While the biggest artificial intelligence companies are spending vast sums to build ever larger models and data centers, Clément Delangue, chief executive of Hugging Face, says the real cost of building useful AI systems is often far lower than widely assumed, arguing that only a handful of frontier labs need billion-dollar training budgets while most real-world applications can be built cheaply by adapting open-source models and fine-tuning them for specific tasks
For much of the past two years, the artificial intelligence industry has promoted a striking idea: that building useful AI systems requires billions of dollars in computing infrastructure and enormous data centers.
That narrative has been reinforced by the spending plans of technology giants racing to dominate the field. Companies such as Microsoft, Alphabet, Amazon and Meta are investing heavily in new AI data centers and advanced chips, often framing the technology as an arms race requiring massive capital.
But according to Clem Delangue, the co-founder and chief executive of Hugging Face, that narrative only reflects the most extreme end of the market.
In a post published on LinkedIn this week, Delangue said the real cost of building many AI systems is dramatically lower than widely assumed.
“They tell you training and running AI model costs billions,” Delangue wrote. “That’s true for a few frontier labs. But for most real-world use cases? Dramatically lower than you think thanks to open-source.”
His comments offer a rare glimpse into the economics of AI development at a moment when the technology is becoming central to the global economy.
Developing the most advanced AI systems remains extremely expensive. Training large language models such as GPT-4 requires enormous computing resources, including tens of thousands of specialized graphics processing units operating inside large-scale data centers. Some analysts estimate that training next-generation models like GPT-4.5 could cost roughly $300 million in computing power alone.
Those costs have fueled an unprecedented infrastructure buildout across the technology industry, as companies compete to develop ever larger and more capable AI models.
Yet Delangue argues that most companies building AI applications do not need that level of investment. According to estimates shared in his LinkedIn post, the cost of training specialized AI models can be far lower when developers use open-source systems and adapt them for specific tasks.
“Real examples from Hugging Face’s latest analysis,” Delangue wrote, show that fine-tuning a text classification model can cost less than $2,000. Training a leading image embedding model can cost under $7,000, while building an optical character recognition system similar to DeepSeek’s can be done for less than $100,000. Even a high-quality machine translation system can be trained for under $500,000, he said.
The difference stems largely from the use of existing models that developers refine for narrow applications rather than training massive systems from scratch.
Hugging Face has become a central hub for open-source AI development, hosting thousands of models that researchers and companies can adapt for specific tasks. In his post, Delangue compared the industry’s obsession with giant models to using a racing car for everyday errands.
“The truth is that you don’t need a Formula 1 car to pick up groceries,” he wrote. “Most tasks are solved just as well by smaller, efficient, targeted models.”
He added that many organizations approach artificial intelligence the wrong way.
“The mistake everyone makes?” Delangue wrote. “Starting with ‘what’s the best AI model?’ instead of ‘what do I need to do?’”
The debate highlights a widening divide in the AI ecosystem. On one side are a handful of frontier labs building massive general-purpose models that require enormous financial resources and specialized infrastructure. On the other are developers building smaller, task-specific systems that can be trained relatively cheaply.
Delangue believes the latter trend will play a growing role in the industry’s future.
“The future of AI is not just bigger models,” he wrote in the LinkedIn post. “It’s cheaper, more customized, open models solving specific problems.”
That shift could make artificial intelligence far more accessible to startups, universities and organizations around the world that lack the resources of Silicon Valley’s largest companies.
Even as the race to build ever larger models continues, Delangue’s remarks suggest that the next wave of AI innovation may come not only from the biggest systems but also from smaller, more specialized ones designed to solve specific problems.
Get the latest news and insights that are shaping the world. Subscribe to Impact Newswire to stay informed and be part of the global conversation.
Got a story to share? Pitch it to us at info@impactnews-wire.com and reach the right audience worldwide
Discover more from Impact AI News
Subscribe to get the latest posts sent to your email.
