AI efficiency replaces scale as industry's next priority

The efficiency imperative

After years of pursuing ever-larger AI models, the technology industry is confronting a new challenge: making these systems affordable and efficient enough for widespread deployment. The shift reflects growing concern about the economic sustainability of current AI approaches, particularly as companies scale up agent deployments and face mounting infrastructure costs.

Sara Hooker, cofounder and CEO of AI lab Adaption, described the problem during Fortune's Brainstorm Tech conference on Tuesday. Most contemporary AI models are "monolithic," she explained—their knowledge and capabilities remain fixed after training. When circumstances change or users provide valuable feedback, that information doesn't integrate into the model automatically.

"You need models that can evolve," Hooker said, "otherwise you end up with massive inefficiencies."

Why it matters

The economics of AI deployment have become a critical bottleneck for enterprise adoption. Companies deploying AI agents at scale are paying repeatedly for the same computational errors because their systems don't learn from mistakes. This dynamic drives up API costs, infrastructure expenses, and energy consumption—potentially limiting which organizations can afford to implement advanced AI capabilities and slowing the technology's practical impact.

The scale paradox

Rodrigo Liang, CEO of AI chip company SambaNova, acknowledged that large models aren't disappearing soon, though he expects "plenty of room for more efficient models to come in." For now, customers struggle with the costs of scaling models, energy-intensive infrastructure requirements, and securing adequate computing resources.

Hooker argued the industry has reached "an inflection point with massive urgency to change that curve" of model size. She noted that most people intuitively understand not every problem requires the same computational approach. "Probably 90% of problems are very easy—many things that you do in bulk processing, for example, you shouldn't be throwing a massive model at."

Future AI systems will need to adapt continuously to new information and modify their behavior dynamically, rather than relying on repeated calls to static models, she said. Current enterprise deployments often lack this learning capability, forcing companies to pay repeatedly in compute resources, API calls, and infrastructure for identical errors.

Hardware as a bridge solution

While researchers like Hooker focus on building more adaptive AI architectures, hardware providers are working to make today's massive models economically viable. Liang said trillion-parameter models remain too expensive and power-hungry for most real-world applications.

SambaNova's approach centers on faster inference with lower power consumption through hardware designed specifically for large-model workloads. "We're getting two to 3x better than the [Nvidia] Blackwells [GPUs] on the exact same models, and so we think that at scale that's the way to at least bring the cost down," Liang said.

These details were first reported by Fortune from the Brainstorm Tech conference.