AI · Quantum · Infrastructure
AGI Needs New Infrastructure, Not Just Bigger Models
Every few months, the industry announces a larger model and calls it progress toward general intelligence. Parameters grow, benchmarks tick upward, and the underlying machine stays the same: classical accelerators, brute-force statistics, energy budgets that double like clockwork. This essay argues the uncomfortable version of events — the binding constraint on AGI is the computing substrate, not model size — and sketches what infrastructure-first thinking looks like in practice.
The cost curve is the argument
Reasoning is expensive. Every chain-of-thought, every agentic loop, every verification pass multiplies inference cost. If intelligence means sustained, reliable reasoning over long horizons — planning a supply chain, simulating a mission, modelling a molecule — then today's economics break before today's architectures do. A system that can reason but cannot afford to is not general intelligence; it is a demonstration.
This is why I spend my time on infrastructure. At AGI Quantum Cloud we work from a simple premise: the winners in machine intelligence will be decided as much by what intelligence runs on as by how it is trained.
What "new infrastructure" actually means
Three layers, none of them science fiction, all of them under-built:
- Hybrid compute orchestration. Classical CPUs and GPUs for what they do best, quantum processors and simulators for the workloads with genuine quantum structure — optimisation, sampling, certain simulations — scheduled as one fabric, not three silos.
- Reasoning-native runtimes. Agents need memory, tools, verification and rollback as primitives — the way databases once needed transactions. Most "agent stacks" today are scaffolding taped onto chat endpoints.
- Simulation as a first-class workload. Trajectory planning, molecular modelling, logistics digital twins: the problems where machine intelligence earns its keep are simulation problems. Infrastructure should treat them that way.
Where quantum computing honestly fits
Quantum computing is not a faster GPU. It is a different kind of machine for a different kind of question.
Candour matters here, because the field is full of hype I refuse to repeat. Quantum hardware today is early, noisy and narrow. Its honest near-term role is specific: certain optimisation and simulation workloads where quantum algorithms have structural advantage, accessed through hybrid pipelines where classical systems do everything else. My position is that founders should build the hybrid muscle now — simulators, algorithms literacy, orchestration patterns — so that as hardware matures, the software is already waiting. That is the work, unglamorous and necessary.
The economics test
Any infrastructure claim should survive one question: does this make a unit of reasoning cheaper, more reliable, or newly possible? Bigger models improve capability while worsening cost. Infrastructure plays improve the ratio. History favours the ratio — every computing revolution that mattered was, at bottom, an economics revolution: mainframes to PCs, servers to cloud, and now, perhaps, classical-only to hybrid.
What I'm building toward
A mission-engineering platform where an engineer describes an objective — a trajectory, a molecule, a logistics network — and hybrid infrastructure returns an answer that classical-only systems could not reach affordably. We are early. The platform is a concept under active development, not a finished product, and I state that plainly. But the direction is set: intelligence follows infrastructure, and infrastructure is what I build.