Subquadratic claims an LLM bottleneck fix, but experts say the receipts still don’t add up

An AI startup says it cracked a bottleneck that has haunted large language models for almost a decade. Subquadratic came out of stealth last month with a big claim: it solved a mathematical bottleneck by slashing the number of computations transformers need to generate answers, producing an LLM that is faster, cheaper, and uses far less energy than any other model on the market.

That is the headline claim, and it matters because compute is the tax that keeps getting paid, again and again, by everyone building and funding LLMs. But the story also comes with a built-in skepticism: many experts remained unconvinced, even as Subquadratic started to share the receipts. This is the moment where boards should stop treating “efficiency” as marketing, and start treating it like an investment-grade metric.

So what is the bet? The pitch is not a new dataset, or a new tool for prompting. It is a structural change to how much computation a transformer system needs to carry out to generate answers. In plain English, if you can get the model to do less work per response, you can lower cost per query, improve latency, and reduce energy burn. That combination tends to be a straight shot to better unit economics. In AI, the unit economics question never stays quiet for long, because inference at scale is where margins get won or lost.

Still, “faster and cheaper” is not automatically “better” in the way that enterprise buyers care about. The MIT Technology Review piece points out that skepticism persists, and that some researchers have not been convinced yet. The reason this debate is worth your attention is that the definition of “breakthrough” can quietly shift. A method can reduce computations yet trade off accuracy, reliability, or robustness in ways that show up only under tough evaluation. Or it can deliver gains in a narrow setting that looks impressive in a demo, but less so at production scale. The article flags that Subquadratic has started to share the receipts, which is exactly what executives should demand when claims challenge existing assumptions.

Meanwhile, in a very different corner of technology, brain-computer interface trials are taking off in a way that should make operators think about the pace of clinical validation and commercialization. The Download’s other lead, from The Checkup, highlights a case: Casey Harrell, a man with ALS, who is described as the first power user of a brain implant. The device has enabled him to maintain an income, reconnect with friends and family, and read to his daughter. In that same segment, Harrell is quoted as calling the experience “nothing short of revolutionary.” That is not a scientific paper, but it does illustrate what “getting to market” means in human terms.

The key trend is that the number of BCI trial volunteers has soared over the past couple of years. And this year, China became the first country to approve a BCI for medical use. That regulatory milestone is huge because it signals something markets and investors track closely: the pathway is real. Technology Review also notes that advances in technology are allowing engineers to provide more features than ever. In other words, the space is not just growing in patient count, it is improving in capability. For decision-makers, the second-order implication is that timelines for adoption can compress once regulators clear initial use cases, especially when devices demonstrate tangible day-to-day value for patients.

Now connect these two stories, because they rhyme in how they test credibility. In AI, you have an efficiency claim aimed at a mathematical bottleneck that researchers say has held back large language models for almost a decade. In BCIs, you have an approval milestone that turns experimental systems into regulated medical tools. Both involve skepticism and proof requirements. Both require careful evaluation of performance, scalability, and real-world outcomes, not just lab results.

The “receipts” concept should apply to both. In AI, the board question is whether compute reductions translate into dependable performance across the situations where customers actually use models. In BCI, the question is how trial growth and regulatory approval map onto manufacturing, clinical support, and longer-term outcomes. And for peers watching all of this, the strategic stake is clear: if you are betting on the future of compute-heavy AI, you need a serious answer to what changes when someone claims a new efficiency ceiling. If you are tracking health-tech hardware, you need to understand how quickly clinical pathways can turn into product lines.