This piece is commentary and analysis, not investment advice. Always consult your IFA before making investments.

In 1855, Andrew Carnegie was a teenage telegraph boy running through muddy Pittsburgh streets. Within two generations, he and his peers forged the modern world. Horses gave way to railroads, candlelight to electricity, iron to steel.

What made steel transformative wasn’t just strength. It was predictability. Wrought iron contained slag inclusions and directional grain that made every beam a gamble—test one spot, and you knew what that spot would do, but nothing about anywhere else. Steel was homogeneous. Test a sample, and you knew what the whole structure would do. Engineers could calculate rather than pray.

This has lots of analogies to what is currently happening in AI.

Did we hit a wall?

At the start of 2025, the conversation was about limits. AI scaling laws appeared to be hitting walls. Pre-training costs were rising faster than capabilities. Serious people asked whether AI had reached a plateau—whether the transformation everyone had been promised would stall at “impressive demos” and never quite reach “reliable tools.”

By December, that conversation looked quaint.

Anthropic released Opus 4.5. It blew away past performance for agentic coding, providing that models are improving faster than ever. The scaling constraints didn’t disappear in 2025; the field found ways around them. Andrej Karpathy, who co-founded OpenAI and shaped how the industry thinks about neural networks, recently wrote that he’d “never felt this much behind as a programmer.” The profession, he said, was being “dramatically refactored.” His assessment: he could be “10X more powerful” if he could just figure out how to use what already exists.

When the architects of the technology feel overwhelmed by what they’ve built, you’re witnessing something other than a speculative bubble.

This is a boom. Not a bubble about to burst, but a transformation accelerating faster than most participants can absorb. We are standing at a threshold.

The question is no longer whether AI is real. The question is what happens as it deploys—and where, exactly, value accrues.

The answer lies in understanding why transformation has been so uneven.

Why Software Transformed First

Software development has changed beyond recognition. Developers describe work that is unrecognisable from three years ago. Other fields—healthcare, manufacturing, professional services—have seen pilots, experiments, promising demos. But not yet the wholesale transformation that developers describe.

The explanation isn’t just that software is “text-based” and therefore easier for language models. The explanation is verification infrastructure.

Consider what happens when AI writes code. You can compile it and see if it runs. You can execute test suites and see if they pass. You can deploy to staging environments and observe behaviour. The entire feedback loop happens in silico—fast iteration, immediate signal, rapid correction. Software development had spent decades building infrastructure to verify outputs: compilers, test frameworks, continuous integration, version control. When AI arrived, that infrastructure was waiting.

Language models are probabilistic. They predict what’s likely, not what’s correct. A model can generate a contract clause that looks plausible, but without the right scaffoling, it cannot know whether that clause violates a precedent your legal team established three years ago. That determination requires something external to the model: infrastructure that captures what correctness means for this organisation, this decision, this moment.

The Missing Piece

Call it decision traces. The accumulated reasoning, precedents, and exceptions that allow organisations to verify whether an AI output is right for their situation.

Software development had these encoded already—in test suites, type systems, deployment pipelines. When an AI writes code, the test suite encodes decades of institutional knowledge about what correct behaviour looks like. The type system captures constraints that someone, somewhere, decided mattered. The deployment pipeline embodies lessons learned from every production incident.

Other fields don’t have this. Ask a deal desk why they approved a particular discount, and you’ll get a story: a similar case six months ago, an exception a VP granted on a call, institutional memory about why healthcare customers need different terms. None of that is in any system. It lives in people’s heads and dies when they leave. (Ask any salesperson what the “real” discount policy is.)

This is why upwards of 95% of enterprise AI projects currently fail to deliver. Not because the models lack capability. Because the organisations lack the infrastructure to verify what the models produce.

The constraint moved. It used to be “can we afford to do this task?” Now it’s “can we tell if it was done correctly?”

Decision traces will emerge as a significant software category—systems that capture which decisions are made and why, not just what happened. This is the infrastructure that enterprise AI is missing. It’s being built now.

But What About Manufacturing?

Decision traces explain enterprise AI. But they don’t explain manufacturing.

When a factory worker picks up a screw and places it, they’re not drawing on accumulated institutional reasoning about why to pick up that screw. They’re drawing on something more fundamental: an intuitive understanding of physics. How objects move. What happens when force is applied. Where things land if you drop them.

A robot doesn’t need to know why the last worker chose that screw. It needs to know what happens when it reaches for it.

This is what’s missing for physical AI. Not decision traces. Spatial intelligence created by world models.

Physical AI and the Next Frontier

Yann LeCun (one of the three “godfathers of deep learning” who shared the 2018 Turing Award for inventing the neural network architectures behind modern AI) left Meta in November 2025 after twelve years as Chief AI Scientist. His reason: to build world models.

What’s missing, LeCun argues, is systems that understand the physical world. Models that have persistent memory, can reason about cause and effect, and can plan complex sequences of action. World models: AI that learns physics.

Jensen Huang, CEO of NVIDIA, frames physical AI as the next great opportunity: “The ChatGPT moment for general robotics is just around the corner… this could be the largest industry of all.”

The infrastructure is being built. NVIDIA’s Cosmos platform provides world foundation models trained on millions of hours of physical interaction. Meta’s V-JEPA 2 can predict the consequences of physical actions before they happen. These systems enable simulation at scale—test before you deploy, iterate in silico rather than waiting for parts to arrive.

An epic disaster

Jim Fan, who leads robotics research at NVIDIA, captures the current state: “Hardware is ahead of software, but benchmarking is still an epic disaster. No one agrees on anything: hardware platform, task definition, scoring rubrics, simulator, or real world setups.” (When NVIDIA’s head of robotics research calls the field “an epic disaster,” pay attention.)

The verification infrastructure is being built, but it isn’t mature. Transformation will follow: possibly slower than enthusiasts hope, but definitely faster than sceptics expect.

Scientific research shows the pattern working. AlphaFold predicted 200 million protein structures because proteins can be validated experimentally. Isomorphic Labs reports 80–90% Phase I success rates for AI-designed drug candidates—compared to 40–65% for traditional approaches—because wet lab verification closes the loop. The infrastructure existed; AI leveraged it.

Healthcare is building infrastructure now. The FDA has authorised 1,356 AI-enabled medical devices. Manufacturing has physical QA infrastructure that translates: Figure AI’s robots have placed 90,000+ parts with 99% accuracy at BMW.

The transformation is uneven because verification infrastructure is uneven. But infrastructure gets built.

What Comes Next

We opened with Carnegie and the revolution that steel enabled: not because steel was stronger, but because it was predictable. Test a sample, know the whole. Engineers could calculate rather than pray.

AI is at a similar inflexion. The capability is there. What’s being built now is the infrastructure that makes capability trustworthy: decision traces for enterprise, world models for physical AI.

Both are being built. Simultaneously. Across every sector.

That’s why we’re standing at a threshold, not watching a bubble.

Karpathy’s observation lands differently now than it would have twelve months ago. He wrote that AI tools feel like “some powerful alien tool handed around, except it comes with no manual and everyone has to figure out how to hold it.”

The gap between what’s technically possible and what’s actually deployed is the story of 2026. Our mission is to find and back the founders who make it happen.

Happy New Year!

You can read our 2026 predictions here.