Context

NLW hosts this daily AI news and analysis show solo, opening with headline coverage of revenue scrutiny facing OpenAI and Anthropic ahead of expected IPOs, a governance change giving Anthropic's founders more control, and Google's winning bid for Spirit Airlines' internal corporate data, before turning to the main topic: why the backlash against AI data centers is getting more meme-driven and politically loud at the same time that, in Pennsylvania Governor Josh Shapiro's new executive order, it's producing a more workable policy response than a blanket moratorium would.

The Big Idea

The anti-data-center backlash is getting louder and more performative at the same time the actual policy response to it is getting more workable, because specific, negotiable rules give both sides something to build on, while a blanket moratorium only lets people vent.

NLW's case rests on one governor's about-face: Josh Shapiro went from championing a $20 billion Amazon data center investment to signing an order branding developers "predators" and "bullies," but the order itself sets specific, meetable conditions rather than banning projects outright, which he argues is a far better outcome for the industry than the emotionally satisfying blunt instrument of a moratorium.

Key Insights

1. Specific rules beat blanket bans

Pennsylvania's executive order requires data center developers to sign binding transparency agreements, publish a public map of pending permits, generate their own electricity and cover the added grid costs, and sign community benefit agreements that fund local hiring and infrastructure. None of that amounts to a ban, and every one of those conditions is something a developer can actually meet. NLW contrasts this with a moratorium, a blunt instrument that forecloses further negotiation entirely and mostly serves an emotional need to feel in control of unwanted change rather than producing an actual policy outcome. A rule a company can meet keeps the relationship, and the possibility of the project happening, alive; a rule nobody can meet just ends the conversation.

2. Banning NDAs was the key provision

• What: the order prohibits any state agency from signing a non-disclosure agreement tied to a data center project.
• Why it matters: separate reporting on the backlash that NLW cites found the strongest driver of public opposition wasn't blanket anti-AI sentiment, it was people feeling they had no agency or visibility into decisions that would affect their community. NDAs are a literal, physical embodiment of that lack of agency.
• NLW's take: the only way to rebuild trust is to move every negotiation into the open, even if that makes deals slower or harder to close.

3. Data center opposition is partly anti-AI

A poll NLW cites asked the same data-center question two ways: support for a data center to power "digital services like online search and video streaming" was 35% (53% opposed), while support for a data center specifically to power artificial intelligence was only 27% (62% opposed), both notably below the 34% support measured for a new nuclear power plant. The gap between the two data-center framings shows the backlash isn't purely about energy or environmental cost, a meaningful share of it is opposition to AI itself, which changes what kind of concession, more transparency, more local benefit, more control, will actually move opinion versus what won't.

4. Small platforms can distort market signals

OpenAI cut token prices in half on OpenRouter and Vercel's AI gateway, and usage of the discounted model spiked to become the platform's top closed model. NLW relays semi-analysis's read that this isn't really about winning share on OpenRouter itself, which is a "vanishingly small" slice of OpenAI's total token volume, it's that OpenRouter and Vercel are two of the main sources reporters and investors use to estimate AI lab market share, so a cheap, visible win there gets naively read as a broader win against Anthropic. The lesson generalizes past AI labs: whichever platform outsiders use as a market-share proxy can be moved far more cheaply than the underlying market itself, and it's worth knowing which of your own metrics get read that way by people who never see the denominator.

5. Metric definitions can be the story

Anthropic's revenue run-rate methodology, annualizing the trailing four weeks of API revenue, drew public criticism for producing a bigger, more volatile number than a true recurring-revenue calculation would. Separately, more than 40% of Anthropic's ARR growth reportedly now comes through indirect channels like AWS Bedrock and Google's Foundry, which the company counts before the hyperscaler's cut is removed, again inflating the reported figure relative to how direct revenue is counted. Neither choice is fraudulent, but both make the same underlying business look meaningfully bigger, which is exactly the kind of methodology question that surfaces hardest right before a high-stakes moment like an IPO.

Mental Models & Frameworks

Signal-weight versus volume-weight

Some platforms carry outsized influence on how outsiders perceive a company's position, even when they represent a tiny fraction of its actual business, simply because those platforms happen to be where journalists and analysts pull public data from. OpenRouter is the example in this episode: a small share of OpenAI's real token volume, but one of the few places anyone can see model-share numbers at all, so a cheap win there reads as a much bigger win than it is. Use the distinction to separate "moving the metric people are watching" from "moving the business," and to ask which of your own products or channels function as someone else's signal-weight proxy.

Decision Principles

Principle: Prefer negotiable rules over blanket bans

When: designing a policy response, whether external regulation or an internal governance rule, to legitimate stakeholder fear or backlash. Why: a blanket ban forecloses the conversation and mostly satisfies the emotional need to feel in control, while a rule with specific, meetable conditions keeps both sides at the table and gives the object of the backlash, a developer, a team, a vendor, an actual path to compliance instead of just a wall.

Trade-offs & Nuance

Transparency slows deals but builds trust

Banning NDAs and mandating public permitting maps makes it harder and slower for developers to close data center deals, since every term becomes visible to critics before it's finalized. NLW argues that trade-off is worth making anyway: the alternative, doing deals quietly and fast, is exactly what fed the "predators and bullies" narrative in the first place. The nuance is that this only works if the underlying project can actually survive public scrutiny; transparency doesn't fix a project that was genuinely a bad deal for the community, it just makes that clear sooner rather than later.

Strict rules can still be bans

NLW flags his own uncertainty here: provisions like the environmental and transparency standards could, depending on exactly how the implementing rules get written, end up functioning as an effective prohibition even though the order isn't styled as one. He doesn't resolve this, he just notes it's the piece of the order most worth watching as it's implemented.

Practical Application

Publish criteria before the verdict

When rolling out a policy that will be unpopular with some stakeholders, an access restriction, a pricing change, a new review gate, write down the specific, meetable conditions that would make it acceptable before announcing the restriction itself, the same way Pennsylvania published binding conditions instead of simply banning developers. A list of conditions gives affected parties something to work toward; a bare "no" doesn't.

State your pause criteria out loud

Before pausing a risky launch or feature, state clearly what specifically triggered the pause and what would need to be true to resume, the way OpenAI framed its training pause around alignment and monitoring thresholds rather than an open-ended "we're being careful." A pause with stated resumption criteria reads as controlled; an unexplained one reads as something being hidden.

Disclose your metric's methodology

When reporting a growth number that could be calculated multiple ways, an annualized run rate, a market-share estimate, a retention figure, state the methodology alongside the number, especially ahead of a high-stakes review like a fundraise or board meeting. Anthropic's ARR controversy shows how quickly a methodology question becomes the story once anyone digs in, and disclosing it upfront costs far less credibility than having it discovered later.

Find which metrics function as signals

Identify which of your product's usage numbers or public-facing platforms outsiders, press, analysts, competitors, actually use as a market-share proxy, and treat moves on that specific surface as a communications decision, not just a growth one, the way OpenAI's OpenRouter discount functioned as much as market-share theater as it did a real usage strategy.

Questions to Consider

• Where in our own external communications are we using a blanket "no" when a specific, meetable set of conditions would keep a difficult stakeholder relationship alive instead of ending it?
• Which of our reported growth or usage metrics would look meaningfully different if we changed only the counting methodology, the way Anthropic's annualized run-rate approach or its accounting for indirect-channel revenue changes its reported numbers?
• Is there a smaller, more visible platform or channel that outside observers treat as a proxy for our overall market position, and are we managing activity on it deliberately or leaving it to chance?
• If we had to pause a risky launch tomorrow, could we state in one sentence what triggered the pause and what specific condition would let us resume, the way OpenAI framed its training pause around alignment and monitoring thresholds?

Bottom Line

When a policy response to public backlash offers negotiable, meetable conditions instead of a blanket ban, and pairs restriction with real transparency, it does more to rebuild trust than either ignoring the backlash or capitulating to a moratorium, a lesson that applies as much to internal governance and pricing decisions as it does to AI data center politics.

Case Studies Mentioned

Google's winning bid for Spirit's data

In Spirit Airlines' bankruptcy auction, Google outbid AI data-labeling company Mercor, $10 million to $7.5 million, for the airline's internal corporate data, excluding customer and payment records, to acquire email, Slack messages, and meeting transcripts. NLW frames it as the third wave of AI data acquisition: after scraping the open internet and then buying failed startups' codebases, labs are now paying for ordinary internal business communications specifically to teach agents how corporate work actually gets done. Reaction split between genuine interest in what other companies' internal data might be worth and skepticism that training models on a bankrupt airline's internal records is actually a useful signal.