Arrow Forum 2026 · Germany · July 16, 2026
A ~40-minute channel keynote. The main-stage version of the argument published a few days later as the blog post How to build an AI strategy that survives the bubble pop.
This talk was not recorded. The summary below is reconstructed from the slides.
You can't predict what a worker looks like in 2031, but you can identify what's true across every plausible path there and build for that. The spine of the talk is a running list of things we know about the next five years, revisited and revised as each assumption gets stress-tested:
Two clocks. Capabilities: what AI can do, still climbing. Diffusion: how fast organizations absorb it, slow. The space between the curves is the diffusion gap. Nearly eight in ten companies report using gen AI, yet just as many report no significant bottom-line impact — the problem isn't capability, it's absorption.
Knowledge work is about 20% visible (emails, documents, meeting transcripts, chats) and 80% invisible (thinking, reasoning, judgment). Traditional IT only ever operated in the visible 20%; the invisible 80% was "not our problem." AI changes the mandate. IT after AI has to reach the invisible 80%, aiming to make 100% of knowledge work visible and supportable.
A futurist does not predict the future. A futurist works with probabilities and thinks about all possible futures. So take the assumptions built into each "thing we know" and stress-test them. Which ones hold across every scenario?
Two assumptions hide inside that line. First, technical progress will continue — probably, not guaranteed. Second, the most advanced models will be available — but only if the US government allows them, only if they still exist after a bubble pop, and only if the Chinese government allows them. Frontier access has become a policy variable, not a given, and the comforting line "even if the bubble pops, the technology still exists" is itself an assumption that may not hold for AI the way it held for railroads and dark fiber.
The one thing that survives every asterisk: open-weight models. GLM-5.2 is already released, MIT-licensed, and in the Opus 4.8 / GPT-5.5 class. So the assumption gets rewritten: Sonnet-class AI is real, and guaranteed to exist. That's the reliable floor to plan on — no government and no bubble can take it away. And it's enough, because most real enterprise AI ROI today comes from AI orchestrating administrative and work processes, which is exactly what Sonnet-class models are best at.
Diffusion has a mechanism for change: forward-deployed engineers. The whole industry is now pouring billions into them — Palantir (US commercial up 133% year over year, credited to the FDE model), OpenAI's Deployment Company ($4B+, 150 FDEs on day one), Anthropic and DXC (tens of thousands of FDEs), AWS ($1B for its own FDE org), and Microsoft's "Frontier Company" ($2.5B, 6,000 FDEs).
How do you get an FDE to close your diffusion gap? Option A: hire one of those firms. Option B: figure out what an FDE actually does, and do it yourself. What they do maps onto the 20/80/100 picture — they go into an organization and turn the invisible 80% into visible, encoded, usable knowledge. So the assumption gets rewritten: AI diffusion is important, and FDEs can speed it up.
Closing the gap starts with understanding how a worker's AI use evolves, with the context vault at the center from Stage 3 on:
At the pod stage, three worker types emerge: cognitive owners (context plus judgment, the source of expertise), cognitive operators (run the agent fleets), and cognitive curators (maintain the context vaults and skill libraries).
Per-worker daily token consumption climbs by stage: 100K at faster search, then 1M, 10M, 100M, 1B, and 10B at the self-running pod. Moving a workforce up the stages is a 10x-per-stage compute event, which is why token management becomes a first-class problem.
Today's EUC model assumes one person, one screen, one set of apps, one set of hours. AI changes every one of those. But most of EUC transitions over rather than disappearing:
The throughline: the words change (agent, cognitive, skills, context) but how and why we do the work doesn't. The core IT skills transfer, and the job gets bigger.
Workers still need to work. Work still needs to happen somewhere. Somebody has to make that somewhere work — safely, observably, cost-effectively. That somebody is EUC/IT, and the job is bigger than it has ever been.
What next? Make the jump to Stage 3. Build your own brain. "You have to feel it before you can govern it. You have to have one before you can manage thousands."
Page one is building your own second brain. Here's the starter prompt:
Build your own AI second brain (GitHub gist)