I believe right now we are still in the phase of “how can AI help engineers write better software”, but are slowly shifting to “how can engineers help AI write better software.” This will bring in a new herd of engineers with completely different views on what software is, and how to best go about building computer interactions.

Vibe coding or LLM accelerated development is going to turn this on its head. Everyone will be able to afford custom software to fit their specific needs and preferences. Where Salesforce currently has 150,000 customers, imagine 150,000 customers all using their own customised CRM. The scope for software expansion is unbelievably large right now.

Jevons paradox would be if despite software becoming cheaper to produce the total spend on producing software would increase because the increase in production outruns the savings

Jevons paradox applies when demand is very elastic, i.e. small changes in price cause large changes in quantity demanded. It's a property of the market.

My view is actually the opposite. Software now belongs to cattle, not pet. We should use one-offs. We should use micro-scale snippets. Speaking language should be equivalent to programming. (I know, it's a bit of pipe dream)

In that sense, exe.dev (and tailscale) is a bit like pet-driven projects.

As for the average quality: it’s unclear.

My intuition is that agents lift up the floor to some degree, but at the same time will lead to more software being produced that’s of mediocre quality, with outliers of higher quality emerging at a higher rate than before.

I honestly think this is ideal. Video games aside, I think one day we'll look back and realize just how insane it was that we built software for millions or even billions of users to use. People can now finally build the software that does exactly what they've wanted their software to do without competing priorities and misaligned revenue models working against them. One could argue this kind of software, by definition, is higher quality.

Another one could be Bitwarden, although I don't host my own password manager personally. Or netbird. You get the point

I hope this wasn't interpreted towards exe.dev. That really is a cool service!

The key point is the partner companies. Almost nobody is actually running their own clouds the way they would with various 365 products, AWS or Azure. They buy the cloud from partners, similar to how they used to (and still do) buy solutions from Microsoft partners. So if you want to "sell cloud" you're probably going to struggle unless you get some of these onboard. Which again would probably be hard because I imagine a lot of what they sell is sort of a package which basically runs on VM's setup as part of the package that they already have.

International visitors might tell us more about benefits of non EU, US or UK nexus companies/legal/rights.

People use Kubernetes for way too small things, and it sounds like you don't have the scale for actually running Kubernetes.

Scale vertically until you can't because you're unlikely to hit a limit and if you do you'll have enough money to pay someone else to solve it.

Docker is amazing development tooling but it makes for horrible production infrastructure.

The irony is that "DevOps" was supposed to be a culture and a set of practices, not a job title. The tools that came with it (=Kubernetes) turned out to be so complex that most developers didn't want to deal with them and the DevOps became a siloed role that the movement was trying to eliminate.

That's why I have an ick when someone uses devops as a job title. Just say "System Admin" or "Infrastrcutre Engineer". Admit that you failed to eliminate the siloes.

Maybe those devops folks only pay attention to k8s clusters and you're flying under their radar with your single debian VM + Kamal. But the same thinking that results in an overtly complex, impossible to debug, expensive to run k8s cluster can absolutely result in the same using regular VMs unless, again, you are just left to your own devices because their policies don't apply to VMs, yet.

The problem usually is you're one mistake away from someone shoving their nose in it. "What are you doing again? What about HA and redundancy? slow rollout and rollback? You must have at least 3 VMs (ideally 5) and can't expose all VMs to the internet of course. You must define a virtual network with policies that we can control and no wireguard isn't approved. You must split the internet facing load balancer from the backend resources and assign different identities with proper scoping to them. Install these 4 different security scanners, these 2 log processors, this watchdog and this network monitor. Are you doing mtls between the VMs on the private network? what if there is an attacker that gains access to your network? What if your proxy is compromised? do you have visibility into all traffic on the network? everything must flow throw this appliance"

And I’m building and happily using Uncloud (https://github.com/psviderski/uncloud) for this (inspired by Kamal). It makes multi-machine setups as simple as a single VM. Creates a zero-config WireGuard overlay network and uses the standard Docker Compose spec to deploy to multiple VMs. There is no orchestrator or control plane complexity. Start with one VM, then add another when needed, can even mix cloud VMs and on-prem.

All of this just adds so much extra complexity. If I'm running Amazon.com then sure, but your average app is just fine on a single VM.

at 19€/mo are you subsidizing it given the sharp rise of LLM costs lately?

or are you heavily restricting model access. surely there is no Opus?

Using fio

Hetzner (cx23, 2vCPU, 4 GB) ~3900 IOPS (read/write) ~15.3 MB/s avg latency ~2.1 ms 99.9th percentile ≈ ~5 ms max ≈ ~7 ms

DigitalOcean (SFO1 / 2 GB RAM / 30 GB Disk) ~3900 IOPS (same!) ~15.7 MB/s (same!) avg latency ~2.1 ms (same!) 99.9th percentile ≈ ~18 ms max ≈ ~85 ms (!!)

using sequential dd

Hetzner: 1.9 GB/s DO: 850 MB/s

Using low end plan on both but this Hetzner is 4 euro and DO instance is $18.

Hey wait a minute!

Our exe.dev web UI still runs on AWS. We also have a few users left on our VM hosts there, as when we launched in December we were considering building on AWS. Now almost all customer VMs are on other bare metal providers or machines we are racking ourselves. We built our own GLB with the help of another vendor's anycast network. You can see that if you try any of the exe.xyz names generated for user VMs.

We would move exe.dev too, but we have a few customers who are compliance sensitive going through it, so we need to get the compliance story right with our own hardware before we can. It is a little annoying being tied to AWS just for that, but very little of our traffic goes through them, so in practice it works.

(Percentages cited above are tongue-in-cheek, actual numbers are probably different)

https://exe.dev/docs/regions

VMs have a built-in gateway to cloud providers with a fixed url with no auth. You can top that in via the service itself. No need for your own keys.

So likely a good tool for managing AI agents. And "cloud" is a bit of a stretch, the service is very narrow.

The complete lack of more detailed description of the regions except city name makes it really only suitable for ephemeral/temporary deployments. We don't know what the datacenters are, what redundancy is in place, no backups or anything like that.

So having everyone use the same deployment model (and that’s typically k8s) saves effort. I don’t like it for sure

My app is fairly simple node process with some side car worker processes. k8s enables me to deploy it 30 times for 30 PRs, trivially, in a standard way, with standard cleanup.

Can I do that without k8s? Yes. To the same standard with the same amount of effort? Probably not. Here, I'd argue the k8s APIs and interfaces are better than trying to do this on AWS ( or your preferred cloud provider ).

Where things get complicated is k8s itself is borderline cloud provider software. So teams who were previously good using a managed service are now owning more of the stack, and these random devops heros aren't necessarily making good decisions everywhere.

So you really have three obvious use cases:

a) You're doing something interesting with the k8s APIs, that aren't easy to do on a cloud provider. Essentially, you're a power user. b) You want a cloud abstraction layer because you're multi-cloud or you want a lock-in bargaining chip. c) You want cloud semantics without being on a cloud provider.

However, if you're a single developer with a single machine, or a very small team and you're happy working through contended static environments, you can pretty much just put a process on a box and call it done. k8s is overkill here, though not as much as people claim until the devops heros start their work.

If that's true, I wonder if this is a deliberate decision by cloud providers to push users towards microservice architectures with proprietary cloud storage like S3, so you can't do on-machine dbs even for simple servers.

Business 101 teaches us that pricing isn't based on cost. Call it top down vs bottom up pricing, but the first principles "it costs me $X to make a widget, so 1.y * $X = sell the product for $Y is not how pricing works in practice.

Edit: I posted this before reading, and these two are the same he points out.

If you're doing anything complicated, Excel just doesn't make sense anymore. it'll still the be data exchange format (at least, something more advanced than csv), but it's no longer the only frontend.

"No one uses" is no longer the insult it once was. I don't need or want to make software for every last person on the world to use. I have a very very small list of users (aka me) that I serve very well with most of the software that I generate these days outside of work.

I could see maybe more customization of said software, but not totally fresh. I do agree that people will invent more one-off throwaway software, though.

But I came across Mythic Beasts (https://www.mythic-beasts.com/) yesterday, similar idea, UK based. Not used them yet but made the account for the next VPS.

Does a regular 20-something software engineer still know how to turn some eBay servers & routers into a platform for hosting a high-traffic web application? Because that is still a thing you can do! (I've done it last year to make a 50PiB+ data store). I'm genuinely curious how popular it is for medium-to-big projects.

And Hetzner gives you almost all of that economic upside while taking away much of the physical hassle! Why are they not kings of the hosting world, rather than turning over a modest €367M (2021).

I find it hard to believe that the knowledge to manage a bunch of dedicated servers is that arcane that people wouldn't choose it for this kind of gigantic saving.

The technology itself in its current form is not valuable

I'm not surprised even in the slightest that DevOps workers will slap k8s on everything, to show "real industry experience" in a job market where the resume matches the tools.

I am primarily a backend developer but I do a lot of ops / infra work because nobody else wants to do it. I stay as far away from k8s as possible.

This is why you get many folks over-thinking the solution and picking the most hyped technologies and using them to solve the wrong problems without thinking about what they are selling.

You don't need K8s + AWS EC2 + S3 just to host a web app. That tells me they like lighting money on fire and bankrupting the company and moving to the next one.

Docker Compose is good for running things on a single server as well.

Docker Swarm and Hashicorp Nomad are good for multi-server setups.

Kubernetes is... enterprise and I guess there's a scale where it makes sense. K3s and similar sort of fill the gap, but I guess it's a matter of what you know and prefer at that point.

Throw on Portainer on a server and the DX is pretty casual (when it works and doesn't have weird networking issues).

Of course, there's also other options for OCI containers, like Podman.

If you wanted to, you could make an argument about the principal-agent problem - that as hunter-gatherers or subsistence, farmers, our quality versus quantity decisions only affected us, whereas in a market economy, you could argue that one person’s quality versus quantity decision affects someone else.

But dismantling capitalism will not solve this problem. It just moves the decision-making to a different group of people. Those people will face the same trade-offs and the same incentives. After the Revolution, even the most loyal comrade will have to contend with the fact that they can choose to provide the honourable working class with more of a thing if they drop the quality.

Could you rephrase that / elaborate on that? Isn't Tailscale's selling point precisely that they do identity-based networking?

EDIT: Never mind, now I see the sibling comment to which you also responded – I should have reloaded the page. Let's continue there!

But if its use was confined to this use case, pretty much nobody would be using it (unless as a customer of the organization's infra) and barely would be talking about it (like how there isn't too much talk about Borg).

The reason k8s is a thing in the first place is because it's being used by way too many people for their own goods. (Most people having worked in startups have met too many architecture astronauts in our lives).

If I had to bet, I'd wager that 99% of k8s users are in the “spin a few containers to run your web app” category (for the simple reason that for one billion-dollar tech business using it for legit reasons, there's many thousands early startups who do not).

RS 1000 G12 AMD EPYC™ 9645 8 GB DDR5 RAM (ECC) 4 dedicated cores 256 GB NVMe

Costs 12,79 €

Results with the follwing command:

fio --name=randreadwrite \ --filename=testfile \ --size=5G \ --bs=4k \ --rw=randrw \ --rwmixread=70 \ --iodepth=32 \ --ioengine=libaio \ --direct=1 \ --numjobs=4 \ --runtime=60 \ --time_based \ --group_reporting

IOPS Read: 70.1k IOPS Write: 30.1k IOPS ~100k IOPS total

Throughput Read: 274 MiB/s Write: 117 MiB/s

Latency Read avg: 1.66 ms, P99.9: 2.61 ms, max 5.644 ms Write avg: 0.39 ms, P99.9: 2.97 ms, max 15.307 ms

For agents, declarative plans are still valuable because they are reviewable. The interesting question is whether exe.dev changes the primitive: resource pools for many isolated VM-like processes, or just nicer VPS provisioning.

- I'm building a server farm in my homelab.

- I'm doing a small startup to see if this idea works.

- We're taking on AWS by being more cost effective. Funding secured.

Instead they make the default "meager IOPS" and then charge more to the people who need more.

And yes, IO typically happens in 4kb blocks, so you need a decent amount of IOPS to get the full bandwidth.

It certainly is for lots of businesses, otherwise they go out of business.

There is something called 'revenue' which they need to make from customers which are their 'users', and that revenue pays for the 'operating costs' which includes payroll, office rent, infrastructure etc.

This just means that it is important than ever to know what to build just as how it is built. It is unrealistic for a business to disregard that and to build anything they want and end up with zero users.

No users, No revenue. No revenue, No business.

I need to fix our transfer pricing. (In fact I'm going to go look at it now.) I set that number when we launched in December, and we were still considering building on top of AWS, so we put a conservative limit based on what wouldn't break the bank on AWS. Now that we are doing our own thing, we can be far more reasonable.

Managing servers is fine. Managing servers well is hard for the average person. Many hand-rolled hosting setups I've encountered includes fun gems such as:

- undocumented config drift.

- one unit of availability (downtime required for offline upgrades, resizing or maintenance)

- very out of date OS/libraries (usually due to the first two issues)

- generally awful security configurations. The easiest configuration being open ports for SSH and/or database connections, which probably have passwords (if they didn't you'd immediately be pwned)

Cloud architecture might be annoying and complex for many use-cases, but if you've ever been the person who had to pick up someone else's "pet" and start making changes or just maintaining it you'll know why the it can be nice to have cloud arch put some of their constraints on how infra is provisioned and be willing to pay for it.

You can use block storage if data matters to you.

Many services do not need to care about data reliability or can use multiple nodes, network storage or many other HA setups.

If you'd know Kubernetes, you know not to use it. I say that as someone who used to do consulting for it.

The reality is that yet again "making money" completely collides with efficient, quality, sane productive work.

For me one of the main reasons to leave that space is that I couldn't really deal with the fact that my work collides with a client's success. That said I have helped to get off that stuff and other things that they thought they needed, that just wasted time and money. It just feels odd going into a company that hired you to consult on a topic only to end up telling them "The best approach for you is not doing that at all". Often never. Like some people thought "Well, if we have hundreds of thousands or even millions of users" and the reality was that even in these scenarios if you went away from that abstract thought and discussed a hypothetical based on their product they realized that they'd still be better off without it. Besides the fact that this hypothetical often was in a future that made it likely that they said they'd likely have completely different setup so preparing for that didn't even make sense.

I think a big thing related to that was/is the microservice craze where people end up moving to a complex architecture for not many good reasons and then they increase complexity way faster than what they actually deliver in terms of the product, because it somehow feels good. I know it does, I've been there. When in reality the outcome often is just a complex mess with what could have been a relatively simple monolith. And these monoliths do work. And in the vast majority of cases they are easy to scale, because your problem switches from "how do we best allocate that huge amount of very different services across our infrastructure" to (for the most part) "how do we spin up our monolith on one more server" which tends to be a way easier to tackle service.

And nothing stops you from still using everything else if you want. Just because it's a monolith doesn't mean you need to skip on any of the cloud offerings, etc. For some reason there seems to be that idea that if you write a monolith you are somehow barred from using modern tooling, infrastructure, services, etc. Not sure where that comes from.

The price is what the customer will pay, regardless of your costs.

Whether or not cloud is viable for a company is very individual. It's very hard to pin point a size or a use case that will always make cloud the "correct" choice.

But there is middleground in form of VPS, where hardware is managed by the provider. It's still way way cheaper than some cloud magic service.

It is like 4 lines of config for Postgres, the only line you need to change is on which path Postgres should store the data.

I mean, I worked with people who were suprised that you can run more applications inside ec2 vm than just 1 app.

This is certainly the case from all the third person accounts I hear. Online. I never actually met a single one that is like that, if anything, those same people are the ones that are first to tell me about their Hetzner setups.

I am sure it's luck but we have few hetzner VPSes in both German locations and in last 5 years afaik they've never been down. On our http monitor service they have 100s of days uptime only because we restarted them ourselves.

Tags permanently erase the user identity from a device, and disable things like Taildrop. When I tried to assign a tag for ACLs, I found that I then could not remove it and had to endure a very laborous process to re-register a Tailscale device that I added to Tailscale for the express purpose of remotely accessing

IOPS: read 325k, write 139k

Throughput: read 1271MB/s, write 545MB/s

Latency: read avg 0.3ms, P99.9 2.7ms, max 20ms; write: 0.14ms, P99.9 0.35ms max 3.3ms

so roughly 100 times iops and throughput of the cloud VMs

Radboud University recently announced they're rolling it out for managing containers across the faculty which is the most "serious install" I know about, but there could be other: https://cncz.science.ru.nl/en/news/2026-04-15_uncloud/

Using a Netcup VPS 1000 G12 is more comparable.

read: IOPS=18.7k, BW=73.1MiB/s

write: IOPS=8053, BW=31.5MiB/s

Latency Read avg: 5.39 ms, P99.9: 85.4 ms, max 482.6 ms

Write avg: 3.36 ms, P99.9: 86.5 ms, max 488.7 ms

OP is not saying they push new versions at such a high frequency they need checks every one minute.

The choice of one minute vs 15 minute is implementation detail and when architected like this costs nothing.

I hope that helps. Again this is my own take.

Maybe change the filesystem?

Almost every VC rejected us when we went to get seed funding for Tailscale, we knew none of them. Friends of friends of acquaintances got us meetings. Fundraising is very possible for you if you are committed to building a business. Most important thing is don't think of fundraising as the goal, it is just a tool for building a business. (And some businesses don't need VC funding to work. Some do.)

The biggest challenge is personal: do you want to build a business or do you want to work with cool tech? Sometimes those goals are aligned, but usually they are not. Threading the needle and doing both is difficult, and you always have to prioritize the business because you have to make payroll.

To be fair though, that's true for every profession or skill.

> I mean, I worked with people who were suprised that you can run more applications inside ec2 vm than just 1 app.

I've seen something similar where people were surprised that you can use an object storage (so effectively "make HTTP requests") from every server.

It kinda is, but obscured by GP's formula.

More simply; if it costs you $X to produce a product and the market is willing to pay $Y (which has no relation to $X), why would you price it as a function of $X?

If it costs me $10 to make a widget and the market is happy to pay $100, why would I base my pricing on $10 * 1.$MARGIN?

But to quote someone: "you are not Google".

Here are some "Regular Performance" shared resource stats

Hetzner CPX11 (Ashburn, 2 CPUs, 2GB, 5.49€ or $6.99/month before VAT)

read: IOPS=36.7k, BW=144MiB/s, avg/p99.9/max 2.4/6.1/19.5ms

write: IOPS=15.8k, BW=61.7MiB/s, avg/p99.9/max 2.4/6.1/18.7ms

Hetzner CPX22 (Helsinki, 2 CPUs, 4GB, 7.99€ or $9.49/month before VAT)

read: IOPS=48.2k, BW=188MiB/s, avg/p99.9/max 1.9/5.7/10.8ms

write: IOPS=20.7k, BW=80.8MiB/s, avg/p99.9/max 1.8/5.8/10.9ms

Hetzner CPX32 (Helsinki, 4 CPUs, 8GB, 13.99€ or $16.49/month before VAT)

read: IOPS=48.3k, BW=189MiB/s, avg/p99.9/max 1.9/6.2/36.1ms

write: IOPS=20.7k, BW=81.0MiB/s, avg/p99.9/max 1.8/6.3/36.1ms

The trouble is that we are literally expected to do this everywhere we go. I've personally advocated for approaches which use say, a pair of dedicated servers, or VMs as in GPs example. If you want it outside of AWS/GCP/Azure, you're regarded as a crazy person. If you don't adopt "best practices" (as defined by vendors) then management are scared. Management very often trust the sales and marketing departments of big vendors more than their own staff. Many of us have given up fighting this, because what it comes down to is a massive asymmetry of information and trust.

Which it won't be, if at every turn you choose the hyperscaler.

For example I calculated the cost of a solar install to be approximately: Material + Labour + Generous overhead + Very tidy profit = 10,000€

In practice I keep getting offers for ~14,000€, which will be reduced to 10,000€ with a government subsidy and my request for an itemized invoice is always met with radio silence.

But that is an equilibrium result, and famously does not apply to monopolies, where elasticity of substitution will determine the premium over the rental rate of capital.

I think this is the crux of the matter. Also, "everybody is doing it, so they must be right" is also a very common way of thinking amongst this population.

They're getting kickbacks from cloud vendors. Prove me wrong.

Around the time of the pandemic, a company wanted to make some Javascript code do a kind of transformation over large number of web-pages (a billion or so, fetched as WARC files from the web archive). Their engineers suggested setting up SmartOS VMs and deploying Manta (which would have allowed the use of the Javascript code in a totally unmodified way -- map-reduce from the command-line, that scales with the number storage/processing nodes) which should have taken a few weeks at most.

After a bit of googling and meeting, the higher ups decided to use AWS Lambdas and Google Cloud Functions, because that's what everyone else was doing, and they figured that this was a sensible business move because the job-market must be full of people who know how to modify/maintain Lambda/GCF code.

Needless to say, Lambda/GCF were not built for this kind of workload, and they could not scale. In fact, the workload was so out-of-distribution, that the GCP folks moved the instances (if you can call them that) to a completely different data-center, because the workload was causing performance problems, for _other_ customers in the original data-center.

Once it became clear that this approach cannot scale to a billion or so web-pages, it was decided to -- no, not to deploy Manta or an equivalent -- but to build a custom "pipeline" from scratch, that would do this. This system was in development for 6 months or so, and never really worked correctly/reliably.

This is the kind of thing that happens when non-engineers can override or veto engineering decisions -- and the only reason they can do that, is because the non-engineers sign the paychecks (it does not matter how big the paycheck is, because market will find a way to extract all of it).

One of the fallacies of the tech-industry (I do not mean to paint with too broad a brush, there are obviously companies out there that know what they are doing) is that there are trade-offs to be made between business-decisions and engineering-decisions. I think this is more a kind of psychological distortion or a false-choice (forcing an engineering decision on the basis of what the job market will be like some day in the future -- during a pandemic no less -- is practically delusional). Also, if such trade-offs are true trade-offs, then maybe the company is not really an engineering company (which is fine, but that is kind of like a shoe-store having a few podiatrists on staff -- it is wasteful, but they can now walk around in white lab-coats, and pretend to be a healthcare institution instead of a shoe-store).

Personally, I believe that the tech industry sustains itself via technical debt, much like the real economy sustains itself on real debt. In some sense, everyone is trying to gaslight everyone else into incurring as much technical debt as possible, so that a way to service the debt can be sold. Most of the technical debt is not necessary, and if people were empowered to just not incur it, I suspect it would orient tech companies towards making things that actually push the state of the art forward.

Teams are free to use EKS internally.

https://uncloud.run/docs/getting-started/install-cli/#instal...

But yeah, let's just spin-up a shadow IT VM with Debian like GP said, it's easy!

The challenge is convincing people that "golden images" and containers share a history, and that kubernetes didn't invent containers: they just solved load balancing and storage abstraction for stateless message architectures in a nice way.

If you're doing something highly stateful, or that requires a heavy deployment (game servers are typically 10's of GB and have rich dynamic configuration in my experience) then kubernetes starts to become round-peg-square-hole. But people buy into it because the surrounding tooling is just so nice; and like GP says: those cloud sales guys are really good at their jobs, and kubernetes is so difficult to run reliably yourself that it gets you hooked on cloud.

There's a literal army of highly charismatic, charming people who are economically incentivised to push this technology and it can be made to work so- the odds, as they say, are against you.

That’s literally how they sold AWS in the beginning.

Cloud won not because of costs or flexibility but because it allowed teams to provision their own machines from their budget instead of going through all the red tape with their IT departments creating… a bunch of shadow IT VMs!

Everything old is new again, except it works on an accelerated ten year cycle in the IT industry.