It’s All Math
A startup says it cracked the bottleneck holding back AI
https://thenextweb.com/news/subquadratic-subq-sparse-attention-llm-bottleneck
Miami startup Subquadratic claims its SubQ model breaks the 'quadratic attention' bottleneck. Independent tests back much of it, but doubts remain.
Filed June 21, 2026 at 11:17 am
A Miami startup called Subquadratic says it cracked the math problem that's kept AI models slow and power-hungry for almost a decade, a claim loud enough that one engineer called it either the biggest breakthrough since the transformer or "AI Theranos."
The bottleneck is real. Since 2017, transformers have compared every word in a text to every other word, so doubling the length quadruples the work. Subquadratic's model, SubQ, keeps only the word pairs that matter and picks them based on content. When an outside firm ran the tests, the numbers held: 56 times faster than a leading method, 89.7 percent on a hard coding benchmark, and a long-context test that costs around $2,600 on Anthropic's top model done for eight dollars.
The caveats are pretty obvious... benchmarks aren't real use, almost nobody has access yet, and SubQ wasn't built from scratch but grafted onto an existing open-weight model. They may have something real, but the evidence doesn't yet justify the bigger story. Their CEO seems to be betting everyone follows anyway.