AI’s future could give people more choices about how they use it and who provides it, with open models that run on their own devices and a wider range of companies competing to serve them. David Gray Widder, an assistant professor at the iSchool, studies what it would take to make that possible and the forces that stand in the way.
From the Seoul AI Policy Conference to the Mila-Quebec AI Institute, and from the pages of Nature to MIT Technology Review, Widder has been speaking and commenting on how AI is developed and governed. His research connects a critical examination of the industry’s growing concentration with an alternative vision of more competitive markets and open- and device-based AI systems that are less reliant on data centers.
AI on Your Own Device
One way to expand those choices is through "local AI", with models that run directly on a laptop or phone rather than back and forth with a data center. Widder sees an opportunity to separate the resource-heavy work of training a model from how people adapt and use it.
"There's not a way to train a competitive AI model locally," he says. "Training is still done in the cloud. But if that model is then released openly, other people can take it and find ways of making it smaller."
That combination of centralized training and decentralized adaptation and use could support a less data-center-intensive AI future. Widder intends to examine ways to use AI locally; for example, in an in-progress paper, Widder and PhD student Kevin Woohyeuk Lee have studied llama.cpp, an open-source community built to make large models like Meta's Llama run on ordinary, "broadly available, low-powered" hardware.
The llama.cpp example shows how community contributors and academics can help models work across different kinds of hardware, if the models are sufficiently open. AI openness claims versus reality is another focus of Widder’s scholarship. The tools that let people move between models and devices matter, too.
"For AI to not be controlled by any one big player," Widder says, "we need interoperability layers that are themselves not controlled by any one big player."
The Barriers to Greater Choice
The challenge is creating room for these alternatives in an AI industry dominated by a handful of cloud providers. In a report co-authored with Nathan Kim, “How Big Cloud Becomes Bigger,” Widder examines how Google, Microsoft and Amazon’s investments reinforce that position.
“Google, Microsoft and Amazon have two-thirds of cloud compute market share between them,” Widder says. “They rule the market. This means that if you want to build and use AI, there aren't many routes that don't involve them.” Smaller cloud providers also face a persistent backlog for the chips they need to compete.
The big three can also subsidize their customers, inducing demand in ways smaller competitors cannot hope to imitate. Together, Google, Microsoft and Amazon invest in other companies roughly 100 times more often, and pour in about four times as much capital, as the rest of "Big Tech" combined. Increasingly, that investment is focused on cloud-dependent AI startups, tying the industry’s growth to the infrastructure these companies provide.
Cloud Investments and Bubble Fears
That pattern raises questions about both the direction of AI development and the durability of its investment boom. Widder has watched closely for signs that the boom is at least partially a bubble. “The bubble concern is that these three companies are investing a lot in building more cloud compute, building data centers, but then they're also investing in the people that are supposed to create the demand for that compute,” he says. “It's hard to differentiate what of this is actually organic versus spurious demand.”
If underlying demand for cloud compute is weaker than it appears, speculative investment could increase the danger of a bubble. The consequences would reach beyond the technology industry. “If a bubble popped quickly rather than deflate slowly, that would hurt a lot of people’s retirement plans,” Widder notes. “People have a lot of their investments in AI companies."
For Widder, these investments also raise a question about whose priorities AI infrastructure serves. Demand generated by the largest cloud providers’ investments may steer the supply chain toward the needs of the startups they finance, rather than the broader needs of consumers and businesses. As cloud dependency becomes more deeply embedded, changing direction gets harder.
Giving the Public a Voice
The public may have a role in pushing for that change. Widder’s research considers how people’s views could influence the industry’s direction, particularly as its physical footprint grows. Polling suggests that attitudes toward AI remain cautious, with enthusiasm concentrated mainly among people earning more than $200,000 a year or those who build AI systems for a living.
"We're seeing people increasingly concerned about data centers being built," Widder says. “I share that concern."
New research Widder has begun at UT with his PhD student Ali Alkhatib, a Harrington Fellow, and Haley Triem a third year in the department looks at data center construction across Texas, where opposition cuts across party lines. He sees that concern as one of the few bipartisan flashpoints in American politics.
Questions about concentrated power in AI have also brought together unlikely voices: Biden administration FTC chair Lina Khan, right-wing strategist Steve Bannon, and venture capitalist Marc Andreessen have all expressed concern, even as they disagree about nearly everything else. For Widder, these converging concerns suggest that public pressure could help change the direction of the AI supply chain.
Bringing These Questions into iSchool Classrooms
At the iSchool, Widder is helping students examine the same forces shaping AI’s future. Building on his popular “Political Economy of AI Supply Chains” seminar, Widder debuted a new course this semester, "AI, Hype, and Capitalism," that quickly became one of the most enrolled electives in its concentration. The syllabus includes a broad variety of perspectives on the topic, from Andreesen to Pope Leo XIV to highly technical papers on how GPUs work. He sees this as evidence that UT students are hungry for tools for thinking critically about how AI systems are built, financed and governed.
Widder argues that AI development does not fully reflect the range of public perspectives, and young people may sense that gap especially clearly. He hopes their questions, alongside the rigorous, human-centered research pursued at the iSchool, can help UT’s future leaders imagine and develop AI systems that better reflect public priorities. "How can we reimagine AI ecosystems," he asks, "so that they serve Americans' interests, and the public interest?"
Find out more about Widder’s research and thought leadership at his personal website, https://davidwidder.me/