The AI industry has a cost problem no one wants to talk about plainly.

Foundation model companies are burning billions annually to stay operational. The more capable their models become, the more expensive each inference call gets. The market expects token prices to fall. The physics of Transformer architecture says they cannot, at least not sustainably.

That tension is where the most interesting opportunity in AI sits right now.

Transformer architecture was designed for translation tasks. It processes the entire network on every inference call. Every neuron activates for every token, regardless of relevance. As tasks get more complex, compute requirements scale non-linearly.

Chain-of-thought reasoning, which powers models like OpenAI's o1 and o3, is not a solution to this problem. It is a workaround. Structured reasoning bolted onto an architecture that was never built to reason natively. It works. It is also extraordinarily expensive to run.

In September 2025, researchers at Pathway published a paper titled "The Dragon Hatchling: The Missing Link Between the Transformer and Models of the Brain." The paper introduces BDH, an architecture that operates the way the brain actually does: through sparse, selective neuron activation based on relevance rather than full network activation on every inference step.

The research reveals two properties that matter for enterprise deployment. First, efficiency. Think of a Transformer like a factory where every single worker shows up for every job, even if only two people are needed. BDH only calls in the workers relevant to the task. The paper measured this: only around 5% of the model is active at any given moment. The other 95% stays idle. That is where the cost reduction comes from.

Second, transparency. With a Transformer, you see the answer but you cannot trace how the model got there. With BDH, you can. The paper identified specific neurons that activate only when a currency is mentioned and others that activate only for country names. The model has essentially learned to file concepts into dedicated mental folders. You can open those folders and read them. For a bank, a hospital, or a law firm deploying AI in a regulated environment, that auditability is what makes deployment legally defensible.

BDH also learns continuously without forgetting. Competing architectures show dramatic performance degradation when trained on new tasks. BDH does not. This is validated in the paper across multiple experiments.

A 95% reduction in active parameters per token should, in theory, threaten GPU demand. Yet in December 2025, AWS and Nvidia jointly announced an infrastructure collaboration with Pathway at AWS re:Invent. BDH now runs on Nvidia AI infrastructure through AWS's cloud stack, with AWS as Pathway's preferred cloud provider and Nvidia Hopper architecture identified as the optimal hardware environment for BDH's workloads.

The logic is Jevon's Paradox. When a resource becomes more efficient to use, total consumption typically increases because efficiency unlocks use cases that were previously uneconomical. Cloud computing did not reduce server demand. It multiplied it. If BDH makes enterprise AI viable for verticals that Transformer pricing currently prices out, the total GPU market grows even as per-token costs fall. By ensuring BDH runs optimally on Nvidia hardware first, AWS and Nvidia are positioning their stack as the natural platform for the architecture transition they believe is coming.

Regulated industries have been slow to adopt AI at scale. The obstacle is not capability. It is trust. Black box reasoning creates liability. Hallucinations create liability. BDH's observability addresses this directly. If you can read which neurons activated and why a conclusion was reached, the black box risk disappears. Pathway's existing enterprise customers include NATO, a Formula 1 racing team, and Intel.

The BDH paper was the number one paper of the day on HuggingFace, the day after its September 2025 publication. Forbes covered the research independently. The architecture is now live on Nvidia AI infrastructure and available through AWS.

The Transformer is not going away. What the paper proves mathematically is that it is not the end state of AI architecture. If Jevons' Paradox holds here the way it has everywhere else, the firms positioned on both sides of that transition, the architecture and the infrastructure beneath it, are the ones worth watching. 

Bashar Aboudaoud
Managing Member, UpRound