Electricity, not models, is the binding constraint on AI now

Roughly half of all planned US data center builds for 2026 are delayed or canceled. High-power transformer lead times have stretched from about two years pre-2020 to five years today. AI deployment cycles want eighteen months. The math doesn't work, and no amount of capital fixes it — you can't print transformers.

The same week, the Trump FY2027 budget proposed $1.5T for defense (a 42% increase, the largest single-year jump since WWII) and explicitly redirected $15B from clean energy to AI supercomputers. Congress historically waters down domestic cuts and funds the military line items. So treat the defense and AI-compute numbers as the durable signal and the rest as negotiating posture. The federal government just became the marginal buyer of AI infrastructure at the exact moment the private buildout stalled.

This is the story of the day. Everything else — Anthropic's tool lockout, the Mercor breach, the chatbot lawsuits, the eval blind spots — is either downstream of compute scarcity or a separate accountability tremor running on a parallel track. Both deserve attention. But if you only have time for one mental model this week, make it this: physical infrastructure is the gating function, and the people who priced their roadmaps assuming current cloud rates through 2027 are wrong.

What scarcity actually does to your roadmap

The second-order effects are already visible. Anthropic blocked third-party agentic tools from flat-rate Claude subscriptions on April 4, forcing per-token billing and absorbing OpenClaw's features into Claude Code. OpenAI moved Codex to usage-based pricing in the same window. The framing was "compute and engineering strain" — which is partially true and entirely consistent with a vendor whose supply curve has gone vertical. Flat-rate pricing was the economic substrate for iterative agent loops. It just got pulled, and any pipeline that retries fifteen times now costs fifteen times more overnight.

This is what platform consolidation looks like when the underlying resource is constrained. Expect more of it. Every flat-rate AI subscription in your stack is now a pricing model that can be revoked with a week's notice. If your LLM provider abstraction is config-swappable rather than code-swappable, you're fine. If it isn't, that's the work of the sprint.

The hedge arrived on schedule: Google released Gemma 4 under Apache 2.0 — first time, fully permissive, commercial use allowed. Benchmark it against your most cost-sensitive workloads this sprint, not next quarter. The point isn't that Gemma 4 is better. It's that you need a credible second option on file before the next vendor makes the next move.

Efficiency is the real strategic high ground

Three research results landed in the same cycle, all pointing the same direction. Polar-coordinate KV cache compression: 2-bit quantization, claimed 99% accuracy retention, 8x memory reduction. Self-distillation lifting a 7B model to 60.4% on HumanEval — matching dense models 10x its size. Mercury Edit 2 claiming 10x faster code generation via diffusion. The KV cache result is the most validated; the diffusion claim needs independent reproduction before you bet on it.

Validate them on your actual workloads. The polar-coordinate "99% accuracy" metric is undefined in the paper — perplexity, downstream task accuracy, and needle-in-a-haystack are very different claims. But the direction is unambiguous: if scale is constrained, efficiency wins. The conventional wisdom that bigger always beats smaller is inverting underneath us. A team that can serve comparable quality on a quarter of the compute owns the strategic high ground when capacity is rationed.

The accountability tremor running in parallel

Separately and just as consequentially: Jay Edelson — the plaintiff's attorney who extracted settlements from Facebook — is filing chatbot-specific lawsuits. Microsoft's own ToS describes Copilot as "for entertainment purposes only," which means Microsoft's legal team assessed standing behind Copilot for work and decided not to. Sam Altman went on camera in a documentary admitting OpenAI's plan for existential risk amounts to trusting governments. Anthropic disclosed "functional emotions" in Claude that measurably influence output behavior.

Fold these together and the shape is clear: every UX decision in a customer-facing AI product — anthropomorphization level, disclaimer copy, persona framing, guardrail design — is now legal discovery material. The team that treated guardrails as launch-blocking annoyances last year is about to wish they'd treated them as moats. The Microsoft ToS finding belongs in every competitive battle card you ship this quarter, before Microsoft quietly updates the language.

Meanwhile DeepMind published the first structured taxonomy of six agent exploit categories. Combine that with the finding that reasoning models choose tools in their first few tokens — pattern-matching on prompt surface, not deliberating — and your red team has a concrete checklist. Test whether reordering the first ten tokens of a prompt changes tool selection. If it does, your reasoning chain isn't driving the decision; it's narrating one.

What to do this week

One specific thing, because this is where the piece earns its keep: pull your AI infrastructure roadmap and stress-test it against cloud pricing 25–30% above current rates, six-month GPU lead times on specific SKUs, and at least one major vendor revoking flat-rate access on your second-largest dependency. If the roadmap survives all three, you're ahead of most teams. If it doesn't, the gap between what you've committed to and what you can actually ship in 2027 is the number you need on a slide before your next planning cycle.

The rest of the year is going to reward operators who treated electricity, vendor contracts, and guardrails as load-bearing infrastructure rather than someone else's problem.