The 2030 Consensus: Compute Metrics, Energy, and Shifting AI Timelines

Key takeaways

 

• The AGI conversation has shifted from philosophy to hard limits: compute, memory bandwidth, and the power grid
• World models need continuous, multi-modal processing — a different bottleneck from text-only LLMs (VRAM and latency, not just raw FLOPs)
• Energy constraints are pushing workloads to the decentralised, off-peak edge
• For business: build "generational" structured assets that survive the next four years of change

 

The Mundane: the grid feels it first

The hum of a substation on the edge of Milton Keynes is louder than it used to be. The abstract race toward more capable AI lands, eventually, as a very physical question: can the local grid carry it? Heavy copper, winter demand surges, a datacentre drawing megawatts — the future of intelligence is colliding with the wiring of the present.

The Machine: past FLOPs, into bandwidth and watts

Judging an AI by raw FLOPs no longer captures the picture. Simulating physical reality in a continuous latent space means processing high-definition video, depth and kinetic vectors at once — which shifts the bottleneck from processor speed to ultra-high-bandwidth memory (VRAM) and low-latency edge clustering. Running a persistent, real-time physics simulation of the built environment is power-hungry, and centralised infrastructure is bumping into national grid limits. That friction is exactly what accelerates the move to decentralised collectives: spread the workload across local networks and tap off-peak power instead of spiking one region.

Forecasts about when this converges (often pointed at around 2030) rest on compute-scaling curves and energy availability rather than certainty. Treat any specific threshold — for example a "10^26 FLOPs" figure — as a forecast, not a fact.

The Digital Eye: a region as a balancing act

To the old grid, demand is something to react to. To a regional digital twin it's a forecast: continuous simulation of weather fronts, industrial load and solar-panel occlusion across Northamptonshire and Milton Keynes, projecting need hours ahead, signalling batteries to store off-peak, and — when soot build-up drops solar output — flagging a cleaning job to restore yield. The same physics that predicts the grid predicts the maintenance.

The Mindset: build generational, not patchwork

The old game was reaction — fix the fault after it fails, re-plan month to month. In a world running on continuous prediction, value moves to generational systems architecture: assets — property, workflows, service networks — built within clear long-term parameters that adapt automatically. And the moat is the same one this series keeps naming: as automation commoditises generic information, proprietary value belongs to whoever captures and structures localised physical data. Document your degradation metrics and operational edge cases in machine-readable form now, and you stay an essential node in the infrastructure rather than a replaceable one.

Try this, this week

Audit your operational timeline: name one dependency in your business that is vulnerable to a fast technology shift over the next three years (a platform, a skill, a single supplier). Write down what a "generational" version of that dependency would look like — owned, structured, portable. You don't fix it today; you just stop being blind to it.

Common questions

Is 2030 a real AGI deadline?

No — it's a forecast horizon drawn from compute-scaling and energy trends, not a fixed date. Treat timeline figures as projections, not promises.

Why does energy matter to AI now?

Continuous, real-time world-model processing is power-intensive, and centralised datacentres are hitting grid limits — which is pushing workloads toward decentralised, off-peak, edge computing.