“The AI factory is no longer a place. It is becoming a network.”
The most consequential development this morning is not another increase in frontier-model size. It is where AI is beginning to run.
Anthropic has signed its first Australian data-centre agreement at a campus planned for 2.16 GW of capacity near Brisbane, with first operations expected from 2027 — a facility intended for AI inference rather than model training. At almost the same moment, MediaTek has launched a 2-nanometre smartphone processor designed explicitly for always-on Agentic AI, capable of supporting models up to 30 billion parameters on-device. And European inference-chip company Axelera AI says its processors are moving into Dell and Supermicro systems and EU AI Factories, with more than 600 customers already using its technology.
Taken separately, these are product announcements. Taken together, they indicate something more important: training remains concentrated, inference is beginning to diffuse.
Where Intelligence Runs
training core, inference periphery
-
Anthropic has secured its first Australian data-centre agreement at a proposed campus roughly 250 kilometres from Brisbane, with planned capacity of 2.16 GW. The facility is expected to begin coming online in 2027, designed to use renewable electricity and closed-loop air cooling. Reuters reports the site is intended for inference rather than training; regulatory approval is still required. Anthropic already has an Australian policy relationship, having signed a March agreement with the government covering AI safety research and information sharing.
The distinction between training and inference is becoming economically structural. Training favors enormous, concentrated clusters; inference derives value from running existing models repeatedly for millions of businesses and billions of users. The scarce resource is no longer simply compute capacity — it is deployable compute in the right geography. signal: This is not the end of hyperscale concentration; it is the beginning of a second infrastructure layer beneath it.
-
MediaTek's new Dimensity 9600 Pro is built on a 2-nanometre process with an architecture optimized for Agentic AI running locally on smartphones. Its NPU delivers 51% higher LLM prefill performance, 55% more token generation per watt, and a 40% reduction in power consumption for always-on AI versus the previous generation. It supports models of up to 30 billion parameters, with first devices expected this quarter.
The next compute architecture is unlikely to be cloud versus edge. It will be cloud + regional inference + device — the cloud creating and coordinating intelligence, regional infrastructure scaling deployment, devices increasingly executing it. That architecture could ultimately generate far more inference than training. signal: The smartphone becomes sensor array, personal data store, inference engine and AI agent at once.
-
AI coding-agent company Factory raised $200 million at a $5 billion valuation, roughly tripling its valuation. Its objective is enabling enterprises to delegate increasingly long-duration software-development tasks to autonomous agents. The unit of automation changes from the task to the workflow.
Software engineering matters because software is a production input across virtually every modern industry. If agents reduce the marginal cost of software development, they indirectly reduce the cost of automation everywhere else. Still, there is not yet enough audited productivity evidence to declare a broad macroeconomic inflection — that baseline remains unchanged. signal: Investors increasingly see Agentic AI as an enterprise-production layer, not merely a user-interface improvement.
Capital Begins Following Conversion
the Great Split
Who Pays for AI Infrastructure?
energy, grids and geography
The Assurance Economy Acquires Institutions
who controls the brake
Europe: No Policy Inflection, but a Commercial Signal
Europe 2030
The Inference Localization Law — NERC Hypothesis 07
three geographic layers
Training Core
Gigawatt-scale AI factories creating frontier intelligence, gravitating toward locations offering extraordinary power and infrastructure scale.
Regional Inference Grid
Distributed data centres executing models near economies and users, valuing proximity, data jurisdiction, latency, political alignment and energy availability.
Intelligent Edge
Vehicles, robots, smartphones and industrial devices running AI locally — where the majority of long-run AI tokens may eventually be generated.
who wins the next chapter
- The hidden pattern extends the chain: BUILD → CONVERT → PROTECT → ASSURE → ABSORB → DISTRIBUTE → COMPOUND.
- Realized AI Value = Model Capability × Inference Density × Proximity × Energy Efficiency × Assurance.
- A new Great Split emerges: Intelligence Producers vs. Intelligence Distributors — semiconductor firms, regional data-centre operators, telecom networks, sovereign-cloud providers, edge-AI platforms, industrial automation, mobility systems and device manufacturers.
- For Europe, this is potentially good news: it need not win every frontier race to build a powerful position in inference silicon, industrial AI, telecom infrastructure, sovereign cloud, mobility, robotics and edge deployment.
- For investors, metrics like tokens per watt, tokens per euro, utilization, latency and inference revenue become increasingly important.
- For policymakers, sovereign AI is not only sovereign frontier models — it is the ability to operate critical inference locally, under domestic rules, on resilient infrastructure.