1 week Ago By Tobias Mann
Analysis Microsoft AI CEO Mustafa Suleyman has extolled the virtues of playing second fiddle in the generative-AI race. In a TV news interview last week, Suleyman argued it's more cost-effective to trail frontier model builders, including OpenAI that has taken billions from the Windows giant, by three to six months and build on their successes than to compete with them directly. Our strategy is to play a very tight second, given the capital intensiveness of these models "Our strategy is to play a very tight second, given the capital intensiveness of these models," he told CNBC on Friday.
In addition to being cheaper, Suleyman said the extra time enables Microsoft to optimize for specific customer use-cases. While the strategy might seem unusual for a corporation at the beating heart of the GenAI movement, it reflects the position in which Microsoft - and now Suleyman - finds itself. As you may recall, Suleyman made a name for himself as the co-founder of DeepMind, which was acquired by Google in 2014.
Suleyman joined Microsoft last year after a brief stint as CEO of Inflection AI. While his former employer at the Chocolate Factory is competing directly with the likes of Anthropic and OpenAI to build ever more capable and feature-rich models, Microsoft has yet to launch a frontier model of its own. Instead, Redmond's strategy is closely tied to OpenAI, to which it furnishes a not inconsiderable amount of Azure cloud compute in exchange for the right to use the startup's GPT family of models in its growing suite of Copilot -branded AI services.
This relationship may well explain Suleyman's approach. There's not much sense in investing the massive quantities of capital necessary to build frontier models that may or may not be successful in the market when your buddy Sam Altman over at OpenAI will do it for you. Having said that, Microsoft isn't putting all of its eggs in one basket.
While the GPT series is at the heart of many familiar Windows and Microsoft cloud Copilot services, it's not the only model collection out there. The Excel giant notably develops a line of permissively licensed small language models under the Phi codename. Compared to something like GPT-4.
5, these open models are minuscule, usually weighing in the single to double-digit billion-parameter range, making them appropriate for use on edge devices, including laptops, as opposed to multi-million-dollar GPU clusters. The models have also generally lagged behind OpenAI's top-tier offerings in terms of features, such as multi-modality or mixture of experts (MoE) architectures. In this vulture's personal experience, Microsoft's Phi family of models are generally quite competent given their size, even if they don't tend to be all that exciting feature-wise, relatively speaking.
And their small size brings with them certain advantages. At 14-billion parameters, Phi-4 for instance can operate on a single high-end GPU while maintaining acceptable generation rates. This makes these neural networks comparatively cheap to run next to models several times larger, which often require multiple GPUs, if not GPU servers, to achieve acceptable performance.
While Suleyman might not be interested in competing directly with OpenAI or Anthropic any time soon, Microsoft's reliance on OpenAI may not last forever. It's absolutely mission critical that long term we are able to do AI self-sufficiently at Microsoft "It's absolutely mission critical that long term we are able to do AI self-sufficiently at Microsoft," he told CNBC. But while Phi may be a precursor to achieving this goal, it appears Redmond's tie up with OpenAI will last at least another five years.
"Until 2030, at least, we are deeply partnered with OpenAI, who have [had an] enormously successful relationship for us," he added. Suleyman downplaying concerns over Microsoft's relationship with OpenAI follows the super lab's Stargate collaboration with Oracle and Softbank, which was announced last year. As part of that deal, Microsoft was no longer OpenAI's exclusive cloud partner.
However, it should be noted that Microsoft isn't the only one playing this game. Several other cloud providers have found success in this follow-the-leader strategy. Amazon Web Services arguably falls directly in this camp.
AWS is heavily invested in OpenAI rival Anthropic, to which it contributes an astronomical amount of compute, such as its Project Rainier cluster announced back in December. At the same time, AWS has been quietly building a family of language models of its own, codenamed Nova. However, unlike Microsoft, AWS appears to be keeping a tighter leash on its project.
Nova is proprietary, while Microsoft's Phi models are MIT-licensed and freely available on model hubs including Hugging Face. The case can also be made that Chinese e-commerce and cloud giant Alibaba has employed a similar strategy with its Qwen team. The Qwen family of models garnered considerable attention for many of the same reasons as Microsoft's Phi.
The models, while not necessarily ground-breaking technologically, often punch well above their weight class, achieving performance comparable with LLMs several times their size. Qwen's QwQ 32B preview made its debut in late November, a little over two months after OpenAI's o1 preview popularized the concept of "thinking" aka reasoning models. It took another three months of polishing before Alibaba released the final model, three months after the o1 was finalized.
A similar argument can be made for DeepSeek. With the concept of reasoning language models confirmed, the Chinese AI startup could focus on iterating and optimizing the concept in order to vastly reduce the compute requirements of creating and running such a model. Along with being cheaper, Suleyman's strategy also means Microsoft can focus more of its energy on building applications and other systems around large language models rather than finding new ways to wrangle neural nets.
While a lot of attention has gone into the models themselves, as we've previously discussed , integrating them into enterprise systems in a way that's actually valuable can be a rather tricky proposition. Alongside its Phi models, Microsoft has steadily pumped out research and software frameworks designed to make integrating these models easier and more compute efficient. For example, the IT titan developed Autogen, a framework for orchestrating multiple AI agents.
Meanwhile, last month, Redmond detailed its work on KBLaM, which aims to reduce the computation and complexity associated with extending a language model's knowledge using structured data. And last week, Microsoft introduced VidTok, an open-source video tokenizer for converting video into tokens in order to make it easier for machine learning models to comprehend video content. ®.
