The Gist: Compute & Infrastructure Brief

Thursday, March 05, 2026

Top Line

• Nvidia signals strategic shift on AI investments — CEO Jensen Huang ruled out further large-scale investments in OpenAI and Anthropic, while competitor Broadcom projects $100B in AI chip sales by 2027, intensifying competition for Nvidia's dominance. Bloomberg, Bloomberg

• White House data centre pledge signed but lacks enforcement teeth — Seven tech giants committed to building dedicated power infrastructure for AI data centres, but energy experts warn the voluntary agreement won't prevent electricity rate increases for consumers. Politico, The Verge

• Physical infrastructure becomes warfare target — Three Amazon data centres suffered damage in drone strikes during Iran conflict, exposing vulnerability of concentrated compute infrastructure and forcing strategic rethink on facility locations. Bloomberg

• TSMC capacity allocation shifts as export controls bite — Nvidia is redirecting TSMC production away from H200 chips intended for China toward its latest Vera Rubin products, revealing how geopolitical restrictions are reshaping semiconductor manufacturing priorities. Financial Times

• Meta announces plans for custom AI training chips — Despite recent deals with major chipmakers, Meta is developing proprietary processors for training future AI models, joining the ranks of hyperscalers seeking to reduce dependence on Nvidia. Bloomberg

Key Developments

Nvidia's Strategic Retrenchment and Rising Competition

Nvidia CEO Jensen Huang explicitly ruled out a $100 billion investment in OpenAI, contradicting earlier reports about the chipmaker's maximum pledge to the AI startup. Huang said Wednesday that Nvidia's investments in both OpenAI and Anthropic "will likely be its last," though his explanation left key questions unanswered about the strategic shift. TechCrunch, Bloomberg

This retrenchment comes as Broadcom CEO Hock Tan projects his company's AI chip sales will exceed $100 billion in 2027, marking significant encroachment into Nvidia's territory. Broadcom's forecast signals intensifying competition in custom AI silicon, particularly for hyperscale customers seeking alternatives to Nvidia's standard offerings. Meta's announcement that it plans to develop custom processors for training future AI models reinforces this trend toward vertical integration. Bloomberg, Bloomberg

Why it matters: Nvidia's dominance in AI compute — and its 80%+ gross margins — faces its first credible threat from both established semiconductor firms and customer-developed alternatives, potentially reshaping pricing power across the supply chain.

What to watch: Whether Broadcom can deliver on its $100B projection, and how quickly Meta's custom training chips reach production scale.

Data Centre Energy Pledge: PR Exercise or Infrastructure Solution?

Seven tech giants — Google, Meta, Microsoft, Oracle, OpenAI, Amazon, and xAI — signed a White House "rate payer protection pledge" committing to build dedicated power infrastructure for their AI data centres. President Trump framed the voluntary agreement as shielding consumers from electricity rate increases driven by AI infrastructure buildout. The Verge, Politico

Energy market experts immediately challenged the pledge's effectiveness. The agreement contains no enforcement mechanisms or penalties for non-compliance. Grid interconnection requirements mean even dedicated generation facilities will impact regional electricity markets. Critics note that companies were already planning significant power investments regardless of the pledge. As Trump acknowledged at the signing event: "Data centers... they need some PR help." Wired, Politico

Why it matters: The pledge acknowledges bipartisan political pressure on AI infrastructure expansion but offers no binding commitments, leaving actual rate impacts uncertain as data centre power demands accelerate.

What to watch: State-level regulatory responses and whether utilities begin imposing stricter interconnection requirements or cost allocation mechanisms.

Geopolitical Supply Chain Realignment

Nvidia is redirecting TSMC production capacity away from H200 chips originally intended for the Chinese market toward its latest Vera Rubin products, according to sources. US export controls have effectively stalled China sales, forcing reallocation of advanced node capacity. This shift reveals how quickly geopolitical restrictions can reshape semiconductor manufacturing priorities at the world's leading foundry. Financial Times

The supply chain vulnerability extends beyond production allocation. A bipartisan group of six US senators pressed Intel for information about its relationship with ACM Research, whose subsidiaries remain on a Commerce Department blacklist for national security reasons. The inquiry highlights ongoing concerns about semiconductor equipment supply chains even within ostensibly domestic manufacturing. Bloomberg

Meanwhile, tech industry groups representing Google and Apple are urging Trump to reverse the designation of Anthropic as a national security risk, warning of "detrimental ripple effects" across the industry. The designation threatens to complicate supply chains and partnerships for AI infrastructure. Bloomberg

Why it matters: Export controls and security designations are actively reshaping semiconductor allocation and AI supply chains, creating winners and losers independent of market fundamentals.

What to watch: Whether TSMC can maintain production efficiency while rapidly rebalancing customer allocations, and the scope of potential semiconductor equipment restrictions.

Physical Infrastructure as Warfare Target

Three Amazon data centres suffered damage in drone strikes during the ongoing Iran conflict, marking the first significant attacks on civilian cloud infrastructure during modern warfare. Analysts warn that data centre concentrations increasingly represent strategic targets, forcing reconsideration of facility locations and hardening requirements. Bloomberg

The vulnerability extends beyond physical attacks. India's security agencies have requested private space startups develop "bodyguard satellites" to protect space-based infrastructure assets, reflecting growing concerns about orbital compute and communications infrastructure. The development suggests sovereign nations are preparing for potential attacks on satellite networks that support AI training and inference. Bloomberg

Why it matters: The assumption that data centres represent protected civilian infrastructure is collapsing, potentially driving geographic diversification costs and insurance premiums sharply higher.

What to watch: Whether hyperscalers announce facility hardening investments or geographic rebalancing away from concentrated regions, and insurance market responses.

Signals & Trends

Memory storage crisis hits consumer devices amid infrastructure boom. Nintendo Switch 2 users face storage capacity constraints as memory prices surge, driven by competing demand from AI infrastructure. The consumer electronics squeeze suggests data centre buildout is creating cascading effects across the memory supply chain, potentially forcing prioritisation decisions between infrastructure and consumer applications. Bloomberg

European sovereign compute ambitions accelerate. Deep tech startups focused on AI, fusion, and space technologies are attracting significant capital in Europe's push for "digital independence" from US infrastructure dependencies. The funding shift, documented in the Financial Times' European startup hubs analysis, suggests European governments are backing indigenous compute capacity rather than relying on hyperscaler buildout. Germany's defence tech ecosystem centred in Munich demonstrates how sovereign infrastructure investments are creating regional compute clusters. Financial Times, Financial Times

Quantum computing enters public markets via SPAC. Pasqal Holding agreed to merge with a blank-check firm at a $2 billion pre-money valuation, joining other quantum startups choosing public listings. While far from commercial viability, the SPAC route suggests investors are positioning quantum computing as the next infrastructure buildout wave after AI, despite unproven economics and unclear timelines to practical applications. Bloomberg