Artificial intelligence is progressively emerging as a formidable instrument for major organizations to attain their profit objectives, prompting substantial investments in AI research and development. Microsoft, Alphabet (Google), and Meta have experienced a significant surge in cloud revenue due to the incorporation of AI technologies into their services. Nonetheless, to obtain that substantial revenue, the investment must likewise grow.
For instance, in the most recent quarter, Microsoft reported $14 billion in capital expenditures, largely driven by AI infrastructure investments, a 79% increase compared to the year before. Alphabet spent $12 billion, a 91% increase, and expects to continue at that level as it focuses on AI opportunities. Meta increased its annual capital expenditure estimate to $35-$40 billion, driven by investments in AI research and development. This rising cost of AI has caught some investors by surprise, especially as stock prices fell in response to higher spending.
The primary factors contributing to the substantial expense of investing in AI are outlined as follows:
AI models: AI models are getting bigger and more expensive to research.
Data centers: The worldwide demand for AI services necessitates the construction of numerous additional data centers to accommodate it.
Large language models get larger
The AI products currently attracting significant attention, such as OpenAI’s ChatGPT, are driven by extensive language models. These models depend on vast datasets—comprising books, papers, and online comments—to deliver pertinent responses to consumers. Prominent AI firms are concentrating on the development of increasingly larger models, since they contend this will enhance AI capabilities, potentially surpassing human performance in certain activities.
Constructing these larger models necessitates substantially greater data, computational resources, and training duration. Dario Amodei, CEO of Anthropic, states that existing AI models require approximately $100 million for training, however forthcoming versions may necessitate up to $1 billion. By 2025 or 2026, these expenses may escalate to $5 to $10 billion.
The mind behind AI: where creativity meets technology. 🌈🧠
Chips and computing costs
A significant portion of AI's elevated expenses arises from the specialized processors required for model training. AI firms use graphics processing units (GPUs) instead of the conventional central processing units (CPUs) found in most computers, as GPUs can analyze extensive data rapidly. These GPUs are highly sought after and exceedingly costly. The most sophisticated GPUs, like Nvidia's H100, are regarded as the benchmark for AI model training, with an estimated cost of $30,000 per, and some resellers demanding higher prices. Meta intends to procure 350,000 H100 chips by year-end, signifying a multi-billion-dollar expenditure.
Companies have the option to lease these chips rather than purchase them; however, leasing is also costly. Amazon's cloud division levies over $100 per hour for a cluster of Nvidia H100 GPUs, in contrast to roughly $6 per hour for conventional processors. Last month, Nvidia unveiled the Blackwell GPU, a chip that significantly outperforms the H100 in speed. Training a model equivalent to GPT-4 necessitates 2,000 Blackwell GPUs, in contrast to 8,000 H100 GPUs. Notwithstanding these advancements, the pursuit of larger models may undermine these cost reductions.
Powering AI with precision-engineered chips. 💻🔧
Data centers
To meet the increasing demand for AI, IT businesses require additional data centers to handle GPUs and other specialized hardware. Meta, Amazon, Microsoft, Google, and others are competing to construct new data centers, comprising arrays of CPUs, cooling systems, and electrical infrastructure. Companies are projected to allocate $294 billion for the construction and outfitting of data centers this year, in contrast to $193 billion in 2020. A substantial portion of these charges is associated with the elevated prices of Nvidia GPUs and other AI-related components.
Currently, there are more than 7,000 data centers globally, an increase from 3,600 in 2015. The typical dimensions of these facilities have expanded considerably, indicating the heightened demand for AI computing capabilities. This expansion is propelled by the increase of digital services such as streaming and social media, as well as the necessity to facilitate the AI surge.
Where data meets design: the engine room of AI. 🔗📊
Deals and talent
Besides chips and data centers, AI firms are investing millions in acquiring licenses for data from publishers to train their models. OpenAI has established agreements with multiple European publishers, compensating them with tens of millions of euros to obtain news stories for training purposes. Google has entered into agreements, including a $60 million contract to lease data from Reddit, while Meta has contemplated acquiring book publishers.
The rivalry for AI expertise is consequently escalating expenses. Organizations are providing substantial remuneration to entice proficient workers. Netflix promoted a position for an AI product manager with a compensation of up to $900,000. The intense competition for talent is driving up labor prices throughout the business.
Ideas in action: the blueprint of innovation. ✍️📄