A startup claims it broke through a bottleneck that’s holding back LLMs

The claim by Miami-based AI startup Subquadratic that it has broken through a fundamental mathematical bottleneck in large language models (LLMs) is more than just another industry press release—it could signal a turning point in how we scale artificial intelligence. At its core, the bottleneck in question has been the computational cost of training ever-larger models. For years, the exponential growth in model size has outpaced improvements in hardware efficiency, leading to ballooning energy consumption and diminishing returns. If Subquadratic’s solution—which it describes in abstract terms as resolving a "mathematical bottleneck"—holds up, it could democratize access to advanced AI by slashing the resources needed to train state-of-the-art models. The challenge isn’t new. The so-called "scaling laws" that govern LLM performance suggest that bigger is better, but the cost of training a model like GPT-4 or its successors has become prohibitive even for tech giants. Alternatives like sparse activation methods or mixture-of-experts architectures have offered incremental gains, but none have fundamentally altered the underlying math of attention mechanisms, which remain the computational linchpin. Subquadratic’s announcement implies it may have found a way to reduce the complexity of these mechanisms without sacrificing performance—potentially unlocking smaller, cheaper, and more efficient models. Yet skepticism is warranted. Breakthrough claims in AI are notoriously difficult to verify without peer-reviewed evidence or independent replication. The startup’s secrecy—coming out of stealth mode with little technical detail—raises questions about whether this is a genuine innovation or a marketing gambit. If real, the implications could ripple across the industry, forcing incumbents to rethink their infrastructure or risk obsolescence. It might also shift the balance of power in AI development, allowing startups and research labs with limited resources to compete with hyperscalers. The broader trend this hints at is the growing realization that raw compute power alone won’t sustain AI’s next phase. Efficiency breakthroughs, whether in algorithms, hardware, or system design, are poised to become the defining race of the decade. Whether Subquadratic’s claim withstands scrutiny could set the tone for how the industry approaches this inflection point.