Physix FrontierCallosum raises $100M: The strategic path behind the funding
Title: Callosum is talking about raising $100M; what I'm looking at is the path behind it
The most valuable info in this article is that Callosum is negotiating another massive funding round. This London company claims it will challenge Nvidia's monopoly, but what's truly interesting isn't how much they plan to raise, but the route they're taking. It happens to be the opposite direction of the redundant construction worries I mentioned in my previous post about Anthropic's valuation.
Let me give some background. Callosum's founders are two neuroscientists from Cambridge. They reportedly raised a round previously, with investors including the UK's ARIA, that government agency dedicated to high-risk, high-reward projects. Now, according to Electronics Weekly, they are negotiating a new funding round in the $100 million range. What they do can be summed up in one sentence: write software to schedule AI workloads across different chips, without being tied to Nvidia.
This gets interesting. There are currently two paths competing for money in the market. One is making the chips themselves—Groq, Neocloud, including the LPU I just tried recently—all making their own cards, trying to grab Nvidia's share with faster inference speeds. The other is Callosum's path: don't make chips, make the scheduling layer, letting existing cards in others' hands become usable.
From an asset allocation perspective, the risk-reward ratio of these two paths is completely different.
Making chips is heavy-asset. You dump billions upfront for tape-outs, betting that the next-gen architecture can outperform the CUDA ecosystem. I've run LPUs hands-on; they are indeed fast, but speed is just a ticket. What makes Nvidia truly terrifying is that its ecosystem is deeply embedded in customers' engineering teams. Switching to a different card means redoing the entire toolchain. I've tested this migration cost, and it's much higher than most people expect. I also configured a Neocloud instance recently, spent nearly an afternoon tinkering, while Nvidia is ready-to-use out of the box.
Callosum's path is smart because it doesn't force a choice. Its core assumption is that data centers will always have Nvidia cards and always have other vendors' cards. Instead of making you choose one, help you mix and match all resources. Which chip runs this task more cost-effectively? Schedule it there. Simply put, they aren't selling compute power; they're selling the efficiency of using compute power.
This logic reminds me of a phrase investment circles loved a few years ago: The person selling shovels might not make money, but the person selling maps makes steady money. They aren't betting on any single chip manufacturer winning; they're betting that the market will remain fragmented forever. As long as fragmentation exists, the scheduling layer has value.
But I must say, this path has an unavoidable problem: wolves ahead, tigers behind. Ahead is Nvidia itself doing similar things; its software stack is getting thicker, internalizing scheduling capabilities. Behind are cloud providers; AWS and Azure have always been doing multi-cloud management. If they want to swallow this layer, Callosum's survival space will be squeezed very narrow. Moreover, documents mention they've received investment from the UK Sovereign AI Fund, carrying government color, meaning their expansion path might not be fully market-driven.
So my judgment is this. Betting on Callosum itself is risky at this stage; $100 million in the AI infrastructure battlefield is honestly not enough to make waves. But the route it represents is directionally correct. If mixed multi-chip deployment becomes standard for data centers in the next two years, companies like Callosum in the scheduling layer, even if they don't succeed independently, will be acquired at a high price by larger players. This is essentially betting on the industry's evolutionary direction, not betting on one company's life.
While reading this news, I kept thinking of a comparison. Apple didn't do chip manufacturing; they did chip design, yet still controlled the entire supply chain. What Callosum wants to do, in the AI infrastructure chain, is also a designer role rather than a manufacturer. I approve of this positioning, but the barrier for designers is: your blueprints must be perceived as better than what others assemble themselves.
I plan to wait until this funding round is officially announced, then put in a small amount to follow along. Not because I'm sure they'll succeed, but because investing at this position means losing is tuition, and winning is early-positioning return. One thing I am fairly certain of: future AI infrastructure won't belong solely to Nvidia.
📌 This article is compiled from Bloomberg Tech. Original: https://www.bloomberg.com/news/articles/2026-08-20/ai-startup-callosum-raises-100-million-to-make-ai-tasks-cheaper
All rights reserved by the original author. This is a compilation and independent analysis based on public reports.
