The guy who built Meta's brain-computer interface just left to build AI chips that think like neurons—and investors are already pricing in a future where your agents run on milliwatts, not megawatts.
The Summary
- Thomas Reardon, former lead on Meta's Neural Band brain-computer interface, is raising funds for Flourish, a startup targeting energy-efficient AI at a $2.5 billion pre-revenue valuation
- The bet: neuromorphic computing can solve AI's energy crisis by mimicking how biological brains process information
- Signal for builders: if this gets funded at $2.5B, the market believes the current GPU paradigm is economically unsustainable at scale
The Signal
Reardon isn't a typical AI founder. He spent years translating neural signals into digital commands at Meta. Now he's reversing the equation: building digital systems that compute like brains. The startup is called Flourish, and it's going after the most expensive problem in the agent economy—keeping the lights on.
Current AI infrastructure burns power like it's free. Training GPT-4 took an estimated 50 gigawatt-hours. Running inference for ChatGPT costs OpenAI roughly $700,000 per day. Scale that to a world where every person has multiple AI agents running 24/7, and you hit an energy wall fast. Data centers already consume 1-2% of global electricity. Agents push that toward 5-10% by decade's end.
"The brain runs on 20 watts. A GPU cluster running equivalent inference burns 20 kilowatts. That's not a rounding error."
Neuromorphic computing promises a different physics. Instead of moving data between memory and processors millions of times per second, it processes in place using analog circuits that spike like neurons. Intel's Loihi 2 chip runs certain AI workloads at 1000x better energy efficiency than GPUs. IBM's TrueNorth processes sensor data on microwatts. But neither scaled to general-purpose AI. That's the gap Flourish is betting it can close.
The $2.5 billion valuation tells you where the smart money thinks this goes. For context:
- Cerebras, which makes wafer-scale AI chips, raised at $4B in 2021
- Graphcore, the British AI chip unicorn, peaked at $2.8B before stumbling
- SambaNova, another AI chip startup, hit $5B valuation last year
Reardon's edge is cross-domain expertise. Building brain-computer interfaces meant solving latency, power, and signal processing under biological constraints. You can't just throw more transistors at a neural implant. You work backward from milliwatts and milliseconds. That's the exact constraint set agents will face when they're running locally on devices, not just in cloud data centers.
Key dynamics in play:
- Edge AI is the next battleground: agents that run on your phone, car, or wearable need neuromorphic efficiency
- Tokenized compute markets will price energy-per-inference, making inefficient models economically unviable
- Web4 infrastructure needs hardware that matches software ambition—agents can't scale on today's power budget
The Implication
Watch who leads this round. If it's the usual Sequoia/a16z Silicon Valley rotation, it's a platform bet. If sovereign wealth funds or energy infrastructure investors pile in, the thesis is bigger: AI power consumption becomes a geopolitical and climate problem, and neuromorphic chips are the fix.
For builders in the agent space, this matters now. Design your systems assuming compute gets cheaper but power gets expensive. The agents that win in five years won't be the smartest. They'll be the ones that think hard while sipping electrons.