A San Francisco startup says it has found a way to make AI video generation faster and cheaper by borrowing a page from biology rather than semiconductor design. On September 22, 2026, The Biological Computing Co. (TBC), working with Amazon Web Services (AWS), announced what the companies are calling a “neuron-derived” AI video model: a tiny software adapter built from measurements of real, lab-grown nerve cells that plugs into existing diffusion-based video generators. According to The Decoder and corroborating reporting from XenoSpectrum, the layer adds less than 0.1% to the size of the underlying video model, yet TBC claims it delivers video up to 5x faster and cuts inference costs by roughly 80%, numbers that place the September 2026 announcement squarely at the center of a fast-moving biological computing push.
It’s a strange-sounding claim on first read: biological neurons, grown in a lab, somehow informing the math inside a text-to-video AI model. But the underlying mechanics, as described in the announcement and reported by The Decoder, are more mundane than the headline suggests, and more interesting for what they reveal about where AI infrastructure companies are willing to look for efficiency gains as GPU costs keep climbing.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What TBC and AWS actually announced
The Biological Computing Co., previously known as Biological Black Box, was founded by two neurosurgeons, Alex Ksendzovsky and Jon Pomeraniec, according to The Decoder’s reporting. The company is based in San Francisco and has spent the past two years working on a proprietary technique that translates signal patterns from lab-grown cortical nerve cells into software components it calls “adapters.” Those adapters, TBC says, can be dropped into existing AI models to change how they behave, without retraining the whole model from scratch.
The AWS partnership is the commercial vehicle for the idea. Per the announcement, the resulting model will run on AWS Trainium chips, be served through Amazon SageMaker AI, and be listed in the AWS Marketplace so enterprise customers can license it directly. That distribution path matters: AWS is not just hosting a research demo, it’s putting a biologically-derived AI component into its standard enterprise machine learning stack, alongside its own custom silicon.
TBC raised $25 million in February 2026, The Decoder reported, money that funded the lab work behind the adapter technology. The company has not disclosed which open-source text-to-video model it used as its base for the AWS announcement, a gap that matters more than it might initially seem, since every performance claim TBC has made is relative to that unnamed baseline rather than to any named, checkable rival.
How lab-grown neurons turn into a software layer
The process TBC describes works like this, according to The Decoder’s account: the company grows cortical nerve cells on chips wired with 4,096 electrodes. Researchers stimulate the cells electrically and record how activity propagates across the culture. Real biological neurons process and route information differently than artificial neural network layers do, and TBC’s bet is that some of those patterns are useful priors for how an AI model should organize information over time, which is exactly the kind of thing a video model needs to get right to keep a scene coherent from one frame to the next.
Those biological recordings are then converted into what TBC calls adapters: compact software modules that slot into an existing diffusion model’s architecture. Crucially, the company says the adapters do not replace GPUs, does not require new hardware, and doesn’t touch the base model’s core weights in any drastic way. It’s installed as an add-on layer that runs on the same infrastructure the base model already used, which is one reason AWS was comfortable wrapping it into Trainium and SageMaker AI rather than building new hardware around it.
One data point gives a sense of scale: The Decoder reported a related, separately-described version of the adapter approach used just 156,000 parameters bolted onto a 600-million-parameter video model, extending how long the video stayed visually coherent to roughly twice the baseline while adding less than 0.5% model overhead. That’s a useful illustration of the core pitch: instead of retraining a multi-billion-parameter model, TBC argues you can graft on a tiny, biologically-informed module and get outsized gains in efficiency and quality.
The performance claims, and the catch
TBC’s headline numbers, repeated in both The Decoder’s and XenoSpectrum’s September 2026 coverage, are a 5x speed-up in inference and an 80% reduction in inference cost compared with its unnamed open-source base model, plus a claim of improved video quality, though neither outlet reported a standardized quality metric to back that part of the claim. Those two top-line figures, speed and cost, are not actually independent measurements. As The Decoder’s own analysis pointed out, five times faster and 80% cheaper describe the same underlying effect stated two different ways, since inference cost on cloud infrastructure scales directly with compute time. That doesn’t make the claim false, but it does mean TBC is presenting one improvement as if it were two.
There is more granular data from a specific benchmark: TBC tested its adapter against Oasis, a Minecraft-style AI world model, and reported a 19% improvement in differential entropy compared with the unmodified base model, a 15% improvement over conventional fine-tuning at the same parameter budget, and a 5% improvement over LoRA-based adaptation. On the same test, frame rate reportedly rose from roughly 2 frames per second to about 10 frames per second. Separately, an external evaluation by infrastructure provider Bluesky Compute in July 2026 found approximately 4.4 times lower inference costs and more than three times as much coherent video generated before quality or semantic consistency broke down, according to The Decoder.
Those Oasis and Bluesky Compute numbers are more specific and more useful than the round 5x and 80% headline figures, because they come from a named benchmark and a named external tester rather than TBC’s own marketing framing. But even here, the comparisons are against TBC’s own base models and fine-tuning setups, not against any commercially available AI video product a reader would recognize.
What’s missing from the announcement
The most important caveat in this story is what TBC and AWS have not said. The Decoder’s reporting is explicit that there are no independent public benchmarks sufficient to fully verify TBC’s headline performance claims, and that the company has not disclosed which open-source base model it used for its AWS comparison. Without a named base model, there’s no way for a third party to reproduce the 5x and 80% figures, and there’s no direct, controlled comparison anywhere in the public record between TBC’s system and any of the AI video models consumers or enterprises actually use today, such as OpenAI’s Sora, Google’s Veo, Runway, or Kling.
That’s not a small gap. Readers who compared frontier text-to-video systems in our own Veo 3.1 vs Sora 2 vs Kling breakdown know how much these tools already differ in cost, quality, and speed even without any biological computing involved. TBC’s numbers, as reported, tell us how its adapter compares to its own unnamed baseline model, not how a TBC-enhanced video generator stacks up against the tools currently shaping the market.
Why AWS is interested at all
AWS’s motivation is easier to read than TBC’s science. Cloud providers are under sustained pressure to cut the cost of running generative AI workloads, and video generation is one of the most compute-intensive categories in the field, often orders of magnitude more expensive per output than text or image generation. AWS has already built its own custom silicon strategy around Trainium chips to reduce dependence on Nvidia GPUs for training and inference, a strategy visible in deals like Uber’s adoption of AWS Trainium3 and Amazon’s $5 billion Trainium investment tied to Anthropic. A software layer that promises to cut inference cost on top of that hardware, without requiring new chips, fits neatly into that broader push, regardless of how well the biological framing holds up under scrutiny.
Putting TBC’s model into SageMaker AI and the AWS Marketplace also lets Amazon test the idea with real enterprise customers at low risk. If TBC’s adapter genuinely delivers meaningful cost savings on video workloads, AWS gets a differentiator against Google Cloud’s Veo-linked Vertex AI offerings and Microsoft Azure’s OpenAI-powered video tools. If the claims don’t hold up under production traffic, AWS has lost little beyond a marketplace listing, since the adapter runs on existing infrastructure rather than dedicated new hardware.
The broader biological computing landscape
TBC is not the only company exploring biological or brain-inspired computing as a path to more efficient AI. The field, sometimes called synthetic biological intelligence, includes a handful of startups working on lab-grown neural tissue, organoid intelligence, and neuromorphic approaches that borrow structural ideas from biological brains without literally growing cells. The best-known example is Cortical Labs, which launched its CL1 biological computer, combining lab-grown neurons with silicon, back in March 2025; Live Science reported in May 2025 that CL1 units sold for around $35,000 apiece, after the company raised $10 million in funding in April 2025, according to Wikipedia, and began shipping its first 115 units later that year, per Boing Boing. What sets TBC’s announcement apart is the direct commercial packaging: rather than publishing academic research about brain organoids computing simple tasks, the company is shipping a product listing on a major cloud marketplace, aimed squarely at engineers who need to cut video generation costs today.
That commercial framing is itself notable in a market where AI infrastructure costs have become a recurring headline. Our coverage of world model comparisons like Genie 3, Marble, and NVIDIA Cosmos shows how much of the current AI video race is really an efficiency race: every lab is trying to generate longer, more coherent video for less compute, whether through better architectures, distillation, or, in TBC’s case, an entirely different source of design inspiration.
Historical context: efficiency layers are not new
The idea of bolting a small, specialized adapter onto a large pretrained model isn’t new. Techniques like LoRA (Low-Rank Adaptation) have been standard practice in the AI industry for years, letting engineers fine-tune huge models by training a small number of additional parameters rather than the entire network. TBC’s own benchmark against LoRA on the Oasis test, where it claimed a 5% edge, positions its adapter as a direct competitor to that established technique rather than something categorically different in kind, even though the inspiration behind it is unusual.
What’s changed is the search for where new efficiency gains can come from. Years of scaling laws pushed the industry toward bigger models and more GPUs, a trend visible in everything from Nvidia’s data center revenue growth to the scale of deals like the shifts in OpenAI’s Sora strategy. As GPU capacity gets more expensive and harder to secure, cheaper roads to the same output, even unconventional ones sourced from cell cultures, are getting serious commercial attention rather than being dismissed as lab curiosities.
Market impact: what this means for AI video pricing
If TBC’s cost claims hold up even partially in production, the immediate effect would be pressure on AI video pricing across the board. The AI video generation market has already seen sharp price cuts this year, illustrated by moves like Gemini Omni 1.1 Flash’s cut to $0.03 per second of generated video. An 80% inference cost reduction, if it were verified and generalized beyond TBC’s own benchmarks, would be large enough to reshape how competitively cloud providers can price video generation products, and would put pressure on AWS’s rivals to find comparable efficiency wins of their own.
The more measured read, though, is that this remains a single vendor’s claim about a single, undisclosed base model, tested mostly by the vendor itself with one external check from Bluesky Compute. Enterprise buyers evaluating the AWS Marketplace listing will need their own benchmarks against their own actual workloads, on their own actual video models, before anything like an 80% cost cut becomes something they can put in a budget line item.
Data table: TBC’s reported performance claims
| Metric | Reported result | Comparison baseline | Source |
|---|---|---|---|
| Inference speed | Up to 5x faster | Unnamed open-source base model | TBC / The Decoder |
| Inference cost | ~80% lower | Unnamed open-source base model | TBC / The Decoder |
| Adapter size overhead | Less than 0.1% of base model size | Base video model | The Decoder |
| Oasis world model: differential entropy | +19% | Unmodified base model | TBC / The Decoder |
| Oasis world model: vs. fine-tuning | +15% | Conventional fine-tuning, same size | TBC / The Decoder |
| Oasis world model: vs. LoRA | +5% | LoRA adaptation | TBC / The Decoder |
| Oasis frame rate | ~2 fps to ~10 fps | Base model frame rate | TBC / The Decoder |
| Independent test (Bluesky Compute, July 2026) | ~4.4x lower cost, 3x more coherent video | Base model | Bluesky Compute / The Decoder |
Data table: TBC vs. established AI video infrastructure players
| Company | Approach | Distribution | Independent verification |
|---|---|---|---|
| The Biological Computing Co. | Adapter derived from lab-grown neuron signal patterns | AWS Marketplace, SageMaker AI, Trainium chips | Partial (one external test, Bluesky Compute) |
| OpenAI (Sora) | Large-scale diffusion transformer, trained end-to-end | Direct consumer and API access | Publicly used and reviewed at scale |
| Google DeepMind (Veo) | Diffusion-based video model integrated into Vertex AI and Gemini | Google Cloud, Gemini app | Publicly used and reviewed at scale |
| Runway | Proprietary diffusion video models plus creative tooling | Direct web and API access | Publicly used and reviewed at scale |
| Kling | Diffusion-based video generation model | Direct web and API access | Publicly used and reviewed at scale |
Competitive comparison: an unproven layer versus proven products
The fairest way to frame TBC’s position is as an infrastructure layer, not a finished consumer product. Sora, Veo, Runway, and Kling are all directly usable text-to-video tools with millions of generated clips behind them and public track records for quality, coherence, and pricing. TBC’s adapter, by contrast, is a component meant to sit underneath somebody else’s video model, similar in spirit to how LoRA sits underneath a base language or image model rather than competing with ChatGPT or Midjourney directly.
That means the interesting comparison isn’t “TBC vs. Sora,” which the current public record doesn’t support anyway, but rather “TBC’s adapter vs. other efficiency techniques,” such as LoRA, quantization, and distillation, all of which compete to squeeze more output per dollar out of the same underlying video models. On that narrower comparison, TBC’s own Oasis benchmark numbers, a 5% edge over LoRA and a 15% edge over standard fine-tuning, are plausible-sounding but, again, self-reported.
Skepticism from the AI research community
The framing of “neurons” in a marketing announcement invites more scrutiny than a typical software efficiency claim, precisely because it sounds more dramatic than it may turn out to be. The Decoder’s own analysis of the announcement was notably cautious, pointing out the mathematical overlap between the speed and cost claims and explicitly flagging the absence of independent, reproducible benchmarks. That kind of pushback from technology press covering the announcement itself, rather than from a rival company with a competing product to sell, is usually a signal that a story deserves a second look before its headline numbers get repeated uncritically.
None of this means TBC’s approach is fraudulent or without merit. Biological neurons genuinely do route and process signals differently than artificial layers, and there is a real, if small, body of academic research on organoid and cell-culture computing to back that up. Boing Boing reported in February 2026 that Cortical Labs demonstrated its CL1 system, using just 59 electrodes to interface with lab-grown neurons, playing Doom, and that DishBrain neurons learned to play Pong in about 5 minutes, versus roughly 90 minutes for standard deep reinforcement learning on the same task. The gap here is narrower and more mundane: it’s the gap between a company’s own benchmarks on its own undisclosed base model, and the kind of standardized, third-party verification that would let engineers and researchers actually trust the 5x and 80% numbers without caveats.
What enterprise AI teams should do next
For engineering teams evaluating the AWS Marketplace listing, the practical path is straightforward: treat TBC’s numbers as a starting hypothesis, not a guarantee, and run the adapter against your own video generation workload on your own base model before committing budget to it. Because the technology is described as a drop-in layer that runs on existing Trainium and SageMaker AI infrastructure, the switching cost to trial it appears low, which is likely part of why AWS agreed to list it in the first place. Teams already running Trainium-based inference, similar to the setups described in coverage of AWS’s broader Trainium rollout, are the most natural early testers.
It’s also worth tracking whether TBC eventually discloses its base model. That single piece of information would let outside researchers replicate or refute the 5x speed and 80% cost claims directly, turning an unverifiable marketing number into a testable one. Until that happens, the responsible reading of this announcement is “a promising efficiency technique with unusually interesting biological framing and an unverified headline claim,” not “AI video generation just got five times faster for everyone.”
Predictions: where this story goes next
- Expect independent benchmarking firms and AI research labs to attempt to replicate TBC’s claims once the AWS Marketplace listing goes live, especially if TBC discloses or leaks its base model.
- Watch for AWS to expand the SageMaker AI integration to other AWS-hosted generative media workloads beyond video if early enterprise trials show even a fraction of the claimed cost savings.
- Expect competing cloud providers, including Google Cloud and Microsoft Azure, to highlight their own efficiency techniques for video generation in response, whether or not they engage directly with TBC’s biological framing.
- Expect continued skepticism and scrutiny from AI research outlets until TBC publishes a peer-reviewable, named-baseline benchmark rather than relying on self-reported comparisons.
- Expect renewed general interest in biological and organoid computing startups as investors look for TBC’s funding and AWS partnership as a signal that the category has moved from academic curiosity toward commercial products.
The bottom line
This is a genuinely novel piece of AI infrastructure news: a cloud giant is putting a product built from lab-grown neuron measurements into its standard enterprise machine learning marketplace. That alone makes it worth tracking. But the specific numbers driving the headlines, 5x faster and 80% cheaper, describe one effect rather than two, rest on an undisclosed base model, and have only one outside verification data point so far. The technology may well prove out. For now, it sits exactly where The Decoder placed it: an intriguing, commercially packaged idea with plenty of numbers and, so far, comparatively little independent proof.
FAQ
What is The Biological Computing Co.?
The Biological Computing Co. (TBC), formerly known as Biological Black Box, is a San Francisco startup founded by neurosurgeons Alex Ksendzovsky and Jon Pomeraniec. The company develops software adapters derived from measurements of lab-grown cortical nerve cells, aimed at improving the efficiency of existing AI models, according to The Decoder.
How does a “neuron-derived” AI video model actually work?
TBC grows cortical nerve cells on chips with 4,096 electrodes, stimulates them electrically, and records how activity propagates across the cell culture. Those patterns are translated into small software modules called adapters, which are integrated into existing diffusion-based video models without retraining the entire model.
Does this technology replace GPUs?
No. According to reporting on the announcement, the adapter is a proprietary software layer that runs on conventional AI infrastructure, including AWS Trainium chips. It does not replace GPUs or require new specialized hardware.
How much faster and cheaper does TBC claim its model is?
TBC claims up to 5x faster inference and roughly 80% lower inference costs compared with its unnamed open-source base video model. Note that these two figures largely describe the same underlying speed improvement, since cloud inference costs scale with compute time.
Has TBC’s technology been independently verified?
Only partially. Infrastructure provider Bluesky Compute ran an external test in July 2026 and reported roughly 4.4 times lower inference costs and more than three times as much coherent video before quality degraded. However, The Decoder notes there are no independent public benchmarks sufficient to fully verify TBC’s headline claims, and TBC has not disclosed its base model.
How does TBC’s model compare to Sora, Veo, Runway, or Kling?
There is currently no public, controlled, head-to-head benchmark comparing TBC’s adapter-enhanced system to OpenAI’s Sora, Google’s Veo, Runway, or Kling. TBC’s reported gains are measured against its own unnamed open-source base model, not against any of these commercially available AI video tools.
How is AWS involved?
AWS is partnering with TBC to run the neuron-derived video model on AWS Trainium chips, serve it through Amazon SageMaker AI, and list it in the AWS Marketplace, giving enterprise customers a direct path to license and test the technology on their own workloads.
How much funding has TBC raised?
The Decoder reported that TBC raised $25 million in February 2026. Other funding figures reported elsewhere have not been independently confirmed and are not treated as fact here.
