The Investor

What Happened

The Trump administration released its FY2027 budget proposing $1.5 trillion for defense — a 42% increase over the current Pentagon allotment and the largest single-year military spending increase since World War II. The budget includes specific line items for munitions, ships, an Iron Dome-inspired missile defense system, and higher military salaries. Critically, it also explicitly redirects $15 billion from renewable energy and clean air programs to fossil fuels and AI supercomputers.

This lands during active US military conflict with Iran — two American aircraft were shot down this week — giving the proposal a wartime urgency that historically overrides Congressional budget opposition.

Why This Is Different

The last time Congress saw a Trump budget, they approved the military boost but rebuffed most domestic cuts. The directional signal for defense spending is high-conviction regardless of final Congressional math. But the $15B AI supercomputer redirect is new: it explicitly declares AI compute as a national security asset, creating a government-backed demand floor for AI infrastructure that didn't exist 48 hours ago.

AI compute is now co-equal with missile defense as a national security priority — funded by the same budget, under the same wartime logic.

This matters because it arrives at the exact moment private AI infrastructure is stalling. With ~50% of data center builds facing delay or cancellation (detailed in the next brief), the government just became the marginal buyer of AI infrastructure. Companies positioned at the intersection of defense contracts and AI compute — sovereign AI platforms, classified-capable compute, defense-grade AI systems — are operating in a TAM that expanded by government fiat.

Sector Impact Matrix

Critical caveat: domestic cuts are negotiating positions. The defense and AI supercomputer numbers are the structural signal.

Cross-Source Validation

This budget arrives alongside macro data showing Nasdaq -5.86% YTD, S&P -3.84%, and Bitcoin -23.53%. The labor market added 178K jobs in March but February was revised down an additional 41K, and unemployment declined partly due to labor force shrinkage. Healthcare leading job gains is a defensive rotation signal, not growth. In this environment, government-backed demand visibility is exceptionally valuable. Defense tech startups in autonomous systems, electronic warfare, and AI-enabled defense are entering a golden procurement window.

The New Data

We flagged grid delivery constraints last week via the solid-state transformer thesis. Today's intelligence puts a devastating number on it: approximately 50% of all planned US data center builds in 2026 face delay or cancellation. The bottleneck is electrical equipment — specifically high-power transformers, switchgear, and batteries. Lead times have stretched from ~2 years pre-2020 to as long as 5 years today, while AI workloads demand deployment cycles under 18 months. The math doesn't work.

The geopolitical layer makes it worse. China controls 40%+ of US battery imports and ~30% of transformer and switchgear categories. Any trade escalation doesn't just affect chips — it threatens the physical layer that runs them. This is a structural mismatch, not a cyclical delay.

Why Hyperscaler Capex Guidance Is Mispriced

Even as builds stall, announced commitments keep accelerating: Microsoft committed $10B to Japan, Nebius committed $10B to Finland, and Oracle is laying off thousands to fund its AI pivot. OpenAI confirmed another mega-funding round. These are 3-5 year capital lockups premised on infrastructure that may not be physically deliverable on their timelines.

Every hyperscaler's capex guidance implicitly assumes the transformer bottleneck resolves on their timeline. It won't.

The utilization risk is the 2000 fiber-optics parallel: if application-layer revenue doesn't materialize to fill these data centers — and today's evidence on AI unit economics isn't encouraging — the math becomes brutal. This isn't alarmist; it's the base case you should stress-test against.

Where the Alpha Sits

Three cross-source signals point to the same investable thesis:

1. Domestic electrical equipment manufacturers — transformers, switchgear, modular power distribution. These companies face a structural 3-5 year demand tailwind with supply that cannot be quickly replicated. Add the $15B government AI supercomputer redirect and these become dual-catalyzed.
2. Modular and prefab data center solutions — anything that compresses the 5-year deployment gap to 18 months. The companies solving the time-to-capacity problem are more valuable than the compute providers themselves.
3. Alternative power generation — SMRs, on-site gas turbines, distributed solar+storage. The grid can't keep pace; the solution is going off-grid.

Critically, these are picks-and-shovels plays where you're betting the capex gets spent — not that utilization rates justify it. The $20B+ in announced builds, plus $15B in government redirect, means the capital deploys regardless.

Efficiency Gains Complicate the Picture

A contrarian data point worth noting: AI model efficiency is improving faster than expected. Self-distillation enables 7B parameter models to match 70B accuracy (50.6% → 60.4% on HumanEval), diffusion-based code generation runs 10x faster than autoregressive models, and KV cache compression achieves 8x storage reduction at 99% accuracy. If these gains compound, the total compute demand curve may flatten — undermining the very infrastructure thesis the market is pricing in.

The resolution: infrastructure demand is real and the supply gap is structural, but the specific compute mix and intensity may look very different than current projections. Favor flexible infrastructure plays over fixed-architecture bets.

The Framework You're Missing

While capital pours into AI model companies and agentic tooling, a critical counter-narrative is emerging from enterprise deployment data: most US enterprises are stuck at L0-L1 on a 6-level AI maturity ladder, and the bottleneck isn't model capability — it's that organizations literally cannot describe their own workflows in machine-readable terms. The hardest transition in the entire framework isn't deploying AI agents (L3→L4, described as "the most overhyped transition"); it's making tacit knowledge explicit (L1→L2, "the hardest transition").

Pilots that skip the legibility layer look impressive in demos and then disappear within six months.

This has direct portfolio implications. A bookkeeping automation project required six weeks of pre-deployment process documentation before any AI code shipped — because the business had never made its workflows explicit. One developer built 224 commits of working logic, with critical data mappings living entirely in one person's head. This isn't an edge case; it's the norm outside Silicon Valley.

Why This Reprices Enterprise AI Valuations

Cross-referencing this organizational readiness gap with the platform consolidation signals from other sources reveals a compounding problem. Anthropic is locking out third-party tools and enforcing per-token billing. OpenAI shifted Codex to usage-based pricing. Both moves reflect compute cost pressure — but they also assume customers can generate consistent, high-volume usage. If most enterprises can't even describe their workflows to machines, the usage curves that justify platform pricing aren't materializing outside tech-forward buyers.

The implication for portfolio companies: any enterprise AI vendor showing strong new-logo growth may be masking a retention crisis. Segment NRR by customer AI maturity level, not just vertical or company size. A portfolio company with 140% NRR in tech customers but 70% NRR in traditional enterprise is not a $5B company — it's a niche tool with an overstated TAM.

Where the Undervalued Layer Sits

A contrarian but investable insight: SMB AI platforms may offer better unit economics than enterprise. If AI compresses competitive distance and smaller organizations can achieve workflow legibility faster, the conventional wisdom that enterprise contracts are always superior may invert. This is early-signal territory, but the structural logic is sound.

The Political Resistance Factor

One underappreciated adoption barrier: AI forces organizational transparency that surfaces hidden power dynamics, budget manipulation, and knowledge hoarding. Making workflows machine-readable means making them visible to leadership. This creates political resistance that has nothing to do with technology and everything to do with organizational incentives. Companies that account for this in their go-to-market — selling change management alongside software — will retain better than those selling AI features.