At $500 million, Flourish is the largest bet ever placed on the idea that copying the brain's wiring — not scaling GPUs — is the only path out of AI's energy crisis. Jeff Bezos committed close to $100 million of that himself. The startup has no product, no revenue, and no published benchmarks. The valuation: $2.5 billion.
It raised $500M at a $2.5B valuation from Bezos, Lux Capital, GV, and Catalio Capital to build Cortex AI, a synthetic intelligence system targeting 20–50 watts of operating power — the same range as the human brain.
The company's method is connectomics: mapping real neurons and their connections at electron-microscope resolution to extract a compact "core algorithm" that explains how three pounds of biological tissue outperforms megawatt-scale data centers at reasoning, generalization, and continuous learning.
If Cortex AI delivers even a fraction of its efficiency target, it reshapes the unit economics of every downstream AI product and undercuts the infrastructure rationale behind the current $650B hyperscaler capex cycle.
TIMELINE: Connectomics → Cortex AI
─────────────────────────────────────────────────────────────
2009 2019 2024 2026 2027+
HCP CTRL-labs FlyWire Flourish Cortex AI
starts → Meta connectome $500M Target:
$38.5M Reardon published Bezos 20–50W
human sells (Nature) backs silicon
brain BCI ~140K connectomic inference
mapping startup neurons AI startup ?
◉ PAST ◉ PAST ◉ PAST ◉ NOW 🔮 NEXT
From the Human Connectome Project to the startup's Cortex AI. The thread of reverse-engineering biological intelligence.
AI Has a Power Problem That More GPUs Won't Fix
The arithmetic no longer works. A single server-grade GPU draws more than 1,000 watts under load. A frontier AI training run consumes megawatts over weeks. The collective capital expenditure on AI infrastructure from Alphabet, Amazon, Meta, and Microsoft alone exceeds $650 billion in 2026 guidance. And The bottleneck is no longer compute. It is grid capacity, cooling, and the 18-month lead time to bring a new data center online.
The industry's answer so far has been to build more power. More GPUs per cluster. More nuclear PPAs. More gas peaker plants near server farms. But the company's founding thesis, shared with an increasingly vocal minority of researchers, is that the industry is optimizing at the wrong layer. The real efficiency gains are not in the chip. They are in the architecture. And the best reference architecture in existence runs on 20 watts, fits inside a skull, and learned to generalize from a few hundred thousand examples before its owner could walk.
The Man Who Built Internet Explorer Went Back to Study the Brain
Thomas Reardon's career reads like three separate lives. At 19, he created the project that became Internet Explorer, turning Microsoft into an internet company and triggering the most consequential antitrust case in technology history. In between, he helped deliver the first implementation of CSS and served on the founding board of the World Wide Web Consortium.
Then he left software and earned a PhD in neuroscience from Columbia University. In 2015 he founded CTRL-labs, a brain-computer interface startup that read electrical signals from the forearm muscles to control computers. Meta acquired it in 2019 for an estimated $500 million to $1 billion, and Reardon went on to direct neuromotor interface research at Meta Reality Labs. His co-founder, Rob Williams, is a former Amazon S-team executive who helped shape strategy at one of the world's largest technology organizations.
The startup was incorporated in 2024. It operated in stealth until April 2026, when Bloomberg reported that Reardon was raising $500 million at a $2.5 billion valuation. The round closed in five weeks. Bezos initially committed $50 million, then nearly doubled his stake after Lux Capital, GV (Alphabet's venture arm), and Catalio Capital joined. The final check from Bezos was close to $100 million.
What Cortex AI Actually Is
The startup is not building a chip. It is building what it calls Cortex AI, a synthetic intelligence system whose design is derived directly from biological neural tissue rather than from the transformer architecture that powers every major AI model today.
The method is connectomics: mapping individual neurons and their synaptic connections at sub-micron resolution using electron microscopy. The field has produced increasingly detailed wiring diagrams — the FlyWire project published the complete connectome of the adult fruit fly in 2024 (140,000 neurons, 50 million synapses). Eon Systems used that data in March 2026 to produce the first embodied whole-brain emulation, a milestone we wrote about in July. The company is taking the same fundamental approach but aiming higher: it wants to find the brain's "core algorithm", a compact, reusable computational principle that explains why biological networks learn, generalize, and adapt on a power budget that no artificial system approaches.
The company has hired roughly two dozen neuroscientists and AI researchers, and is equipping an in-house lab with electron microscopes capable of resolving structures far smaller than what optical instruments can see. One stated focus is cortical columns — the repeating neural circuits, roughly 0.5 millimeters across, that many researchers believe are the fundamental processing units of the neocortex. Understanding how a cortical column computes could provide a blueprint for a silicon implementation that preserves the brain's efficiency while operating at digital speeds.
Cortex AI targets 20 to 50 watts, roughly the power draw of a laptop. A single H100 GPU, by comparison, draws 700 watts. A Blackwell B200 draws 1,000. The gap is not a factor of two or ten. It is a factor of hundreds.
Where the Startup Fits in the AI Efficiency Race
| Parameter | Flourish | Groq | Cerebras | Etched |
|---|---|---|---|---|
| Approach | ✔ Architecture (connectomics) | ◐ Silicon (LPU) | ◐ Silicon (wafer-scale) | ◐ Silicon (transformer ASIC) |
| Power target | ✔ 20–50W | ◐ ~100W per card | ✗ ~15kW per system | ◐ ~50–100W per card |
| Commercial product | ✗ None | ✔ Shipping | ✔ Shipping | ✔ Shipping |
| Funding raised | $500M | $750M+ | $700M+ | $500M |
Groq, Cerebras, and Etched are all building specialized silicon for faster AI inference. It works at the model and architecture level, using the brain as its reference design. This makes it harder to compare directly. If Cortex AI works, the relevant benchmark is no longer the same one Nvidia competes on.
The Bet That Could Reshape AI Infrastructure
What makes the startup different from the dozens of neuromorphic computing startups that preceded it is not the technology. It is the scale of the bet. A $2.5 billion valuation for a pre-product connectomics startup signals that seasoned investors treat brain-inspired computing as a genuine competitor to transformer scaling, not a research curiosity.
If Cortex AI holds at even 200 watts — ten times its stated target — it would still represent an order-of-magnitude improvement over current inference hardware. That changes the unit economics of every AI product. A model that delivers comparable quality at a fraction of the power changes who can afford to serve inference at scale, which incumbents must defend their margins, and what categories of AI product become viable for the first time.
A continuously learning model running on a consumer device — a phone, a wearable, a robot — is not viable today because the cloud dependency and power cost make the unit math work against it. A model that learns on-device at single-digit watts would reset that math entirely.
The risks are equally large. Neuroscience-inspired AI has been in development for decades without displacing the approaches that emerged from empirical scaling. Translating biological insights into working silicon systems has a poor commercial track record. And the startup is competing for talent — computational neuroscientists and silicon engineers — against hyperscalers and chip companies that offer more near-term certainty and compensation.
The transformer is not a solved architecture. It is a 2017 paper that happened to scale well. If the brain's wiring contains a different computational principle, and if connectomics can extract it, the payoff is not a slightly better GPU — it is a different category of intelligence.
The company publishes its first benchmark or peer-reviewed paper — proof that connectomics-derived architecture produces measurable efficiency gains.
A hyperscaler (Google, Microsoft, Amazon) announces a partnership or investment in connectomics-based AI — institutional validation of the approach.
Cortex AI achieves a real inference workload at under 200W — demonstrating the efficiency gap is real, not theoretical.
A major competitor (Groq, Cerebras, Etched) announces an architecture pivot toward brain-inspired design — the market following the startup's thesis.
