Today in New York, I met with a friend who works in neurology, currently developing brain-machine interfaces.

She surprisingly said that in the academic community, Musk's Neuralink is considered outdated technology.
Biologists look down on its use of press releases to present scientific claims; Neuralink is seen as "engineering-led, science-cutting corners."
After a three-hour drive from DC, I thought it was just a catch-up, but we ended up talking all afternoon about brain-machine interfaces, AI, and storage anxieties.
They can design their own chips after raising funds, and order 22-nanometer versions from European foundries.
22 nanometers is the golden node for low-power analog chips, capable of operating at 0.5V, making it especially suitable for implantable neural recording devices.
Currently, they are researching how weight-loss injections affect the brain.
No one knows how different weight-loss drugs influence the brain's appetite suppression, or how the hypothalamus and reward circuits (dopamine pathways) specifically work.
So they drill holes in mouse brains, connect chips, and need to measure neural signals in real time.
Because the data volume is enormous and the sampling frequency is very high, they need 10,000 TB, or 10 PB of storage.
This is the first time I’ve heard of storage units in PB!
Neural signals need to be sampled at 30k Hz (each electrode measuring 30k times per second) to capture neural pulses that last only 1 millisecond.
Each data point is just 2 bytes, but multiply that by hundreds of electrodes, and continuous 24-hour recording, one probe per hour results in 80GB.
She said a 1PB enterprise-grade disk costs more than her monthly salary, and long-term maintenance of PB-level storage can cost millions of dollars over five years.
The lab doesn’t have that much funding, so they buy only one PB at a time and use it carefully.
To give a sense of scale, Harvard and Google can achieve full resolution of a 1 cubic millimeter human brain, which is about 1.4 PB.
An entire mouse brain probably requires about 1 EB (1,000 PB).
It’s not because the data per session is large, but because the sampling frequency is so high.
And storage anxiety isn’t unique to biomedicine either.
In astronomy, the SKA radio telescope generates 700 PB of data annually.
CERN’s collider is already handling data at the EB level.
Storage determines the ceiling of cutting-edge science.
When data gets large, communication bandwidth becomes a bottleneck.
So we talked about optical communication.
The earliest chip connection material was aluminum; in 1997, IBM replaced it with copper (which has 40% lower resistance).
Now they’re switching from copper to optical.
This echoes what Huang Renxun discussed at GTC: in March 2025, NVIDIA released silicon photonics + co-packaged optics (CPO) switches to expand AI data centers to the million-GPU scale.
Why must copper give way to light?
Because the faster the speed, the shorter the transmission distance for copper (at high frequencies, skin effect and loss increase sharply).
At 1.6 Tb/s, a single copper wire can’t span the height of a server rack.
So signals must be converted into lasers.
Besides storage, the cost of lab animals is also very high.
A lab monkey costs between $35k and $50k; considering years of special care, surgeries, veterinary services, and institutional fees, a full cycle can easily exceed $100k.
A mouse costs $80, and the price reflects its genetic purity.
I asked: why use mice and monkeys, not rabbits?
She said mice have a complete set of genetic tools (24k existing strains, gene knockouts, optogenetics), and monkeys are used because their brain structure is most similar to humans.
Rabbits have moderate intelligence and are not used because they don’t fit the bill.
Many AI companies are now recruiting researchers like her who do brain research, for two reasons.
The first is efficiency.
The human brain is the most energy-efficient computer on Earth.
It handles vision, language, movement, reasoning—all in parallel—using only about 20 watts, roughly the power of a dim light bulb.
High-end AI chips, by contrast, consume 300–700 watts, and training large models can require several megawatts or gigawatts.
The difference is that computers operate on binary (0/1), with transistors switching at GHz frequencies.
The brain is analog and sparsely firing; neurons only light up when needed, consuming energy only then.
AI companies want to learn from this efficiency.
So, the most sought-after positions in AI companies now are neuroscientists.
The second, more subtle reason:
We actually don’t understand how the brain works at all, and AI faces the same dilemma.
Anthropic found that when AI answers your questions, it’s actually internally reasoning through its own set of ideas, and the reasons it gives may not be truthful.
For example, researchers secretly feed Claude an incorrect answer as a “prompt,” and it will follow that answer to produce a plausible reasoning, only admitting 25% to 39% of the time that it used the prompt.
Anthropic’s CEO said that when AI summarizes a document, “we don’t really know, at a specific and precise level, why it makes those choices.”
AI’s “thought process” is a black box.
Their goal is to build an “MRI” for AI, and by 2027, they hope to understand what’s happening inside.
Finally, we talked about surgery.
Since the brain itself has no pain sensation, people can be awake during craniotomies.
Mice can have their skulls opened, chips implanted, and then be sewn up, continuing to live normally for months.
She joked that she could help me have my brain opened, but hoped I’d never need it.
After a whole afternoon of talking, my biggest takeaway is:
Currently, our understanding of the human brain and AI is still very limited.
But precisely because of that, I increasingly feel that we are living in the best era.
Fully autonomous driving is already on the roads, quantum computing is advancing rapidly, humans are seriously preparing for the first landing on Mars, and brain-machine interfaces are beginning to truly understand the brain…
Today in New York, I kept seeing AI and crypto ads: a taxi top that says “It’s happening with Ripple,” a BNY Mellon Bitcoin ETF on a bus ad, and an OpenAI Codex on the station platform.
Things that were once unimaginable are happening all at once in the U.S., and we are fortunate to be living in this most extraordinary time.