Why Artificial Intelligence Is Expensive
Today, artificial intelligence technologies are in a phase of explosive growth. The pace of its adoption is accelerating around the world and covering more and more industries, including healthcare, education, finance, media and marketing. The competitive advantage of artificial intelligence in automating processes and personalizing services is forcing companies to quickly integrate these technologies into their business processes.
By forecasts, in 2024 the artificial intelligence market will reach $184 billion, which is 35% more than in 2023. With an annual growth rate of 28.46%, the market size is expected to exceed $826 billion by 2030.
However, according to more optimistic estimates annual growth will be 35.7% from 2024 to 2030, and the market size will increase to $1.339 trillion by 2030.
But as technology grows, so do costs. And there are two key reasons for this: artificial intelligence models are becoming increasingly large-scale and expensive to develop, and global demand for artificial intelligence services requires an increasing number of data centers and infrastructure to support them.
By data Epoch AI – an institute that studies trends in the development of artificial intelligence, 47-67% of all costs are accounted for by artificial intelligence chips, server components and peripheral equipment. R&D personnel costs account for 29-49%, and the remaining 2-6% costs are electricity consumption.
Let's take a closer look at what the costs of artificial intelligence consist of and try to estimate the magnitude of these costs.
Large Language Models (LLM)
Today's best-known artificial intelligence applications, including ChatGPT, are based on large language models (LLMs) – systems that process huge amounts of data, including books, articles and online comments, videos, to produce the best answers to user queries. Leading software companies are betting that the path to more sophisticated artificial intelligence—perhaps even AI systems that can outperform humans—is to make large language models even bigger.
And although in April 2023, OpenAI CEO Sam Altman said: “I think we’ve reached the end of the era of giant models. We’ll make them better in other ways,” so far only in terms of increasing computing power, the amount of data used, and the exponential growth in training costs.
By data Stanford University's cutting-edge AI models are becoming increasingly expensive, and the cost of training today's AI models has reached unprecedented levels. For example, in GPT-4 from OpenAI, computing resources worth about $78 million were used for training (according to other sources, more 100 million $), while Google's Gemini Ultra cost $191 million.
According to research Epoch AI The computing power of the computers used to train the latest AI models increases by 4-5 times every year.
If cost trends continue, the cost of training advanced AI models will exceed $1 billion by 2027.
In an interview with CNBC in early April, Dario Amodei, CEO of AI startup Anthropic, said: “I think next year we'll see model training costs around $1 billion. And then in 2025, 2026, we're going to increase it to $5 billion or $10 billion. And I think there's a chance that this amount could exceed $100 billion.”
Chips
In 2023, the market for artificial intelligence chips has reached volume of $53.6 billion. This figure is projected to increase to $71.3 billion by the end of 2024 as excitement around artificial intelligence applications continues to grow, and longer-term forecasts suggest the market will top $90 billion in 2025.
The rapid growth of artificial intelligence is accompanied by a shortage of chips, which directly affects their cost. Market leader, Nvidia, whose share have to from 60 to 70 percent of the world's supply of chips for artificial intelligence, recently became the largest company in the world with capitalization more than $3.3 trillion, ahead of Microsoft and Apple
By data Raymond James analysts say Nvidia is currently selling H100 GPUs for $25,000 to $30,000 apiece, and Blackwell's next-generation AI GPU will cost 30 – 40 thousand $. At the same time, the number of GPUs required to train artificial intelligence is already in the millions.
Industry leaders Microsoft and OpenAI plan to create a unique supercomputer “Stargate” for artificial intelligence research, using millions of specialized server chips. It will be the centerpiece of a five-phase plan to create a series of supercomputers that companies plan to build over the next six years. Stargate is scheduled to launch in 2028.
In January of this year, Meta CEO Mark Zuckerberg in his message on the social network Instagram* (*banned in the Russian Federation) statedthat the company's future AI roadmap calls for it to build a “massive computing infrastructure” that will include 350,000 Nvidia H100 GPUs by the end of 2024. If Meta were to pay at the lower end of the price range, the cost would be about $9 billion. The total number of different models of GPUs from Meta by the end of 2024 will be almost 600 thousand.
Elon Musk reported investors that his artificial intelligence startup xAI plans to build a supercomputer to run the next version of its artificial intelligence chatbot Grok. The Grok 2 model required approximately 20,000 Nvidia H100 GPUs to train, while the Grok 3 and later models will require 100,000 Nvidia H100 chips.
In addition to the absolute industry leader Nvidia, which introduced the GH200 GPU in March 2024, and traditional chip manufacturers Intel and AMD, all major technology companies involved in artificial intelligence are participating in the AI chip race.
In a sign that the race leaders need alternatives to expensive Nvidia GPUs, statement Meta, OpenAI and Microsoft last December about using a cheaper AI chip from AMD – the Instinct MI300X.
Data centers
The development of artificial intelligence increases the need for computing power, data storage and cloud services, which provides a powerful impetus for the modernization and construction of new data centers. Today all over the world work there are about 11,000 data centers, the top three are: USA – 5381, Germany – 521, UK – 514, and Russia ranks 9th with 297 data centers.
In the USA, the absolute leader in the number of data centers, in 2023 there were built 46% more than in 2022.
In August 2023, Nvidia CEO Jensen Huang predictedthat in 4 years $1 trillion will be spent on modernizing data centers for AI. “The cost of data centers is about $1 trillion, which is a quarter of a trillion dollars in capital expenditure per year,” he said. Most of these costs will likely fall on the leading cloud providers – hyperscalers Amazon, Microsoft and Google, joined by Meta, and other large technology companies.
Dell'Oro Group Researchers calculatedthat in 2024 companies will spend $294 billion on building and equipping data centers, compared to $193 billion in 2020. These costs include the transition of companies to GPU accelerated servers – accelerated servers equipped with graphics processors and custom accelerators. In the 1st quarter of 2024 to a share GPU accelerated servers accounted for more than half of all server sales. According to forecasts, in 2024 the GPU accelerated servers market will increase by more than 80%.
In addition to the construction of new data centers in existing ones there will be increase rack density of hard drives, processors, cooling systems, with a compound annual growth rate of 7.8%. By 2027, the average rack power is expected to reach 50 kW, exceeding the current average of 36 kW.
Electricity
Average for processing a request in ChatGPT required almost 10 times more electricity than searching on Google. And this is one of the main reasons for the increase in electricity consumption around the world in the near future. All those neural networks trained on Internet data have an insatiable appetite for electricity, as do the cooling systems needed to keep them from overheating.
Existing data centers are capable consume a combined 508 terawatt-hours of electricity per year if they were to operate continuously. This is more than the total annual electricity production of Italy or Australia. By 2034, global energy consumption by data centers is expected to exceed 1,580 terawatt-hours, roughly equivalent to the consumption of all of India.
By estimates Goldman Sachs Research, by 2030:
– demand for electricity for data centers will increase by 160%.
– electricity consumption in data centers will increase from 1%-2% of the global total to 3%-4%.
– in the USA, which accounts for 49% of data centers, the growth will be even greater – from 3% to 8%.
– about 20% of the total energy consumption of data centers will come from artificial intelligence.
This increase in energy consumption requires non-standard approaches to solve the problem of electricity generation.
In May 2023, Microsoft announced power purchase agreement with Helion Energy, which plans to begin generating nuclear power through nuclear fusion by 2028 (several working prototypes have already been built). A little later Microsoft published opening for a nuclear technology program manager whose responsibilities include developing a strategy “to power the data centers that house Microsoft Cloud and artificial intelligence.” Microsoft is expected to focus on using microreactors and small modular reactors, which are much cheaper to build and operate than large nuclear reactors.
Personnel
The competition for AI talent has never been fiercer, with every company in the field competing for a very small pool of researchers and engineers. Paying AI professionals today is a battleground. Recognizing that in-demand artificial intelligence professionals may receive multiple job offers, many companies are offering six-figure salaries, plus bonuses and stock grants, to attract more experienced workers, recruiters say.
The heads of companies are faced with the task of not only attracting a highly qualified employee to their side, but also keeping him from moving to a competitor.
For example, there is a well-known case where Google co-founder Sergey Brin personally called to an employee who was considering leaving the company to join OpenAI. Brin's phone call, along with other promises and additional compensation, convinced the employee to stay at Google.
By forecastsBetween 2022 and 2032, the number of computer and information science professionals will grow by 23%, much faster than the average for all occupations.
By data According to the Glassdoor portal, the average annual salary of an ML engineer (machine learning engineer) is $131 – $210 thousand. The most common salary range for engineering positions is listed on the OpenAI website ranges from $200,000 to $370,000, with more qualified job openings ranging from $300,000 to $450,000. In addition to salary, there is also bonuses, which are called “Profit Participation Units” (PPU). For example, with a salary of 300 thousand, they can amount to $500 thousand per year.
The introduction of artificial intelligence has led to an unprecedented increase in the need for specialists in a variety of industries. Companies like Netflix and Walmart are actively seeking employees to develop innovative solutions, improve customer experiences, optimize supply chains, and make decisions based on artificial intelligence. Netflix is hiring for a Product Manager for Machine Learning Platform in 2023. offered applicants received a salary + bonuses of up to $900 thousand per year, which caused a stir on social networks.
But the size of the salary does not decide everything. How difficult is it to hire a highly qualified specialist with generative artificial intelligence skills? told Aravind Srinivas, CEO of AI startup Perplexity: “I tried to hire a senior researcher from Meta, and you know what he said? “Come to me when you have 10 thousand H100 GPUs.”
Legal expenses
When talking about the cost of artificial intelligence, one cannot fail to mention legal costs. It's no secret that all available information is used to train AI – news, works of fiction, messages on message boards, Wikipedia articles, computer programs, photographs, podcasts and videos. And at the same time, copyrights are very often violated.
Today, a number of lawsuits against AI startups and their investors for copyright infringement are pending in the courts. The plaintiffs include both leading news portals like the New York Times and ordinary writers – book authors, artists, photographers – who believe that their works were used for profit without any compensation.
And although the lawsuits do not specify the specific amount of monetary claims, they are talking about “billions of dollars in established and actual damages” associated with “illegal copying and use of unique texts.”
And to the claims already filed are added newThis week, the world's largest record labels Sony Music, Universal Music Group and Warner Records sued two AI startups Suno and Audio for alleged copyright infringement. They are seeking $150,000 in damages for each piece.
In order not to lose in AI courts, startups are forced to expand their legal departments, hiring leading lawyers from BigTech companies, and attract large law firms, whose services are very expensive. At the same time, the question of whether they will win in court remains open.
