Every time someone in your company asks an AI tool to summarise a document or generate a report, a data centre somewhere burns a significant amount of electricity to produce that answer. That cost is mostly invisible to you now because the big AI companies are absorbing it to build market share. It will not stay invisible.
Data centre electricity demand grew 17% in 2025, and the International Energy Agency projects the total will double by 2030. The five largest tech companies spent over $400 billion on data centre infrastructure last year and are set to spend 75% more this year. In parts of the US, data centres now consume over a quarter of all state electricity. Utility bills for ordinary households are rising partly as a result. This is the infrastructure problem that AI currently refuses to acknowledge.
Flouish is a startup that wants to solve it by going back to the drawing board. The human brain operates on roughly 20 watts, similar to a standard LED bulb, while a single chip inside a large AI training cluster uses more than 30 times that. Current AI models also have to be trained on enormous amounts of text before they can do anything useful, and once that training is done they stop learning. A child learns a language from a few hundred thousand heard sentences; today's large AI models need to process essentially everything humans have ever written, multiple times over.
Flouish's argument is that these two failures, energy waste and inability to learn on the go, come from the same root cause: AI software was loosely inspired by biology decades ago, then diverged completely. The company wants to study how real brains are actually wired, down to the level of individual neuron connections, and use those findings to build better AI architecture. The goal is an AI system that runs on 50 watts or less and adapts continuously from new information without needing a full retraining cycle.
Thomas Reardon is leading the company. He built Microsoft's first web browser in 1994, later completed a doctorate in computational neuroscience at Columbia, co-founded a startup that built a wristband capable of reading nerve signals from the wrist, and sold that company to Meta for between $500 million and $1 billion. The wristband now ships with Meta's smart glasses. His co-founder is Rob Williams, a former executive at Amazon who ran products including Alexa. That combination of technical depth and large-scale product experience is rare, and it is a large part of why investors are willing to fund a company with no product at a $2.5 billion valuation.
Jeff Bezos committed $50 million after reading a two-page pitch document, then nearly doubled his stake. Lux Capital and Google Ventures are anchoring the round. A senior researcher from Google's DeepMind division, who runs one of Google's most advanced AI assistant projects, joined as a part-time adviser. The company had hired around two dozen neuroscientists and AI researchers by the end of March, working in a New York office with lab equipment arriving in stages.
Flouish is not alone in this direction. At least 28 companies are now working on brain-inspired computing, including established projects at Intel and IBM. Several are attacking the hardware layer, building new chips rather than new software. Flourish is focused on the software architecture layer, which means its findings could in theory run on existing hardware and benefit a wider range of businesses faster.
The near-term commercial path involves a memory system inspired by the hippocampus, the part of the brain that handles new learning, which would let AI models absorb new information without requiring a full retraining. The team is also working on a model designed to run on devices you carry in a pocket, and is in discussions with a chip manufacturer about embedding the approach into silicon.
The honest answer about whether any of this will work is: unknown. The adviser from Berkeley who sits on Flourish's board says directly that he is not convinced it will succeed. But the energy trajectory of current AI is not sustainable, and any operator using AI tools at scale should be watching this space closely. If efficient, continuously learning AI becomes real in the next five to ten years, the cost curve for using AI in your business changes substantially. That is worth understanding now, even if the product does not exist yet.