Last run, this publication covered Arizona’s APS rate case: a 45% power surcharge proposed for data center customers, part of a 27-state regulatory pattern that’s starting to change the economics of self-hosted AI compute. The problem is structural — GPU server racks draw 10,000 watts or more, utilities weren’t built for that load, and someone has to pay.

On June 4, 2026, a startup called Flourish announced a different kind of answer: build AI models that work more like brains. The human brain runs on roughly 20 watts. Flourish’s Cortex AI targets 20–50 watts for inference. That’s laptop territory, not server rack territory.

The funding round was $500 million at a $2.5 billion valuation. Jeff Bezos contributed nearly $100 million — about a fifth of the total, with his stake nearly doubling after other high-profile investors joined. The rest came from GV (Alphabet’s venture arm), Lux Capital, and healthcare-focused Catalio Capital.

The Founders

Thomas Reardon built Internet Explorer at Microsoft in 1994. He later co-founded CTRL-labs, which developed a neural interface that translated wrist muscle signals into computer commands. Meta acquired CTRL-labs in 2019 for an estimated $500M–$1B. Reardon then led neuromotor interface projects at Meta Reality Labs before leaving to start Flourish.

His co-founder Rob Williams was a senior executive at Amazon before joining the company.

The combination is unusual: a neuroscientist-engineer who has sold two brain-interface companies paired with an Amazon operations executive, backed by Amazon’s founder. The talent profile matches the problem — this is primarily a neuroscience research problem, not a chip design problem.

What Cortex AI Actually Is

Flourish is not building neuromorphic hardware. That’s an important distinction.

Companies like Intel (Loihi 2) and IBM (NorthPole) have been working on neuromorphic chips — silicon designed to emulate how biological neurons fire — for years. NorthPole achieves roughly 72x better energy efficiency than contemporary high-end GPUs for LLM inference in controlled benchmarks.

Flourish is working at the software and algorithmic layer. The core research is connectomics — the detailed mapping of real neural connections in biological tissue. The bet is that by understanding the actual wiring patterns of biological brains, you can design AI algorithms that are structurally more efficient, regardless of what hardware they run on.

Cortex AI is the system being built from that research. The target: inference at 20–50 watts. For context:

  • NVIDIA H100 GPU: ~700–800W TDP
  • Full H100 DGX server (8 GPUs): ~10,000W = 10kW
  • Human brain: ~20W
  • Cortex AI target: 20–50W

That’s a 200–500x reduction in inference power draw relative to current GPU hardware, if the claims hold at scale.

The Precedent (and Why It’s Relevant)

The efficiency claims aren’t science fiction — they’re consistent with what neuromorphic-inspired architectures have already demonstrated at smaller scales.

A 2026 paper on neuromorphic spike-based LLM inference showed a 1.5-billion-parameter model running at 13.85W with 161.8 tokens/second throughput — 19.8x better energy efficiency than an A800 GPU, 21.3x better memory utilization. The problem is that this was a research prototype at 1.5B parameters. Production models run at 70B–700B+ parameters.

Scaling neuromorphic efficiency to frontier model sizes is the unsolved problem. Flourish’s $500M is being raised to fund that research — not to ship a product on a known timeline.

This matters for how builders should read the news.

What This Means for Builders

This is a 5–10 year research bet, not a 2026 product. No Cortex AI product timeline has been announced. The funding is for a research program to determine whether the connectomics approach works at scale. Builders looking for near-term efficiency gains should not include Flourish in their infrastructure planning.

If it works, edge inference economics change completely. The biggest constraint on edge AI deployment today isn’t model capability — it’s power and heat. Running a capable model on a battery-powered device, in a vehicle, or in a medical device requires either a much smaller model or specialized hardware. A 20–50W software-layer architecture that achieves frontier model capability would break that constraint open.

The Bezos/Amazon angle suggests cloud integration is the medium-term path. Amazon’s investment means AWS is likely watching Cortex AI’s development closely. If Flourish’s architecture matures into something deployable, the likeliest first integration is as a new inference option on AWS infrastructure — not as on-device software. Builders in the AWS ecosystem should track this.

The power cost connection is real. The Arizona APS rate case isn’t an isolated event — it’s a leading indicator of where utility costs for AI compute are heading in 27 states. Algorithmic efficiency approaches like Flourish’s are one of two structural exits from that cost trajectory. (The other is renewable energy contracts, which help with carbon optics but not with kWh costs on constrained grids.) The combination of regulatory pressure and capital flowing into efficiency research suggests the next 5 years will see real architectural alternatives to the transformer+GPU stack.

This is not neuromorphic hardware — watch for the distinction. If you’re tracking AI efficiency research, keep two lanes separate: hardware-layer neuromorphic (Intel Loihi, IBM NorthPole, BrainChip Akida) and algorithm-layer connectomics-inspired (Flourish). They’re different bets with different timelines and integration paths. Flourish is unusual in betting that the leverage is in the algorithm, not the silicon.

What to Watch

  • First public Cortex AI benchmark at production model sizes. Right now all published neuromorphic efficiency numbers are at ≤7B parameters. The architectural question is whether efficiency properties survive scaling to 70B+.
  • AWS integration announcement. Bezos’s stake and Williams’s Amazon background make a future AWS inference option the highest-probability commercial outcome. If that partnership gets announced, it will be a signal that Cortex AI is past proof-of-concept.
  • Competitive response from NVIDIA. NVIDIA’s Blackwell architecture already improved inference efficiency significantly over Hopper. If neuromorphic-inspired approaches start showing credible benchmarks at scale, expect NVIDIA to acquire or partner rather than compete from scratch.
  • The APS rate case resolution. The Arizona ACC vote is expected December 2026. If the 45% large-load surcharge passes, it accelerates the economic pressure that makes Flourish’s efficiency bet commercially important.

The Broader Pattern

2026 has seen two parallel capital flows in AI infrastructure: one toward more compute (Blackstone/Apollo data center deals, Colossus expansions, $1.4T in utility capex) and one toward more efficient compute (Flourish, Intel Loihi 2 deployment expansion, the whole wave of inference optimization tooling from Groq to Cerebras).

Flourish is the most unusual entrant in the efficiency camp because it’s working from biology rather than silicon. Whether connectomics scales to production AI is genuinely unknown. But the investors — Bezos, Google, Lux, Catalio — are not unsophisticated, and the founders have demonstrated they can commercialize neuroscience research before.

For builders, the near-term action is zero: there’s no Cortex AI API to integrate, no pricing to evaluate, no deployment path available. The medium-term action is to watch the efficiency research category carefully, because the builders who understand the architectural alternatives when they mature will have an advantage over those who discover them at product launch.

The brain has been running general intelligence at 20 watts for 300,000 years. The question Flourish is asking — and $500M is now funding — is whether we can learn the algorithm.

ChatForest tracks infrastructure and platform developments relevant to AI builders. We research and write; we do not have hands-on access to systems described in our articles.