Two years after ChatGPT captured the world’s imagination, generative artificial intelligence (AI) seems to have hit a wall. The energy costs of building and using larger AI models are skyrocketing, and breakthroughs are becoming increasingly difficult. But fear not, researchers and entrepreneurs are racing to find solutions, pushing the boundaries of AI and reshaping the landscape of the industry.
Large language models, the backbone of generative AI, are voracious energy consumers. Training OpenAI’s GPT-4 model, for example, could have powered 50 American homes for a century. And as models grow bigger, the energy demands increase exponentially. Estimates suggest that today’s largest models cost a staggering $100 million to train, with the next generation projected to cost a billion dollars, and the one after that, a whopping $10 billion.
But that’s not all. Using these models to answer queries, known as “inference”, also comes at a significant computational cost. Summarizing the financial reports of the world’s 58,000 public companies, for instance, could cost anywhere from $2,400 to $223,000. These inference costs can eventually exceed the cost of training, raising serious concerns about the economic viability of generative AI.
This situation is understandably alarming for investors who have poured billions into the AI sector. Companies like Nvidia, which designs the chips commonly used for AI, have seen their market capitalization skyrocket by $2.5 trillion in the past two years. Venture capitalists and others have invested nearly $95 billion in AI startups since the beginning of 2023 alone. OpenAI, the creator of ChatGPT, is reportedly seeking a valuation of $150 billion, which would make it one of the world’s largest private tech firms.
However, there is no need for panic. History teaches us that even seemingly insurmountable limits can be overcome with human ingenuity. The challenges of sending humans into space spurred innovations that revolutionized Earth-based technologies. The oil crisis of the 1970s fostered energy efficiency and alternative energy sources, including nuclear power. And decades later, fracking enabled the extraction of previously inaccessible oil and gas reserves, turning America into the world’s leading oil producer.
The AI industry is already demonstrating this spirit of adaptation. Companies are developing specialized chips designed specifically to run large language models, achieving greater efficiency than general-purpose processors like those manufactured by Nvidia. Tech giants like Alphabet, Amazon, Apple, Meta, and Microsoft are all designing their own AI chips. In fact, AI-chip startups have received more funding in the first half of this year than in the previous three years combined.
Software developers are also making strides. The trend is shifting away from massive, brute-force models towards smaller, specialized systems tailored to specific tasks. OpenAI’s latest model, o1, is designed to excel at reasoning rather than generating text. Other developers are utilizing less computationally intensive methods, optimizing chip usage and efficiency. By combining different models, each suited to a particular type of problem, researchers have drastically reduced processing times.
These developments will undoubtedly reshape the AI industry. The dominance of large, established players is no longer a given. While Nvidia currently holds an impressive four-fifths share of the world’s AI chip market, specialized competitors are emerging, poised to challenge its stronghold. Google’s AI processors are already the third most widely used in data centers globally.
OpenAI may have pioneered the large language model, but other giants like Anthropic, Google, and Meta are rapidly catching up. While a gap still exists between these top players and second-tier models like France’s Mistral, it’s likely to close as the trend towards smaller and specialized models gains momentum. This could lead to a constellation of diverse models, rather than just a few dominating superstars.
This shift brings uncertainty for investors. Their bets on current leaders seem less certain. Nvidia could lose ground to other chipmakers, and OpenAI’s dominance might be challenged. Big tech companies are aggressively acquiring talent, hoping to build the devices that will connect consumers to their AI assistants. However, competition among them is fierce, and few have a clear strategy for turning a profit from generative AI.
Even if the industry ultimately consolidates under a single winner, it’s impossible to predict who that will be. Governments, too, need to adapt their thinking. Their traditional focus on industrial policy, often centered around subsidies, is no longer sufficient. Progress in AI requires more than just capital and computing power; it requires a thriving ecosystem nurtured by talent and innovation.
European and Middle Eastern countries may find that cultivating ingenuity is as crucial as acquiring computer chips. America, on the other hand, boasts a strong foundation with its world-class universities, established clusters of talent in San Francisco and Silicon Valley, and a thriving chip industry.
However, America’s attempt to restrict China’s access to cutting-edge chips, in an effort to maintain a technological advantage, has backfired. This restriction has inadvertently spurred China’s research system to excel at finding workarounds and overcoming constraints.
In a world where ingenuity trumps brute force, the best way for America to solidify its lead is to attract and retain top researchers from around the globe. This can be achieved through policies like easier visa regulations.
The AI era is still in its nascent stages, and the future remains uncertain. But the breakthroughs needed for AI to truly flourish will come from fostering ideas and talent within our borders, not from attempting to stifle competition abroad. It’s time to embrace innovation, nurture creativity, and allow the AI revolution to unfold its full potential.
