The $725B Lockout: How Hyperscaler Capex Is Splitting the AI Stack Into Two Tiers

The Core Tension

The most important number in AI infrastructure right now is not a model benchmark. It is $725 billion. That is the combined capital expenditure the four largest US hyperscalers — Amazon, Microsoft, Alphabet, and Meta — are committing to AI infrastructure in 2026. Combined capex from the top four hyperscalers is projected to reach $725 billion in 2026, representing a 77% increase over the $410 billion recorded in 2025. When you include Oracle, these five companies alone plan to spend roughly $660–690 billion on infrastructure in 2026, the vast majority directed at AI compute, data centers, and networking. Subsequent earnings revisions pushed the consensus to $725 billion.

This level of concentrated spending colonizes supply chains. Companies outside this group cannot compete on allocation. They pay a rising floor price for everything that remains.

What $725B Actually Buys — and Who It Locks Out

Roughly 75% of the aggregate expenditure is targeted directly at generative AI workloads, GPU-driven clusters, fiber optics, and advanced power management solutions. That concentration matters because semiconductor supply is not elastic on a two-year horizon. TSMC cannot double advanced node throughput by next quarter.

The four hyperscalers are now projected to increase capital expenditures by more than 60% from the historic levels reached in 2025, as they load up on high-priced chips, build new facilities, and buy the networking technology to connect it all. Meta's own earnings release was explicit about the mechanism: Meta cited "expectations for higher component pricing this year" as a primary driver of its increased forecast. Supply is tight, prices are rising, and the hyperscalers are absorbing the increase.

For mid-tier AI companies, inference-focused startups, and regional cloud providers, the situation is tighter. HBM4—the memory architecture required for frontier training and high-throughput inference—is reported to be fully allocated under multiyear contracts with SK Hynix, Samsung, and Micron. New entrants cannot solve this with money alone; the slots do not exist. Mature-node foundry pricing is rising 5–15% quarter over quarter, compressing margins for anyone doing inference at scale who cannot pass those costs upstream. All the hyperscalers report that their markets are supply-constrained, rather than demand-constrained. The constraint flows downhill.

Blackstone's Japan Bet Is a Cost-Floor Play

Blackstone plans to invest roughly $30 billion in artificial-intelligence data centers in Japan over the next three to five years, as its president and COO Jonathan Gray told Nikkei. To date, Blackstone has developed data centers in Japan with a total capacity of over 500 megawatts and is evaluating new facilities exceeding one gigawatt.

Blackstone is not betting that Japan displaces US capacity. It is betting that offshore capex in a market with lower labor costs and available land lowers the global cost floor enough to matter on returns. Gray told Nikkei the main risk is not a speculative bubble in the sector, but a future shortage of computing capacity. That framing aligns with the hyperscaler view: undersupply is the existential threat, not oversupply.

Blackstone's presence in Japan is already established via AirTrunk: in March 2026, a record ¥191.6 billion (~$1.24 billion) green loan was secured to expand its Tokyo AI data center campus. The $30 billion commitment accelerates an existing position rather than a greenfield bet—a distinction that matters for execution risk.

If Blackstone demonstrates that Japanese capacity delivers comparable compute at lower total cost, it creates pressure on Western hyperscalers to diversify. Japanese capex does not threaten US dominance in the near term, but it begins to erode pricing power in the supply chain.

Etched: Architecture Arbitrage, Not Yet Proven

Etched came out of stealth announcing a working chip and over $1 billion in signed customer contracts, along with $800 million raised across multiple financing rounds, the latest being $500 million at a $5 billion post-money valuation.

The thesis is specific. Etched built its Sohu chip as a custom ASIC designed to run transformer model inference as fast as physically possible. Etched plans to produce its inference chips using TSMC's N4P process, an enhanced version of the five-nanometer node that provides 11% better performance than the original. The company claims a single 8-chip Sohu server processes around 500,000 tokens per second running Meta's Llama 70B model—a figure that, if it holds in production, represents meaningful cost-per-token improvement over H100 clusters.

The most recent investment, which closed in December, valued Etched at $5 billion and included participation from VentureTech Alliance, a startup fund associated with TSMC. A TSMC-linked fund backing a TSMC-fabbed chip is a signal about production access, not just capital. That relationship reduces, but does not eliminate, foundry risk.

Etched has not named its customers, disclosed contract terms, or explained when those agreements turn into revenue. The company needs to ship working racks by summer 2026, hit the performance benchmarks it has been advertising, and convince customers that building their inference stack around a single-purpose chip from a startup is worth the switching cost. The $5 billion valuation reflects desperation for alternatives, not demonstrated production economics. The $1 billion in contracts is real commercial interest, but not yet shipped systems.

Training large models once dominated hardware spending, but as companies deploy chatbots and recommendation engines, inference workloads now account for the majority of compute costs. Nvidia's general-purpose GPUs handle both training and inference, but specialized inference chips can slash energy use and latency. That is the market Etched targets. The window exists. Whether Etched executes before hyperscalers lock the next procurement cycle is open.

The Second-Order Effect Nobody Is Pricing

If inference costs remain structurally high due to HBM scarcity and mature-node price inflation, fewer AI applications reach commercial viability outside hyperscaler use cases. The $725 billion buildout simultaneously creates infrastructure for AI at scale and concentrates those economics in five balance sheets. In roughly 18 months, the aggregate annual AI infrastructure commitment has increased from approximately $380 billion in 2025 to a projected $660–690 billion in 2026, driven by a shared conviction that AI workloads will consume every available unit of compute capacity.

That conviction may be right. Or it may produce a 2027 where mid-tier operators are squeezed between rising inference costs and customer price sensitivity, forcing consolidation that further concentrates the market.

What to Watch

1. Mature-node foundry pricing, Q3 2026: Whether the 5–15% quarterly price increases continue or demand softens is the leading indicator for inference margin pressure across the industry. Watch TSMC's Q3 earnings commentary on N3 and N4P utilization rates.

2. Etched's first rack shipments, summer 2026: Named customers, independent benchmarks, and yield data from initial deployments will determine whether this is a real architecture wedge or a well-funded lab result.

3. Second-tier hyperscaler moves: Whether ByteDance, Alibaba, or Mistral pursue independent foundry partnerships to escape allocation pressure on HBM and advanced packaging would confirm that the two-tier market is forcing alternative procurement strategies. Any announced direct foundry deal from this cohort signals structural divergence.

4. Blackstone Japan pace and cost structure: Track AirTrunk's Tokyo campus utilization and whether Blackstone's stated $30 billion deployment accelerates or slips. The cadence reveals whether Japanese costs are genuinely competitive or balance-sheet positioning ahead of a US capacity crunch.