I Built an Open-Source Alternative to AWS — Daalu.

Added: 2026-05-11

[00:00:00]If you're running anything on a cloud service today, there's a very good chance you're paying 10 times what you actually need to. And with AI driving up the demand for compute, that gap is only getting wider, not smaller. So, I spent some time building Daloo, an open-source alternative to traditional cloud services. Because while everyone is focused on AI models and the software ecosystem around it, not too many are paying attention to the underlying infrastructure that everything depends on. Daloo runs on hardware you can buy off the shelf, costs 90% less for the workloads most people actually run and ships the same APIs and the same developer experience. In this video, I'll walk you through why the hyperscala model is broken for the vast majority of workloads. the real numbers on what owning compute actually costs and what changed for me when I stopped renting.

[00:00:56]Then over the next series of videos, I'll show you step by step how to build your own production grade cloud using Daloo and how to run real production applications on it end to end. If you're done overpaying for cloud and you want control back, you're in the right place.

[00:01:11]Here's the reality. The real bottleneck in AI isn't ideas or how to implement them. There are millions of ideas and plenty of good tools to bring them to life. The real bottleneck is compute resources at the foundation of building those ideas and the people who own that compute capture the value everyone else is trying to create. So the question becomes, do you want to rent that power forever or do you want to own it? So here is something I learned the hard way. Most cloud users are massively overpaying for cloud infrastructure right now. Not by 10%, not by 20%, but by 70, maybe even 80%.

[00:01:50]I found this out by doing a detailed audit of my resource utilization over a period of of time. I analyzed CPU and GPU utilization, memory utilization, and realized I was at just below 20% utilization, meaning I was using only 20% of what I was paying for. And it's easy to miss this because when you look at your cloud bill, you see the total cost of your infrastructure.

[00:02:19]What you don't see is how much of that infrastructure you're actually using.

[00:02:23]For most people running steady workloads, the answer is shockingly low.

[00:02:28]10%, maybe 15, which means you're paying for a massive amount of capacity that just sits there idle.

[00:02:35]>> I know this because I was doing the exact same thing. I was paying $3,500 a month for a server that was barely doing anything. CPU sitting at 10 to 15%.

[00:02:47]Memory underutilized.

[00:02:49]Network nowhere near saturated.

[00:02:51]And yet, every single month, the bill came in exactly the same, very high. It wasn't a grand vision that made me decide to build my own cloud. It wasn't some big startup idea. It was a spreadsheet. I was staring at my AWS dashboard looking at the numbers and something didn't make sense. If I was actually using the full capabilities I needed to actually use the AI models I was working on to serve production workload and do inference for my real users, the price would be closer to 16,000.

[00:03:25]That's when it clicked. This isn't pricing based on usage. This is pricing based on convenience and lock in. So, I did something that honestly sounds a little crazy. I decided to build an open-source alternative to the hyperscalers in 6 months, and it actually works. It wasn't one thing that pushed me here. It was a thousand small frustrations.

[00:03:49]I was running predictable workloads, longunning services, data pipelines, early AI experiments. Nothing bursty, nothing hypers scale, just steady, consistent compute. And yet every month I was paying for flexibility I never used. Then one night something small broke. A minor scaling issue triggered autoscaling.

[00:04:13]And it didn't just scale, it scaled badly. Costs spiked almost instantly.

[00:04:20]And when I tried to debug it, I couldn't really see what was happening.

[00:04:24]Everything was abstracted away behind APIs. That was the moment everything changed. I realized I don't control cost. I don't control performance and debugging. I don't control the processing of my most important asset, my data. I was renting a black box. And the rent was due every month whether I used it or not. So I sat down and did the math. Real uncomfortable math.

[00:04:50]$3,500 a month. That's $42,000 a year for a workload that could realistically run on a single decent on-prem server. So, I priced out the hardware. A super micro 36bay 4U server with a SAS3 backplane about $1,000.

[00:05:10]A GPU like the Nvidia RTX Pro 5000 Blackwell with 72 GB of RAM about $4,000. So, roughly $5,000 total for a very powerful machine. And the thing is all my existing applications could run on that. Not just run, run comfortably with plenty of headroom. But then I thought about something important.

[00:05:33]Redundancy, high availability, failover.

[00:05:37]Because the argument cloud sales people make is yeah, but we have availability zones in different geographical locations which makes it reliable. While this is a fair argument, it is not difficult to meet that requirement. So instead of one server, let's say you deploy three three identical GPU nodes in different locations. Now you have proper failover, proper redundancy, but your total hardware cost isn't $5,000 anymore. It's $15,000. Now amotize that over 5 years. That's $3,000 per year.

[00:06:11]Add power, maintenance, rack space.

[00:06:14]Let's say $500 per node per year. That's another $1,500 per year. So now you're at roughly $4,500 per year all in for a redundant production grade setup.

[00:06:27]Compare that to $42,000 per year on a hyperscaler. That's still a difference of over $35,000 every single year. Let that sink in for a second. Over $35,000 saved every year. and you still have all the reliability, redundancy, and other features that make a production-grade cloud service. When the cost difference is that large, it forces you to look at what cloud pricing is really based on.

[00:06:56]We're not just paying for compute, we're paying for convenience, for someone else to worry about the reliability of the infrastructure our applications run on.

[00:07:04]And that is genuinely valuable. But here's what most people don't realize.

[00:07:09]You don't have to give that up to save money. You can own your hardware and have it professionally managed. You can still hit 99.999% of uptime or whatever SLA you actually need without renting everything forever.

[00:07:25]The trade-off isn't binary. It's not hyperscala or chaos. There's a middle ground. Most people just never explore it. So at that point I kept thinking surely someone has already solved this.

[00:07:37]Surely there's a clean developer friendly self-hostable cloud that feels like a hyperscaler but runs on your own hardware. There wasn't. What I found instead were pieces. Kubernetes for orchestration, Seph for distributed storage, OpenStack for compute and virtual machines, Metal Cubed and Tinkerbell for server provisioning and life cycle management. Keycloak for identity and authentication.

[00:08:06]powerful tools but completely disconnected. No clean opinionated system that just works end to end. And at the same time, people kept asking me the same questions. How do I run this cheaper? Can I host this myself? How do I avoid egress costs? How do I run AI without burning money? Can I get hyperscalerike APIs without the hyperscaler? That gap is still wide open. Now, not everyone agreed with me.

[00:08:35]Someone I respect told me, "Why are you doing this? Just use a hyperscaler.

[00:08:41]You're basically trying to rebuild a trillion dollar company's product." And honestly, they weren't completely wrong.

[00:08:48]Hyperscalers solve real problems. Scale, reliability, convenience. But they were wrong about one key thing. They assumed the cost of cloud made sense for the value you're getting out of it. It doesn't. They assumed every workload needs hypers scale elasticity. Most don't. And they ignored something that's becoming more important every single day. Control. Control over your infrastructure. Control over your data.

[00:09:16]Control over your costs. The insight is simple. Hyperscalers are optimized for flexibility. What I'm building with Daloo is optimized for efficiency. Two completely different games. I'm not trying to replace hyperscalers. I'm building an alternative for a specific group of people. People running predictable workloads, people tired of overpaying, people who don't need to rent a Ferrari just to drive to the grocery store. And once you see it this way, the business model is obvious. If a service runs at steady state and you can reasonably expect it to keep running for the next 5 years, then it is worth taking a closer look at how you can save on this recurring cloud cost. Because the value isn't in more features, it's in efficiency. At the core of all this is a simple belief. Cloud shouldn't mean renting computers forever. For steadystate workloads, AI inference, data pipelines, longunning services. The hyperscaler model just doesn't make sense. You're paying a premium for optionality you don't use. The real opportunity is this. give people the same power, the same APIs, the same developer experience, but without the cost structure. So, here's what I want you to do. Go look at your cloud bill.

[00:10:32]But don't just look at the bill. Also, take a look at your actual utilization of the cloud resources you are paying for. This can be done by going into your cloud provider metrics console. Look at your CPU utilization, your memory usage, your GPU utilization. If you're running AI workloads and don't just check a single moment, look at it over time, over hours, over days, because that's where the truth is. For example, if you paid for 100 hours of GPU, how many of those hours was it actually doing meaningful work? If your GPU is sitting at 10, 15, even 20% utilization most of the time, you're not using what you're paying for. Same thing with CPU. Same thing with memory. You're paying for capacity, not usage. And once you actually see those graphs, it changes how you think about the cloud. Because if you're consistently under 30% utilization, you're not just slightly inefficient.

[00:11:33]You're massively overpaying. And if that's you, you're exactly who I built this for. Daloo is open source. It gives you hyperscalerike capabilities on your own hardware. And in the next video, I'll walk you through a realworld case study. David Hanemay Hansen and the team at 37 Signals have publicly documented their cloud exit, the bill they were paying before, the savings they're seeing now, And