Andrew Ng just made an argument that should unsettle every company racing to build bigger AI models: stop scaling compute and start fixing your data. His thesis, which he has been developing since around 2021-2022 and has continued to refine, is that the next wave of meaningful AI improvement for most real-world applications comes from data-centric approaches and agent orchestration, not from training ever-larger foundation models on ever-bigger GPU clusters. He has referred to aspects of this shift as moving away from the “bigger is always better” mindset. The timing remains pointed.
The data-centric thesis
The conventional AI playbook has been: take a large model, throw more compute at it, and performance improves. Ng’s counter-argument is that for most production and enterprise use cases, systematically improving the quality of the training data often produces better results at lower cost than simply scaling the model. Clean labels, consistent annotation, targeted data iteration, and tight feedback loops between deployment and retraining are the “boring” parts of machine learning, and Ng argues they are frequently the highest-leverage parts for real systems.
This is not purely theoretical. Landing AI, Ng’s company, works extensively with manufacturing and healthcare organizations where datasets are typically small, messy, and highly domain-specific. In these settings, a well-engineered smaller model trained on high-quality data can outperform a much larger model trained on noisier data. The implication is that the massive compute arms race driving demand for HBM and filling data centers may be delivering diminishing marginal returns for many of the use cases that actually pay the bills.
Agent orchestration: the next layer
The second half of the argument focuses on agents. Instead of building one giant model that attempts to handle everything, build systems of smaller, specialized components that hand off tasks intelligently. One agent handles input parsing, another retrieves relevant context, a third generates output, and an orchestrator manages the overall workflow. This is the broader shift from chatbots to agents that gained significant momentum in 2025. Still, Ng emphasizes that the individual agents do not need to be massive; they need to be well-orchestrated and reliable.
Recent moves validate parts of this direction. OpenAI’s acquisition of Ona (a platform focused on secure, persistent cloud environments for long-running agents) strengthens the infrastructure layer for agentic systems. On the financial side, Coinbase has been building tools and infrastructure (including Agentic Wallets and agent-friendly trading and payment capabilities) that enable autonomous agents to operate in crypto markets. The infrastructure conversation is shifting from “how big is your model?” toward “how well do your agents coordinate and execute over time?”
What this means for the hardware stack
If Ng’s broader thesis holds, the implications for hardware are meaningful. The dominant investment narrative, that AI compute demand is effectively infinite, and the winners will be those who build the largest GPU clusters, assumes that scaling laws will continue to deliver proportional gains indefinitely. Ng is betting that, for many enterprise and production workloads, the returns diminish and that better data and smarter orchestration can deliver strong results with less brute-force compute.
This does not threaten the need for massive training clusters for frontier models. However, it does suggest that the much larger inference market could increasingly be served by smaller, well-tuned models running on more cost-effective hardware. Early examples already exist: chips like the Snapdragon 8 Gen 3 can run 10-billion-parameter-class models locally for tasks such as summarization, photo enhancement, and voice processing. If data quality and orchestration improve, the hardware requirements for many practical applications need not continue to grow.
Ng’s perspective is uncomfortable for an industry that has become heavily invested in scale. The highest-leverage improvements may increasingly come from cleaner data pipelines and better agent systems rather than from the next larger chip. That is a less exciting narrative for investors chasing exponential hardware demand, but it may prove to be the more accurate one for where most economic value is actually created.