The year is three-quarters over and it is looking like the most consequential one of the artificial-intelligence boom so far. Agents went mainstream for enterprise users during the first half of the year, and if Meta Platforms’ Muse is any indication, consumers aren’t far behind. At the same time, cheap open-weight models—which anyone can download for free and run on any capable hardware—have gotten quite good. They are putting pricing pressure on the two leaders in AI model-making, OpenAI and Anthropic, even as they have rapidly pushed the AI frontier forward this year.
As the last 50 years of technology development have taught us, software eventually wins. AI is proving to be no different. For example, early PCs were expensive toys for hobbyists until the first spreadsheet software, VisiCalc, came out in 1979 and turned them into professional tools. It turned the Apple II into a hit, and launched what is now one of the world’s most valuable companies. We see the same thing in the mature smartphone market. The iPhone competes against flagship Android phones at similar price points with similar hardware. Customers are choosing between the operating systems and the selection of software for each.
In the first phase of the AI boom, the focus was on data center hardware and models. Nvidia, the AI chip leader, has seen its stock go up 1,263% since ChatGPT was released in late 2022. The average Wall Street analyst price target predicts the stock could rise another 44% from its current level. During the same period, OpenAI and Anthropic have gone from unknown start-ups to potential trillion-dollar-plus valuations.
Hardware will remain supply-constrained for the foreseeable future and the rise of agents just means demand for more chips. Agents like Muse and OpenAi’s Dots can have conversations with AI models at a pace that no human can match, and they burn through cloud computing.
More agents means more cloud computing, which requires more GPUs and more custom AI chips from Alphabet, Meta, Amazon.com, and others. But agents themselves are software, which run on CPUs from the likes of Advanced Micro Devices, Intel, Arm, and others. Estimates for 2030 server CPU sales have rocketed up from $50 billion at the end of last year to nearly a quarter trillion dollars today. More agents will also worsen the memory and storage chip shortage.
Meta’s Muse agent cloud software runs on AMD CPUs. The model behind Muse, called Muse Spark, runs on Nvidia servers and Meta’s own custom MTIA chip made in conjunction with Broadcom. Both types of chips are manufactured at Taiwan Semiconductor Manufacturing.
In this second phase models are becoming commoditized, and we can see this in recent price cuts from OpenAI and Anthropic. They still have the best models, but there are also other powerful and less expensive options now. Many customers are using software that routes only the most difficult jobs to the pricey models, reserving the rest for models that can be 50 times cheaper, including, in some instances, open-weight ones.
Muse’s early success shows how important the software layer already is, especially in the consumer market. Released late last year, OpenClaw is an agent that works much the same way Muse does, but the software was designed for techies, not normal people, and thus its appeal and uptake have remained limited. Anthropic has had a personal agent available since January, but hasn’t had much success with it.
Anthropic and the open-source engineers behind OpenClaw aren’t skilled at app design, but Meta is. Muse breaking through emphasizes how crucial user-friendly app design is to successful mass-market AI products. Usability and approachability are only going to become more important going forward. And the more AI agents come to market, whether for consumer or enterprise use, the more AI infrastructure will be needed.
Write to Adam Levine at adam.levine@barrons.com