“Following IBM’s earnings announcement earlier today, I had a lot of conversation and without fail, DeepSeek came up. It’s clear this moment is resonating across the industry,” Krishna wrote on LinkedIn Wednesday, hours after the company’s Q4 earnings report.

“For too long, the AI race has been a game of scale where bigger models meant better outcomes. But there is no law of physics that dictates AI models must remain big and expensive. The cost of training and inference is just another technology challenge to be solved,” Krishna said.

DeepSeek has dominated headlines since releasing its open-source model R1—a model it claims is far cheaper and more efficient than OpenAI’s O1, yet has outperformed it on several reasoning benchmarks. Krishna said the pattern is not new. “In the early days of computing, storage and processing power were prohibitively expensive,” he wrote. “Yet, through technological advancements and economies of scale, these costs plummeted. AI will follow the same path.” 

IBM stock were up more than 12% on Thursday, a day after IBM beat analysts’ expectations. While IBM’s total revenue increased modest 1% year over year to $17.6 billion, Big Blue said its software revenue increased 10% and pointed to robust demand for its generative AI products – with the company posting $5 billion in generative AI bookings. 

IBM has been “steadfast” in its belief that “open is key to innovation” and that “we’ve been on this journey for years,” Krishna wrote.

IBM has long-advocated for open-source AI and co-founded the AI Alliance with Meta, AMD, and others to advance open AI development. The company does not open source all of its AI models, but in May 2024 it released a series of models under the Apache 2.0 license, making them freely available to the public, designed for tasks such as code generation and bug fixing. In October 2024, IBM introduced the latest open-source version of its Granite family of models aimed at business applications. And its watsonx platform offers access to a variety of both open source and proprietary AI models.

On the company’s earnings call Wednesday, Krishna called IBM’s AI strategy “a comprehensive platform play” with an ecosystem of AI models, software, middleware and consulting services, as well as partnerships with companies including Adobe, AWS, Microsoft, Meta, Mistral, Salesforce and SAP. 

But in answering an analyst question on the call, Krishna also drilled down further into the decision to open source IBM’s Granite models. “There was a thesis out there about a year-and-a-half ago that maybe one, maybe two extremely large models are going to run away with the bulk of the market share,” he said. “We always felt that was both technically and economically infeasible” because of the large size and high cost of the biggest models. 

“If you can drop the model size, you can drop all of that by 90%,” he said, pointing out that smaller, specialized models–such as English-language only, or just for programming–may be much more efficient for enterprise businesses. And open-sourcing the models mean clients can fine-tune them on their own data for their unique needs. “Open sourcing our models under the Apache license gives our clients the freedom that what they add onto our underlying open model, they can keep to themselves,” he said.