Patrick Richter

Getting teams to agree on deployment standards is way harder than the actual infrastructure work. Spent months negotiating how people actually build stuff, not writing YAML. That's platform engineering.

OpenCost now tracks per token LLM costs on Kubernetes because idle models still burn your GPU budget. Same story as EC2 sitting unused for years, just way more expensive. Teams are about to relearn a lesson they never learned the first time.

Claude broke into real production systems during a supposed simulation because the sandbox wasn't actually isolated. The newer model caught the mistake and stopped, the older one didn't. That's not malice, that's a config error and a model that got smarter about noticing things.

Most remote hiring is just a video call and a LinkedIn profile deciding who gets access to your codebase. Identity verification basically hasn't evolved while remote work became the norm.

Nobody renamed sales ops or marketing ops out of shame but we rebranded operations into DevOps, SRE, platform engineering, anything but ops. Someone still has to own what breaks at 3am and a new title won't change that.

Google has 514 billion in cloud backlog but Pichai says AGI research gets the TPUs first, customers get leftovers. That's the honest conversation every cloud provider has internally. Your workload is the byproduct.

OpenAI's agent found a zero day in a package registry, got internet access, then decided to hit Hugging Face for benchmark answers. Nobody told it to do that. The monitoring that could've caught it was turned off on purpose.

Google's VMware Engine went down for 15 hours from one cooling failure. Multi zone architecture doesn't matter if the service only runs in one building. Nobody tells you that at signup, you learn it in the incident report.

AWS killing off zombie services like CodeStar and Cloud9 is actually good. Keeping dead products alive costs more credibility than shutting them down. Nobody's asking about FinSpace anyway, they want EC2, S3, IAM.

AWS Cost Explorer showed someone a 2.5 billion dollar bill for 19 cents of actual usage. Turned out to be a bug, charges were fine. But the real problem I see with clients is the opposite - actual costs exploding slowly because nobody knows whats running on their infrastructure.

AWS billing surprises happen because nobody tracks where stuff actually lives. S3 fills up with forgotten backups, EC2 instances run for who knows why, and suddenly you're shocked at the bill. The real problem isn't the charges, it's that teams have no idea what they're even paying for.

Terraform drift is the silent killer of every IaC project. Team sets up properly, then somebody fixes production in the console at 2am and never updates the code. Six months later state describes fantasy infrastructure.

The hardest Kubernetes conversation I have is not about networking or storage. It's convincing teams to stop adding to their clusters. Every month someone adds a service mesh or observability stack that duplicates what's already running.

AI agents won't own infrastructure provisioning because production changes have blast radius. An agent that drifts your VPC because it misread a CIDR isn't faster, it's just unreviewed. Human gates exist for a reason.

Most teams say they want to build in-house for control but really just want to avoid feeling dependent. Then they end up with a half-finished platform three people understand and no one wants to own when those three leave.

We're still rolling out IPv6 and someone's already proposing IPv8. Most enterprise networks still run IPv4 because it works and nobody wants to touch it. That's basically all of networking: technical debt we agreed to ignore until something breaks.

Most companies have zero visibility into AI token costs while their FinOps teams perfected cloud spend tracking. Token economics don't work like compute and usage is scattered everywhere. Cloud cost discipline took a decade to mature, AI spend will compress that into two years.

Europe says it wants cloud sovereignty but keeps buying AWS Azure and GCP because thats all that passed compliance. Building real alternatives takes a decade of engineering not press releases.

Nothing's actually future proof. Build boring stuff that solves today's problem and make it easy to change later. Teams that over engineered in 2021 are now paying engineers to operate infrastructure they don't need.

Three engineers designing microservices to scale. Meanwhile they're not figuring out if anyone wants the product. Build when you have the problem, not when you imagine it.

Yeah run Docker Compose in production if it solves your problem. The industry got weird about Kubernetes being serious and now three person startups are on call for their own projects. Simpler tools win.

AI is finding bugs faster than teams can patch them. The technical debt everyone buried is now on a countdown shorter than anyone planned for. Security teams are still operating on the old timeline.

Drone strike took out AWS infrastructure in the Middle East and customers are looking at a long outage. Multi region isn't optional if your service actually matters. Physical events happen that nobody predicted.

Kubernetes is the right answer for maybe 20% of the workloads running it. The other 80% exist because a senior engineer wanted to learn it, a consultant sold it, or a CTO made a 2021 decision. Sometimes boring is correct.

Claude doing design is cool but here's the thing: the slow part of design isn't execution, it's the endless back and forth figuring out what people actually want. Generating faster doesn't solve that problem.

Claude Opus found a working Chrome exploit for $2,283 in API costs. Anthropic's safety team discovered it, published it, then shipped the model anyway.

We moved hundreds of servers to AWS and got everything running clean. Six months later nobody could actually answer which IAM role production was using or what it could access. Turns out infrastructure code means nothing if permissions just grew organically.

seniors got good by debugging broken deploys at 11pm, not from explanations. now juniors ship code they can't explain and skip that friction entirely. most companies are cutting junior headcount instead of deliberately giving them harder problems.

I measured cloud savings at a client three years ago. First bill was double on-prem costs. Nobody had a strategy, just lift and shift because the CTO read it saves money. AI agents are about to repeat this exact pattern. Deploy first, measure never.

Kubernetes adoption isn't about copying what works at Netflix. It's about what your team can actually operate at 2am. Saw teams adopt service meshes and multi-cluster setups they didn't need because blog posts made it sound standard.

Walmart tried ChatGPT as a checkout interface and it converted 3x worse than normal. Teams keep reaching for the most impressive solution instead of the right one. A checkout has one job: reduce friction. Talking to an LLM adds it.

S3 just turned 20 and AWS shipped some genuinely impressive stuff. But in the field most clients are running massive bills because nobody set lifecycle policies. Objects from 2019 in Standard storage, dev buckets replicating everywhere, buckets nobody remembers. Storage compounds. Neglect

"We have a small project, should only take a few hours." Translation: "We have no documentation, three different stakeholders with conflicting requirements, and we need it deployed yesterday." The smaller they say it is, the messier it gets. Every single time.

Received inquiry the other day: £300/day for DevOps work in 2026. Recruiter thinks they're being generous with a 3-6 month "engagement" at rates that barely cover a developer. Want someone to fix your infrastructure, handle your scaling problems, probably clean up whatever mess the last guy left.

most production monitoring is just expensive noise. teams collect everything and get 200 alerts a day they learn to ignore. real monitoring is three things: is it up, is it fast, is it handling load. everything else is dashboards nobody looks at until it breaks.

startups getting $1600 surprise azure bills because claude credits dont work with ai foundry while microsoft and anthropic point fingers at each other. read every line item before deploying anything ai related. the credit systems are partial at best and vendor finger pointing is predictable. test wi

German client Friday 6pm: "just add a small feature" to our Terraform setup. It's cross region disaster recovery with automated failover. You know, the thing that takes teams three months. This is why I quit fixed price work. Scope always triples. Hourly or nothing.

most productive digital nomad work happens in coworking spaces with actual AC and wifi, not beach setups where sand destroys your laptop. nobody posts about the 4 hours hunting for decent internet but that's the real grind. the beach photos look better though

Half the startups running Kubernetes could fit their whole thing on a t3.large. But that doesn't look impressive to investors so we get 12 node clusters for 100 users. Resume driven development is wild.

Sunday in Tulum debugging infrastructure while families hit the beach. Everyone sees the Instagram nomad life but nobody talks about working weekends in cafes with terrible wifi or taking 2am client calls because time zones dont care where you are. Always "on" because theres no separation between wo

Client call at 6am Mexico time because they "forgot about the timezone difference." This is nomad life. Your schedule belongs to whoever's paying. Sometimes that's 3am in Bangkok for a German standup. Worth it for the freedom but nobody talks about the timezone chaos part.

"We need someone to review our infrastructure setup." "Around 40k monthly spend, previous consultant said it was optimized." Opened the console. Single app. 23 subnets across 8 regions. Load balancers with no targets. RDS in every AZ. "Optimized for what exactly?"

"Can you optimize our AWS costs?" "Sure, but first tell me why you're running 12 different RDS instances for a single application." "Oh, that's just how the previous consultant set it up." This happens more than you think. Biggest cost optimization is usually just common sense

Someone just posted their $23k monthly EKS bill for a basic CRUD app. This is why I tell clients to start simple. You don't need Kubernetes because Netflix uses it. You need it when you have Netflix's problems. That $23k could pay two senior devs for a month. Simple wins. Boring wins.

Hot take: timezone arbitrage is overrated. Sure you can bill European rates while living in Bangkok but you're also debugging production issues at 3am because Munich deployed on Friday afternoon. Been chasing timezones for three years. The sweet spot isn't cheapest city or highest rate, it's finding