Progressive engineer: fight AI's infrastructure, not its users

The author, a full-stack engineer since 2009 who uses AI daily, opens by acknowledging the core progressive critiques: AI training on artists' work without consent is extraction; hyperscale datacenters burden neighborhoods with higher power bills, stressed grids, and water depletion, while companies walk away with tax abatements that shift costs onto residents; and the technology concentrates wealth. He concedes these harms are real and communities fighting datacenter projects are right to do so.

However, he argues progressives are making a "strategic mistake" by treating AI monolithically. A single slogan—"AI is theft and it is destroying the planet"—cannot distinguish between a corporation dropping a trillion-dollar datacenter and a disabled person using an assistive tool to speak, or between a frontier model training run and an engineer generating boilerplate code on a Tuesday. He compares this to "protesting climate change by yelling at people who ride the bus": individual usage sits downstream of the real decisions. When he runs a query, he did not commission a datacenter or training run; the marginal cost of his usage is tiny and nowhere near where leverage exists.

To illustrate, he describes AntiFreeze, an app with thousands of users that lets people report and track ICE activity for neighborhood safety. The backend, security model, threat model, and data handling were hand-written by him carefully—because a mistake could result in someone being detained. AI compressed the tedious work: frontend scaffolding, translation strings for non-English speakers, and boilerplate code. This approach allowed the tool to ship in weeks instead of months. He also co-built LatticeNet (latticenet.ai) with another community organizer, a publishing layer where AI agents can post and read what others publish, like Substack for agents. Both projects required two humans with ideas, design, and execution; the tools made implementation faster.

On the economics: OpenAI posted a negative 122 percent operating margin in early 2026—losing about a dollar and twenty-two cents for every dollar taken in. The company is projecting a fourteen billion dollar loss for the year and does not expect to turn a profit until 2029 or 2030. A big chunk of that loss comes from roughly 900 million free users. So using ChatGPT's free tier costs the company money, money covered by Microsoft, Nvidia, Amazon, and SoftBank competing for market share. The individual user is a rounding error, not the revenue engine. His final call is precise: progressives should fight the datacenter placement, tax abatements, ratepayer protection, community consent, and payment for training data—winnable fights about who pays and who decides. Stop running purity tests on community organizers and volunteers using tools to build power faster. The disabled person, the civic tech volunteer, the engineer—these are not enemies; they are using leverage the way leverage works.

The author's core claim is that progressive opposition to AI commits a category error: it flattens industrial harms (datacenter buildouts, uncompensated training on artists' work, corporate concentration) into a single villain, then aims that anger at individual users downstream of those decisions. He supports this argument by distinguishing the marginal cost of one person's query from the leverage points where real decisions are made: the infrastructure, the training runs, the companies that control them, and the tax incentives that subsidize them. The economics reinforce his point: because OpenAI loses money on consumer-scale usage (negative 122 percent operating margin in early 2026, projected $14 billion（約2.2兆円） annual loss), individual users are a rounding error absorbed by megacorps competing for market share, not the revenue engine. This inversion—that free-tier users actually drain rather than enrich the company—reframes the progressive concern that AI use "feeds billionaires" as misidentified leverage.

His two project examples (AntiFreeze and LatticeNet) are offered as proof that community members can build things of value by using existing tools judiciously—without commissioning datacenters or training runs. In both cases, the author emphasizes that human judgment and security expertise went into the parts that matter most; AI handled scaffolding. This distinction parallels the larger argument: the question is not whether AI should exist, but where in the stack decisions require human accountability and where they do not. His final prescription—fight the datacenter, the tax abatement, and consent for training data—reorients progressive anger toward targets he considers winnable and aligned with what progressives are "actually good at": asking who pays and who decides.

No comments yet. Be the first to share your thoughts!