| The RatioA weekly newsletter on reliability economics | |
|
The Number
0.77 median ratio, the 50,000–99,999 employee band
The benchmark's biggest group, companies with 50,000–99,999 employees, spends under $1 preventing failures for every $1 fixing them
The benchmark's largest cohort, 45 responses, posts the lowest median ratio of any size band: 0.77. For comparison: 100,000+ employees sit at 5.63. The 10,000–49,999 band hits 8.88. The 1,000–9,999 group lands at 4.48. Only the 50K–99,999 band falls below 1.0.
The Ratio Take
You'd expect big companies to overspend on reliability. This group proves otherwise. It piles up risk faster than it builds prevention, and more staff won't fix that. Think of it like a hospital that keeps adding ER doctors but never funds diagnostic imaging, every patient still arrives in crisis because nobody caught anything upstream. The likely cause: treating reliability spend as a fixed slice of the IT budget instead of tying it to what failures actually cost. It's not the smallest companies that under-invest in reliability. It's one specific size band, and it's a big one.
|
|
This Week in Reliability
AI Observability Gets Real
The enterprise AI infrastructure stack is maturing past the proof-of-concept phase, forcing teams to confront operational realities: debugging AI agents in production, connecting LLMs to telemetry data, and measuring reliability when deployments jump from weekly to 1,000× monthly. The prevention vs. reaction budget conversation is shifting from 'can we build this?' to 'can we keep it running?'
|
| 1 | The New Stack · Primary evidence—deployment velocity forcing tooling rethink AI teams are deploying at 1,000× the cadence of traditional software teams, creating pipeline bottlenecks and observability gaps that weren't designed for this velocity. The Ratio Take This is the deployment frequency inflection point that breaks traditional reliability tooling. When your pipeline assumes weekly releases and suddenly faces hourly model updates, every manual gate becomes a budget leak—either you staff 24/7 reviews or you automate the risk assessment, and most teams are bleeding OpEx on the former while their AI competitors chose the latter six months ago. 1,000 deploys/month turns every manual gate into a staffing crisis. |
| 2 | Datadog · Vendor response—tooling adapting to AI-first operations Datadog's MCP Server connects AI agents directly to observability data—metrics, logs, traces—so agents can query telemetry during debugging workflows within governance controls. The Ratio Take This is observability vendors recognizing that the next customer isn't a human in PagerDuty at 3am, it's an AI agent trying to self-heal before the human wakes up. The prevention ROI flips when your telemetry becomes machine-readable in context: instead of paying engineers to correlate five dashboards during an incident, you're paying once to structure the data so agents can do it autonomously. Telemetry's new customer is the AI agent, not the on-call engineer. |
| 3 | Microsoft / The New Stack · Strategic signal—hyperscaler pivot to reliability as differentiator Microsoft's Foundry platform prioritizes AI reliability over raw capability, arguing enterprises care more about consistent uptime than bleeding-edge model performance. The Ratio TakeMicrosoft is signaling that the next enterprise AI budget fight isn't model selection—it's SLA enforcement, and they're positioning reliability as the moat. Enterprise AI competition just shifted from capability to uptime SLAs. |
| 4 | Lorin Hochstein · Conceptual foundation—why AI ops is harder to instrument Modern systems hide complexity through abstraction and decomposition, but failures emerge in the gaps between components where no single team has visibility or ownership. The Ratio TakeThis is the reliability tax that scales with architectural complexity—AI systems amplify it by adding non-deterministic components that fail in ways your monitoring can't predict. Failures live in the gaps your abstractions were designed to hide. |
| 5 | Charity Majors · Cultural tension—the prevention vs. velocity tradeoff at org level AI advocates rush to deploy before the opportunity closes; AI skeptics fight to prevent technical debt and operational chaos from accumulating faster than teams can remediate. The Ratio TakeBoth sides are optimizing different failure modes, and neither has budget to lose—this is the reliability culture split that determines whether your AI investment becomes a profit center or a perpetual ops subsidy. Speed vs. entropy: pick which failure mode you're optimizing for. |
| 6 | SRE Weekly · Adjacent signal—incident recurrence as reliability debt metric SRE Weekly's latest issue curates reliability engineering articles, including repeat incident analysis and tooling for surfacing root causes instead of symptoms. The Ratio TakeI know, I know. But the sponsorship message stopped me: 'Your team solved this incident last month. Why is it back?' That's the entire AI reliability problem in one sentence. Repeat incidents mean you fixed the symptom, not the system. |
| 7 | KORE1 · Economic baseline—labor cost context for AI reliability investment SRE salaries in 2026 range from $130K to $210K base depending on level and location, with big-tech outliers pushing higher; includes cost analysis for hiring teams. The Ratio TakeIf you're deploying AI at 1,000×/month velocity and your SRE headcount plan still assumes pre-AI hiring ratios, this is your budget wake-up call—reliability engineering labor costs are the denominator in your prevention ROI calculation. SRE salary bands set the floor for your prevention budget. |
|
The Challenger — Contrarian Take
Popular belief: "A prevention ratio below 1.0 means under-investment."
Under-investment relative to what? The U-curve says total cost, prevention plus failure, bottoms out at some optimal ratio. The benchmark measures the ratio. It doesn't locate the optimum. And nothing puts the sweet spot at 1.0. That's a round number, not a derived one.
A cohort with older, more entangled systems may have a genuinely lower optimum, since each prevention dollar covers less surface. If so, 0.77 isn't a deficit. It's close to right.
Here is the thing. Underestimate failure cost, and the real optimum sits higher. Then the deficit returns. Below 1.0 isn't under-invested. It's unmeasured.
|
|
The Ratio is a weekly newsletter by Florian Hoeppner. Early edition, you are one of the first.
Forwarded this? Get The Ratio in your inbox every week.
|
Take the assessment → reliabilityeconomics.com/benchmark Reply to this email with your take.
How this newsletter gets better: Every link is tracked and every vote is counted. Sections that get 👎 get replaced. Topics that get 👍 stay. No personal data is stored — just click and vote counts per section. Your feedback shapes the next edition.
|
|
|