Stop Chasing the Biggest Model

There’s a conversation happening in every boardroom right now about AI strategy, and it almost always ends up in the same place: which frontier model should we be on?

GPT-5 or Claude? Gemini or Llama? Who has the best benchmark? Who’s winning on reasoning tasks? The assumption underneath all of it — that bigger, newer, more powerful is automatically better for your business — is one of the most expensive default settings in enterprise technology today.

The companies quietly outperforming their peers aren’t running the biggest models. They’re running the right ones.

The Scaling Assumption Is Breaking Down

For most of the last decade, the rule was simple: more compute in, more capability out. The relationship between training investment and model performance was predictable enough that you could essentially bet on it. That predictability is gone.

Sara Hooker, one of the researchers who helped build the scaling paradigm at Cohere Labs, recently published work arguing that the field has been optimizing the wrong variable. The compute-to-performance relationship is no longer reliable. We’re past the point where throwing more resources at a model produces proportional gains — and founders who haven’t updated their mental model here are paying a premium for diminishing returns.

This matters because frontier model decisions aren’t just technical choices. They’re financial ones. Enterprise contracts for top-tier model access are expensive, the compute costs compound with usage, and the organizational overhead of keeping up with rapid model releases — retraining staff, updating integrations, managing breaking changes — creates drag that doesn’t show up in any vendor pitch deck.

The Moat Isn’t the Model

Here’s the more uncomfortable truth: even if you do get access to the best frontier model available, so does everyone else. Model capability is commoditizing faster than most people realize. The question that actually matters for competitive advantage isn’t which model you’re on — it’s what you’ve built around it.

Harvard Business Review put it plainly: as model capability becomes table stakes, the real moat shifts to your workflows, your domain knowledge, and your embedded organizational judgment. The organizations investing in proprietary context and data infrastructure now are building something that won’t be replicable when the next model releases and everyone else catches up on raw capability.

Bloomberg GPT is the clearest case study for this. Bloomberg didn’t build the biggest model. They built a model trained on decades of proprietary financial data, tuned specifically for the kinds of tasks their analysts actually do. The result outperforms far larger general-purpose models on the work that matters to their business. That’s not a model advantage — it’s a context advantage. And context is a lot harder to copy.

Only 14% of enterprises have a documented AI strategy, according to a recent survey of 500+ Global 2000 executives. The majority defaulted to cost reduction as the primary rationale for AI deployment — which means they’re optimizing for the wrong thing from the start. Cost reduction is a floor, not a ceiling. The organizations that will win the next five years are the ones building proprietary AI capability on top of proprietary organizational knowledge, not just running the same frontier model as their competitors and hoping for the best.

The Signal You Might Have Missed

The clearest sign that the calculus has shifted came from an unlikely source. NVIDIA — the company whose GPU revenue depends almost entirely on demand for frontier model training — recently published research arguing that small language models are not just viable but comparably more suitable than large models for the majority of agentic use cases.

The numbers: 10 to 30 times cheaper to run, faster update cycles, stronger task-specific performance. The recommendation: reserve frontier models for tasks that genuinely require broad semantic reasoning, and use smaller, purpose-built models for everything else.

When the company that profits most from frontier model demand tells you to use smaller models for most of your work, that’s not a technical footnote. That’s a strategic signal.

AI strategy isn’t model selection. Model selection is the easy part — and increasingly, it’s also the least important part. The hard part is building the organizational context, the workflow integration, and the governance infrastructure that turns a capable model into a durable advantage.

The best model isn’t the biggest one. It’s the one that fits the job, runs on sustainable economics, and is surrounded by the organizational knowledge that makes it actually useful to your specific business.

Most companies haven’t started building that yet. Which means the window is still open.

Charles Costa, MLIS is a researcher, strategist, and founder of Lexora Labs, where he works on AI adoption, knowledge management, and the future of expert work.