# I can haz networks, plz? 

> Networks are more important than ever. Why are so few people building new ones?

It's not uncommon today to find new clouds with upwards of a million users routing all their traffic through a single availability zone. For all the progress in distributed systems, the OG global-scale system (the Internet!) is little more than a footnote for most new operators.

I think that’s crazy, and networks have never mattered more! Consider what's happening: 

- Companies are spreading data across dozens of new providers; 
- Agents and humans alike are shipping far more code than ever (👋 GitHub Actions!);
- Security incidents are through the roof; 
- Zero trust and authorization models are being rewritten and pushed out to the edge; 
- Inference is the fastest-growing workload on the planet by a long shot. 

I'm biased, of course. But as I flipped through a [Cisco report on the impact of AI ](https://www.cisco.com/c/dam/en/us/solutions/collateral/artificial-intelligence/mass-scale-infrastructure/ai-network-traffic-report.pdf) on the wide area network (don't judge, this stuff is mad cool!), one stat jumped out: AI agents generate up to 450% more network traffic than humans. Talk about noisy neighbors, eh?

For these reasons and more, we should see a rush to build new networks akin to the ISP and hosting boom. Back in the early 2000s, everyone was standing up networks, buying routers and switches, and learning how to splice fiber. Not today.

So what gives? Three things, from my point of view.

## Long live the VPC

The first reason we don't see more new networks being born is that hyperscale cloud networking may just have spoiled us.

While VPCs can be complicated and expensive, it's hard to argue with their value. Across every public cloud you can access a robust suite of networking primitives just as readily as you can spin up compute, storage, or the latest open source model. Alongside automation, VPC’s come with the creature comforts that platform teams expect: observability, robust permissions, tons of features with a button click, etc.

The ubiquity of VPC constructs is a double-edged sword. On the one hand, developers don't need to ask networking experts for help (yay!), but abstraction away from the harder core networking makes it far less obvious when the time comes to consider building something of your own.

## It's harder than ever to DIY

Today we announced a [strategic partnership with DataBank](https://www.einpresswire.com/article/934819679/datum-deploys-its-open-network-cloud-at-databank-to-connect-distributed-ai-infrastructure). 

The premise is simple: DataBank has some of the best interconnected datacenter assets in the US (especially for AI), and we're building a platform that brings similar value to the higher-level primitives that AI and agent workloads need. Racks, power, and physical ecosystem access on their side; DNS, load balancing, TLS, proxy-based routing, SRv6 overlays (and more) on ours.

I've worked with much of the executive team at DataBank for 20 years, and we’ve been itching to collaborate on how to evolve the  internet for AI. Furthermore, I’m a longtime datacenter nerd who is super familiar with what DataBank sells. And yet it still took serious work (plus a bit of good luck and a few well-timed text messages to my old friend Raul) to land the racks we needed in Ashburn, Salt Lake City, and Dallas.

Here's why: space and (above all) power are super scarce resources right now. AI training and inference buildouts have absorbed capacity, and enterprise demand hasn't slowed to make room. At the same time, a network upgrade cycle from 400G to 800G  is steadily eating the power budget from the inside. So we’re competing for scarce power twice over, once against everyone insatiable growth, and once against the next-generation of core internet gear.

If it's this hard for two companies that know each other well, are aligned on outcomes, and aren't fighting over price and terms, imagine how it looks to a team that's never done it before!

Welcome to building a new network in 2026.

## Finally, the “i can haz” generation is thinning out 

The group of humans who built networks in the early 2000s (folks who immediately understood the title of this post) is shrinking fast. There are a lot more grey hairs at the NANOG meetings, and it's not just from late nights troubleshooting BGP or squashing TCAM limits in routers!

![I-Can-Haz.png](https://grateful-excitement-dfe9d47bad.media.strapiapp.com/I_Can_Haz_6b91e1f8f2.png)

This is partly a talent thing. While the cloud era produced an enormous number of sophisticated infrastructure engineers, networks (and the experts who built and maintained them) were often replaced by VPC abstraction. But there's also an important cultural element at play: the human relationships and spirit of generosity that helped the Internet scale. While it still exists in pockets, how it feels to participate in (and contribute to) that kind of community is becoming more rare.

## Why you must haz network

I realize I’m a network fanboy, but the advantages of controlling who and how you connect aren’t theoretical, especially in the AI era.  Take a look at Equinix's latest [Global Interconnection Index](https://www.equinix.com/gxi-report) and you’ll see that private interconnection is forecast to grow at roughly 2x the rate of the public internet and is already an order of magnitude larger in aggregate. This isn't a niche practice; it's how the digital economy actually works at scale.

Financial trading firms interconnect directly to shave microseconds off execution. Gaming and streaming companies build out delivery networks to put content a short hop from players and viewers. Banks wire directly into payments, fintech, and capital-markets ecosystems. Each of these industries figured out the same thing: once you need to solve for the economics of data, latency, proximity to partners, it is clear that global reach, direct interconnection, edge networks, and traffic shaping are hugely important tools in your toolbox.

Now hold that up against a typical fast-growing AI company. A million users, exploding agent traffic, and inference workloads that benefit from interacting at the edge — and yet everything is funneled through a single availability zone in Virginia?! The capabilities that help hyperscalers, trading firms, and streamers use to compete are the same ones today's most disruptive companies are going without.

So the real question isn't "why don't more people build networks?" It's: why shouldn't every disruptive AI company get the same internet superpowers the bug guys use?

## What we're doing about it

At Datum, our goal is to help thousands of companies unlock those superpowers. Making that happen comes down to a few key things:
- **Aggregate capacity.** Pool physical and logical assets so a single company doesn't have to negotiate racks in a dozen cities (and fight for constrained power in each) just to get started.
- **Developer experience.** The network should be as easy to reason about and operate as any other part of the modern stack.
- **Infrastructure plus services.** Not just primitives, but the services and relationships that make it all go.

There's a fourth thing, because it ties back to those NANOG “beers with peers” gatherings: you shouldn't have to spend years in the trenches to get here. The whole point is to open the door to a new generation (the builders who grew up with VPCs not routers), and help them wield internet superpowers anyway. Networks shouldn't be reserved for people who happened to be in the room in 2003!

I like to say that if Stripe lets you incorporate a company and embed financial infrastructure in your app in minutes, why can't we do the same with global networks? 

We know we can, and that's our vision. Happily, our partnership with DataBank puts some key physical layers in place, so the next thousand companies don't have to spend months getting to the starting line.