictable AI agent behavior.” Empirik’s traffic-cop model — where high-risk changes are escalated to human engineers rather than automatically blocked — reflects the same design philosophy. The product is not a full autonomous gate but a triage and escalation layer.

Empirik has not disclosed the specific technical architecture underlying its dependency-graph inference or the LLM stack it uses for change-risk reasoning, nor the specific error rates achieved in its Sequoia production deployment. For enterprise buyers evaluating a tool at this stage, those are the questions that will determine whether a pilot converts to a contract.

Frequently Asked Questions

What makes Empirik different from Datadog, Splunk, or existing observability tools?

Existing observability platforms collect telemetry — logs, metrics, and traces — and alert engineers when something breaks. They answer “what broke?” Empirik operates before failures occur, analyzing each change entering enterprise infrastructure and predicting how it will affect dependent services. It acts before the alert fires rather than in response to one. This is a fundamentally different position in the infrastructure engineering stack, not a substitute for existing monitoring.

Is AI actually causing more infrastructure outages, not fewer?

Yes, based on current research. Spacelift’s 2026 State of Infrastructure Automation report found that 93% of organizations have already experienced an AI-caused infrastructure incident. Gartner separately predicts that 90% of organizations will experience an AI-caused outage by 2029 — even as they continue using AI SRE tools — because the speed and volume of AI-generated changes to infrastructure outpaces human review capacity. Empirik is specifically designed to address that gap.

What is the cost of an enterprise IT outage in 2026?

According to Splunk and Cisco’s 2026 “Hidden Costs of Downtime” report, the Global 2000 collectively lose $600 billion per year to unplanned outages — a 50% increase in two years. The median enterprise downtime event costs $9,000 per minute, according to ITIC’s 2025 Hourly Cost of Downtime Survey. For 41% of large enterprises, a single hour of downtime costs between $1 million and $5 million or more.

How did Empirik end up being built inside a venture capital firm?

Avon Puri, who joined Sequoia Capital in 2020 as chief digital and information officer after more than a decade running infrastructure at Rubrik and VMware, identified around 2023 that large language models had reached a capability level where they could reason about infrastructure dependencies. He and fellow Sequoia IT leader Sudheer Dhurjati built an initial version of Empirik inside Sequoia’s own IT environment, which served as both a proving ground and first customer. Sequoia then recruited Kartik Chandrayana as CEO and spun the company out as an independent entity.