Two days at Goldman Sachs’ 11th Disruptive Technology Symposium in London (Plumtree Court’s auditorium is properly cinematic) felt less like a conference and more like a compressed download of where tech is actually heading - from frontier AI through to cyber, quantum, defence, and the very unsexy bottleneck that keeps coming up: power.

What struck me most wasn’t the volume (over a thousand attendees) or even the calibre of speakers - it was the discipline of the agenda. This wasn’t “AI theatre”. It was a deliberately cross-stack view: capital formation, compute, deployment, regulation, security, and geopolitics - in the same room, in the same sentences. That’s the point: AI is no longer a feature. It’s becoming infrastructure. And infrastructure always becomes political.

The big shift: AI moves from copilots to “co-workers” - but reliability is the choke point

The most honest thread running through the AI sessions: agentic systems are not deterministic enough yet to be left alone. We’re still in the “copilot makes you 10–30% faster” era for many workflows. But the ambition is unmistakably “digital co-workers” running end-to-end tasks - and that transition hinges on trust: governance, evals, security, auditability, and knowing when a human must be in the loop.

One provocative implication (that several speakers danced around but didn’t fully say): the workforce is going to bifurcate. Fewer people doing the production. More people doing two things:

1. Setting direction, and

2. Validating outputs.

If your job is “produce the first draft”, you should assume the floor is moving under you.

Distribution beats model quality (and everyone knows it)

There was a recurring, slightly uncomfortable consensus: the best model doesn’t necessarily win - the model with the best distribution, partnerships, and narrative momentum often does. If you’re building an AI company, the differentiator isn’t “we used the latest architecture”; it’s:

• Proprietary data (and the network effects around it),

• A wedge into a real workflow,

• And a go-to-market engine that compounds.

The second-order effect: we’ll likely see a wave of consolidation, because buying something “enterprise-grade” and plugging it in is often faster than building it - especially as incumbents wake up and start purchasing their way back into relevance.

Everyone’s rediscovering hardware - because the real world is the TAM

A theme that kept reappearing: the future isn’t just more software, it’s software eating physics. The most compelling case studies were where AI collides with atoms: engineering, aerospace, industrial optimisation, drones, robotics, edge inference.

The PhysicsX story is emblematic: if you can make hardware innovation move at software speed, you don’t just build better products - you reallocate billions of capex and change national competitiveness.

Which leads to a view I increasingly believe: Europe’s best shot isn’t winning “the model race” - it’s winning the deployment race in regulated + industrial markets, where trust, safety track record, and real-world integration matter more than hype cycles.

Energy isn’t a subplot - it’s the plot

If you want one line that explains the next decade: compute is becoming an energy problem. Big Tech’s AI capex is now being discussed in the hundreds of billions per year range – the big 4 destined to spend $700B this year up from $400B last. The direction matters more than the exact figure: we’re in an industrial buildout phase, and the grid, permitting, cooling, and supply chains are now strategic constraints. This isn’t just “tech”. It’s national infrastructure with corporate balance sheets.

Pricing models are breaking (and that’s good news if you’re brave)

Seat-based SaaS pricing was repeatedly called into question - not because software is dying, but because we’re about to consume far more of it through agents and automation. The direction of travel is clear: outcome-based pricing and value capture tied to measurable results. The reason it matters: if AI compresses marginal cost, the winners will be those who can price to value delivered - without making procurement feel like a maths exam.

Cyber, quantum, defence: the adult supervision is back

Cyber in the age of AI is exactly what you fear: the cost barrier to sophisticated attacks collapses, the volume of attacks rises, and “embedded security” stops being optional. Defence and drones weren’t treated as fringe topics either - they were presented as the leading edge of physical AI deployment under real constraints.

Quantum, meanwhile, felt refreshingly sober: progress is real, but utility is still about fidelity, scaling paths, and matching algorithms to architectures. The most investable insight wasn’t “quantum will change everything” - it was “this will create specialist capital and consolidation dynamics that don’t look like classic software VC.”

The most provocative paragraph (courtesy of 20VC energy in the room)

Harry Stebbings’ take was brutally simple: in modern venture, being right matters less than getting in. He unapologetically reframed venture as a market access game, not an “I’m cleverer than you” picking contest - selection is table stakes; securing allocation is the edge - and in 2025 that edge is built on relationship density, speed, and brand.

Harry has been saying for a while that VC now sits at the intersection of venture + media, and you could feel that in the room: distribution isn’t just what startups need, it’s what funds need too. He breaks “winning” down into a fairly ruthless operating system - sourcing, selecting, securing, servicing - and the subtext is that if you don’t build a machine (and increasingly, a tech-enabled one), you get outcompeted by those who do. Hearing Antler founder Magnus Grimeland talk about their tougher-than-Harvard machine of turning c.160,000 applications into c.500 investments in 54 programmes a year across 27 cities only reinforced the point: this era rewards systems.

The spicier bit - very on-brand 20VC - is the implication that classic Series A “good companies” are less interesting than multi-billion-dollar valuation hyper-scalers: if the best AI-native businesses can hit escape velocity faster than the old SaaS playbooks, then the return profile increasingly concentrates in the outliers and the breakouts, not the (relatively) steady grinders. That’s where the “triple-triple-double-double is dead” energy comes from: not as a meme, but as a signal that expectations have shifted toward exceptional growth (eg 0-£20m ARR in a year) and a clearer path to category ownership.

And then the properly provocative Europe take: he basically argues Europe’s opportunity is that the average is lower - not talent, but venture quality and ambition - which makes it an unusually good market to stand out if you’re relentless. The “Trump is the best CMO for Europe” line is extreme, but it’s a forcing function: Europe doesn’t need sympathy; it needs swagger and execution. Even the Revolut-style punchline was deliberately uncomfortable: culture isn’t the input, it’s the output - winning, learning fast, and compounding momentum creates culture, not the other way around. Whether you agree or not, it’s a very 2025 message: capital is abundant, competition is brutal, and storytelling + distribution + speed increasingly decide who gets to be in the round and who gets to be the winner.

If distribution is the moat and speed is the weapon, then the obvious next question is: what does a truly backable founder look like in this cycle?

Across the Antler + 20VC angles, the “most backable entrepreneur” profile came through pretty clearly – spike, drive and grit:

Truly investable founders tend to have a spike - a world-class edge in something specific (technical, commercial, product, domain, distribution). Not “well rounded”; unfairly strong in one area. Drive is not just enthusiasm, but relentless execution energy: the person who keeps shipping, keeps selling, keeps recruiting, keeps pushing when it’s messy and uncomfortable. And Grit is the ability to absorb hits, keep going, and outlast the market. Because in practice, it’s rarely the best product that wins; it’s the team that refuses to die long enough to find the compounding loop.

20VC’s spicier overlay on top of that was more psychological / pattern-based: founders who often show unusual formative experiences that correlate with edge - early entrepreneurial behaviour, comfort navigating uncertainty, and an ability to assimilate fast into new environments (his examples were things like moving a lot as a kid, an uncomfortable relationship with a parent, building “clans”/communities early, etc.). Whether or not you buy the causality, the point is consistent: the best founders are forged in volatility - and they perform when conditions are ambiguous.

Quick takeaways: the biggest themes (if you only read one bit)

• AI is becoming infrastructure, so governance, trust, and geopolitics move from “nice-to-have” to core strategy.

• Agents are coming, but reliability + evals + security are the gating factors before true autonomy.

• Distribution > model quality in the market: partnerships, workflow wedges, and narrative momentum are decisive.

• Energy is the hard constraint: the AI boom is an industrial buildout with grid-level consequences.

• Hardware is back: physical AI (robots, drones, industrial optimisation, edge compute) is where defensibility gets real.

• SaaS pricing resets: seat-based models erode; outcome-based/value-based models expand.

• Cyber + defence are front-page again: AI lowers the cost of attack and accelerates real-world deployment cycles.

• Europe can win - but it needs to play to its strengths (regulated markets, industrial depth, finance, healthcare) with louder ambition.

And finally… this was the view from the room

Across a couple of the sessions, Kahoot pulse checks added colour from the room, turning “panel opinions” into a real-time market signals. The room’s view of AI winners was telling: existing platforms led (37%), with hardware close behind (25%), while application-layer challengers still pulled a meaningful 23% - and frontier models themselves came last (15%). That’s a pretty clear statement that distribution and infrastructure beat model purity in most investors’ minds.

The macro read-through was equally interesting. A majority thought Mag 7 concentration falls over the next five years (57%), but the room also leaned to “not peak US tech dominance” (56%) - suggesting people expect broadening winners without necessarily expecting the US to lose its structural edge. And in Europe specifically, what excited people most was AI application-layer innovation (37%), but with a very notable second-place vote for robotics/hardware (25%) and semis/quantum/compute (23%) - i.e., Europe’s opportunity is being framed less as “another SaaS wave” and more as physical AI + sovereign compute.

Wrap up

Leaving this symposium, I’m more convinced than ever: we’re not heading into “an AI era.” We’re heading into a re-architecture of how work, companies, and national advantage are constructed - and the winners will be the ones who can marry software speed with real-world deployment under real constraints. The open question isn’t whether AI changes everything - it’s who captures the value: incumbents, AI-native startups, or customers DIY-ing it themselves.