Editor’s note: This work is part of AI tracker, Atlantic‘an ongoing investigation into the production-AI industry.
As they scramble to get their systems online, AI companies are making things expensive for the rest of us. Large language frameworks like ChatGPT and Claude are so resource hungry that tech companies can buy them. 70 percent of the world’s supply of advanced computer memory, leading to shortages. As a result, prices for computer memory and storage are skyrocketing: The hard drives I bought for reporting two years ago for $350 each were $800 when I checked two weeks ago, and are now out of stock. The prices of some computers have gone up to 50 percentageand low-cost computers are the most affected. Low-end PCs may “disappear by 2028” according to one prediction. And memory shortage is expected continue for years.
Memory is stored in data centers, which technology companies are expanding at an incredible rate. They plan to multiply the total US data center capacity by a the eighth factor in the next few years. The demand for electricity at these sites is already so great that some companies are restore jet engines to strengthen them.
The problem isn’t just that AI is being deployed widely or quickly. Other computer technologies have seen similarly dramatic growth without spurring massive increases in electricity or shortages of computer components: Video and music are now streamed around the world, accounting for many terabytes of Internet traffic each day; The growth of the smartphone required the production of billions of devices that now transfer large amounts of data; billions of home devices are also now part of the Internet of Things; and entire industries have moved their operations to cloud software, which is not used in space but in, yes, data centers.
The problem with generative AI, in the industry’s own jargon, is that it doesn’t scale. The cost of growing from, say, a thousand users to a million is an important factor that venture capitalists examine when evaluating a startup. They want to see that additional cost each new user decreases over time, so that the company can support millions of users and earn increasing profits. This is achieved in part through careful engineering of computer systems that can better handle more users wanting to post photos, Uber greetings, or stream music.
For generative AI, the task of building efficient, scalable systems is not done. And the problem is aggravated by bigger models to generate AI, which has increased from 175 billion parameters in 2020 to more than 1 trillion today, according to independent. estimates (the exact sizes of the models that enable products like Claude and ChatGPT are secret). The big in example of a large language it should not be a selling point. But the industry’s observation that larger designs tend to perform better than smaller ones has led to greater confidence in “Scaling Laws” which suggests any problem can be solved by making the models larger. “Maybe with 10 gigawatts of computing, AI can figure out how to cure cancer,” OpenAI CEO Sam Altman he wrote on his blog in September.
Still the income is decreasing. The bigger the AI model, the better it gets with each parameter added, and so it must be made bigger at a faster rate just to keep up steady progress. I asked a few AI researchers if they could name any other real-world software that ranks so badly. None of them could think of anything. Even outside the world of software, it is difficult to find a comparable model, due to that economies of scale it’s the principle that has made light bulbs, cars, and clothes so cheap. By economic and engineering measures, generative AI may be the worst technology ever used.
But with the huge investment behind the current bloated approach, there may not be much will to change. Ilya Sutskever, co-founder and former chief scientist of OpenAI, he said in a November interview that companies take a leveraged approach “because it gives you a risk-free way to invest your resources.” It is more difficult, he said, to invest in research that can recreate a product that is currently available. trillion appreciation. Those who suspect we are in AI driven bubble economy and said that the profit of these companies still open questionlargely because of the high cost and inefficiency of the technology.
Efficiency is a fundamental principle of computer science. One of the first things that graduates learn is that writing a program that sorts a list of 50 words is easy. But if you give that program 50 million words, it will run out of memory or take several hours to finish. Much of computer science is studying clever coding techniques that prevent this from happening. Many of these techniques take advantage of repeating patterns in the data so that as the program receives more input, it takes less time or memory to process each additional bit. That efficiency is one reason that modern smartphones and computers are so powerful and so cheap. This is called logarithmic add, and it looks like this when you draw it:
Large language structures do not scale logarithmically. As they are given more words to process, they slow down and use more memory—time and resources increase quickly as the input grows. In technical terms, the level of LLMs four times. Any student of computer science knows that this is very bad.
Epoch AI, an organization that attempts to quantify the operating costs of AI models, published a graph last year, which is reproduced here with permission. It shows the skyrocketing costs of generating more “tokens” — words that users type into chatbots — and several public AI models.
AI doesn’t have to be built this way. TraditionallyThe goal of AI was to solve problems in ways that mimicked human mental processes. The researchers observed their own thoughts and tried to implement their mental behavior into principles. This method has often been abandoned, in part due to the difficulty of deciphering and explaining the principles of human thought, but it had the advantage of using very few resources and data.
Today’s approach to AI does not attempt to explain the principles of human thought; instead, it gives the computer millions of models to simulate. The large number of models is one reason that older models can do better than younger ones when it comes to producing language, visuals, and music—they have more material to draw from. Some researchers they want to bring back an old, more efficient approach and combine it with a modern approach, but so far these projects haven’t attracted as much attention or funding as the designs that are driving the buzz.
Chatbot companies know that their products are useless. Some have discovered method for improvg performancebut they haven’t made much profit. Time and time again, companies claim to have made a breakthrough – Anthropic CEO Dario Amodei has he called them “compute multipliers”—but usually defined internally unclear termsand there is no evidence that the basic problems of quadratic scaling and exploding model scaling have been eliminated. (Anthropic declined to comment on the record when I reached out to inquire about this.)
Other researchers are working on very small models that require less data and less computing power. I spoke to Alexia Jolicoeur-Martineau, an AI researcher at Microsoft who independently designed one of these, and asked her about the industry’s brutal approach. “It’s a little crazy,” he told me. “At some point you have to learn to be more efficient.”
Last year, Jolicoeur-Martineau won $50,000 award for his paper on a “small recursive model” that does not use a large amount of computing resources. “The idea that one has to rely on large, multi-million dollar trained models by some large organization to achieve success at difficult tasks is a trap,” he said. he wrote. Its model isn’t a replacement for the LLM—it’s designed to solve logical problems in fields like biology and electrical engineering, rather than providing language—but it can perform some of the tasks that larger AI models are currently used for.
We still seem to be stuck with the LLM, perhaps because they are marketed so strongly. Now they are added to everything, whether you want them or not. In 2024 and 2025, they were combined in both Windows and macOSwhich means that more computing power is needed to run a basic personal computer. Smartphones are also sold by improved equipment as companies expect new AI elements. Ineffective AI is also being added to common applications such as Adobe Photoshop and Microsoft Wordmeaning that computers need to be more powerful to run this software.
This is all too bad because computers are no longer improving at the speed they used to. Since the 1950s, manufacturers have learned to make microchips faster, smaller, and cheaper, a trend commonly known as Moore’s Law. But in the past few years, devices have become so small that manufacturers have run into molecular-level limitations on shrinking them further, which has slowed development to a great extent.
Instead of declining features, manufacturers have focused on developing new AI-powered devices. This has produced constant improvements in performance, but none have come close to keeping up with the exponential curve of increasing AI demand.
Finally, inefficiency can be of little concern to people within the technology industry who believe me that they imitate the mind itself. There is an almost religious belief among many in Silicon Valley that something like intelligence can emerge from LLMs, which are ultimately just software to make a language of statistics—this, despite the software powerlessness remember the basics, its lack of the common mind, and its complete difference from the biological brain. Even Yann LeCun, one of the “godfathers” of AI, he told it The New York Times recently that “LLMs are not a path to high intelligence or even human-level intelligence.” But the mythological appeal of AI is so strong that many engineers believe that nothing should stand in their way. Even the basic task of writing an efficient program.
