A New York startup with no product, no benchmark and no shipped model just raised half a billion dollars. Flourish AI closed a $500 million round at a $2.5 billion valuation in early June 2026, backed by Jeff Bezos, Lux Capital, Alphabet’s GV and healthcare-focused fund Catalio. The premise is straightforward and ambitious: today’s large language models burn far too much electricity, and the only system that runs intelligence efficiently is the human brain. Flourish wants to copy how the brain works.

What Flourish AI is building

Flourish is developing AI models inspired by biological neural networks, with a system internally called Cortex AI. The target is a power budget of 20 to 50 watts, roughly the draw of a laptop or a couple of LED bulbs. A standard server-grade graphics card, by contrast, uses about 30 times more energy than the human brain to process the same kind of information. Closing that gap, even partially, would reshape the economics of every hyperscaler running AI workloads.

The company is not building chips. It is working one layer above the silicon, on the architectural and mathematical structures that define how a model learns and reasons. That distinguishes Flourish from the neuromorphic computing efforts at Intel and IBM, which have focused on specialised hardware. Flourish argues that the bottleneck is software design, not transistor design, and that better architectures can run on existing and future chip platforms without requiring new fabrication processes.

Cortical columns and connectomics

The scientific anchor of the work is connectomics, the mapping and analysis of neural connections in biological brains. Flourish researchers are studying cortical columns, neuron structures believed to act as the canonical computational units of the cortex. The lab is being equipped with electron microscopes capable of resolving structures that ordinary optical instruments cannot reach. Those instruments cost millions of dollars apiece and use beams of electrons rather than light, giving them resolutions many orders of magnitude finer than a standard microscope.

The thesis behind this work is that biological neural networks differ from transformer-based AI in fundamental ways. Brains are sparse, with most neurons not connected to most others. They process information asynchronously, firing only when a threshold is crossed rather than computing on every input at every step. They are extraordinary at routing information through local processing and hierarchical abstraction. Translating those properties into algorithms is the technical bet.

Continuous learning and memory

Beyond raw efficiency, Flourish wants to solve a problem current frontier models cannot: continuous learning. Once a large language model finishes training, it stops learning. A child acquires language from a few hundred thousand utterances. An LLM needs to ingest essentially every book ever written, repeatedly. Flourish is developing a hippocampus-inspired memory mechanism that would let models keep learning without expensive retraining cycles, and it has built a continuously learning prototype. The company is reportedly in talks with an unnamed chipmaker to ship a processor capable of running its model on consumer-grade devices.

Who is behind Flourish AI

The company was founded by Thomas Reardon and Rob Williams. Reardon has an unusual résumé. He dropped out of the University of New Hampshire at fifteen, joined Microsoft as a teenager, and led the team behind Internet Explorer in 1994. He served as a founding board member of the World Wide Web Consortium and delivered the first implementation of CSS in a browser. He later earned a PhD in computational neuroscience from Columbia University and co-founded CTRL-Labs, a neural interface startup that built a wristband translating electrical signals from the brain into computer commands. Meta acquired CTRL-Labs in 2019 for an estimated $500 million to $1 billion, and the technology now powers the Meta Neural Band, the wristband used to control Meta’s smart glasses.

Rob Williams is a former Amazon S-team executive who worked on Alexa. According to reporting from the funding process, Williams used the Amazon ritual of writing a press release for a product that does not yet exist and presenting it to Bezos for a yes-or-no decision. Bezos approved the pitch in December 2025, weeks after Williams left Amazon.

The scientific team

By the end of March 2026, Flourish had hired around two dozen senior neuroscientists and AI researchers into a ten-storey building in West SoHo with a built-in data centre. Co-founder Joshua Vogelstein recently co-authored a paper showing that a fruit fly’s neural network is roughly ten times more efficient than a transformer, the kind of comparative result Flourish wants to convert into a design principle. Jacob Vogelstein, who helped launch the Open Connectome Project, is connecting the lab’s brain-imaging work to the modelling team.

Greg Wayne, the DeepMind researcher who heads Google’s Project Astra, joined as a senior adviser after DeepMind CEO Demis Hassabis agreed to let Wayne spend 20 per cent of his time at Flourish while keeping his Google role. Ben Recht, a UC Berkeley computer scientist, is also advising on the technical side.

The investors and the $2.5 billion valuation

Jeff Bezos contributed roughly a fifth of the round, around $100 million. His initial commitment was about $50 million, but he nearly doubled the check after other high-profile investors joined. Lux Capital and GV led the round, both returning to back Reardon nearly a decade after they led CTRL-Labs’ Series A. Josh Wolfe, Lux’s co-founder, has been a vocal advocate for Reardon’s ability to translate neuroscience into commercial technology. Healthcare-focused fund Catalio also joined.

The valuation is striking. Two and a half billion dollars for a company without a product, a benchmark or a chip partner places Flourish in the same territory as startups with thousands of enterprise customers and material revenue. The investors are pricing the founder, the team and the size of the problem. They are not pricing traction.

Why the timing matters

AI’s power consumption is approaching the scale of national economies. Data centres and AI workloads consumed roughly 460 terawatt-hours globally in 2022, a figure the International Energy Agency expects to nearly double by the end of 2026. Data centres could account for 3 per cent of global electricity consumption by 2030, around twice Germany’s current total. In the United States, the number could reach 9 per cent of national electricity use by the same year. Combined AI infrastructure spending guidance from Alphabet, Amazon, Meta and Microsoft exceeds $650 billion for 2026 alone.

The industry response has been to build more power: nuclear deals, wind farms, even proposals for orbital solar arrays. Flourish is selling the opposite approach. If brain-inspired architectures can cut inference energy by even fifty per cent, the savings for a hyperscaler running millions of GPU-hours a day would be measured in hundreds of millions of dollars per year. The buyers for that intellectual property are precisely the companies currently spending tens of billions on transformer clusters.