Anthropic Refused the Pentagon on Ethics. Two Months Later They Bought Elon's Dirty Compute.

I cheered when my Claude Code limits doubled. 48 hours later I read where the compute came from.

8 min read

Anthropic announces doubled Claude Code limits for all Pro and Max subscriptions. Hitting the limits often enough (dev sessions blocked mid-afternoon, tokens melting on tasks that weren't even that complicated). When the email landed, I smiled. I thought "they're finally listening." I took the gift without asking questions.

48 hours later I was reading Simon Willison's live blog of the Code with Claude event. Buried in the recap, a quiet editorial aside: "That's the same Colossus data center in Memphis with the particularly bad environmental record." I clicked. I read the follow-up he published the next day. Gas turbines running without Clean Air Act permits. Documented increases in hospital admissions. The NAACP in active litigation against the facility since April. And Anthropic leasing 100% of the installation's capacity.

The same Anthropic that had turned down a $200M Pentagon contract two months earlier on ethics grounds. The same Anthropic whose Super Bowl ads reminded me that unlike OpenAI, they weren't monetizing my data with advertising. The same Anthropic that had just become the #1 App Store app because tens of thousands of users had left ChatGPT as a moral protest. I had taken the gift. I hadn't asked where the electricity came from.

TLDR: Anthropic doubled my Claude Code limits. I said thanks. Then I read where the compute actually comes from (a Memphis data center running gas turbines without Clean Air Act permits, currently the subject of a NAACP lawsuit). The same Anthropic that turned down a $200M Pentagon contract on ethics grounds. There's a pattern here, and it's not specific to Anthropic.

Office worker excitedly pointing at monitor while superhero disapprovingly holds pollution lawsuit newspaper behind him
Ethics? Never heard of her. *doubles limits anyway*

The Safety Brand They Sold You

For about 6 months, Anthropic has been executing one of the cleaner brand positioning moves in tech. In March 2026, they walked away from a $200M Department of Defense contract. The stated red lines: mass surveillance and autonomous lethal weapons. The Pentagon deal required accepting "all lawful purposes" as a scope. Anthropic said no.

What happened next is documented. Claude went from rank 42 to rank 1 on the App Store the following weekend. A Reddit post calling users to cancel ChatGPT pulled 30,000 upvotes. Someone drew chalk art in front of Anthropic's SF offices: "you give us courage." Free active users climbed 60% since January, daily sign-ups quadrupled. Annualized revenue moved from $14B to $19B in 2 months.

Axios quoted the internal framing directly: "Trust is a competitive advantage."

I bought it. I wrote about Anthropic crashing cybersecurity stocks by $15 billion in February, positioned them as the adult in the room. The brand was coherent. The red lines were documented. The narrative ("we draw lines, OpenAI doesn't") was converting tens of thousands of ChatGPT migrants. You believed it because it was consistent. Nothing in the posture was obviously wrong.

Where Your Doubled Limits Actually Come From

TITLE "Colossus 1: What Powers Your Doubled Limits" + subtitle "220K GPUs · 300 MW · 0 Clean Air Act permits". Metaphor: industrial cross-section diagram of a factory floor cut open, showing power input on the left and GPU racks on the right, with a gas turbine stack labeled "TEMPORARY" in quotes at the bottom. Style: engineer blueprint aesthetic, technical drafting lines on dark navy background, white and amber labels. Palette: navy #0D1B2A, amber #F4A261, white #F8F9FA, warning red #E63946, concrete grey #6B7280. Content: left column shows "GAS TURBINES" block labeled "classified as temporary to bypass permitting", right column shows "220,000 GPUs" block labeled "H100 / H200 / GB200", center shows power arrow labeled "300 MW" flowing from turbines to GPUs. Bottom bar shows "NAACP LAWSUIT FILED APRIL 14 2026" in warning red. Highlight: the "TEMPORARY" label in quotes with a red warning icon, the NAACP bar pulsing in warning red. Footer: © rentierdigital.xyz. NOT flat corporate infographic, NOT clean tech startup aesthetic, NOT minimalist.
Colossus 1 Data Center Power Infrastructure Breakdown

Colossus 1 is a data center in Memphis, Tennessee. 220,000+ NVIDIA GPUs (H100, H200, GB200). 300 megawatts of power draw. Originally built by xAI, subsequently absorbed into SpaceX when SpaceX acquired xAI earlier this year. Anthropic is leasing 100% of the facility's capacity (not a partial agreement, the whole thing).

The compute case is straightforward. Multi-agent orchestration, rubric-graded iteration loops, out-of-band memory consolidation. None of the keynote announcements from Code with Claude are economically viable without a massive throughput increase behind them. One Redditor documented a single prompt consuming 10% of their session limit versus the 0.5-1% they expected. The limits weren't a policy choice, they were a compute ceiling. Colossus was the fix.

The problem isn't the GPUs. It's what powers them.

The gas turbines installed to run Colossus were initially operated without Clean Air Act permits and without pollution control devices. They were classified as "temporary" (a designation that exists in environmental law and was used here to avoid the standard permitting process). The NAACP filed a lawsuit on April 14, 2026 against xAI and MZX Tech for unpermitted air pollution. According to the NAACP's Director of Environmental and Climate Justice, Anthropic's compute deal is "at best, an uninformed decision, and at worst, a total disregard for the community's wishes and health." The pollutants involved are linked to asthma, respiratory disease, and cardiovascular problems. This is a majority-Black community in Memphis.

Andy Masley is not an activist. He's a datacenter advocate (someone whose default position is pro-infrastructure, pro-buildout, pro-compute). After the Anthropic deal was announced, he said explicitly that he would "simply not run my computing out of this specific data center." When the pro-datacenter people won't defend a specific datacenter, that's the useful signal.

A caveat worth stating directly: Anthropic doesn't operate Colossus. They lease capacity from a facility they don't control. That's a real distinction. It doesn't change the fact that they chose this facility knowing its record.

Safety Is Marketing. Compute Is Operations.

The asymmetry that makes sense of all of this runs like this.

Safety is visible. Refusing the Pentagon gets named in the press. Super Bowl ads run during primetime. Blog posts about red lines get shared by the people who needed to see them. Dario Amodei's public positions on autonomous weapons are documented going back to 2023 and have been consistent. The safety brand is built entirely in daylight, in front of the audience it's designed to convert.

Compute is invisible. Colossus is 800 kilometers from San Francisco. The deal was signed and announced during a product keynote, buried under multi-agent orchestration and infinite context windows. The gas turbines are in Memphis. The NAACP lawsuit was filed in April and got limited tech press coverage until Simon Willison flagged it in a parenthetical on May 6. Nobody put out a press release about the permitting violations.

What you end up with is a company that operates 2 parallel systems that never need to talk to each other. The brand system faces the customer. The infrastructure system faces the constraint (availability, cost, throughput, speed). Ethical fashion brands manufacture in Bangladesh. Apple runs "privacy is a human right" campaigns while assembling hardware at Foxconn. The premium-tier promise is for the customer relationship. The supply chain runs on a different logic entirely, and the 2 systems are structurally isolated from each other by design, not by accident. The customer who cares about the brand doesn't have visibility into the supply chain. The community that has visibility into the supply chain isn't the customer base. This isolation isn't a bug in the model. It's what makes the model work at scale across every industry that has figured it out.

According to environmental analyst Ketan Joshi, there are 184 planned or operational data center projects with a greenhouse gas footprint larger than Colossus. Anthropic isn't an outlier. They're the standard. I think that's actually the uncomfortable part (it would be easier if this were an Anthropic-specific failure you could fix by switching providers).

The Pentagon Refusal Looks Different Now

I want to be clear upfront: what follows is my read, not Anthropic's stated rationale. I'm connecting dots they haven't connected publicly.

In March, Anthropic turns down $200M because the Pentagon's terms require accepting "all lawful purposes" as a use scope. The decision is presented as ethical. The business result is documented: #1 App Store, $5B additional annualized revenue in 2 months, 60% user growth.

In May, Anthropic signs a compute deal with a facility under environmental litigation since April, operated by a company whose founder publicly described Anthropic as a company that "hates Western civilization" (and takes 100% of its capacity).

These 2 events make sense together if you accept that safety is a marketing category and compute is an operations category. The Pentagon refusal created demand. The demand made Colossus necessary. Whether that sequence was calculated or just structurally inevitable doesn't change the outcome. The refusal may also have had a genuine component (Dario's position on autonomous weapons has been consistent for years). Sincerity and business logic can produce the same result. The gas turbines in Memphis still ran without permits either way.

What This Means If You're Paying Anthropic

I'm still using Claude Code. The doubled limits are real, the tool works, and I'm not going to stop using something that genuinely helps me ship because the supply chain is uncomfortable to look at. I've been through enough of these vendor reality checks to know that rebuilding your setup when a vendor breaks the implicit contract is sometimes the right call. Sometimes you just update your mental model and keep shipping.

What I've stopped doing is paying the emotional premium. The safety isn't something you're buying with your Pro subscription. It's a conversion argument with a measurable ROI, and understanding that makes you a more useful customer (you evaluate on technical performance and actual output quality, not on whether the brand story makes you feel good about the invoice).

When an AI company sells you a moral posture, ask where the datacenter is. That information is public. It just requires a click the marketing doesn't encourage you to make.

Today I opened Codex and typed: "add an AGENTS.md pointing to claude.md and ~/.claude/CLAUDE.md". The irony of configuring one AI tool's context files via a competitor's tool wasn't lost on me 😅 Make of that what you will.

Anthropic didn't lie. They sold you half the story (the half that was visible).

Sources

This post may contain affiliate links. If you click them, I might earn a small commission -- costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure.

Anthropic doubled your Claude Code limits—but the compute comes from a data center running unpermitted gas turbines while facing NAACP litigation. When your infrastructure choices contradict your stated ethics, the Demo vs Product Checklist in the welcome kit shows you what actually separates principle from performance theater.

→ Get the welcome kit