The summer of neoclouds
Hi everyone, I hope you’ve had a nice week so far. I’d like to thank Box once again for supporting Hypertext. Feel free to share this email with a friend, it really helps 💌 Sign up here.
The summer of neoclouds
This week, many AI leaders came to Paris for the Raise Summit, an AI-focused conference (it’s called Raise because there’s AI in the middle of “Raise”…). They manage to attract good speakers because they say it takes place in the Louvre and you’re going to the Château de Versailles.
The reality is a bit different as the conference actually takes place in the mall below the Louvre (Carrousel du Louvre) and only VIPs get invited to Versailles (not general attendees). It might be a bit deceptive to some but it does the trick.
Compared to last year, I noticed a big tone shift. Instead of talking about the latest and greatest AI models and applications, the conversation mostly revolved around costs and governance.
Cloud hosting companies focusing on GPUs (so-called “neoclouds”) had a massive presence. They naturally talked about how you could secure data center capacities and lower the cost of serving AI models at scale (so-called “inference costs”).
Baseten, CoreWeave, Crusoe, Fireworks AI, Nebius, Nscale, Together AI and several others had big booths. It suggests that they see Europe as a priority market.
Some of them don’t even have data centers in Europe yet. But as they build out capacity on the continent, they’ll want to secure clients looking to serve European customers from European data centers directly. Winning customers today could lock in years of GPU demand.
This group of companies has suddenly become much more important in the overall AI conversation due to three converging factors:
- Big companies that gave unlimited AI budgets to their software engineering teams realized that they’re now facing bills worth thousands of dollars/euros per employee. Claude Opus 4.8 is expensive. Claude Fable 5 is ludicrously expensive.
- The US Government export restrictions on Fable 5 and GPT-5.6 (that have since been lifted) were a warning sign that, in the AI industry, you can’t take anything for granted. It’s time to implement some fallback mechanisms, especially if you’re using APIs from OpenAI and Anthropic to power some critical parts of your infrastructure.
- Open-source models are getting better. The release of GLM-5.2 in particular was impressive. Software engineers feel like open-source models are now “just a few months behind” the latest models from Anthropic and OpenAI.
So companies are experimenting with replacing OpenAI’s or Anthropic’s models with GLM-5.2 running on a neocloud near you. It’s both cheaper and more resilient.
I feel like we’re still very much in the experimentation phase. Nobody is abandoning OpenAI or Anthropic. They’re still setting the pace on frontier models. But companies increasingly want the freedom to swap models, negotiate costs, and avoid depending on a single provider. That’s why infrastructure providers have become a much bigger part of the AI conversation than it was a year ago.
Box’s State of AI report is here. Grounded in a global survey of 1,640 IT decision-makers, the report reveals that security and compliance remain paramount, with 49% of organisations having already experienced an AI-related data exposure incident. Read the full report to discover how organisations are establishing a governed, secure content layer to achieve trusted automation without increasing risk.
The end of “Bring Your Own AI”
And when you look more closely at the number one use case for large language models these days (coding), switching isn’t as seamless as it sounds.
This week, I spoke with an engineer who has switched from Claude Code to OpenAI’s Codex because he thought it was a superior agentic coding tool. Most of his teammates, however, are sticking with Claude Code because it works well enough.
“If it ain’t broke, don’t fix it” also applies to tech-savvy developers. Even though switching from Claude Code to Codex isn’t too hard, they don’t want to switch tool every other month.
Moving to an open-weight model like GLM-5.2 is an even bigger leap. You need an agent harness such as Pi, a dev-friendly inference provider like OpenRouter. More importantly, you need to choose a model from hundreds of models. And finally, you have to find a way to expense the tokens that you use…
It just doesn’t work at scale… Unless there’s a big corporate policy shift saying that Claude Code and Codex are now banned. I believe it’s a governance decision rather than an individual one.
And it turns out that some companies are doing just that. A few months ago, Microsoft famously told its developers that they have to use the company’s own AI coding tool (GitHub Copilot CLI). More recently, according to Reuters, Alibaba is banning Claude Code. According to The Information, Meta also told engineers to stop using Claude Code and Codex.
Now, we’ll see whether they can keep those bans in place because morale at big tech companies is low right now, as this interesting study from Lenny’s Newsletter highlighted. Not being able to use the best tools for the job sucks. And the biggest constraint on developers may soon be company policy, not model capability.
Have a good day ☀️
Romain