The news comes with just a few lines, almost under the radar, but it strikes at the heart of a tension running through the entire world of Large Language Models: the possibility of achieving transformer-like performance without paying the price – in GPUs, energy, and dependencies – that transformers impose. Pathway, a company operating in the field of alternative neural architectures, has shown that its BDH architecture (a post-transformer evolution, still light on technical details) scales almost identically to GPT-2, covering ranges from 10 million to 1 billion parameters, and does so while training from scratch on ordinary GPUs.

This is not an isolated benchmark: matching such an iconic model family – even if now surpassed by later generations – serves as a proof of concept. And the concept is disruptive because it breaks the assumption that multiplicative attention is the only path to robust scaling laws. Over the past two years we have seen state-space models (Mamba, Hyena, RWKV) and recurrent variants promise linear efficiency in sequence length, but often with trade-offs in generalization on complex linguistic tasks. BDH, according to the shared data, seems to sidestep the problem with an approach that empirically reproduces the perplexity growth dynamics of transformers, without inheriting their quadratic computational complexity.

For on-premise and self-hosted deployment, the message is clear: if architectures like this reach maturity, the rush to enterprise-grade GPUs – from A100s to H100s – may no longer be the only path to training or inference with competitive LLMs. The TCO of an on-prem cluster would plummet, along with the barriers to entry for companies, research institutions, and public administrations that today must choose between cloud and unsustainable hardware investments. From a data sovereignty perspective, the ability to train models comparable to GPT-2 internally, using consumer or at least export-unrestricted hardware, shifts the balance of negotiating power toward those who own the data and wish to keep it local.

Of course, the leap from a demonstration to a market-ready technology is substantial, and Pathway has not yet released details on token context, energy costs, or fine-tuning for specific tasks. But it is precisely in these first cracks in the Transformer supremacy that the next battles of artificial intelligence will be fought. It is no coincidence that NVIDIA is diversifying its stack beyond simple matrix multipliers, and that large cloud providers sense the need to support a more heterogeneous model ecosystem.

In the logic of AI-RADAR, which closely follows every signal that reduces dependence on oversized hardware components, Pathway's announcement is not just a paper: it is a notice that the transformer monolith may not be eternal, and that the economic and technical assumptions on which we built the current wave of LLMs could soon be redefined.