AI Infrastructure Onshores, Governance Fractures, Agents Go Operational

Three developments this week converge on the same structural reality: the geography of AI infrastructure is being actively contested. SK Hynix's Indiana groundbreaking addresses the most exposed supply-chain chokepoint — HBM packaging — by bringing it into US territory ahead of NVIDIA's 2028 platform. Microsoft's A$25 billion Australia commitment is explicitly framed around data sovereignty and is as much a government procurement strategy as a capacity investment. And a UK survey finding that one in five British firms have already offshored AI workloads due to energy costs reveals the flip side: governments that fail to address industrial power pricing actively lose existing AI activity, not just future investment.

The MATCH Act's proposed extension of export controls to DUV lithography equipment adds a further layer, using supply-chain leverage to constrain where advanced chip capacity can be built globally. Taken together, these signals confirm that industrial energy policy, domestic content requirements, data localisation rules, and semiconductor export controls are now the primary axes on which AI infrastructure geography is being determined. The companies and governments that understand this are making multi-billion-dollar commitments accordingly; those that do not risk ceding both the economic and strategic returns from AI adoption to lower-cost or better-positioned jurisdictions.

Google's TPU 8 announcement is the clearest single demonstration this week that full-stack vertical integration is not only technically viable but is now executing at production scale. By deploying separate training and inference silicon, replacing Intel's x86 with in-house Axion Arm cores, and simultaneously embedding Gemini across Workspace and Chrome while signing frontier labs like Thinking Machines onto GCP, Google has eliminated external dependencies across silicon, server architecture, cloud infrastructure, and enterprise software distribution in a single product cycle. This is the architecture Amazon is pursuing with Trainium, Inferentia, and Graviton, and that Microsoft is building toward with Maia and Cobalt — but Google's execution this week is the most complete public demonstration of the model.

SpaceX's reported $60 billion Cursor acquisition extends the same logic into a non-traditional vertical: a space and infrastructure company acquiring the dominant developer coding interface because controlling that layer determines where AI inference runs and which models enterprises adopt. Combined with SpaceX's separately disclosed ambition to develop in-house GPUs, the pattern mirrors the hyperscaler playbook applied to an industrial conglomerate. For NVIDIA, Intel, and independent software vendors, the strategic implication is consistent: the addressable market for merchant silicon and middleware is shrinking as integrated incumbents build inward, with inference — the highest-volume, highest-margin workload — as the primary battleground.

The Anthropic Mythos situation crystallises a governance failure that is systemic rather than isolated. A model capable of finding 271 real vulnerabilities in production software — independently validated by Mozilla — simultaneously leaked to unauthorised parties through a private forum and was withheld from CISA, the agency with the clearest mandate to oversee exactly this capability class. The access controls governing this deployment were designed for early-adopter software previews, not for dual-use tools operating at demonstrated offensive-security capability levels. This is not a one-off: as labs accelerate deployment of specialised models in security, biology, and enterprise workflows, the gap between model power and governance rigour is widening across the industry.

Meanwhile, OpenAI's workspace agent launch and Meta's employee-instrumentation initiative for agent training data illustrate that agentic deployment is now the active product battleground, not a future roadmap item. Meta's approach — generating behavioural telemetry from employee computer activity as imitation-learning data — represents a distinct competitive moat strategy: proprietary workflow data rather than benchmark performance. For enterprises, this creates an immediate governance question that most organisations are not yet asking: the AI tools deployed today are simultaneously generating training data for the next generation of agents, and understanding what data flows to which lab is a material strategic and contractual concern.