A new Cockroach Labs survey found that 83% of global tech leaders expect AI-driven demand to cause their data infrastructure to fail within the next 24 months. A third expect failure within 11 months.
The framing across the industry is urgent: AI is overwhelming our systems. Infrastructure can't keep up. Without massive investment, everything breaks.
Here's what's strange about that narrative. If 83% of companies knew their brakes would fail in 11 months, they wouldn't keep driving. They'd pull over and fix the brakes. Yet these same organizations continue the exact approach that's creating the strain.
That's because the infrastructure panic isn't about AI succeeding too fast. It's about AI strategy failing at scale.
The Pattern No One Is Connecting
Three data points tell the real story.
First: DDN's 2026 State of AI Infrastructure Report found that 65% of organizations say their AI environments are too complex to manage. Over half (54%) have delayed or canceled AI initiatives in the past two years. The environments are complex not because AI is working, but because organizations are running so many fragmented experiments that the overhead alone creates chaos.
Second: Deloitte's State of AI in the Enterprise 2026 confirms that 75% of organizations remain stuck in pilot mode. They haven't moved 40% or more of their AI experiments into production. The infrastructure isn't straining under production AI workloads. It's straining under the weight of endless experimentation.
Third: Harvard Business Review reported this week that 60% of AI-related layoffs are anticipatory, not based on actual AI implementation. Companies are cutting headcount in anticipation of AI gains that haven't materialized, while simultaneously claiming they need more infrastructure to handle AI demands that aren't delivering value.
The pattern: organizations are making major decisions, on infrastructure, on workforce, on strategy, based on predicted AI capabilities rather than demonstrated ones. They're creating the crisis they claim to be preparing for.
Pilot Purgatory Creates Infrastructure Load Without Value
I've written about the dabbling problem before: enterprises launching experiments instead of committing to operational change. The 95% pilot failure rate isn't just wasted investment. It's the source of the infrastructure strain everyone's panicking about.
Consider what happens when an organization runs dozens of AI pilots simultaneously:
Each pilot needs compute resources, even when it doesn't produce results. Each experiment creates data pipelines, storage requirements, and integration points. Each proof-of-concept adds complexity to an environment that 65% already describe as unmanageable. And when 54% of these initiatives get canceled, the infrastructure they consumed doesn't automatically get cleaned up.
The Cockroach Labs survey found that 100% of respondents expect AI workloads to grow, with 60% predicting increases of 20% or more. But growth in what? Production AI generating business value? Or more pilots adding load while delivering nothing?
The same survey found that 77% expect AI to drive at least 10% of all service disruptions in the coming year. That's not a prediction about AI succeeding. That's a prediction about AI chaos spreading.
The Leadership Disconnect Is the Root Cause
Here's the most revealing statistic: 63% of respondents say their leadership teams underestimate how quickly AI demands will outpace existing data infrastructure.
Think about that. Who approved the AI initiatives creating those demands? The same leadership.
Organizations are launching AI projects without infrastructure strategy, watching those projects create strain, then blaming the strain on AI's inherent demands rather than on the lack of strategic coordination. The infrastructure crisis isn't happening to these organizations. They're doing it to themselves.
Spencer Kimball, CEO of Cockroach Labs, captured the dynamic: "Leaders aren't worried about whether AI works. They're worried their infrastructure won't hold up once AI is always on."
But here's the thing: AI isn't always on. For 75% of organizations, it's perpetually experimental. The infrastructure strain comes from running experiments at scale, not from AI delivering value at scale.
Security Caution Is the One Rational Response
The Dynatrace Pulse of Agentic AI 2026 report found that 52% of enterprise leaders cite security, privacy, or compliance concerns as the top blocker to scaling agentic AI. Industry coverage often frames this as excessive caution holding back progress.
Given everything else in the data, it's the opposite. Security caution is rational.
I've written about agentic AI as an insider threat: AI agents with system access can be compromised and turned against you. When 83% of organizations predict infrastructure failure, giving autonomous agents access to those failing systems isn't cautious innovation. It's reckless.
The Dynatrace data shows that 44% of organizations still use manual methods to review AI agent communication flows. Only 13% deploy fully autonomous agents. This isn't organizations being slow to adopt. It's organizations correctly recognizing that they haven't built the foundation for safe autonomy.
If your house is on fire, you don't hand the keys to a robot and hope it figures out the evacuation plan.
The Anticipatory Trap
The HBR research on AI layoffs reveals a pattern that extends beyond workforce decisions. A survey of 1,006 global executives found:
- 39% have made low-to-moderate headcount reductions in anticipation of AI
- 21% have made large headcount reductions in anticipation of AI
- Only 2% have made large reductions related to actual AI implementation
As the researchers put it: "The job losses and slowed hiring are real, even though companies are still waiting for generative AI to deliver on its promises."
Oxford Economics analysis is blunter. Their research suggests "firms don't appear to be replacing workers with AI on a significant scale," and that companies may be using AI as cover for routine headcount reductions. Deutsche Bank analysts predict "AI redundancy washing will be a significant feature of 2026."
This is the same logic driving infrastructure panic. Organizations aren't making decisions based on what AI is actually doing. They're making decisions based on what they've been told AI will eventually do. The anticipation itself becomes the disruption.
What Would Discipline Look Like?
In Navy EOD, we didn't run experiments on live devices. We didn't make decisions based on what we thought a bomb might do. We assessed the actual situation and executed with a plan.
The disciplined approach to AI infrastructure would look like this:
Stop measuring AI activity. Start measuring AI value. The Cockroach Labs survey tracks expected workload growth. But workload from what? If 95% of pilots don't produce ROI, workload growth isn't a success metric. It's a warning sign. Organizations should track production AI workloads separately from experimental ones.
Kill pilots that aren't on a path to production. The 54% cancellation rate in the DDN data isn't a failure. It's the beginning of rationality. But those cancellations come too late, after the infrastructure load has already been added. Set kill criteria upfront. If a pilot doesn't hit specific milestones in 90 days, shut it down and reclaim the resources.
Match infrastructure investment to production value. The hyperscalers plan to spend over $600 billion on AI infrastructure in 2026. Enterprise organizations are following suit, investing in capacity for AI workloads that aren't yet delivering. Match infrastructure scaling to demonstrated production value, not to anticipated experimental growth.
Treat AI agents as the identities they are. The security concerns blocking agentic AI aren't irrational fear. They're appropriate caution. As I've written, AI agents with enterprise access are identities, and like any privileged identity, they require governance. Build that governance before scaling access.
The Real Crisis
The Cockroach Labs report frames infrastructure readiness as a financial risk. They found that 98% of companies estimate one hour of AI-related downtime would cost at least $10,000, with nearly two-thirds projecting losses exceeding $100,000 per hour.
Those numbers are real. But the cause isn't AI succeeding too fast. It's organizational chaos compounding.
The real crisis isn't that AI is overwhelming infrastructure. It's that undisciplined AI adoption is overwhelming organizations that confuse deployment with strategy.
We saw this with Microsoft's panic over Claude Cowork: enterprise AI has failed because it generates content instead of doing work. The infrastructure strain is the systems-level manifestation of the same problem. Organizations are generating AI activity instead of generating AI value. The activity creates load. The value doesn't materialize. And leadership blames the technology.
The Path Forward
The 83% predicting infrastructure failure aren't necessarily wrong about the outcome. They're wrong about the cause.
AI won't overwhelm infrastructure because it's too successful. AI will overwhelm infrastructure because organizations keep launching experiments without operational plans, making anticipatory decisions without demonstrated value, and blaming technology for problems created by strategy.
The organizations that avoid infrastructure crisis won't be the ones who invest most heavily in capacity. They'll be the ones who invest most deliberately in discipline: killing failed pilots, scaling only production workloads, and making decisions based on what AI is actually doing rather than what vendors promise it will eventually do.
The infrastructure panic is real. But it's self-inflicted. And the solution isn't more infrastructure. It's better strategy.
