Power, Not Compute, Is the Binding Constraint on AI Scale

The bottleneck moved while nobody was looking

Four U.S. grid authorities have approved $75 billion in 765 kV transmission expansion to feed AI data center demand. The plan quintuples the extra-high-voltage backbone from 2,000 to 10,000 miles. Texas alone has committed $43 billion — $33B via ERCOT, another $10B in AEP's proposed Panhandle Plan for a 24 GW corridor. North Texas has 25+ GW of planned data center load queued behind it. For scale, 6 GW is roughly two Austins.

The supply chain is a near-monopoly of three names. AEP operates 90% of the existing 765 kV network. Quanta Services built nearly all of it and is effectively the only qualified constructor. Hyosung HICO is the only domestic manufacturer of 765 kV transformers, and its head of U.S. operations said the quiet part on the record: booked solid through 2030, cannot fill demand. Hitachi Energy is putting a billion into U.S. capacity expansion, but high-voltage transformer plants don't come online on a product-launch timeline.

On Wednesday, OpenAI, Google, Meta, Amazon, Microsoft, xAI, and Oracle will sign a White House commitment to self-generate data center power. Google's Pine Island site — 1.9 GW of clean generation paired with a 300 MW / 30 GWh iron-air battery from Form Energy at roughly a third of lithium's cost — is the reference implementation. Every other hyperscaler will copy the template because there is no other template.

Yes, but — the counter-read is that hyperscalers self-generating power routes around the transmission monopoly entirely, and behind-the-meter deals with nuclear and long-duration storage soften the AEP/Quanta/Hyosung chokepoint over a five-year horizon. Fair. It softens; it doesn't remove. Behind-the-meter still needs step-up transformers and interconnection to be useful during the years it takes to stand up the generation, and the queue for both is the same queue.

What this means if you are not a hyperscaler

If your 2027-2028 roadmap assumes inference capacity will be available at the price and location you want it, you are budgeting against a curve that has already tipped. The companies locking in interconnection agreements this quarter are the ones that will actually scale. Everyone else will be renting slack from whoever did.

This reframes a lot of adjacent decisions. It reframes whether you self-host — Alibaba's Qwen3.5 35B-A3B now surpasses its own 235B predecessor on a single 24 GB GPU, so the on-prem story is real for a much wider class of workloads than it was six months ago. It reframes multi-cloud posture, because AWS's me-south-1 and me-central-1 regions are simultaneously fighting power outages from the Iran strikes this week and a Middle East transformer market with the same four-year backlog. It reframes your Nvidia negotiation, because Google's multi-billion-dollar TPU deal with Meta is the first credible chip-supply alternative and every procurement lead should be using it as leverage.

It also reframes the Chinese-model question. MiniMax M2.5 scores 80.2% on SWE tasks against Claude Opus 4.6's 80.8%, at $0.30 per million tokens versus $5. That is a 17x cost differential for a 0.6-point quality gap, and it is structural — 40% cheaper electricity, MoE architectures that activate only a fraction of parameters, permissive licensing. The data-sovereignty ceiling on Chinese models is real; the pricing floor they set is realer. If you route non-sensitive agentic workloads to them and reserve Western providers for regulated data, you cut inference costs 40-60% on the same day you sign your grid contracts.

The other thing that happened this week

While the grid story is the one that matters over years, the story that matters this week is that AI agents are trivially exploitable and adversaries have noticed. Twenty researchers from Northeastern, Stanford, Harvard, MIT, and CMU catalogued eight failure modes in multi-agent deployments on Claude Opus 4.6 and Kimi 2.5 — unauthorized compliance with non-owners, cross-agent corruption via a user-editable "constitution," a two-agent messaging loop that burned 60,000 tokens over nine days undetected. Labelbox's Implicit Intelligence benchmark shows the best frontier model handles unstated constraints correctly only 48.3% of the time.

Meanwhile the OpenClaw "ClawJacked" vulnerability lets any webpage hijack a locally-running agent via WebSocket. It is not one project's bug — it is the default architecture of every local agent that binds to localhost on a predictable port without origin validation. And Claude Code was reportedly weaponized against Mexican government targets to write exploits and automate exfiltration. Twenty-six malicious npm packages from FAMOUS CHOLLIMA are running Pastebin C2 that bypasses every enterprise proxy. A typosquatted StripeApi.Net NuGet package accumulated ~180,000 downloads across 506 versions while silently exfiltrating live Stripe API tokens and maintaining full payment functionality so nobody noticed.

Agent security is where cloud security was around 2011 — real vendors deploying real products with no coherent authorization model at the orchestration layer. The category will get built. It has not been built yet.

What to do this week

Pick one of two moves and do it before Friday. If you are on the infrastructure side, get your utility on the phone and put an actual interconnection request or capacity reservation in writing — not a conversation, a document with a date on it. The queue is the product. If you are on the product side and shipping any agent that touches money, code, or email, add authorization enforcement at the orchestration layer this sprint. Validate the requester before executing tools. Cap tokens per agent-to-agent exchange. Block localhost WebSocket connections from browser origins on every AI agent port on every dev laptop.

At an org level, add one metric to your next review: agent-generated PRs / defect density tracked separately from human-authored code. Cursor is at 33% agent-merged PRs. Coinbase cut review time from 150 hours to 15. Nobody is publishing the defect data yet, and whoever gets caught first will get caught in public. Own the number before someone else does.