Advances in AI reasoning models like a brand new one from Alibaba are coming fast and furious. They rile up investors, and seem very exciting, but the sobering realities of vast computing changes are hard on CIOs and enterprises that want a reliable technology that doesn’t break the bank.
Efficacy ranks ahead of price, one CIO told Fierce Electronics recently when referring to AI broadly.
In other words, does the new product work the way you want it to? And that question has to be answered yes before the evaluation of price and TCO can be finalized. Big concerns could well be revealed around data privacy and bias, one analyst told Fierce.
Alibaba unveils benchmark of its open source QwQ-32B
Some background: China-based Alibaba announced its new open source QwQ-32B on Thursday, claiming it rivals cutting-edge reasoning models like DeepSeek-R1. The Alibaba approach uses 32 billion parameters compared to DeepSeek’s 671 billion parameters with 37 billion of those engaged during the inference stage.
Alibaba claimed “impressive results” with the ability to continually improve performance in math and coding. Results were posted online March 6.
Efficiency is clearly the name of the game, and investors have noticed. Alibaba said revenue growth driven by AI will continue to accelerate, an indication that shoppers of AI tools will be lining up to buy—potentially at unheard of levels. Bill Gates recently told CNN that AI will be bigger than the internet, which probably sounds obvious to CIOs but perhaps has not sunken in with the public, and even some experts at cloud providers. Nvidia CEO Jensen Huang has made a similar point, noting the value of inference to organizations, which of course supports growth in GPUs that Nvidia makes and sells.
Futurum Group CEO Dan Newman told CNBC that LLMs are now becoming “commoditized” with developers of software working to drive down costs and usability. “As we see this more efficiency, this cost coming down, we’re also to see use going off. The training era, which is what Nvidia really built its initial AI boom off, was a big moment. But the inference, the consumption of AI, is really the future and this is going to exponentially increase that volume.”
Newman’s comment seems most valuable for investors in stocks like Alibaba’s, which soared on Thursday by more than 8% on the Hong Kong market, then dropped less than 1% shortly afterwards.
Questions of reasoning models are different for CIOs than investors
However, the questions for CIOs at enterprises and even at the cloud, are actually a little deeper and more nuanced. When OpenAI first hit, some big companies could not buy enough GPUs fast enough to support AI training for LLMs. It seemed the sky was the limit, with companies paying multiples for GPUs that should have started at $60,000. What’s a CIO to do, or plan, or even noodle out, strategically? It could become a field day for IT consultants!
“It’s still pretty early on, and I’d estimate that 80% of enterprise use of AI is still in the early experimental stages with CIOs trying to figure out the technology and how to best integrate it into business workflows,” said Jack Gold, founder and chief analyst at J. Gold Associates.
“This is still an emerging marketplace and as such, there will be many entrants and claims made,” Gold added. “The issue is how practical and useful each model is for a particular task.” Generally, he said the new, smaller models from DeepSeek and the new Alibaba and OpenAi small models are for fairly limited tasks while large models from OpenAI, Atrophic, META and others can be embedded in agentic processes where a wide range of AI skills are needed.
“There will continue to be a large amount of experimentation to see what and where various models are useful and how to make the most cost-effective models as well,” Gold added.
At a more practical level, one analyst advised companies to evaluate how the new Alibaba model works with hosting providers, especially to avoid data bias that could be inherent with Chinese content.
“DeepSeek R1 or Alibaba Qwen AI chatbot services hosted by entities in China could very well pose security and privacy concerns for users and organizations outside of China due to data use policies of these services,” said NeXt Curve executive analyst Leonard Lee in comments to Fierce Electronics. “These Chinese reasoning models are known to have biases as they are largely trained on Chinese data sets.”
Aside from such concerns over bias or privacy, Lee agreed there is still a move to massive commoditization happening in AI supercomputing and all the inference work from cloud to edge that promises to be disruptive. His views are outlined in a recent newsletter.
“There is the dynamic of token prices plummeting that has a real impact on the economics of the ecosystem,” Lee told Fierce.
Gartner analyst Tong Zhang told Fierce that the new Alibaba Qwen reasoning model with 32 B parameters shows the potential of fine AI engineering. "We can achieve more with less," he said. "It would be more affordable for all companies."
However, he added, "The question remaining is whether the new breed model, namely the reasoning model, can be generalized and applied to business use cases with better performance. If they are only good at math and code, the business case would be very limited.
"However, if those reasoning models can resolve complex problems and empower AI agents with some complex planning and reasoning capabilities with lower cost, they would definitely become the favorite."