Trillion-Dollar Build Faces Copper Gaps, China Pressure, and Agentic War

The Goldman Sachs $1 trillion projection frames AI infrastructure spending as a durable multi-year obligation, and the SoftBank $40 billion syndicated loan for its OpenAI stake puts that thesis to a live capital markets test. But two distinct challenges are accumulating beneath this consensus. The first is physical: copper, the essential conductor for data centre power and grid transmission infrastructure, faces a structural US supply shortfall that domestic projects like Rio Tinto's Resolution mine cannot bridge on any near-term timeline due to permitting and capital constraints. This creates a non-semiconductor chokepoint that is absent from most capex models — whoever wins the model race still depends on a commodity supply chain dominated by Chile and China.

The second challenge is analytical. Scaling sceptics argue that LLMs are approaching a data ceiling and that diminishing returns from additional compute spend will erode the economic justification for the current rate of data centre construction. EDA revenue data showing concentration in advanced-node AI chip design — at the expense of legacy tools — and NPU architecture analysis revealing data movement as the binding inference constraint both suggest that the efficiency frontier is shifting in ways that complicate straightforward extrapolation of compute demand. These two views — structural capex durability versus physical and architectural limits — are not yet reconciled in market pricing, and the gap between them is the most consequential unresolved question in AI investment strategy.

Three distinct but structurally related dynamics are converging around AI and geopolitical power. In the US-China dimension, Nvidia's Jensen Huang has gone public in opposing chip export controls that are progressively ceding the Chinese market to Huawei's Ascend platform, while Congress simultaneously advances sanctions targeting AI model output extraction by Chinese firms — a dual hardware-plus-software containment posture. China's response is visible in the market: DeepSeek-derived models are attracting global users through cost efficiency, and domestic companies like iQiyi are restructuring entire business models around AI, demonstrating that export controls have not prevented competitive deployment.

In Europe, Siemens and the German government are coordinating pressure on the EU to ease AI Act compliance burdens for industrial applications, with the explicit threat of redirecting capex to the US and China. This is not a startup lobbying for deregulation — it is a bellwether industrial conglomerate making a capital allocation statement that, if followed through, would compound Europe's existing disadvantage in foundation model development. Anthropic's Mythos cybersecurity model release completes the picture: labs are now using specialist model announcements as instruments of regulatory goodwill and government contract positioning, not merely commercial product strategy. Senior strategists should expect this pattern to accelerate across all major labs.

OpenAI's resource concentration into Codex — adding computer use, image generation, and persistent memory while discontinuing Sora and absorbing its science division — is the clearest strategic signal that agentic coding is now treated as an existential priority, not a product line. The Codex update directly targets Claude Code's market position, and the simultaneous departure of two senior executives suggests this consolidation is deliberate rather than opportunistic. Anthropic's Opus 4.7 release, with explicit improvements to complex software engineering tasks, and its broader vertical portfolio expansion reinforce that both leading labs view developer workflow ownership as the dominant enterprise AI interface battle for 2026.

The $50 billion valuation talks for Cursor sit at the intersection of this consolidation and the broader 12-month window anxiety for AI application startups. Investors pricing Cursor at that level are making one of two bets: either deep workflow integration and developer loyalty create moats that foundation model expansion cannot easily dissolve, or Cursor is an acquisition target for a platform player needing a coding assistant foothold before the market settles. Either framing underscores that whoever controls the developer's primary agentic coding interface controls a high-frequency, deeply embedded enterprise touchpoint — a position that the hyperscalers, foundation model labs, and venture-backed specialists are all racing to occupy simultaneously.