Value moves up the stack: our market view, Q2 2026

Acceleration continues as value moves up the stack

Over the past six to twelve months, concerns about an AI bubble, chip constraints, weak enterprise returns, exhausted training data and slowing progress have grown louder.

We have remained bullish. These are real challenges, but we see them as growing pains, not evidence that AI has plateaued. Recent developments reinforce the thesis in our Commoditized Magic and the Agentic Economy talk: model intelligence and agent performance continue to improve while the cost of intelligence keeps falling.

Recent events provide striking, if sometimes uncomfortable, evidence of this progress. In July, AI models being evaluated for cybersecurity capabilities escaped their intended testing environment, discovered a previously unknown vulnerability and compromised part of Hugging Face's production infrastructure. The models combined multiple vulnerabilities and credentials to pursue their assigned objective. Hugging Face then used AI to analyse more than 17,000 actions, completing in hours work that would ordinarily take days. The incident was a serious warning about safety, but also a concrete demonstration of agents' ability to perform complex, multi-step work in real environments.

Progress in mathematics is equally notable. Claude recently helped produce a verifiable counterexample to an 87-year-old version of the Jacobian conjecture. An experimental Claude system also reportedly made significant progress on a question related to the Riemann hypothesis, without claiming to solve the hypothesis itself. These results require continued expert scrutiny, but they suggest models are beginning to contribute to genuinely new knowledge rather than simply reproducing information found online.

Intelligence is getting cheaper, and enterprises are learning how to use it

Competition, including from Chinese and open-weight models, continues to drive down prices. OpenAI's latest model family was structured around a price-performance ladder, with lower-cost variants reportedly reducing task costs by approximately 50 to 80% relative to the flagship while retaining strong performance.

Enterprise adoption is rising alongside greater financial discipline. Instead of buying one expensive assistant for every employee, sophisticated companies are measuring usage, selecting different models for different tasks and reserving premium intelligence for the problems that justify it. Databricks, for example, shared how it reduced internal AI coding spend by up to 90% in some scenarios while usage kept growing: shifting defaults to cheaper, more efficient models (about 50% of the savings), smart routing (about 30%), user visibility and adaptive budgeting (about 10%), and pruning context bloat and harness tuning (about 10%).

Falling costs should therefore expand the market, but they also reduce supplier lock-in. Our investment in OpenInfer, which brings data-centre-scale AI capabilities to low-end compute and edge devices, and a new investment in AI-driven modernisation of legacy enterprise systems (still in stealth) both reflect this opportunity.

Value is moving from models to agentic systems

The centre of value has already shifted from simple "LLM wrappers" toward proprietary context, tools and workflows. It is now moving toward the agent harness, the surrounding system that gives a model memory, tools, feedback and a process for completing work.

In one software-engineering comparison (Joel Niklaus's analysis), changing the harness moved task success from 23% to 52% for one model and from 15% to 36% for another. A smaller model in the right operating environment could approach the performance of a much larger model in the wrong one.

Harness quality matters today, but we expect frontier models to automate much of this engineering over time. We are therefore cautious about treating handcrafted harnesses as a lasting moat. Durable value will instead come from assets that models cannot easily reproduce: proprietary data, deep expertise, embedded workflows and the ability to execute complex processes in the real world.

The next frontier extends beyond software

The first AI wave focused on language, content and enterprise automation. The next is expanding into scientific discovery, robotics, hardware and manufacturing.

These markets are particularly attractive because real-world constraints create stronger barriers to entry. Improving a factory process, designing a medicine or operating an autonomous machine requires specialised expertise, proprietary feedback loops, safety and deep integration, not merely access to a model.

This is an area in which we are actively investing: Juna applies AI agents to industrial processes, Dreamfold to medicine design, Bench to hardware development and Flink and Levtek to autonomous systems. We set out the argument in AI beyond software.

The market will continue to produce volatility and failed experiments. But beneath that noise, the direction remains clear: AI is becoming more capable, less expensive and more widely deployed. As intelligence and software are commoditised, the largest opportunities move toward defensible systems that apply that intelligence to the world's hardest problems.

That view also determines where we do not invest. Several investors asked where we stand on the new wave of AI labs; we set out the argument in full in The rise of the neolabs: why we looked at a number of these rounds and passed, and why we expect the model layer for general intelligence to commoditise.