Thoughts on the AI Disruption

Over the last few months, the conversations we have across our partner and board network have converged on a single question: what does the AI wave mean for our businesses, and how should we act? The honest answer is that AI is changing the world faster than any technology cycle in living memory — and that nothing, not capital intensity, not regulation, not vendor inertia, is going to stop it. This is how we read the disruption, and why our recent investments are deliberately positioned for the world that is coming, not the one that is passing.

The macro game: a capex super-cycle we don't need to play. At the very top of the stack, foundation-model development has become an industrial-scale capital contest. The largest hyperscalers together lifted their 2025 AI-infrastructure spend above $380 billion, and announced global AI-infrastructure plans now exceed $2 trillion. The energy stack is being rebuilt to match — nuclear power-purchase agreements, restarted reactors, even the first commercial fusion deals — as hyperscalers vertically integrate primary energy for the first time. This is a game whose returns accrue to hyperscalers and a handful of sovereign-scale labs. It is not one European deep-tech venture can, or should, play.

The avalanche: why classic SaaS is being marked down. If the top of the stack is a capex game, the middle is an extinction event. Public software multiples have compressed by roughly 41% from their 2023 peak. The business model under threat is the one that built a generation of venture fortunes: ingest the customer's data, render it back as a dashboard, and charge per seat per month. Generative and agentic AI dissolves that wrapper — once a model can write, query, summarise and act against the same data, the dashboard becomes a feature, not a product. The next category leaders will sell work and outcomes, not seats. Pricing is already migrating from access to output: per resolution, per conversation, per result.

Enterprises are voting with their wallets. Enterprise generative-AI spending reached an estimated $37 billion in 2025 — more than three times the prior year — with the majority flowing to the application layer rather than infrastructure, and most use cases bought rather than built. The disruption is no longer theoretical; it is showing up in revenue.

The opportunity: agents create as much trust deficit as they create value. This is where our thesis turns from defensive to offensive. The agentic wave is not just rolling over SaaS — it is opening the largest trust, governance and oversight gap the enterprise has ever faced. Security incidents are no longer hypothetical: zero-click data exfiltration through prompt injection, AI-security incidents doubling year on year, and over-scoped autonomous agents causing real damage. Regulation has stopped being optional, with the EU AI Act now live and penalties reaching up to 7% of global turnover. And an assurance market is forming around auditability, human-in-the-loop control, and proof of correctness.

In every prior platform transition — client-server, web, mobile, cloud — the durable returns came not from owning the new substrate, but from solving the problems it created at the edge of trust: SSL/TLS, identity providers, device management, zero-trust networking. Agentic AI is now in that exact moment.

Why we invest in critical systems. This is the heart of our thesis. We invest from Seed to Series B in deep tech, with cybersecurity, governance, and the AI of critical systems as a permanent overweight. Take the emerging human-authorisation layer for AI-driven enterprises: before an AI agent executes a sensitive action — a payment, a clinical decision, an insurance adjudication — it routes the action to the right human approver and issues a cryptographically signed, auditable mandate that any regulator or partner can verify. Verifiable human authorisation is to the agentic economy what TLS was to e-commerce. It sits alongside the model-reliability layer — uncertainty quantification and output reliability — and our broader focus on cybersecurity, edge AI, defence and dual-use, and bio-integrated systems: the domains where introducing AI creates opportunity and risk simultaneously, and where getting it wrong is not an option.

What this means in practice. First: don't chase the foundation-model layer — its capital and energy intensity mean the returns accrue elsewhere. Second: re-underwrite legacy SaaS exposure — any company whose value proposition is "ingest your data and show it back to you" should be stress-tested against an agentic alternative that performs the underlying work. Third: lean into the critical-systems stack. The same regulatory and operational pressure forcing every enterprise to deploy AI is also forcing them to govern it — and that gap, from human-in-the-loop authorisation to output reliability, model assurance, agent observability and AI-specific cybersecurity, is where venture-scale outcomes are most likely to compound over the next decade. Equally, true critical systems in aerospace, defence and manufacturing — often with deep hardware integration — are among the most insulated from being made redundant by AI.

A measured note to close. The agentic disruption is real, but incumbents are not standing still, and a portion of legacy software will adapt rather than die. Our position is not euphoric. It is that the direction is settled, the timing is fast, and the value will accrue to whoever solves the hard problems that arrive in the wake of the technology — not to whoever owns the technology itself.