H Company: Why Investors Handed $180 Million to a Startup With No Product

Picture this: you walk into a meeting with investors. You have no finished product, no huge cloud of thousands of GPUs, and not even a website with polished charts. They still write you a check for $180 million. It would sound like something out of the dot-com bubble, except for one catch: the people behind this project literally helped build modern AI inside DeepMind.

H Company, formerly known as Holistic AI, has become the industry's new sensation not because it promises to make GPT-5 faster, but because it questions the entire approach to building intelligence.

The industry has grown used to "scaling laws". The logic is simple: give a model more text, build a bigger cluster, and it gets smarter. But H Company's founders, Charles Le Lan, Laurent Sifre, Karl Tuyls and Julien Perolat, concluded that this path leads to a dead end. They are pushing the concept of "small data". The idea is that a human brain does not need to read the whole internet to learn logic. We learn from examples, grasp cause and effect, and know how to act under uncertainty. That is exactly what today's LLMs lack; they remain merely very advanced statistical parrots.

Andrej Karpathy, a co-founder of OpenAI and former head of AI at Tesla, has been generous with praise for the new team. In an industry where every second company tries to resell an API from Microsoft, H Company's approach looks like an attempt to return to real science. Karpathy notes that the focus on multi-agent systems could be the "silver key" to artificial general intelligence (AGI). Instead of one monolithic model, H Company is building a system in which specialized agents interact, test hypotheses and correct errors in real time.

Why does this matter right now? We have hit a data ceiling. High-quality text on the internet is running out, and training on synthetic data still raises more questions than answers. If H Company proves that "strong intelligence" can be built on compact but perfectly structured datasets, it would completely change the economics of the industry. Development costs would fall tenfold or more, and dependence on NVIDIA would stop being fatal. The investors, who include Bernard Arnault and Eric Schmidt, are clearly betting that the future belongs to efficiency, not brute force.

Skeptics will of course point out that inflated expectations often end in spectacular failures. H Company has nothing yet besides its founders' reputations and an ambitious manifesto. But in a world where Mistral has already shown how a small French team can push aside Silicon Valley giants, H Company deserves serious attention. They are not just building yet another neural network; they are trying to crack the very code of how a machine "thinks".

The bottom line: Will "small" become the new "big"? If H Company delivers on its plan, we are looking at the end of the era of giant LLMs and a shift to compact but truly intelligent agents.