Artificial Intelligence is an Expensive Game

Let us start with chatGPT: it is an instance of the GPT-3 language model which has 175 billion parameters. GPT-3 is one of the largest language models currently available. Its high parameter count enables it to process vast amounts of data and generate highly sophisticated responses to a wide range of inputs. This allows chatGPT to answer questions, summarize information, generate creative writing, and perform many other natural language processing tasks with a high degree of accuracy and fluency.

Training and operating a language model like GPT-3 requires a significant amount of computing resources, including powerful GPUs and a large amount of memory. The exact amount of hardware required to train a GPT-3-sized model can vary depending on a number of factors, such as the configuration of the model and the specific algorithms used.

In the case of GPT-3, it was trained using OpenAI’s custom hardware infrastructure, which includes a large number of GPUs connected by high-speed interconnects. The exact number of GPUs used for training GPT-3 is not publicly disclosed, but it is estimated to be in the thousands of GPUs. The hardware required to train such a large model is significant and requires significant investment in both hardware and energy.

Once trained, running a GPT-3 model requires substantial computing resources as well, including powerful GPUs and large amounts of memory. The exact amount of hardware required to operate a GPT-3-sized model depends on the specific use case and the configuration of the model, but it is generally much less than what is required to train the model from scratch.

The technical requirements for building a hardware infrastructure to train a language model similar to GPT-3:

1. GPUs: Training a large language model like GPT-3 requires a significant number of GPUs to perform the complex computations involved in the training process. These GPUs should be powerful, with high memory capacities and fast interconnects. A typical high-end GPU for deep learning tasks might have 16 GB or 32 GB of memory, with a Tensor Core for acceleration. To train a large language model like GPT-3, you would need several hundred or even thousands of these GPUs, depending on the size of the model and the specific training algorithms used.
2. Memory: In addition to GPUs, much memory is also required for training a model like GPT-3. The exact amount of memory required can vary depending on the size of the model and the specific algorithms used, but it can easily reach hundreds of gigabytes or even terabytes. The amount of memory required for training a language model can vary greatly, but for a model like GPT-3, you would need several terabytes or more of memory.
3. Storage: A large amount of storage is also required to store the data used for training the model, as well as the model itself. This storage should be fast, reliable, and easily accessible to the GPUs. The amount of storage required for training a language model can also vary greatly, but for a model like GPT-3, you would need hundreds of terabytes or more of fast, reliable storage.
4. Networking: The hardware infrastructure should have a high-speed network to facilitate communication between the GPUs and storage. This is especially important during the training process when large amounts of data are being transferred between the GPUs and storage. To ensure fast communication between the GPUs and storage during the training process, you would need a high-speed network with a bandwidth of at least 100 Gbps.
5. Power: Training a large language model like GPT-3 requires a significant amount of energy, so the hardware infrastructure should be equipped with a reliable and scalable power supply. Training a large language model like GPT-3 can consume several megawatts of power, so you would need a reliable and scalable power supply to ensure that the hardware infrastructure remains operational during the training process.
6. Cooling: Training a large language model can generate a lot of heat, so the hardware infrastructure should be equipped with an efficient cooling system to keep the GPUs and other components from overheating. To keep the GPUs and other components from overheating during the training process, you would need an efficient cooling system, such as liquid cooling or immersion cooling.

These are just rough estimates, and the exact hardware requirements for training a language model like GPT-3 can vary greatly depending on the specific use case and the configuration of the model. Nevertheless, building a hardware infrastructure for training a large language model requires a substantial investment in hardware, energy, and other resources, and should be approached with careful consideration. In addition to these technical requirements, it’s also important to consider the software requirements for training a model like GPT-3. This includes a deep learning framework such as TensorFlow or PyTorch, as well as the necessary libraries and tools for data preprocessing, data augmentation, and model evaluation.

Rough estimates for the initial investment and annual operational costs for building a hardware infrastructure to train a language model for different use cases:

1. Research purposes. For a small-scale research project, the initial investment might be on the order of $100,000 to $500,000, including hardware, software, and staffing costs. The annual operational costs, including energy and hardware maintenance, might be on the order of $50,000 to $100,000.
2. Small to medium-sized businesses. For a small to medium-sized business looking to train a custom language model for specific use cases, the initial investment might be on the order of $500,000 to $1 million, including hardware, software, and staffing costs. The annual operational costs, including energy and hardware maintenance, might be on the order of $100,000 to $500,000.
3. Large enterprise. For a large enterprise looking to train a large-scale language model for multiple use cases, the initial investment might be on the order of several million dollars, including hardware, software, and staffing costs. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars.
4. Service providers. For a service provider offering language model training as a service to multiple clients, the initial investment might be on the order of several tens of millions of dollars, including hardware, software, and staffing costs. The annual operational costs, including energy and hardware maintenance, might be on the order of several million to tens of millions of dollars.
5. Weather forecasting. the cost of building a hardware infrastructure to train a language model will depend on the specific requirements of the project. However, here is a rough estimate of the initial investment and annual operational costs: The initial investment might be on the order of $500,000 to $1 million, including hardware, software, and staffing costs. This estimate assumes the need for a medium-sized infrastructure, with several GPUs and a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of $100,000 to $500,000. This estimate assumes the need for several staff members to operate and maintain the infrastructure, and a moderate amount of energy consumption.
6. Forecasting dynamic social processes, The cost of building a hardware infrastructure to train a language model will depend on the specific requirements of the project. The initial investment might be $500,000 to $1 million, including hardware, software, and staffing costs. This estimate assumes the need for a medium-sized infrastructure, with several GPUs and a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of $100,000 to $500,000. This estimate assumes the need for several staff members to operate and maintain the infrastructure, and a moderate amount of energy consumption.
7. Monitoring and predicting cyberattacks for governmental structures. the cost of building a hardware infrastructure to train a language model will depend on the specific requirements of the project. The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with a significant number of GPUs and a large amount of storage and memory, as well as the need for specialized hardware and software. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
8. Fighting against fake news by monitoring all social networks, the cost of building a hardware infrastructure to train a language model will depend on the specific requirements of the project. The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with a significant number of GPUs and a large amount of storage and memory, as well as the need for specialized hardware and software. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
9. Fighting against terrorists by monitoring cyberspace.  The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with a significant number of GPUs and a large amount of storage and memory, as well as the need for specialized hardware and software. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
10. Quantum computing. The initial investment might be on the order of several million to several tens of millions of dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized quantum hardware, as well as a significant number of GPUs and a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
11. AI models in quantum computing. The initial investment might be several hundred thousand to several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized quantum hardware, as well as a significant number of GPUs and a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption
12. AI models in autonomous driving. The initial investment might be several hundred thousand to several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs and a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
13. AI models in national energy management systems and energy infrastructure monitoring. The initial investment might be several hundred thousand to several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs and a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
14. AI models in critical infrastructure for digital twins and predictive maintenance.  The initial investment might be several hundred thousand to several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs and a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
15. AI in traffic management in a large city. The initial investment might be several hundred thousand to several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs, sensors, and cameras, as well as a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
16. AI in safety and security in a city the size of Bucharest. The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs, cameras, sensors, and storage systems, as well as a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
17. AI in safety and security of critical infrastructure of electricity for a country similar to Romania. The initial investment might be on the order of several million dollars to tens of millions of dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs, cameras, sensors, and storage systems, as well as a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
18. AI in the safety and security of critical gas and oil infrastructure for a country similar to Romania. The initial investment might be on the order of tens of millions of dollars to hundreds of millions of dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs, cameras, sensors, and storage systems, as well as a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several million dollars to tens of millions of dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
19. AI in health for a country similar to Romania. The initial investment might be on the order of several million dollars to tens of millions of dollars, including hardware, software, and staffing costs. This estimate assumes the need for a large-scale infrastructure, with specialized hardware such as GPUs, servers, and storage systems, as well as a large amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand dollars to several million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
20. AI in health for a single hospital with 2000 beds.  The initial investment might be on the order of several hundred thousand dollars to a few million dollars, including hardware, software, and staffing costs. This estimate assumes the need for a smaller-scale infrastructure, with hardware such as GPUs and servers, as well as a smaller amount of storage and memory. The annual operational costs, including energy and hardware maintenance, might be on the order of several tens of thousands of dollars to several hundred thousand dollars. This estimate assumes the need for a smaller staff to operate and maintain the infrastructure, as well as a lower amount of energy consumption.
21. AI in agriculture for a country of the size of Romania. The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for large-scale infrastructure, with hardware such as high-performance computers, servers, and sensors. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand dollars to a few million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
22. AI in personalized medicine for the case of Romania. The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for large-scale infrastructure, with hardware such as high-performance computers, servers, and sensors. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand dollars to a few million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
23. AI in food security for a country the size of Romania. The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for large-scale infrastructure, with hardware such as high-performance computers, servers, and sensors. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand dollars to a few million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.
24. A research institute for genomics and stem cells. The initial investment might be several million dollars, including hardware, software, and staffing costs. This estimate assumes the need for large-scale infrastructure, with hardware such as high-performance computers, servers, storage systems, and specialized laboratory equipment. The annual operational costs, including energy and hardware maintenance, might be on the order of several hundred thousand dollars to a few million dollars. This estimate assumes the need for a large staff to operate and maintain the infrastructure, as well as a high amount of energy consumption.

Investing in AI is undoubtedly an expensive game, requiring significant initial investment and ongoing operational costs. As AI is based on data, the cost of implementing AI depends heavily on the cost of collecting and processing data. The cost of building a hardware infrastructure to train a language model varies depending on the specific requirements of the project, but it is always substantial. From small-scale research projects to large enterprises and service providers, the costs of implementing AI can range from hundreds of thousands to tens of millions of dollars. Nevertheless, the benefits of AI are substantial, and the potential for improving businesses, society, and even saving lives is immeasurable. Therefore, while investing in AI may be expensive, the potential benefits of AI make it a worthy investment.

Of course, there are various studies that show the positive effects of implementing AI. For example, a study by PwC found that AI could contribute up to $15.7 trillion to the global economy by 2030. Additionally, a study by McKinsey & Company found that AI could increase global GDP by up to 1.2% per year. Furthermore, AI has already been implemented in various industries, such as healthcare. In 2019, the global healthcare AI market was valued at $2.1 billion, and it is projected to reach $19.9 billion by 2025. AI can assist in early disease detection, drug discovery, and personalized treatment plans. Another industry that has implemented AI is the financial sector. A study by Accenture found that AI could save the banking industry up to $447 billion by 2023. AI can help detect fraud, provide personalized financial advice, and automate tasks, such as loan underwriting. In addition to the economic benefits, AI has the potential to save lives. For instance, AI can aid in medical diagnosis and treatment planning, as well as disaster response and management. A study by the American Cancer Society found that AI can improve cancer screening accuracy, potentially leading to earlier detection and better patient outcomes.

Note from the author of this post

——————

credits: Stelian Brad