Middlegame Weekly

AI infrastructure entered a sharper allocation phase this week. Last week’s note framed the trade around scarce physical capacity — power that clears, memory that ships, metals that arrive, institutional channels that trust the supplier. This week put prices, contracts, and politics around those constraints.

The delta is visible in who gained bargaining power. Micron’s memory scarcity moved from data center abstraction into consumer hardware. Quanta, Powell, Eaton, Vertiv, Hubbell, GE Vernova, HD Hyundai Electric, and ABB showed that electrical infrastructure is becoming a margin pool, not background plumbing. Nuclear moved further from policy debate into procurement, balance sheets, restarts, fuel-cycle expansion, and site preparation. Copper, rare earths, uranium, and even sulfuric acid moved closer to the AI map because the data center is now a power project wrapped around compute.

The headline capex numbers remain enormous — Alphabet, Amazon, Microsoft, and Meta are pointing toward hundreds of billions of dollars of 2026 spending — but the market is learning to separate announcements from conversion. Investors rewarded hyperscalers that can turn capex into cloud growth and punished the ones where spend still reads like future margin pressure. That is the correct frame for this stage of the cycle. Capital is abundant. Deployable capacity is not.

The constraint stack is becoming investable

The cleanest signal this week was the migration of pricing power away from the obvious AI layer and into the machinery underneath it. Memory is the most immediate example. Micron’s reported April move, including a 53% surge tied to AI memory scarcity, was followed by fresh evidence that DRAM allocation is tightening across the broader hardware economy. Apple reportedly discontinued a low-end Mac Mini SKU because DRAM shortages tied to AI data center demand made the economics unattractive. By the end of the week, Micron was still flashing scarcity, with management saying supply was “nowhere close” to meeting demand as Micron gained 137% year to date.

That matters because memory is becoming an allocation market. HBM consumes more wafer capacity per bit than standard DDR5, and the big producers — Micron, Samsung, SK Hynix — control most of the global DRAM base. As capacity shifts toward AI, costs leak into PCs, phones, servers, and networking equipment. The memory bottleneck is not contained inside the rack.

The same pattern is showing up in grid and data center equipment. Quanta Services raised 2026 guidance after reporting a record $48.5 billion backlog. Powell Industries secured its largest-ever contract, more than $400 million for a greenfield data center representing “a couple of gigawatts.” Vertiv reported 83% adjusted EPS growth, with liquid cooling and power management continuing to read as one of the tightest operating layers in AI capacity. Hubbell’s earnings beat and raised price target told the same story from a different corner of the electrical stack.

The phrase to keep in mind is conversion capacity. The bottleneck is the ability to convert announced AI budgets into powered, cooled, connected, permitted, occupied data centers. That conversion layer is becoming the week’s best signal.

Power procurement hardened into strategy

Power remained the center of gravity, but the story broadened from nuclear excitement to a more specific architecture: firm generation, fuel security, grid cost allocation, and bridging assets.

Constellation’s Three Mile Island restart under a 20-year Microsoft deal set the tone early in the week. Brookfield and The Nuclear Company added another restart angle at VC Summer, while Oklo, NuScale, X-Energy, Centrus, Cameco, Vistra, and Constellation kept recurring across the briefings. Cameco’s Q1 profit increase, Centrus’s HALEU expansion at Oak Ridge, and Oklo’s Meta-linked pipeline showed that the nuclear trade is spreading across fuel, enrichment, developers, utilities, and power buyers.

The timing problem remains. Restarts may matter sooner than SMRs; SMRs may matter sooner in valuations than in electrons. But the demand signal is no longer speculative. Data centers already consume 4.4% of U.S. electricity, more than double the 2018 share, and projections for 2028 run as high as 12%. A planned Amazon/Anthropic campus in Indiana requiring power equivalent to more than one million homes is the kind of statistic that turns AI infrastructure into utility politics.

That is where the week changed. Oregon approved a separate billing class for Hillsboro data centers, with a 1-cent per kWh surcharge for grid repairs and local generation. Seattle’s earlier data center resistance no longer looks isolated. The public utility question is becoming explicit: who pays for the grid upgrades required to support private AI load?

Compute is fragmenting around bottlenecks

Compute stayed hot, but the useful read was fragmentation. NVIDIA remains the gravity well. Amazon’s reported order for one million NVIDIA GPUs by end-2027 confirms that accelerator demand still has years of procurement runway. NVIDIA’s own behavior is more revealing: the company has reportedly deployed more than $40 billion in equity investments this year, including capital into data center operators and optical supply.

This is supply-chain finance. Earlier in the week, NVIDIA’s Corning partnership put $300-500 million behind U.S. optical connectivity expansion. If the accelerator ecosystem is constrained by memory, optics, networking, power, or data center shells, the dominant chip company has an incentive to fund the bottlenecks around its own demand curve.

Foundry and equipment constraints added a second layer. TSMC’s AI demand strength, ASML’s equipment dominance, and Applied Materials’ breadth across wafer tools all reinforced the supply chain’s concentration risk. Arista’s warning of “horrendous” pricing and constraints across wafers, memory, CPUs, optics, and silicon, alongside purchase commitments rising from $6.8 billion to $8.9 billion, gave the networking layer its own scarcity signal. Meanwhile, Intel’s possible Apple foundry deal and AMD’s capture of 20,000-30,000 TSMC wafers from a weakening smartphone cycle suggest that capacity is being reallocated toward AI wherever it can move.

Geopolitics is accelerating the split. NVIDIA’s China accelerator share reportedly collapsed to “basically zero” from 66% in 2024 under escalating export controls and revenue-sharing mechanisms. That is a market share statistic, but also a systems-design prompt. The world is moving toward parallel compute ecosystems, and the cost will show up in duplicated capacity, policy risk, and procurement complexity.

Materials and permitting moved into the main frame

Copper had its clearest week yet as an AI infrastructure input. Data center construction spending reached $45 billion annualized, surpassing office construction for the first time. At the same time, the copper market shifted into a 150,000 metric ton deficit, with Chilean permitting timelines stretching to 12 years and average discovery-to-production timelines above 17 years. The copper wire and cable market is projected to grow from $182 billion in 2025 to $324.7 billion by 2032, driven by power transmission and grid modernization.

That is the materials version of the same conversion problem. AI demand can arrive in one budget cycle. Mines, smelters, transmission corridors, transformer plants, and nuclear facilities cannot. The longer the permitting tail, the more valuable existing capacity, brownfield expansion, and already-entitled assets become.

Rare earths entered the week with a cleaner corporate signal. MP Materials’ $500 million partnership with Apple pushes the company toward integrated magnet manufacturing and recycling in California. WisdomTree’s WDIG launch, with exposure to rare earths and strategic metals, is another sign that public-market products are being built around the thesis. The AI infrastructure trade is pulling capital into the old upstream base, but with a new vocabulary: resilience, domestic supply, magnet integration, recycling, strategic inventory.

Names and signals that gained weight

• Micron: the week’s clearest memory bottleneck signal, with AI demand pushing DRAM/HBM scarcity into broader hardware markets.
• Quanta, Powell, Eaton, Hubbell, Vertiv: the electrical conversion layer is becoming the earnings layer of AI infrastructure.
• GE Vernova, Constellation, Vistra, Cameco, Centrus, Oklo, NuScale: nuclear exposure is spreading across generation, fuel, enrichment, restarts, and SMR optionality.
• NVIDIA: acting as supply-chain financier, with investments into operators, optics, and adjacent bottlenecks.
• Arista: an important read-through on AI networking scarcity and component inflation through 2027.
• Intel and AMD: foundry and wafer allocation are becoming more fluid as AI absorbs capacity from weaker smartphone demand.
• Copper and rare earths: moving from supporting cast to strategic input, with permitting duration now part of the investment case.
• Oregon/Hillsboro and Seattle: local cost allocation and permitting politics are now direct AI infrastructure variables.

What to watch

1. Data center rate design: watch whether more utilities create separate customer classes, surcharges, demand charges, or direct grid-upgrade obligations for hyperscale loads.
2. Memory allocation: HBM supply, DRAM price moves, and consumer-device spillovers will show whether memory remains the most immediate constraint after GPUs.
3. Power delivery timing: restarts, gas-to-nuclear hybrids, SMR site readiness grants, and transformer lead times matter more than headline nuclear TAMs.
4. Network and optics pricing: Arista purchase commitments, Corning expansion, and NVIDIA-backed optical capacity are early tells on whether the AI fabric layer tightens further.
5. Copper and permitting: any movement in mine approvals, cable capacity, or high-voltage transformer orders will indicate whether the physical stack can absorb the capex wave.

The week’s message was practical. AI infrastructure is no longer waiting for proof of demand. The demand is here. The question is whether enough scarce inputs can be assembled quickly enough, power, memory, optics, cooling, copper, land, permits, and public consent. The companies that can allocate or unlock those inputs are becoming the new control points of the trade.