Omer Taha Cetin | Anadolu | Getty Images
Chinese AI company DeepSeek had a quick rise to fame in January upon the release of its latest model, DeepSeek-R1, temporarily dethroning OpenAI’s ChatGPT as the most downloaded free app on Apple’s App Store. Behind the scenes, AI enablers — companies building the infrastructure and applications that enable AI — felt the heat of a shaken market.
But despite Nvidia stock falling more than 15% in a single trading day, a result of the market’s realization that AI model development can be done on thinner margins and with lower-quality resources than anticipated, these AI enablers are still creating AI-specific semiconductor chips, building out compute power and developing platforms that foundation models can use to operate.
“On one hand, the DeepSeek approach showed that you can optimize your model building process so that you require much lower compute power. That has a negative impact on Nvidia,” said Mohamed Elgendy, co-founder and CEO of enterprise AI platform Kolena. “However, the obvious thing is now you’ll find a lot of people building foundation models. Foundation models are not going to be just for the top five companies or so that have hundreds of millions of dollars to build the infrastructure.”
Nvidia shares are down close to 9% so far in 2025, though that is after a remarkable run that saw its share price rise close to 500% over the past two years. The chip giant reports earnings after the close on Wednesday, with DeepSeek and the potential hit to future spending by AI “hyperscalers” — companies like Google, Oracle, Amazon and Microsoft — who buy up to half of Nvidia’s AI chips, expected to receive attention from anxious Wall Street analysts and investors.
Elgendy’s reasoning lies in the belief that DeepSeek’s model release marks a shift in the industry, which he believes will now move towards greater democratization, the results of which have already started to be produced, with additional LLMs hitting the market for as little as $50, according to researchers. “The days before DeepSeek are different from days after DeepSeek,” he said.
“While many researchers focus on the critical pursuit of greater computational and data efficiency in AI models, the need for robust infrastructure will remain paramount,” wrote Jad Tarifi, CEO of foundation world model company Integral AI, who formerly led Google’s first generative AI development team, in his 2024 book “The Rise of Superintelligence.” Tarifi went on, “Even as models streamline, anticipated real-world deployments will ensure a growing demand for powerful computational resources.”
Amr Awadallah, CEO of enterprise AI agent company Vectara, has a slightly different view. “I see this as the start of significant margin compression for AI model builders and the large AI enablers that supply them,” he said. “Revenue across the industry will continue to grow, and grow a lot, but the amount of profit that these large companies can extract will go down significantly, so it will create some pressure from that perspective.”
Recent reports that Microsoft was scaling back its AI data center buildout spooked investors for this reason, but Microsoft has disputed the reports, saying it is committed to its stated $80 billion spend, but adding that it might “strategically pace or adjust our infrastructure in some areas.”
Awadallah likens this shift to the history of flash drives, which take a lot of design to make right but are now a common commodity with lower profits than before. DeepSeek was able to train its model on lower-end hardware, without access to the high-end ones major U.S. companies use, effectively commoditizing the market.
DeepSeek itself runs on Intel‘s processors Xeon and Gaudi, which are “helping customers get strong performance at lower costs,” according to an Intel spokesperson. Intel’s Gaudi is also being used for Denvr Dataworks, which offers AI solutions that prioritize both performance and data privacy.
“New AI models bring exciting opportunities, but they also raise important considerations. They can drive innovation, improve efficiency and unlock new possibilities, but scaling AI comes with challenges like cost, energy use and responsible deployment,” an Intel spokesperson told CNBC. For its part, Intel says its products and services take these obstacles into account.
An imperfect model with potential
Despite expectations of improvement over time, DeepSeek’s accuracy issue is no secret. According to testing from Vectara, the DeepSeek-R1 model hallucinates at a rate of 14.3%, compared to a rate of about 2% for OpenAI’s GPT-4 (and it’s even higher than DeepSeek’s own non-reasoning predecessor, Deepseek-V3).
“When we were testing DeepSeek against other models, we noticed that DeepSeek fails in most adversarial attacks, or the jailbreak type of attacks,” said Kolena’s Elgendy. “These were the early failures of GPT-3 that came out years ago. All of that has been solved by all the big providers.”
Still, Elgendy views DeepSeek — or at least what it represents — as a sort of diamond in the rough. “Now we understand that there’s a new approach that is much more efficient to train for large models. And this approach is effective,” he said.
Awadallah thinks DeepSeek’s claim of spending just $6 million to train the model is inaccurate. “Most of us have a consensus that it was way more than that,” he said. “Maybe the final run that produced the model cost $6 million, but you usually have to do many, many runs until you get the model to work well. I would expect that it would cost at least $50 million or more to train this model.” Still, that’s a lot less than Google Gemini’s $149 million even before taking staff salaries into account, which effectively doubles the price.
Ultimately, Elgendy said it’s only a matter of time until more of this type of foundational model built on a relatively bootstrapped budget with comparatively little compute power starts to pop up. “We were operating with the assumption that foundation models require a lot of resources to build. With DeepSeek, we started seeing something that we thought was really down the road. I think this will 10x the number of builders and probably 100x the number of users,” he said.
Specifically, Elgendy anticipates more models that operate in specific domains such as healthcare, research, pharmaceutical, accounting, finance and more. “The infrastructure here will go back to our old AI machine learning era, where you’ll find the specialized AI companies building specialized foundation models, and they all need infrastructure,” he said. “I believe testing and validation will be the biggest part of it, because a lot of the other components have been commoditized so far.”
Regarding the testing element, Elgendy said, “The more of these providers there are out there, the more competition there will be. DeepSeek, the moment it’s out, everybody started testing it, and then it was clear where it hasn’t been tested and where it has been tested.” Competition, he said, will keep things in check, because “the market is now the police.”
Nvidia CEO Jensen Huang said in a pretaped interview last week “The market responded to R1 as in, ‘oh my gosh, AI is finished,’ that AI doesn’t need to do any more computing anymore. It’s exactly the opposite.”
Given the immense potential of DeepSeek and models like it, Awadallah said the competition will get more intense between AI enablers, such as Nvidia, that build the infrastructure that make these models possible. “The ones that can still stand and thrive will only be the ones that are able to sustain themselves despite the margin hit,” he said.
