The data centre industry has a power problem. The serverless hype cycle is settling into something useful. And a CDN company just got a customer satisfaction score that makes you stop and read the fine print. Here is what each of those stories tells me about building tech in mid-2026.
The AI data centre bottleneck is real, and Infinium is trying to hack it
Building a new data centre takes years. Utility interconnection queues are months or years long. Meanwhile, GPU demand keeps climbing. Infinium EdgeSites looks at the problem from a different angle. Instead of pouring concrete and waiting for grid upgrades, they put modular infrastructure into existing commercial buildings that already have electrical capacity.
The core hardware is Vector ONE, a factory-built immersion cooling platform. Each unit delivers 1MW of AI compute and can sit indoors or outdoors. No municipal water connection needed. The company claims a PUE of 1.05 and up to 70% less floor space than air-cooled gear. Multiple modules can stack to 10MW per site.
Robert Schuetzle, the CEO, put it bluntly in the article from Data Centre Magazine. He said the industry has been answering an infrastructure shortage with a construction playbook. Build new facilities. Wait for grid connections. Wait years. He argues the power already exists, the buildings already exist, and the technology now exists to put them to work.
I buy the premise. The electrical capacity in older industrial buildings is often stranded. The wiring is there but nobody uses it. Renting that spare capacity to run GPUs makes more sense than convincing a utility to upgrade a substation.
The timing also lines up with the shift from training to inference. Training needs huge clusters in hyperscale campuses. Inference needs compute closer to users. EdgeSites can drop into warehouses near population centres. That matches what the article says about the growing shift from AI training to inference.
Practical takeaway: If you operate AI workloads, look at inference latency budgets differently now. A distributed model with 10MW sites near metro areas might beat a single 100MW remote campus for applications that need sub-10ms response.
Serverless is good for your workflow, bad for your state
The DEV Community post on serverless is one of those "I wish I had read this before building my last project" pieces. It walks through where serverless helps and where it hurts, using real engineering experience.
Serverless works for event-driven workloads. File uploads, database changes, HTTP requests. You pay per execution, not for idle time. Variable traffic? Serverless auto-scales instantly. Rapid prototyping? You can deploy an API in minutes.
The pain points are also familiar. Long-running processes hit the 15-minute timeout on AWS Lambda. Cold starts add seconds of latency after inactivity. Stateful applications require extra services like Redis or DynamoDB. High steady load costs more than a fixed server.
The post recommends serverless when your workload is event-driven or intermittent, when you want minimal operational overhead, and when you need rapid scaling from zero. Avoid it for long-running or stateful processes, latency-sensitive apps, and constant high-volume traffic.
I have seen teams deploy serverless for everything because "it's the future." Then they spend months debugging cold starts and hitting timeouts. The writer is right to call it a powerful paradigm, not a silver bullet.
Practical takeaway: Map your workload characteristics against serverless constraints before committing. If your application needs persistent connections or sub-100ms P99 latency, look at containerised alternatives. Serverless is glue, not a platform.
Fastly holds a 4.8 out of 5 from actual customers
Fastly was named a 2026 Gartner Peer Insights Customers Choice for Edge Distribution Platforms. The rating is 4.8 out of 5 stars based on 92 reviews, with 95% willingness to recommend.
I have used Fastly for a few projects. The real-time configuration is the main differentiator. You can update VCL or compute code without restarting. The platform also handles WAF and bot protection, which makes it more than a CDN.
The quotes in the article from oneetechavenue include a CTO saying Fastly is "the best CDN on the market today" and an IT security manager praising "strong, reliable, and secure performance." Those are generic but the rating is concrete.
What matters here is that Gartner Peer Insights reviews are verified. Customers have to prove they bought and used the product. A 4.8 out of 5 is unusual. Most platforms in this space hover around 4.2 to 4.5. Fastly's score suggests the platform delivers on reliability and support, not just marketing.
Kim Ogletree, Fastly's Chief Customer Officer, said the recognition comes from the technology and business leaders they partner with every day. That is a careful statement. She did not say "we are the best." She said customers trust them.
Practical takeaway: If you are choosing an edge platform for AI or web delivery, look past the feature comparison tables. Check customer satisfaction scores from verified reviews. High scores correlate with fewer support tickets and faster issue resolution.
A thread connecting the three stories
All three stories are about infrastructure that is optimised for how compute and data travel today. Infinium EdgeSites moves compute closer to users and existing power. Serverless moves execution closer to events. Fastly moves content closer to eyeballs.
Each approach reduces latency by reducing distance. Each trades some complexity for speed. And each works best when you understand the constraints.
My take: 2026 is the year of knowing when to stop abstracting. Serverless is great until you hit a cold start. Edge data centres are great until your workload needs a full GPU cluster. Fastly is great until you need custom proxy logic that hits the edge compute limit. Pick the tool that fits the job, not the tool that fits the hype.