Tomorrow's ☁️ Stack - Issue #1: When Infrastructure Reality Hits
This week: $44.5B wasted on cloud resources nobody's using, AWS outage triggers FTC warnings about cloud concentration, and Gartner predicts 40% of AI agent projects will be canceled by 2027.
💸 $44.5 Billion in Cloud Waste Projected for 2025
Enterprises will waste $44.5 billion on unused cloud resources in 2025 - that's 21% of total infrastructure spend, according to Harness's "FinOps in Focus" report. The culprit isn't lack of visibility. 52% of engineering leaders say the real problem is the disconnect between FinOps and dev teams. Teams know exactly what's wasted - idle instances, oversized databases, orphaned resources. But it takes an average of 31 days to identify and eliminate waste.
Our take: Visibility tools are everywhere. Every dashboard shows the same idle EC2 instances. The bottleneck is execution capacity - knowing about waste and fixing waste are completely different problems. This is the use case for agents that take action instead of generating reports.
⚠️ AWS Outage Triggers FTC Warning on Cloud Concentration
AWS US-East-1 went down for 15 hours on October 20, taking Fortnite, Snapchat, Coinbase, and Robinhood offline. The outage affected roughly 32% of global cloud infrastructure. Former FTC Commissioner Rohit Chopra responded: "The extreme concentration in cloud services isn't just an inconvenience, it's a real vulnerability." The Big Three control 60-65% of the market, making regional failures systemic risks.
Our take: When 32% of the internet depends on one region of one provider, we have an architecture problem. Multi-cloud sounds expensive until you calculate the cost of 15-hour outages.
🤖 Gartner: 40% of AI Agent Projects Will Be Canceled by 2027
Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value, and inadequate risk controls. Security teams blocking production access and lack of governance frameworks are the real killers. The technology works - it's the deployment model that doesn't.
Our take: This isn't an AI failure, it's an architecture failure. If your AI agents can't get past the CISO, you don't have an AI problem - you have an infrastructure problem.
🔐 70% of Leaked Secrets Remain Active Years Later
GitGuardian detected 23.8 million new credentials leaked on public GitHub in 2024 - a 25% year-over-year increase. Worse: 70% of secrets leaked in 2022 are still active today. A U.S. Treasury breach in 2024 was traced to a leaked API key. Despite secret scanning built into GitHub, the problem grows 67% annually.
Our take: Secret scanning is spell-check for security - helpful but insufficient. The real fix is architectural: if infrastructure agents never upload credentials to SaaS platforms, there's nothing to leak. Can't leak what you never send.
🚀 Qovery Adds AI Agents to DevOps Platform
Qovery launched AI DevOps Copilot on November 6 - multiple agents that automate environment provisioning, CI/CD optimization, and FinOps practices through natural language. The agents are trained on 25 million applications and 30 million infrastructure operations. The goal: turn DevOps engineers into orchestrators instead of finishers.
Our take: The shift from "do the work" to "oversee the agents" is real. But only if security approves the deployment model.
🪦 "Kubernetes is Overkill" Goes Viral Again
Multiple posts about companies moving away from Kubernetes hit the front pages this week. The pattern: teams spend more time managing K8s than building features. Alternatives like Docker Compose, Fly.io, and serverless are gaining traction. 84% of organizations are moving away from self-hosted K8s toward managed services.
Our take: K8s makes sense at scale but is overkill for many teams. CloudShip runs Station on K8s not because it's trendy, but because our customers already have it deployed. Leverage existing infrastructure instead of forcing adoption.
📊 CloudZero Ships Kubernetes Cost Visibility at KubeCon
CloudZero launched Kubernetes optimization capabilities at KubeCon North America (Nov 10-13). The platform allocates 100% of K8s spend at hourly granularity by cluster, namespace, label, and pod. It connects costs to business outcomes - cost per customer, per feature, per team.
Our take: K8s cost visibility is table stakes now. The next evolution is automated optimization - agents that rightsize workloads without human intervention.
💭 What We're Thinking About: The Deployment Trust Gap
Four stories this week share a thread: technology that works in demos but fails in production.
Gartner says 40% of AI agent projects will be canceled. $44.5B wasted because teams can't execute on FinOps recommendations. 70% of leaked secrets stay active forever. It's not a capability problem, it's a deployment problem.
Here's the pattern: teams build incredible AI agent prototypes. Demos blow minds. Security shuts it down in production. The agent needs AWS credentials, kubectl access, database permissions. Giving those to a SaaS tool is a non-starter for most CISOs.
The industry's response? Add more controls. Scoped credentials, time-limited access, audit trails. That's fine, but it treats the symptom. The root cause is architectural - we're trying to safely give production credentials to someone else's cloud.
What if the premise is wrong? What if infrastructure agents shouldn't run in a vendor's cloud at all?
Self-hosted runtimes flip the model. Agents run on your servers. Credentials stay on your infrastructure. The SaaS platform (if you use one) only receives structured outputs - never secrets.
If 40% of AI projects fail due to deployment challenges, the answer isn't better controls on SaaS agents. It's not using SaaS agents for infrastructure in the first place.
This week's AWS outage reinforces it. When concentration becomes vulnerability, distribution becomes strategy. Same principle applies to AI agents.
🚀 What We're Building
Station v0.8.1 shipped with improved MCP security controls. We hit 412 GitHub stars. Working on FinOps agent templates that terminate unused resources instead of reporting them.