The AI industry is pivoting from a years-long obsession with bigger models to the urgent task of making those systems efficient enough to deploy at scale. Executives at Fortune's Brainstorm Tech conference argued that the era of "monolithic" models is ending, with Adaption CEO Sara Hooker and SambaNova CEO Rodrigo Liang calling for a fundamental shift in how AI is built and run.
- The Monolithic Model Problem
- Why Efficiency Matters Now
- Hardware Competition Heats Up
- What This Means for the Industry
- Frequently Asked Questions
- Conclusion
The Monolithic Model Problem
For the past few years, the dominant narrative in AI has been "bigger is better." Each new model release pushed parameter counts into the hundreds of billions or trillions. But Sara Hooker, cofounder and CEO of startup AI lab Adaption, argued that this approach has created a fundamental inefficiency.
"Most of today's AI is monolithic — stuck in time," she told the conference. "Once a model is trained, the model's knowledge and capabilities are essentially fixed. If something changes in the world, or if the model learns something useful from users, that knowledge doesn't automatically become part of the model."
Hooker described this as a structural problem that compounds with scale. Enterprises deploying AI agents are paying repeatedly — in compute, API calls, and infrastructure costs — for the same errors, because the models aren't learning from mistakes. "You need models that can evolve," she explained, "otherwise you end up with massive inefficiencies."
Why Efficiency Matters Now
The urgency is rooted in economics. According to Hooker, roughly 90% of problems are simple enough that they don't require massive models. "Many things that you do in bulk processing, for example, you shouldn't be throwing a massive model at," she said.
She described the current moment as an "inflection point with massive urgency to change that curve" of model size. The soaring API bills companies are experiencing are a direct result of using fixed, oversized models for every task. Future AI systems, she argued, will need to adapt continuously to new information and rapidly change their behavior, rather than making repeated calls to a static model.
This shift from "scale for scale's sake" to adaptive, efficient deployment could dramatically reshape the AI industry's cost structure. Instead of spending billions on training ever-larger models, companies would invest in systems that learn on the job and become cheaper to run over time.
Hardware Competition Heats Up
While model developers pursue efficiency through architecture and training methods, hardware companies are taking a different path. Rodrigo Liang, CEO of AI chip company SambaNova, said the immediate challenge is running today's trillion-parameter models affordably enough for real-world deployments.
"Trillion-parameter models remain too expensive and power-hungry," Liang said. His company's strategy is to deliver faster inference with lower power consumption through chips designed specifically for large-model workloads. He claimed SambaNova's hardware achieves "two to 3x better than the [Nvidia] Blackwells [GPUs] on the exact same models," and argued that at scale, that kind of efficiency gain is the most direct path to reducing costs.
Liang acknowledged that the largest models aren't going away anytime soon, but predicted "plenty of room for more efficient models to come in." For now, customers are left to struggle with the cost of scaling, energy-hungry infrastructure, and a shortage of AI talent.
What This Means for the Industry
The macro shift from scale to efficiency carries significant implications for investors, competitors, and the broader tech ecosystem.
For investors: The narrative that "bigger AI models = better returns" is fading. Companies that can demonstrate cost-efficient deployment and real-world ROI will likely command higher valuations. Hardware startups like SambaNova that claim to challenge Nvidia's dominance on efficiency metrics become more attractive, but they face a massive uphill battle against Nvidia's ecosystem lock-in.
For competitors: Nvidia's grip on the AI chip market is based on performance per dollar for training massive models. If the industry pivots to inference-heavy, adaptive systems, the advantage could shift to companies with specialized inference chips. AMD, Intel, and startups offering purpose-built silicon may gain share.
For the tech industry broadly: Enterprise adoption of AI has been hampered by cost overruns. If models become cheaper to run and more adaptive, the barriers to deployment drop. That could accelerate the shift from experimental AI pilots to production systems across finance, healthcare, and enterprise software.
Conclusion
The AI industry is at a turning point. After years of scaling models as fast as compute would allow, the conversation is shifting to sustainability, cost, and adaptability — not just raw intelligence. Leaders like Hooker and Liang are betting that the next wave of AI value creation will come from systems that are cheaper, smarter over time, and tailored to the specific problem at hand. Whether measured by chip efficiency or model architecture, the era of "bigger is better" is giving way to something more nuanced: deployable, affordable, and adaptive AI.
2026 Robot Overflow. All rights reserved.