Rat brain AI stopped being a laboratory curiosity on 23 September 2026, when selected Amazon Web Services customers were given access to a commercial text-to-video model whose optimisations were derived from living neurons. The company behind it, The Biological Computing Co., says the model is up to five times faster than the open-source system it is built on and costs 80% less to run.
The detail that makes this different from every previous biological-computing story is what the customer actually receives. There is no dish of neurons in the loop, no wetware appliance, no new workflow. What ships is a software layer — one the company says adds less than 0.1% to the underlying model — distilled from measurements of how real neurons responded to visual stimulation. The biology sits upstream, in a San Francisco lab, and never touches the inference path.
That design choice is also what makes the claims hard to check. This article sets out what TBC has built and how, what the numbers mean when you do the arithmetic, why the company chose video generation rather than language, how the AWS distribution works, where it sits against the only comparable commercial effort, and the specific question Amazon’s own executive says remains unanswered.
Table of contents
- What the Rat Brain AI Laboratory Actually Does
- The Rat Brain AI Numbers, Measured
- Why the Rat Brain AI Team Chose Video
- How Rat Brain AI Reaches Customers
- The Company Behind the Rat Brain AI Model
- Rat Brain AI Against the Only Real Comparison
- The Question Amazon Itself Has Not Answered
- What Rat Brain AI Still Has to Prove
- The Ethics of Rat Brain AI Nobody Has Settled
- Frequently Asked Questions About Rat Brain AI
- References and Further Reading
What the Rat Brain AI Laboratory Actually Does
The rat brain AI mechanism is less exotic than the headline and more interesting than the press release.
Living cells on an electrode grid
The rat brain AI research programme works with rat brain cells and with neurons grown from human stem cells. The living cells are placed on multi-electrode silicon arrays manufactured by 3Brain, a Swiss biotech company. Each array can both stimulate and record, which is what makes the culture usable as an experimental subject rather than a passive sample.
Stimulate, record, analyse
Researchers send electrical stimulation patterns into the neurons and record how they respond. TBC then analyses that neural activity for computational patterns — regularities in how biological tissue handles information — which can be translated into software and applied to existing models.
The founder’s description
Alexander Ksendzovsky, TBC’s cofounder and chief executive, puts the loop plainly: “We figured out a way to code information, like images for example, to the biological material. We then observe how the biology processes that information, and then we build a tool that mimics that process.” The product is the mimicry, not the tissue.
What rat brain AI is not
This is the point most coverage blurs. No neuron computes anything for a paying customer. TBC’s own materials say the software layer “runs entirely on conventional AI infrastructure” and that “no biological hardware or new customer workflow is required.” Calling the result a rat brain AI model is accurate about its provenance and misleading about its operation.
The Rat Brain AI Numbers, Measured
Three headline figures are doing all the work in the rat brain AI announcement, and they are worth separating by how checkable each one is.
Five times faster, 80% cheaper
TBC claims its optimised text-to-video model delivers 5x faster generation and 80% lower inference costs than the base model, while also improving quality. Run the arithmetic and the two claims describe the same shape from different angles: at five times the speed, a generation takes 20% of the baseline wall-clock time, and at 80% lower cost it consumes 20% of the baseline spend.
A layer that adds almost nothing
The software layer is said to add less than 0.1% to the underlying model. On a model of any realistic size that is a rounding error in parameter count, which is the claim’s rhetorical strength — a near-free addition producing a fivefold speed-up — and also the reason an independent reviewer would want to see the method.
The baseline nobody can see
Here is the gap in the rat brain AI case. TBC will not say which open-source video model it builds on, describing it only as one of the “frontier” video generation systems. A rat brain AI speed-up is meaningless without a named baseline, because a 5x gain over an unoptimised reference implementation and a 5x gain over a well-tuned deployment are entirely different results.
The claims in one view
| Claim | Figure | Source | Checkable now? |
|---|---|---|---|
| Faster video generation | Up to 5x | TBC, 22 September 2026 | No — baseline undisclosed |
| Lower inference cost | 80% reduction | TBC, 22 September 2026 | No — baseline undisclosed |
| Added model size | Under 0.1% | TBC, 22 September 2026 | Partly, once deployed |
| Improved output quality | No figure given | TBC, 22 September 2026 | No metric stated |
| Runs on standard infrastructure | No wetware needed | TBC, 22 September 2026 | Yes, by deployment |
| Total funding raised | More than $50m | WIRED, 22 September 2026 | Company-confirmed |
Why the Rat Brain AI Team Chose Video
The decision to point rat brain AI at video generation rather than language was made for two reasons, one physical and one commercial.
The electrode array is a grid
Ksendzovsky says the physical layout of the multi-electrode arrays shaped the choice. Where each electrode sits affects how it interacts with the neurons, and the company realised images were a natural starting point because visual information could be mapped onto the grid more readily than text or language. The substrate is two-dimensional, so the data that fits it best is two-dimensional too.
A prominent investor’s advice
Jon Pomeraniec, TBC’s cofounder and chief operating officer, credits Jeff Dean — the veteran AI researcher and a TBC investor — with suggesting the company fine-tune video generation models before applying its technique more broadly. “Jeff said if we could do that, there’s the promise of doing more complex, interesting things with our AI models down the line,” Pomeraniec says.
Benchmarks were the real argument
The reasoning behind that advice is the strongest thing in the story. Video generation has accepted industry benchmarks, so a startup with an unfalsifiable-sounding premise can demonstrate early whether its technique produces a measurable gain on problems that are already solved. Picking a domain with existing scoreboards is a discipline, not a shortcut — which makes the undisclosed baseline all the more conspicuous.
Where computer vision comes in
Video is also the branch of the field where the analogy to biology is least strained. Mammalian visual cortex is the best-studied sensory system in neuroscience, and computer vision has borrowed from it repeatedly since the first convolutional architectures. A rat brain AI technique derived from visual stimulation is working in the one domain where the biological reference is genuinely informative. We looked at the opposite end of that idea when the hardware itself becomes the network in our piece on physical AI and neuromorphic substrates.
How Rat Brain AI Reaches Customers
The rat brain AI distribution arrangement is the substance of the Amazon announcement, and it is staged rather than complete.
Three AWS surfaces
TBC says it plans to run its optimised model on AWS Trainium, make it deployable through Amazon SageMaker AI, and pursue commercial distribution in AWS Marketplace. Note the verbs: “plans to”, “make deployable”, “pursue”. What exists on day one is a limited preview for selected customers, with early access by sign-up.
What came before
Until now, TBC’s technology was available only through a neocloud provider called Bluesky Compute. Moving rat brain AI from a specialist compute vendor to Amazon’s marketplace is the difference between a few dozen potential customers and the entire AWS enterprise base, which is the whole point of the deal for TBC.
What Amazon gets from it
Jason Bennett, AWS’s vice-president and global head of startups and venture capital, framed the appeal in terms of efficiency: “Nature solved the computing efficiency problem billions of years ago. TBC’s insight is that we can learn from the original computer — the human brain — to make AI faster, more efficient, and more economical.” For a cloud provider selling inference capacity, a supplier that reduces cost per output is a customer-acquisition story, not a margin problem.
Why AWS picked this one
Deap Ubhi, AWS’s global director of technology for startups, says TBC’s appeal was its restraint. The company takes a “pragmatic approach” — rather than attempting to reinvent the transformer, “they’re working within existing standards of the generative AI space and saying, ‘How can we make the current visual models more efficient?'” That is a lower-variance bet than most of this field offers.
The Company Behind the Rat Brain AI Model
The corporate history behind this rat brain AI model is short, and the funding is larger than previously reported.
Founded by two neurosurgeons
TBC was founded in Baltimore, Maryland four years ago by Ksendzovsky and Pomeraniec, both neuroscientists and neurosurgeons. That background matters to the premise: the founders’ domain expertise is the biology, and the AI engineering is what the company had to acquire rather than the reverse.
Fifty million dollars, in two halves
Earlier in 2026 TBC raised $25 million led by Primary Venture Partners, its first significant round. Shortly after that round closed in March, the company secured an additional $25 million, bringing its total to more than $50 million. The second round had not been reported before WIRED disclosed it on 22 September 2026.
Thirty-five people and two labs
The company is now headquartered in San Francisco, where it opened offices and a small research and development lab last year, and employs around 35 people. Some of them work directly with rat brain cells and human stem cells in the lab; WIRED updated its report the same evening to clarify that not all of them do.
The stated long-term plan
TBC describes the video model as the first commercial product created through its discovery process, with each biological experiment adding to a growing library of neural-response data and candidate algorithms. The company says it intends to apply the same process to other models and architectures, then to other AI workloads, and ultimately to “real-time biological computing, where biology works alongside silicon.”
| When | What happened |
|---|---|
| About four years ago | TBC founded in Baltimore by two neurosurgeon-neuroscientists |
| 2025 | San Francisco offices and a small R&D lab open |
| March 2026 | $25m round led by Primary Venture Partners closes |
| Shortly after March 2026 | A further $25m raised, taking the total past $50m |
| Before September 2026 | Technology available only via Bluesky Compute |
| 22 September 2026 | AWS collaboration announced from San Francisco |
| 23 September 2026 | Limited preview opens to selected AWS customers |
Rat Brain AI Against the Only Real Comparison
TBC’s rat brain AI is not the first biologically derived platform Amazon has listed, and the contrast with the other one is instructive.
Cortical Labs sells the biology itself
Ubhi notes that AWS also works with Cortical Labs, an Australian company that combines lab-grown neurons with silicon chips and calls its offering WaaS, or “wetware as a service.” Cortical Labs additionally sells a physical biological computer — a low-power laboratory device that can reportedly keep neurons alive for around six months.
The CL1 and its economics
Cortical Labs announced the CL1 in March 2025 at $35,000, containing roughly 800,000 human neurons on a grid of 59 input electrodes inside a life-support enclosure, and shipped it that summer. Remote access through its Cortical Cloud service is priced at $300 a week. At that rate a year of rental comes to $15,600, so it takes roughly two years and three months of continuous access to equal the cost of buying the device outright.
The architectural difference
The distinction is not a matter of degree. Cortical Labs puts living tissue in the compute path and sells access to it. TBC extracts patterns from living tissue, ships those patterns as software, and keeps the tissue in the lab. One is a new kind of computer; the other is a new source of ideas for an existing one.
Which is the easier business
TBC’s model is the easier one to sell and the harder one to prove. It requires no change to a customer’s stack, which removes every adoption obstacle — and it also removes any way for the customer to tell whether the biology contributed anything, or whether the same optimisation could have been found by conventional search.
| The Biological Computing Co. | Cortical Labs | |
|---|---|---|
| Where the biology sits | In the lab, upstream of the product | In the compute path |
| What the customer buys | A software layer on a video model | Access to a living neural network |
| Hardware required | Standard AI infrastructure | The CL1 device, or cloud access to one |
| Cell source | Rat neurons and human stem cells | Human brain cells |
| Indicative price | Not published; preview by sign-up | $35,000 outright, or $300 a week in the cloud |
| Primary use case | Text-to-video inference efficiency | Research on cultured brain cells |
The Question Amazon Itself Has Not Answered
The most useful scepticism about rat brain AI comes from the distributor, not from a critic.
Ubhi’s scaling doubt
Ubhi is enthusiastic and explicitly unsure it will hold up. “I think with any model like this, the one thing you have to ask is, if you push it to the extremes, will you still see improvements?” he asks. “If you have customers who want to generate longer-form videos, ten minutes, an hour, what’s the potential loss in fidelity over the course of time?”
Why long-form is the stress test
A video model has to track what has already happened and maintain consistency as the clip grows. Optimisations that trade some internal state for speed tend to degrade exactly there, and they degrade invisibly at short durations. A technique benchmarked on short generations can look flawless and still fall apart at ten minutes.
What the preview is really for
That is a reasonable description of what a limited preview exists to establish. The next phase of the rat brain AI story is not a bigger claim from TBC; it is what selected AWS customers report when they push a neuron-derived optimisation past the length it was tuned on.
What Rat Brain AI Still Has to Prove
Three things would move rat brain AI from an interesting claim to an established result.
Name the baseline
Publishing which open-source video model the comparison uses, at which configuration, would convert an unfalsifiable marketing number into a reproducible one. Nothing about competitive sensitivity prevents naming a public model.
Submit to independent measurement
The company chose video specifically because the field has accepted benchmarks. Running on them publicly is the natural next step, and the argument the founders themselves made for the domain choice.
Show the biology was necessary
The hardest question is counterfactual. If the neural-response data produced an optimisation that a conventional automated search would also have found, the rat brain AI framing is provenance rather than mechanism. TBC’s claim that biology provides “a fundamentally different engine for discovering better optimization strategies” is testable, and nobody outside the company has tested it.
The Ethics of Rat Brain AI Nobody Has Settled
The commercial framing of rat brain AI has moved faster than the governance around it.
Human cells are already in the loop
TBC works with neurons grown from human stem cells alongside rat brain cells, and Cortical Labs’ product contains around 800,000 human neurons. These are cultures, not organs, and nobody serious claims they are sentient. But the field has crossed from thought experiment to procurement without any settled framework for what a cultured human neural network is.
The consciousness question is not the immediate one
The live issue is more mundane and more tractable: consent and provenance for donated cells, disposal, what happens to a culture that outlives a company, and whether a customer buying inference capacity has any interest in knowing biological material was involved upstream. TBC’s design largely sidesteps the last of these by keeping the tissue out of the product.
Regulation has not arrived
No jurisdiction currently regulates commercial biological computing as a category. It falls between medical research rules, which govern the cells, and cloud-services rules, which govern the delivery. For now the practical effect is that a company’s own disclosures are the whole of the public record. New model and platform releases of this kind are tracked on our AI models, tools and releases hub.
Frequently Asked Questions About Rat Brain AI
Does a rat brain run the AI model?
No. The neurons stay in TBC’s laboratory. What ships to customers is a software layer derived from measurements of neural activity, running on conventional hardware.
What does the model actually do?
It is a text-to-video generation model, built on an undisclosed open-source system, which TBC says its neuron-derived layer makes up to five times faster and 80% cheaper to run.
Can anyone use it?
Not yet. From 23 September 2026 it is in limited preview for selected AWS customers, with early access available by request. TBC and Amazon both say wider availability is expected.
Are human brain cells involved?
Yes, in the research. TBC works with neurons grown from human stem cells as well as rat brain cells on its electrode arrays.
Has the performance claim been independently verified?
No. TBC has not named the baseline model, which makes the 5x and 80% figures impossible to reproduce from outside the company.
Who else is doing this commercially?
Cortical Labs is the closest comparison, though its approach differs fundamentally: it puts living neurons in the compute path and sells access to them, rather than extracting patterns and shipping software.
References and Further Reading
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
