A mathematical bottleneck behind expensive LLMs: finally some progress?
Subquadratic, a Miami-based startup, claims to have solved a fundamental mathematical limitation that has constrained language models for a decade. The problem: current transformers consume computing power that explodes with context length (what’s called quadratic complexity). In practical terms, the longer your AI needs to process text, the exponentially more expensive and slower it becomes.
This announcement comes at a time when AI infrastructure costs are becoming the real issue for businesses. Initial skepticism is justified: technical details remain unclear, and revolutionary AI claims come ten a week. But if Subquadratic delivers on its promises, the gains wouldn’t be minor: reducing quadratic complexity means AI APIs 3-5 times cheaper, and models capable of handling 10 times longer contexts without proportional cost increases.
Timing is critical. Finance departments are starting to reject “unlimited” AI budgets. Real infrastructure cost reduction would change the equation for ROI on automation projects.
What this means for your business
Why this matters for you. Today, deploying an AI agent in production is expensive mainly because of infrastructure: the more the model needs to contextualize (read your documents, customer history, knowledge base), the higher the bill climbs. An SMB automating customer support or quote processing quickly sees unreasonable API costs.
If Subquadratic delivers, you can expect: profitable AI projects for use cases that currently don’t pass the ROI filter—especially “low volume / long context” scenarios (document management, contract analysis). More importantly: more honest cost comparisons between building internally and using APIs.
Action: don’t chase the hype, but actively monitor Subquadratic’s benchmarks when they’re published. Your next AI architecture decision will depend on it.
In brief
Executive turnover: instability at OpenAI
Barret Zoph, OpenAI’s VP of enterprise AI sales, is leaving after just five months. He had previously left the company in January to found Thinking Machines Lab with Mira Murati (former OpenAI CTO). This departure reinforces signals of management instability at the market leader.
Kenya and energy: the decentralized model for off-grid communities
With 25% of communities lacking access to the central grid, Kenya is betting on off-grid solar to achieve universal electrification by 2030. A lesson in how to build AI infrastructure in remote regions: decentralize rather than centralize.
Dark matter detection: major breakthrough in underground experiments
Liquid xenon detectors placed beneath the Apennines, the Jinping Mountains, and Dakota mines are advancing direct detection of dark matter. Outside the AI domain, but illustrative of a trend: major scientific breakthroughs emerge from patient, coordinated infrastructure.
Get The AI Brief in your inbox
3x per week, the essentials of AI decoded for business leaders.