Middlegame Weekly

Context: Middlegame Weekly tracks the public signals behind the Agency Era thesis: power, compute, infrastructure, commodities, and capital allocation.

The week's read

AI infrastructure moved deeper into the power stack this week. The strongest signals were AEP planning around 63 GW of incremental load, VoltaGrid financing behind-the-meter power as a product, Cameco and Centrus turning nuclear fuel into a strategic choke point, and transformer lead times stretching into the next planning cycle. Compute demand remains real. The map changed because the scarce asset is increasingly the ability to assemble megawatts, fuel, cooling, optics, transformers, and contracts before capital gets trapped in the interconnection queue.

Signal vs. noise

Signal

• AEP now expects 63 GW of incremental load by 2030, with nearly 90% tied to data centers and hyperscalers, and lifted its five-year investment plan to $78 billion. Utility capex is becoming AI capex.
• VoltaGrid secured $1 billion from Blackstone and Halliburton against a 7.5 GW order book. Behind-the-meter power is becoming a financed product for customers that cannot wait on the grid.
• Centrus pointed to a $3.9 billion backlog through 2040, while Cameco locked in long-duration uranium agreements. Nuclear fuel and enrichment moved from policy detail to AI baseload constraint.
• U.S. transformer demand has risen 274% for generator step-up units and 116% for substations since 2019, with prices up 80% and lead times at four years. The electrical layer is now a deployment governor.

Noise

• Broad AI equity enthusiasm was less useful than the operating signals underneath it. Price targets and index forecasts did not change the thesis unless tied to signed contracts, capacity, or bottleneck relief.
• The weirder architecture stories, including home-node compute and orbital data centers, deserve attention as exception flags, but they are weak evidence until they convert into usable capacity at scale.

The five things that mattered

1. AEP turned utility planning into an AI infrastructure signal

American Electric Power gave the week its cleanest read. The company signed 7 GW of new projects in Q1 and now expects 63 GW of incremental load by 2030, with nearly 90% of that incremental contracted load tied to data centers and hyperscalers. Its five-year capital plan moved from $72 billion to $78 billion in three months.

That is the regulated-grid version of the AI capex cycle. Hyperscaler demand is large enough to change load forecasts, rate-base planning, transmission needs, and the politics of who pays for new capacity. Constellation's 460 MW Pin Oak peaking plant in Texas fits the same pattern. Power adequacy is becoming a state-level economic development issue.

Why it matters: Data-center demand is moving from corporate procurement into utility balance sheets.

Thesis implication: Utilities and grid modernization names deserve higher priority when load forecasts are backed by contracted projects, capital-plan revisions, and interconnection evidence.

2. VoltaGrid made behind-the-meter power look like a product category

VoltaGrid secured $1 billion from Blackstone and Halliburton to expand behind-the-meter systems for AI data centers, with a 7.5 GW order book through 2030 and a target of 300 MW of monthly capacity from its Texas manufacturing expansion.

The customer problem is simple. AI campuses need power faster than utilities can upgrade interconnection, transmission, and substations. A financed, modular power provider can sell speed, certainty, and grid-delay avoidance. Vistra's 20-year agreement to provide Meta with 2.1 GW of nuclear power points in the same direction. Customers are trying to reserve firm capacity for the duration of the AI buildout.

Why it matters: The scarce product is not only energy. It is delivered megawatts on a schedule customers can underwrite.

Thesis implication: Behind-the-meter power, dispatchable generation, and long-duration PPAs are becoming core Agency Era instruments.

3. Cameco and Centrus showed nuclear is a fuel-chain race

Nuclear stayed at the center of the week, but the better signal was upstream. Cameco's reported $80 billion strategic partnership with the U.S. government and Brookfield frames Westinghouse reactors as AI-era infrastructure assets. Cameco also locked in a long-term India supply agreement for nearly 22 million pounds of U3O8 from 2027 through 2035.

Centrus made the constraint explicit. Its $3.9 billion backlog runs through 2040, and the company remains strategically important because it operates the only U.S.-based HALEU centrifuge cascade. Advanced reactors need advanced fuel. Hyperscaler appetite does not solve that bottleneck by itself.

Why it matters: AI has turned nuclear from sentiment into procurement math. Fuel, enrichment, and component availability determine how much baseload ambition can become capacity.

Thesis implication: The nuclear watchlist should privilege fuel, enrichment, life extension, and component suppliers alongside reactor developers.

4. Four-year transformer lead times exposed the conversion bottleneck

The transformer data was blunt. U.S. demand for generator step-up and substation transformers has reportedly risen 274% and 116% since 2019, while prices are up 80% over five years and lead times have stretched to four years.

Four-year waits turn routine procurement into strategy. A hyperscaler can adjust chip mix, financing, or software architecture. It cannot improvise high-voltage equipment at campus scale. Carrier's 500% increase in data-center orders, Vertiv's thermal-management exposure, and ABB's high-density UPS investment all belong in the same conversion layer.

Networking joined the constraint set. Lumentum said demand has overwhelmed capacity, with orders visible through 2028 and a planned $100 million to $250 million Tokyo factory expansion. Arista reported $2.709 billion of Q1 revenue, up 35.1% year over year.

Why it matters: The conversion layer decides whether AI capex becomes usable capacity.

Thesis implication: Electrical equipment, thermal management, and AI networking deserve durable priority because they convert announced capex into operating throughput.

5. Compute scarcity became more contractual and more architectural

The compute signal did not weaken. It changed shape. Axe Compute signed a $260 million contract for a dedicated 2,304-GPU Nvidia B300 cluster over 36 months. Buyers are locking down dedicated capacity rather than waiting for a normal spot market.

At the same time, the architecture question broadened. AMD's agentic-AI argument, with GPU-to-CPU ratios potentially moving from roughly 8-to-1 toward 1-to-1, reframes CPUs as a larger part of the AI infrastructure stack. Apple's reported testing of Intel's 18A-P process and the ASML-Tata MOU for advanced lithography at India's first commercial 300mm fab point to the same response: customers and countries want options when supply is concentrated.

Why it matters: Compute buyers are moving from hope to reservation. The next scarcity set may include CPUs, memory, packaging, optics, and sovereign manufacturing capacity.

Thesis implication: Keep Nvidia central, but widen the compute map to contracted clusters, CPUs, foundry optionality, packaging, and alternative architectures.

Other headlines of note

• Cerebras raised $5.55 billion in the largest global IPO of 2026. The deal confirms that capital markets still want alternative AI compute capacity, but customer concentration and manufacturing dependence keep it secondary to the week's power-and-conversion read.
• Cowboy Space filed for up to 20,000 orbital data-center satellites. The plan remains speculative, but it is a useful exception flag for how binding terrestrial power, land, and permitting constraints have become.
• SPAN pitched 80,000 residential compute nodes and more than 1 GW of capacity by 2027. Treat the cost claims cautiously; the signal is that grid scarcity is pushing compute architects toward stranger topologies.
• U.S. prosecutors alleged $2.5 billion of advanced-chip server smuggling through Thailand. Export-control enforcement is becoming an operating constraint for OEMs, distributors, and AI buyers.
• Copper and lithium stayed in the constraint map. Copper deficits, lithium permitting fights, and domestic resource politics reinforce that AI infrastructure now competes for the same materials base as electrification.

Ticker implications

• AEP: Add - The 63 GW load forecast and $78 billion capital plan make regulated utility capex a direct AI infrastructure read-through.
• VST: Add - Vistra's dispatchable fleet, nuclear exposure, and Meta contract keep it near the center of the power-procurement thesis.
• CCJ: Add - Cameco is becoming a nuclear fuel control point as AI baseload demand meets long-duration uranium contracting.
• LEU: Add - Centrus owns one of the clearest HALEU bottlenecks in the domestic advanced-reactor stack.
• ETN: Watch - Transformer and electrical-equipment scarcity strengthen the thesis, but capacity investment and valuation need monitoring.
• VRT: Add - Liquid cooling and thermal management remain direct beneficiaries of high-density AI deployments.
• ANET: Add - AI networking is moving from supporting cast to scaling constraint as inference and east-west traffic intensify.
• AMD: Watch - Agentic inference could expand the CPU opportunity, but the evidence needs to move from narrative into deployment mix.

What to watch next week

1. Utility load updates and interconnection evidence: signed megawatts, queue positions, substation timing, and capital-plan revisions matter more than generic data-center demand commentary.
2. HALEU and uranium contracting: fuel availability will show whether nuclear AI enthusiasm is becoming executable procurement.
3. Transformer and UPS lead times: any change in delivery windows, pricing, or supplier capex will reveal whether the conversion layer is easing or tightening.
4. Cooling and networking orders: Vertiv, Carrier, ABB, Arista, and Lumentum are better reads on usable AI capacity than broad AI sentiment.
5. Dedicated compute contracts and CPU mix: watch whether more enterprises reserve clusters and whether agentic deployments actually change GPU-to-CPU ratios.

Bottom line

The operating lesson is that AI infrastructure has become an energy-and-conversion regime. Capital can still buy chips, land, and headlines. The advantage is shifting toward firms that can deliver fuel, megawatts, transformers, cooling, optics, networking, and long-duration contracts on real deployment timelines. That is where the Agency Era trade is becoming more industrial, more contractual, and more investable.