The transformer neural network architecture that has powered every major large language model for nearly a decade is facing its most serious challenge yet. A growing number of startups are developing alternatives that promise to be faster, cheaper and more energy-efficient, raising the possibility that the next generation of AI could run on fundamentally different technology.

What You Need to Know

Transformers, introduced by Google researchers in 2017, process text by comparing every word against every other word, a method that becomes computationally expensive as text length grows. Startups like Subquadratic are developing new architectures that reduce this computational load, claiming dramatic efficiency gains. The shift could lower the enormous energy costs of running AI models and enable new capabilities that current transformers struggle with.

The Transformer Bottleneck

The core strength of transformers lies in dense attention, a mechanism that encodes the meaning of text by comparing every token with every other token. This approach works well for short passages but becomes a liability as input grows. A document 10,000 words long can require 50 million multiplications, a process that strains computing resources and consumes huge amounts of energy.

OpenAI, for example, expects to spend $50 billion on computing this year, according to company president Greg Brockman. The International Energy Agency predicts that data center electricity consumption could double by 2030. These costs are not just financial. They also limit what transformers can do. The architecture struggles to maintain large context windows, and reasoning models that rely on chain-of-thought note-taking only add to the data that must be tracked.

Nine years after the landmark paper "Attention Is All You Need" from Google, the industry built on transformers is confronting a fundamental limitation. The technology's key strength has become a bottleneck.

Four New Approaches to AI Architecture

A wave of startups, including Subquadratic led by CEO Justin Dangel, is betting on radically different methods. These new architectures aim to reduce the computational explosion of dense attention while maintaining or improving accuracy. Among the most promising ideas are sparse attention, rolling-summary memory, diffusion-based text generation and worm-brain-inspired networks.

  • Sparse attention: Instead of comparing every word with every other word, this approach selects only the most relevant comparisons, cutting the number of calculations dramatically.
  • Rolling-summary memory: The model keeps a running summary of earlier text, allowing it to reference past information without reprocessing the entire sequence.
  • Diffusion-based text generation: Borrowed from image generation, this method starts with noise and gradually refines it into coherent text, sidestepping the sequential constraints of transformers.
  • Worm-brain-inspired networks: These systems mimic the simple neural circuits of roundworms, which handle complex tasks with minimal computational resources.

Some of these approaches are already showing results. One startup using a non-transformer architecture solved 97% of 250,000 hard sudoku puzzles that top LLMs could not crack, suggesting that the current architecture may actively limit reasoning capabilities.

Why This Matters

The race to replace transformers is not just about speed and cost. It could determine whether AI reasoning models achieve human-level problem solving. If these alternatives succeed, the dominance of companies like OpenAI and Google could weaken, opening the door for smaller players. The energy savings alone could be transformative, reducing the carbon footprint of AI and making advanced models accessible to more organizations. The outcome of this competition will shape the next era of artificial intelligence, determining which architectures power the systems that increasingly mediate our digital lives.

What Comes Next

No single alternative has yet emerged as a clear successor to the transformer. Many of these startups will fail, but the ones that succeed could fundamentally change the economics of AI. LLMs themselves are not going away. The way they are built, however, is very much up for grabs.