Artificial intelligence is not a computer. It is software, made of a program plus a large set of stored numbers, and it runs on a computer the way any other program does. Computers, in the sense Alan Turing formalized in 1936, are machines that carry out stored instructions, and every AI system in use today is a set of instructions and numbers executed by such a machine, from a phone processor to a rack of data center chips.
Key takeaways
- A computer is a machine that follows stored instructions; AI is one kind of program that such machines run.
- A trained model is mostly a file of numbers called parameters, and its size in bytes can be estimated by multiplying the parameter count by the bytes per parameter.
- Meta reports that its Llama 3 models, with eight billion and 70 billion parameters, were trained on more than 15 trillion tokens using 7.7 million GPU hours on H100 chips.
- Chips built for AI, such as Google's Tensor Processing Unit, are add-on processors that receive instructions from an ordinary host computer.
- The common mix-up comes from treating the model, the program around it, and the hardware as one object.
What makes something a computer
The Stanford Encyclopedia of Philosophy entry on the Church-Turing thesis summarizes Turing's 1936 paper as introducing a general-purpose machine that can imitate any single-purpose machine by storing a description of that machine on its tape as a list of instructions. That is the idea of a stored program, and the entry calls the model a bare-bones logical picture of nearly every modern digital computer. The related Church-Turing thesis holds that any procedure carried out by purely mechanical steps can be carried out by such a machine.
By that standard, an AI model is a procedure, not a machine. Given an input such as a photo or a sentence, it applies a long sequence of arithmetic steps and returns an output. A computer performs those steps. Nothing in the definition requires the program to be simple, and nothing about a large program changes what the machine underneath is.
What is inside a trained model
Two things are involved. The first is a program that describes the structure of the computation, which layers to apply and in what order. The second is the set of parameters, the numbers fitted during training. When a developer says a model has eight billion parameters, that is eight billion stored numbers. Training adjusts them using large amounts of example data. Using the model afterward, often called inference, reads them without changing them.
The size of a model on disk or in memory therefore follows from arithmetic. The table below is an illustrative calculation: it multiplies the parameter counts that Meta published for two Llama 3 models by common storage sizes per number. It counts only the stored parameters. Running a model also needs working memory for the text being processed, so real requirements are higher.
| Parameters | Bytes per number | Storage for parameters | Fits in one 80 GB chip? |
|---|---|---|---|
| 8 billion | 2 (16-bit) | 16 GB | Yes, with room left |
| 8 billion | 0.5 (4-bit) | 4 GB | Yes, and small enough for many laptops |
| 70 billion | 2 (16-bit) | 140 GB | No, needs at least two chips |
| 70 billion | 0.5 (4-bit) | 35 GB | Yes, with room left |
The table shows why AI can feel like a different kind of object from ordinary software. The parameters are data, they are very large, and they are produced by training rather than typed by a developer. They are still bytes held in memory.
The hardware behind training
The model card for Llama 3 on Meta's GitHub page gives a concrete scale. Both models were pretrained on more than 15 trillion tokens of text. Training used H100-80GB graphics processing units, with 1.3 million GPU hours for the eight-billion-parameter model and 6.4 million for the 70-billion-parameter model, for a cumulative 7.7 million. A GPU hour is one chip working for one hour, so these figures describe many chips running in parallel for a long period. The card does not state tensor precision, so the table above is an estimate and not a statement about how Meta stores the files.
Training is the expensive step. Inference on a finished model needs far less computation per answer, which is why a model of modest size can run on a laptop while the training run that produced it required a large cluster.
AI chips are still parts of computers
Special-purpose chips blur the picture, so a documented example helps. Norman Jouppi and colleagues at Google described the Tensor Processing Unit (TPU) in a paper presented at the International Symposium on Computer Architecture in 2017. The chip is a custom processor for running neural networks in production, centered on a matrix unit of 65,536 multiply-and-add units that work on 8-bit numbers, with a peak of 92 trillion operations per second and 28 MiB of memory on the chip. The authors report that it ran about 15 to 30 times faster than a contemporary CPU or GPU on their production workloads, with 30 to 80 times better performance per watt.
The same paper shows how such a chip relates to a computer. The TPU was built as a coprocessor on the PCIe bus, plugging into existing servers the way a GPU does, and the host server sends it instructions to execute instead of the chip fetching them itself. The authors compare it to a floating-point unit, an arithmetic helper that older CPUs also attached as a separate part. An AI accelerator is therefore a fast calculator inside a computer, not a replacement for one.
Three layers that are often blurred
- The model: the parameters and the structure of the computation, which only produce numbers.
- The program around it: the code that turns a typed question into numbers, calls the model, and turns the result back into text.
- The machine: the processors and memory that execute both, whether in a phone, a laptop, or a server farm.
Mixing these up produces both common claims, that AI is a computer and that AI is something beyond computers. Both skip the fact that the model alone does nothing without a program and a machine to run it.
FAQ
Is a chatbot a computer program?
Yes. A chatbot combines a trained model with ordinary software that handles input, output, and safety checks. Where that software runs, on the user's device or on a provider's servers, depends on the product.
Does AI need special hardware?
Training large models in practice uses thousands of specialized chips, such as the H100 GPUs named in Meta's Llama 3 card. Running a small or compressed model can work on an ordinary laptop, as the storage arithmetic above suggests.
Is AI a robot?
No. A robot is a machine with sensors and moving parts, and some robots use AI software to decide what to do. The software can also run with no body at all.
Can a computer be intelligent without AI?
The term computer describes the machine, and AI describes a type of software it can run. Calculators and spreadsheets are computers running programs that nobody calls AI, because their steps follow fixed rules.
Related reading
More in AI Basics.
Last verified: October 2026
Sources
- The Church-Turing Thesis, Stanford Encyclopedia of Philosophy, revised 2023
- Turing AM, On Computable Numbers, with an Application to the Entscheidungsproblem, Proceedings of the London Mathematical Society, 1936
- Llama 3 Model Card, Meta
- Jouppi NP et al., In-Datacenter Performance Analysis of a Tensor Processing Unit, ISCA 2017, arXiv:1704.04760