This Startup Claims It Broke the Biggest Bottleneck in AI — and Now There’s Proof

Miami-based AI startup Subquadratic came out of stealth last month with a claim that sounded almost too good to be true: they’d cracked a mathematical bottleneck that has been holding back large language models for nearly a decade.

At first, the details were vague. A handful of self-published test scores. No model for anyone to try. Naturally, the reaction on X was brutal. AI engineer Dan McAteer summed it up perfectly: “SubQ is either the biggest breakthrough since the Transformer … or it’s AI Theranos.”

But now Subquadratic is starting to bring the receipts. They’ve released results from an independent evaluation by Appen, a third-party firm that tests other companies’ models. And those results suggest the hype might actually have some substance.

Subquadratic’s model, called SubQ, is supposedly faster, cheaper, and far more energy-efficient than anything else on the market. They claim it can process up to 12 times as much text at once — think analyzing hundreds of documents or entire code bases in one go — while matching the performance of top models from Google DeepMind, OpenAI, and Anthropic on key tasks like coding.

“We expected healthy skepticism,” says Subquadratic cofounder and CTO Alex Whedon. “In hindsight, releasing the third-party benchmarks alongside the initial announcement would have preempted much of the skepticism.”

Appen’s director of generative AI research, Jeanine Sinanan-Singh, ran the tests herself. “That was really exciting to me, it validated their architecture,” she says. “I was like, ‘Wow, this could be a game changer,’ because models struggle with speed and inefficiency. But when you have kind of shocking results, it’s really not as credible when you say it yourself.”

To understand why this matters, you have to look at how most LLMs work. The core mechanism is a transformer running something called dense attention. When a transformer processes text, it encodes each word (or token) as a number, then multiplies every number with every other number in that text. For a 10,000-word passage, that’s about 50 million individual multiplications. That’s the main reason LLMs are such power hogs.

“If you want to summarize The Great Gatsby, you have to look at the first word and the last word together, and then you have to look at every other combination,” says Subquadratic CEO Justin Dangel.

Subquadratic’s breakthrough is about replacing that dense attention with something far more efficient. They haven’t disclosed the exact math, but the results speak for themselves: SubQ can handle much longer contexts without the computational cost exploding.

SubQ won’t replace existing top models across the board. But for tasks where speed and cost matter — and that’s most real-world applications — it could be a huge leap. Dangel is even bolder: “We hope we’re kicking off a new age of efficiency. We don’t think anybody will be building on transformers in a few years.”

That’s a strong statement. Transformers have been the foundation of every major LLM since Google’s 2017 paper “Attention Is All You Need.” But if Subquadratic’s architecture holds up under wider scrutiny, we might be looking at the beginning of the post-transformer era.

Of course, caution is warranted. The model still isn’t widely available. And one set of third-party benchmarks, however promising, isn’t the same as real-world deployment at scale. But for the first time, Subquadratic has evidence that goes beyond their own claims. That’s more than most vaporware startups ever deliver.

I’ll be watching closely to see who gets early access and what the independent replications look like. If SubQ is as good as Appen’s tests suggest, this could be the most important AI efficiency breakthrough since — well, since the transformer itself.