Welcome to the second edition of Level.UP, brought to you by UP.Labs.

Much of today’s AI coverage focuses on what models can do. We’re more interested in what organizations actually do with them, and whether those deployments create a measurable advantage.

This week, we look at why physical industries are pulling ahead by moving beyond language models, and what a Harvard study of 200 employees reveals about the hidden cost of deploying AI without a clear objective. 

Plus: Hot takes on the deals and signals worth watching across infrastructure, fintech, and creative industries.

Think someone else needs this? Forward it to a friend/colleague navigating the same terrain.

Most enterprise AI deployments today are, at their core, text operations: summarizing, drafting, and classifying. They’re useful, but also structurally limited. 

As AI researcher Yann LeCun has argued, “language alone is not a sufficient substrate for intelligence.” A model trained on text inherits the shortcuts, abstractions, and blind spots embedded in human description, not the underlying physics.

World models take a different approach. Instead of predicting the next word, they simulate cause-and-effect. They ingest sensor data, spatial inputs, and real-world feedback to build internal representations of how systems behave under constraints.

This difference is enormous in physical industries. A wrong prediction in an LLM may produce a bad paragraph. A wrong prediction in a factory can lead to downtime, waste, or a multimillion-dollar redesign.

We’ve been tracking the rise of Vertical AI closely at UP.Labs.

PepsiCo's recent collaboration with Siemens and NVIDIA, announced at CES 2026, illustrates the shift. Rather than deploying AI across knowledge work, the company focused on one of its most capital-intensive challenges: how to reconfigure factories faster, with less risk. 

Using physics-accurate digital twins of its U.S. manufacturing and warehouse facilities, the company can simulate machines, conveyors, and pallet routes before touching a production line. Early pilots reportedly identified up to 90% of potential issues before implementation, increased throughput by 20%, and reduced capital expenditures by 10-15%.

The most durable enterprise deployments aren’t horizontal copilots layered onto generic workflows. They’re vertical systems built around the structured operational data and specific constraints of an industry: regional nuances, customer behavior, financial records, and countless other byproducts of doing business at scale. 

That proprietary history, when embedded into a world model, is what transforms AI from a generic tool into a genuine competitive advantage.

The model is rarely where the difficulty lives. The real leverage sits in the surrounding architecture: data integrity, the incentives governing how outputs are used, and whether the system can learn recursively.

Organizations can start laying the foundations by building digital twins, decisioning models, and agentic tools to empower humans. 

Those who skip this may end up with pilots that perform well in isolation but stall the moment they meet real operations.

A recent Harvard Business Review field study tracked 200 employees at a U.S. tech company over eight months, and their findings should give every leader pause: AI adoption didn't reduce workload. It intensified it.

Task boundaries blurred, coordination overhead grew as specialists corrected output that was technically complete, but strategically misaligned. And because starting anything became as simple as writing a prompt, work expanded to fill every available pocket of time. 

The rise of “vibe-coding” offers a useful parallel. It arrived promising that anyone could describe what they want, and the software would just appear — no technical background required. But as a16z partner, Justine Moore has noted, the people actually shipping production-ready apps are still developers, founders, and designers. 

For everyone else, the barriers remain: security vulnerabilities, setup complexity, and the “imagination problem.” If you've never built software, you often can't envision what’s possible.

This is what happens when capability arrives before clarity.

The HBR findings are particularly instructive because the costs are largely invisible. Task expansion, coordination drag, and attention fragmentation rarely show up in utilization dashboards. They surface later as burnout, uneven quality, and the sense that everyone’s moving faster, but not necessarily in the same direction.

Effective AI adoption answers "why" before "how." The teams navigating this well are not the ones deploying the most tools. They’re the ones defining what they’re optimizing for before rollout: the specific friction to remove, clear ownership of outputs, and a feedback loop that connects AI-assisted work to measurable outcomes.

We’re living this ourselves. At UP.Labs, AI tools have become powerful assets for rapid prototyping and first-draft thinking: building sample products and getting testable concepts in front of clients without ever engaging the development team. 

But the value is concentrated in exploration, not execution. That distinction matters. What gets built quickly in a vibe-coding session requires considerable human guidance and QA before shipping.

The managers who understand this use AI to accelerate creativity, not to bypass building. Those who don’t create a subtler problem: pressure for teams to ”just use AI” and ship what was only ever meant to be a sketch.

At the leadership level, this is what separates teams moving fast in the right direction from those just moving fast.

Ready to work smarter? Here are the tools we're using to actually get more done:

The AI Infrastructure Race Has a Memory Problem. Micron just committed $200B to new chip fabs with its entire high-bandwidth memory output already sold out through 2026. But beyond supply constraints, there’s a deeper bottleneck emerging: orchestration. DRAM prices have jumped 7x over the past year. As inference costs fall, the advantage will shift to operators who can intelligently allocate, cache, and structure memory across their systems. Compute gets cheaper. Coordination doesn’t. → Read more

Mistral Is No Longer Just a Model Company. The French AI lab acquired serverless infrastructure startup Koyeb in its first-ever deal. The signal: Mistral is building toward a full-stack AI cloud, not just competing on model quality. With $400M in ARR, a $1.4B data center investment in Sweden, and now an infrastructure acquisition, the company is positioning itself as Europe's sovereign alternative to U.S. AI infrastructure — a timely move amid a US-led tariff war. → Read more

Sony Built a Copyright Radar for AI Music. Sony has developed a system that can identify original recordings embedded in AI-generated music and estimate the extent to which each source contributed to the final track. The learning: They’re building the governance infrastructure before deciding whether to ban or embrace AI-generated content, rather than waiting for regulation to force the question. → Read more