How can an AI engine run without using models?
Most of what is called AI today is a model. It is a very large network of numbers, fixed by training on a vast collection of past data, on a large number of graphics processors in a data centre, and then queried. Senua AI takes a completely different route, with a long scientific history behind it.
Instead of a trained network, Senua AI learns the structure of each stream it reads, from the stream itself: which situations the system can be in, and what tends to follow each one. Physicists call these the causal states of a process. James Crutchfield and Karl Young introduced them in 1989 [1]. Cosma Shalizi and Crutchfield later proved that they are the smallest summary of the past that predicts the future as well as anything can [2]. Causal states are found in the data, not stored in billions of trained weights. So there is no GPU training run, no network to host, and nothing that has to live in a data centre.
That difference matters most where AI meets the physical world. A robot, a wearable or a machine on a factory floor does not need to know what is typical of the internet. It needs to know the state of the one system in front of it, what usually comes next, and the moment something changes. That is exactly what causal states describe. Because they are learned from the device's own stream, they run on the device or right beside it, on ordinary low end CPUs: no round trip to a data centre, no connection needed, and nothing that has to leave the device. In the Senua AI Robotics Core, the Mind reads a robot's sensors as streams, learns their causal states, and turns each moment into a decision with a confidence [3]. That is why we believe this is the better architecture for physical AI.
And a different kind of company
The same architecture makes a different kind of company possible. Look at how most AI companies are built and the same chain repeats. They rent compute to train and run large networks. Renting compute costs more than any early company earns, so they raise large venture rounds to pay for it. Raising large rounds means selling preferences, board seats and a clock. And the clock means the only acceptable ending is a very large one.
Senua AI inverts every link of that chain. Not as a slogan: it is how the company was designed from the start, and each part protects the next. This post sets out the standard path, with the evidence for it, and then the inversion.
The standard path, and what it costs
- Compute comes first. Andreessen Horowitz, one of the largest investors in AI, wrote in 2023 that it had seen AI companies spend more than 80% of their total capital raised on compute [4].
- Compute keeps costing. The same firm found that AI companies' gross margins were often 50 to 60%, against 60 to 80% or more for comparable software businesses. The difference is the cloud bill, which grows with every customer [5].
- So the money follows. In 2025, close to half of all global venture funding went to AI companies, up from 34% the year before [6].
- And the money comes with a clock. A venture fund is a partnership with a fixed life, which William Sahlman of Harvard identified as one of the main ways investors in the fund keep its managers disciplined (Journal of Financial Economics, 1990) [7]. Every company the fund backs inherits that clock.
Each link makes sense on its own. Together they lock a company into one ending, chosen before it has a customer. We set out what that does to founders, with the research and the court cases, in our previous post [8].
The inversion
| the standard AI path | the Senua AI inversion | |
|---|---|---|
| Technology | Large neural networks, trained or fine tuned | No models. Our own engine, written in our own language, Simplex |
| Compute | Rented GPUs and cloud inference, billed by the hour or the token | Ordinary, low end CPUs. No GPU, no data centre, no token bill |
| Funding | Large venture rounds early, to pay for compute | Bootstrapped. No funding accepted at the time of writing |
| Capital | Venture funds | Customer revenue, and Australian programmes that cost no equity [9] |
| The ending | A listing or a very large sale, on the fund's clock, with investors holding preferences and board seats | The ending that fits deep tech: an acquihire or a licence and hire, with the outcome kept by the people who built it [8] |
We do not use models. No LLMs. No SLMs. Of any kind.
No GPU. No data centre. No AI token costs, ever. Senua AI runs on ordinary, low end CPUs.
Each part protects the next
- The code protects the cap table. Our engine needs no GPU and no data centre, so our running costs are small. Small running costs mean we never have to raise money to keep the lights on. Raising becomes a choice, not a deadline.
- The cap table protects control. No raise means no preferences that are paid first, no investor board seats, and no consent rights over the next financing. Those are the terms that, in the cases on the public record, cost founders their companies.
- Control protects the technology. The language, the engine and the products stay in a company we control. If a larger company wants the technology, we can license it, nonexclusively if that suits us, and keep selling it to everyone else.
- Control protects the ending. With no preference stack to clear and no fund clock running, any offer is judged on its own merits. A buyer meets a company that could carry on without them, and that is the position from which every deal is better.
The links run in both directions. Break the first one, by building on rented compute, and every other link fails with it.
An inversion has to be proved
A company built on "no GPU" is only a story until the technology does the work. So we publish the measurements, and anyone can check them:
- Cyber security, live on our own sites. Our enquiry forms answer a visitor in about a quarter of a second, have every request checked by Senua AI Cyber, and cost about a dollar a month to run at a thousand enquiries [10]. Shield's benchmark and its method are public [11].
- Physics, measured against the largest labs' own engines. The Senua AI Robotics Core has held Meta's SuperDex [3] and Google DeepMind's MuJoCo [12] to the laws of physics. Each deep dive rests on studies registered before they ran, and every rule and run is public.
Build on the engine
The inversion is not only for us. We license the Senua AI engine to other companies, including AI companies already building on large language models, whether a corporate service or an open source one.
Many AI products use a language model for work it was never built for: reading logs, watching security events, interpreting sensor and machine telemetry, and noticing when something has changed. That is stream work, and it is where the Senua AI engine replaces the language model.
- It runs on your own hardware. Ordinary CPUs, on your servers or your devices. No GPU to buy or rent.
- It fits into your product like any other component. Your platform talks to it through a command line and an API.
- Your data stays with you. Nothing has to leave your infrastructure, and the engine does not call home.
- Licensed, not metered. The cost does not grow with every request your customers make.
- It shows its working. Every decision comes with a confidence, and every figure we publish names its dataset and its method.
It is not a chatbot, and we do not sell it as one. If your product writes prose, keep whatever writes it. If your product reads streams and has to decide, talk to us about licensing the engine.
What the inversion is not
It is not a rejection of investors, or of the companies building large AI. Several of our deep dives are on their open source engines. When we raise, it will be to speed up something we have already shown works, on terms that keep the inversion intact.
For a small deep tech company, the strongest position is not the biggest round. It is a company that does not need one.
References
- J. P. Crutchfield and K. Young, Inferring statistical complexity, Physical Review Letters 63 (1989) 105 to 108. Peer reviewed.
- C. R. Shalizi and J. P. Crutchfield, Computational mechanics: pattern and prediction, structure and simplicity, Journal of Statistical Physics 104 (2001) 817 to 879. Peer reviewed.
- Senuamedia Blog, The Senua AI Robotics Core, complete deep dive: Meta's SuperDex Physics, 15 October 2026.
- G. Appenzeller, M. Bornstein and M. Casado, Navigating the high cost of AI compute, Andreessen Horowitz, 27 April 2023. The 80% figure is the authors' observation from their own dealings, not a published dataset.
- M. Casado and M. Bornstein, The new business of AI (and how it's different from traditional software), Andreessen Horowitz, February 2020.
- Crunchbase News, The year's big funding trends, in charts, end of 2025.
- W. A. Sahlman, The structure and governance of venture capital organizations, Journal of Financial Economics 27(2) (1990) 473 to 521. Peer reviewed.
- Senuamedia Blog, What is giving up early equity really worth?, 18 October 2026: the research on founders and control, the court cases, and the acquihire and licence and hire deals, each with its references.
- Australian Taxation Office, Rates of R&D tax incentive offset: for companies with turnover under A$20 million, a refundable offset of the company tax rate plus 18.5 percentage points. Changes announced in the 2026 to 2027 Budget start on 1 July 2028.
- Senuamedia Blog, A web form, protected end to end, 17 October 2026.
- Senua AI, Cyber benchmarking.
- Senuamedia Blog, What Senua AI found and fixed in Google DeepMind's MuJoCo, 18 October 2026.