The “party” of brute-force scaling is winding down, and the industry is finally starting to sober up. For the past few years, the dominant narrative for CEOs and Product Leaders was simply “bigger is better.” We were told that more parameters, more compute, and more data would inevitably lead us to the promised land of AGI. But as we move into this current phase of the cycle, we are hitting a wall that isn’t just technical—it’s economic and structural.
I’ve seen this movie before. During my time at Nokia, we saw the peak of the “more is more” hardware race before the market pivoted toward the elegance of integrated ecosystems. Today, we are seeing the same “vibe check” in AI. The era of the general-purpose, “one model to rule them all” approach is giving way to a more pragmatic, fragmented, and ultimately more useful reality. We are moving from the hype of the demo to the hard work of the deployment.
The Age of Research: Beyond the Scaling Laws
For a decade, we’ve relied on the belief that scaling laws were the ultimate cheat code. If you make it 100 times larger, it gets smarter. But even OpenAI and Meta researchers are now admitting that pretraining results are flattening. We are entering what many are calling the “Great Filtration,” where the labs that simply throw more GPUs at the problem will be separated from those innovating on architecture.
As leaders, this means we must stop waiting for a smarter model to solve our product flaws. The next leap isn’t coming from a larger version of what we already have; it’s coming from “world models” and systems that understand spatial reasoning. Startups like World Labs and Runway are already shifting the focus toward models that learn through experience—physics and interaction—rather than just predicting the next word in a sentence.
The “Small Model” Advantage
While the headlines focus on the giants, the real winners in the enterprise space are the Small Language Models (SLMs). I’ve always argued that value for users drives value for business, and in 2026, user value is found in speed and privacy. AT&T and other major players are increasingly moving away from out-of-the-box LLMs in favour of fine-tuned SLMs. Why?
- Precision: A model trained specifically for your domain outperforms a generalist every time.
- Cost: Inference costs for general models are falling, but the efficiency of running a small model on the edge is unbeatable.
- Enshittification Guardrails: By owning and fine-tuning your own smaller models, you insulate your product from the “Rot Economy” where third-party APIs slowly degrade in quality while increasing in price.
The trend is clear: we are moving toward “Physical AI.” Intelligence is being embedded into devices—wearables, drones, and industrial robotics—where a 200-billion parameter model is a hindrance, not a help. If you aren’t looking at how to deploy intelligence locally, you’re building on borrowed time.
From Autonomy to Augmentation
We’ve been promised for years that AI agents would handle end-to-end tasks autonomously. The reality has been more “slop” than substance. However, the introduction of the Model Context Protocol (MCP)—often described as the “USB-C for AI”—is finally providing the connective tissue needed. By standardising how agents talk to databases and APIs, we are moving agents out of pilot workflows and into the real world.
But here is the critical part for leaders: 2026 is the year of the human. The most successful implementations I’m seeing aren’t those trying to replace a department, but those using agentic workflows to augment human experts. We’ve reached the limit of what “brute force” automation can do without human empathy and ethics. As TechCrunch recently noted, the focus is shifting toward designing systems that integrate cleanly into human workflows rather than trying to bypass them.
The Strategic Pivot
If you are a CEO or a Product Leader, your strategy for the coming year cannot be “more AI.” It must be “more useful AI.” This requires a shift in how we fund and build product teams. We need to move away from the “project” mentality—where we ship an AI feature and move on—toward a “product” model of continuous iteration and ethical oversight.
We must guard against the enshittification of our own platforms. When Google turns Gemini into a personal shopper that bypasses retailer websites, as reported by The Register, they are prioritising their own rent-seeking over the health of the digital ecosystem. Don’t fall into the same trap. Use AI to solve a genuine user friction, not to build a wall between you and your customers.
The long view of technology cycles tells us that every boom is followed by a period of “sobering up.” We are in that period now. The companies that will thrive are not those with the largest models, but those with the most empowered, autonomous product teams who understand how to use these new tools to create tangible, human-centric value. The tools are getting smaller, more local, and more integrated. It’s time our strategies did the same.
Leave a Reply