~/vision · design constraints, not afterthoughts
A human-centric way to build large-scale technology
I believe we can build the most advanced AI systems in the world — and still get this right for people, for the planet, and for the parts of the world without a seat at the table. That means treating five commitments as design constraints, not afterthoughts.
Everything I write comes back to a single claim: the constraints below are not a tax on ambition. They are the design brief. Treated as afterthoughts they become expensive retrofits; treated as constraints from the start they produce better systems than the unconstrained version would have been.
Each commitment has its own page, because each one should be arguable and citable on its own terms rather than only as part of a manifesto.
Built for, and by, humans
Systems judged by the real action they take and the value delivered to a person, not by leaderboard position or parameter count.
Sustainable by design
We have one planet. Data-centre electricity demand is heading toward 950–1,300 TWh by the early 2030s — efficiency has to be an architectural decision, not a caveat.
Grows the next generation
Over a million additional semiconductor workers are needed globally by 2030. Progress that doesn't grow the talent underneath it is borrowing against its own future.
Doesn't quietly raise the cost of living
US electricity prices rose 6.9% in 2025 — more than double headline inflation. Compute demand that outruns the grid is a cost every household ends up paying.
Closes gaps, doesn't widen them
Open standards — RISC-V, OpenTitan, Caliptra, OCP — exist at every layer of the stack precisely so that capability isn't gated behind the few companies who can out-spend everyone else.
Aligned to European sentiment
Sovereignty, open standards and honest, sourced argument over assertion — a European way of building the same advanced technology, not a smaller version of it.
The framework behind the argument
Three imperatives the industry is not clearly winning, and the four fault lines in chip and system design that follow from them. This is the spine of the ESSERC 2026 keynote and of the journal series.
The industry is not clearly winning any of them.
- resources: Energy, compute, capital and materials — demand is outrunning grids, fabs and packaging lines.
- talent: The people who design and operate the systems beneath every model — quietly short of supply.
- progress: Solving bigger problems without assuming "bigger model" is the only path there.
What current chip and system design hasn't fully closed.
- resources: Scale and the power wall — chiplet, package, interconnect and cooling as one coupled system.
- progress: General intelligence versus intelligence-per-watt — there is still no adopted metric for it.
- trust: Reliability and privacy have to run silicon to system, not get bolted on at either end.
- value: Context is AI's real currency — priced by hardware that still treats it as an afterthought.