Edition 2

A $1.1 million fine is supposed to read like accountability. To me it reads more like a line item—an after-the-fact fee that still dodges the only questions that matter: how much ran, how often, what came out of the stacks, and who is actually keeping score.

Ars Technica reports that New Jersey ordered DataOne to pay a $1.1 million fine after investigators concluded the company secretly installed and operated gas generators at its planned Vineland data center in violation of the state’s Air Pollution Control Act Ars Technica. The catalyst, Ars says, was outside reporting that shared thermal drone footage indicating that 45 of 62 generators were operating. Ars adds that the state said the equipment required permits above a 37-kilowatt threshold, while the units were operating at far higher capacity.

Here’s my claim: “the AI race” has become a denominator-killer. It’s a narrative that inflates the urgency of building compute while shrinking the obligation to quantify tradeoffs in public. We get big numerators—62 generators, a seven-figure fine—but not the denominators a community needs to judge what’s being normalized: operating hours, emissions over time, how frequently the site plans to bypass the grid, what “temporary” really means on a construction timeline, and who carries the health risk while the paperwork catches up.

Ars reports why New Jersey tracks these generators at all: state officials noted they emit carbon dioxide, nitrogen oxides, carbon monoxide, and other combustion-related pollutants, and the agency linked those pollutants to asthma impacts, heart risks, and early deaths. That’s not abstract. That’s the physical invoice for the “progress” story.

And the 62 generators are the perfect example of why denominators matter. “Backup power” is elastic language. It can mean rare emergencies, or it can mean routine parallel operation that quietly turns a data center into its own off-grid plant. Without operational numbers, the label does the work. Meanwhile, the people living near the facility are stuck asking the most basic accounting questions. Ars relays residents’ concerns—what have we been breathing, what emissions were released, what’s the air quality near homes and schools, and has it ever been measured? Denominator questions, all of them.

The drone detail is the tell. It suggests an enforcement regime that’s episodic and reactive: someone notices something in imagery; journalists and advocates push; the state shows up; a penalty is assessed. Ars describes how the initial alarm came after a farmer spotted equipment on satellite imagery and how inspectors later found the “tractor-trailer-size” generators impossible to miss once they visited. That is not what durable oversight looks like. If the only audit tool that reliably bites is a drone flight and a news cycle, the incentive is obvious: build first, negotiate later.

Ars also reports that the enforcement action—described by the DEP commissioner as the largest ever taken against a data center in New Jersey—doesn’t immediately relieve residents’ concerns. DataOne was given 45 days to apply for permits or else cease operations, but Ars notes the company can keep operating while seeking those permits. A local environmental group leader told Ars they viewed that as insufficient and argued the site should stop operating during the permit window. You can call that anger. I call it a plain reading of incentives: if the business can keep meeting customer commitments while “compliance” is negotiated, then the community is the one forced to live inside the uncertainty.

This is where “AI race” rhetoric stops being mere hype and becomes a governance problem. Keep the conversation at 30,000 feet—national competitiveness, inevitability, speed—and local friction starts getting framed as sabotage. Permits become “red tape.” Monitoring becomes “anti-growth.” People asking for measurements become “anti-tech.” The race story is a convenient way to treat oversight as optional and delayable.

Ars reports that DataOne told The New York Times it would apply for permits for “temporary generators,” and that it said it would eventually phase out gas generators while transitioning to “low-emission, quiet fuel cells”. Maybe that transition happens. Maybe it’s real. But “temporary” and “transitioning” are exactly the kind of soft words that should trigger hard questions. Temporary for how long? Transitioning on what schedule? Low-emission relative to what, and at what operating conditions? The point isn’t to assume bad faith; it’s to refuse to accept adjectives as a substitute for quantities.

Ars also describes a broader arc of distrust: complaints about noise, public health violations for excessive noise, and residents who feel government has prioritized corporate timelines over community well-being. My argument isn’t that every data center is uniquely villainous. It’s that the structure is tilted: the operator holds the logs, the testing data, and the vocabulary. The neighborhood gets what it can see, hear, and smell—and what can be inferred when somebody finally looks from above.

So when I see “$1.1 million” in a headline, I’m less interested in whether it’s big or small in the abstract. I’m interested in what it buys: time, ambiguity, and a sense that accountability has been satisfied without answering the underlying denominator questions. A fine can be punishment. It can also be a price for keeping the math private.

If you want me to believe the “AI race” story is about the future we’re building—rather than the costs we’re hiding—then stop selling numerators as proof. Publish the denominators. Put the operational reality on the record in a form ordinary people can understand and verify.

Because the real question isn’t whether “AI” is good or bad. It’s whether we’re going to keep letting the word “race” substitute for math. Who measured? Using what baseline? Over how many hours? And who benefits from the gap between what gets promised and what gets built?

The cleanest AI demo is always the same trick: it lets you believe “intelligence” ships on a calendar.

What actually ships is reliability. Reliability has a physical footprint, a regulatory footprint, and—if you build enough of it—a neighborhood footprint. If you want “always-on” AI at mass scale, the bottleneck isn’t just model quality. It’s the substrate: power delivery, cooling, redundancy, and the permissions to run the equipment that makes uptime real.

A clarifying story this month wasn’t about a new benchmark. It was about a data center in Vineland, New Jersey.

Ars Technica reports that New Jersey ordered data center operator DataOne to pay a $1.1 million fine after the company installed and operated 62 gas generators without the permits required under the state’s Air Pollution Control Act. Ars says the fine followed an investigation by The Guardian and Floodlight News that shared thermal drone footage showing 45 of the 62 generators running. The state’s DEP, per Ars, flagged that the generators were above the threshold that triggers permitting requirements and that the emissions from this type of equipment include carbon dioxide, nitrogen oxides, carbon monoxide, and other combustion-related pollutants with serious health impacts. Ars also reports the state framed the action as its largest ever against a data center, and that the order gives DataOne 45 days to apply for permits or else cease operations—while still allowing the generators to run during the permitting process. Ars Technica

That’s the factual spine. Here’s my read: those generators are a receipt for what “AI later this year” implies when it collides with reality.

Consumer AI gets marketed like the hard part is inference quality and UX polish. But the moment you try to make AI feel like an appliance—ask anytime, get an answer now—you’ve committed to an uptime story. And uptime isn’t a vibes problem. It’s a design constraint that bleeds outward into land use, noise, emissions, fuel logistics, maintenance schedules, and, eventually, permit files.

The Ars reporting doesn’t describe a quiet site with dormant emergency equipment. It describes dozens of gas generators installed and used without the required approvals, with enough simultaneous operation that thermal imagery could observe it. Ars also characterizes DataOne as notably the first data center operator in New Jersey to rely on gas generators to create its own off-grid supply. You can debate motives—temporary bridge, schedule pressure, a bet on enforcement timing—but the mechanism is straightforward: when the grid can’t deliver the certainty your service requires, someone will attempt to manufacture certainty on-site.

And “manufacture certainty” is where the romance dies.

AI roadmaps are quietly becoming power roadmaps. At small scale, you treat availability like a cloud checkbox. At mass scale, availability becomes a capital plan. If you train millions of people to use an assistant habitually, usage doesn’t distribute politely. It clusters: mornings, commutes, breaking news, outages, sports finals, the moments when everyone asks at once. That means you don’t engineer for the average hour—you engineer for the worst few minutes you can’t afford to lose.

Engineering for the worst few minutes is how you end up paying for the stuff that doesn’t show up in model cards: redundant power paths, redundant network paths, redundant cooling, “wasted” headroom, and operational staffing that can absorb incidents without the product visibly wobbling. It’s warehouse-sized obligation behind pocket-sized expectation.

The New Jersey case shows what happens when that obligation leaks into public view. Regulators don’t primarily care that the load is “AI.” They care that you’re running combustion equipment above certain thresholds without permits. Neighbors don’t care about your latency target; they care about noise, local air, and whether the rules are enforced only after outside reporting and aerial footage make inaction look indefensible. Ars relays residents’ concerns about pollution, delays in enforcement, and questions about air quality—framed through coverage by outlets including The New York Times.

A million-dollar fine sounds large until you map it onto the economics implied by hyperscale compute. Ars quotes locals describing the penalty as insufficient—a “slap on the wrist” and “pocket change” compared to the capital behind data center development. My opinion: they’re pointing at the right risk. If the upside is a long-term contract to serve a major customer, penalties that don’t interrupt operations can start to look like a cost of doing business, not a deterrent.

The more operationally interesting lever, to me, is the one Ars puts in plain sight: DataOne was given 45 days to apply for permits, but could keep operating during that window. I’m not litigating what the law should be. I’m saying incentives matter. If “apply within 45 days” doesn’t mean “stop now,” then compliance becomes a parallel track to uptime rather than a gate in front of it.

Ars reports that DataOne told The New York Times it would apply for permits for “temporary generators” and ultimately transition to “low-emission, quiet fuel cells.” Maybe that transition happens. Maybe it doesn’t. The pattern I’ve seen is that urgent compute demand arrives first, and the clean, quiet, community-compatible version arrives later—after deployments, after contracts, after momentum.

So when you hear confident timelines for ubiquitous AI—everywhere, always available—ask the unglamorous questions that decide whether the promise survives contact with the physical world. Where is the power margin coming from? What’s the plan when the grid is stressed? What equipment does that plan require, how loud is it, what does it emit, and what approvals does it trigger? And what happens when the permit timeline and the product timeline disagree?

Because in the real world, “AI later this year” often means: AI, assuming the power shows up—and assuming the rules and the neighbors let you run what you need to run.

Microsoft dropping “Copilot+ PC” branding from its newest Surface launches isn’t a retreat from the assistant-centric future. It’s a recognition that you don’t win consumer behavior by issuing a hardware ultimatum. You win by making the assistant the path of least resistance—and that puts an old, unglamorous product requirement back at the center of the story: a real, durable, legible “off.”

Ars Technica reports that Microsoft’s newly announced Surface Pro 12-inch (2nd Edition) and Surface Laptop 13-inch (2nd Edition) don’t use the Copilot+ PC label—even though Microsoft’s Surface lead told Windows Central they still meet the requirements previously associated with it, and the machines’ Qualcomm NPUs are specced well above the earlier bar (Ars lays out the prior requirement set, including 16GB RAM, 256GB storage, and an NPU rated at 40 TOPS or better) Ars Technica. Ars also notes that a Microsoft marketing page once pushing Copilot+ PCs prominently now routes users toward a more generic “performance PCs” framing where the Copilot+ label appears less centrally.

Those are the facts; here’s what I think they mean.

“Copilot+ PC” was an attempt to turn an interaction paradigm into a device category. The idea was simple enough: local AI needs local silicon, so consumers should know which laptops are “real” AI laptops. But categories don’t stick just because marketing says so. They stick when buyers can feel the difference, when they know what it’s for, and when the upgrade cycle lines up.

Ars suggests multiple pressures that could be pushing Microsoft away from the branding. One is that the Copilot+ PC story got tangled up with the security and privacy alarm around Recall, a feature that has remained controversial among privacy advocates. Another is plain market reality: Ars cites IDC research manager Jitesh Ubrani saying interest in AI PCs has been wavering given that cloud-based options are broadly available and on-device AI use cases have been limited. Ars also points to Dell de-emphasizing AI PC branding in its 2026 lineup after earlier pushing it. And there’s a competitive marketing logic too: Nvidia’s incoming RTX Spark laptop GPUs may become a more compelling headline than a Copilot-optimized badge, especially with Microsoft itself confirming a “Surface Laptop Ultra” with RTX Spark later this fall.

Taken together, this looks less like “AI is over” and more like “the wedge strategy changed.” If you can’t make people buy the assistant future as a distinct premium tier, you move the assistant into the ambient background and let distribution do the work.

That’s why I’m not especially soothed by the disappearance of a label. Removing the branding doesn’t remove the incentive. Assistants remain an incredibly attractive control point: they sit on top of search, writing, file discovery, meeting notes, notifications—everywhere intent is expressed. If a platform provider can make the assistant the default way users ask questions, open documents, summarize threads, or find settings, it becomes a habit. Habits beat badges.

And defaults are where consent tends to quietly die.

The subtle danger isn’t that an assistant exists. Many people will like it; some will rely on it. The danger is the slow conversion of “choice” into “assumption”—achieved through setup flows, upgrade prompts, “recommended” toggles, and UI placement that’s optimized for adoption, not for informed preference. The assistant doesn’t need you to love it. It needs you not to notice the moment it became unavoidable.

So if we’re serious about “user control,” we should stop treating “off” as a troubleshooting step and start treating it as a first-class product feature.

A serious “off” has three properties.

First, it’s singular and understandable. Not “off for the icon but on for background services.” Not “off for suggestions but on for data collection.” Not three settings pages with three different meanings. One switch, one meaning.

Second, it’s durable. If an update can re-enable the assistant, then “off” is just a temporary state the vendor tolerates. Durability means staying off through feature updates, UI refreshes, account sign-ins, and the inevitable rebranding cycles that turn yesterday’s controversial feature into tomorrow’s “improved experience.”

Third, it’s respected in the design. Turning something off shouldn’t feel like opting into a degraded, nag-filled experience. If “off” is punished—extra prompts, constant reminders, reduced functionality unrelated to the assistant—then it’s not really off. It’s coercion by friction.

This is the part that matters more now that the Copilot+ PC badge is fading. A label is a front door: it tells you what you’re buying. Defaults are plumbing: they decide what happens when you’re not paying attention. When the strategy shifts from “buy this class of AI machine” to “AI is just what the OS does,” the ethical burden shifts from marketing honesty to product governance.

Ars reports that Microsoft still “leans into the narrative” of edge AI and hybrid approaches even without the Copilot+ terminology. That’s a reasonable technical direction: some tasks should run locally; others will remain cloud-based; most users won’t care where the compute happens as long as it’s fast, cheap, and doesn’t surprise them. But “doesn’t surprise them” is doing a lot of work there.

Recall is the cautionary tale embedded right in Ars’ framing: if a flagship AI feature becomes synonymous with persistent surveillance risk in the public mind, the vendor doesn’t stop wanting AI—it learns to ship AI in ways that are harder to isolate and harder to debate as a single, named thing. That may be good product strategy; it’s also exactly why “off” must be durable and legible.

The unresolved question isn’t whether Microsoft will keep pursuing assistants—it will. The question is whether refusing the assistant will remain a supported, stable state of the platform, or become a perpetual game of whack-a-mole against new defaults and renamed experiences.

If Microsoft can’t mandate assistants through a badge on a laptop box, it will default them through the operating system. That doesn’t have to be dystopian. But it does mean the most important AI feature on your PC might not be a new model or a faster NPU.

It might be an off switch that stays off.

The scandal here isn’t that an algorithm might get a medical call wrong. It’s that “deny” is being treated like a feature you can deploy at scale—without deploying an equally real, equally fast way for the people harmed by that denial to stop it.

Ars Technica reports that the Trump administration launched a Medicare pilot in January that introduces prior authorization—pre-approval—to services that, in traditional Medicare, didn’t require it before, and that the program uses AI and machine learning as part of how approvals and denials happen Ars Technica. In the months after rollout, Ars writes, providers described technical failures, long waits, confusing denials, and patients stuck in pain while decisions lingered. Federal documents obtained by the Electronic Frontier Foundation through litigation—again as relayed by Ars—include provider feedback describing weeks-long delays and an inability to reach a human being.

Those are the reported facts. Here’s my point: this is governance-by-interface.

Prior authorization isn’t new in American healthcare. What’s new—and politically potent—is shifting a “care happens unless a human stops it” system into “care doesn’t happen unless you clear the gate,” and then making the gate feel procedural, neutral, and therefore unarguable. You don’t debate it as policy. You experience it as a workflow.

The WISeR model (Wasteful and Inappropriate Service Reduction) is explicitly framed as protecting taxpayers by reducing fraud, waste, and abuse, and it has been rolled out in six states with plans to run through 2031, according to the documents described by Ars Technica. It currently applies to roughly a dozen services—pain interventions like epidural steroid injections and nerve stimulation are among them—so the kinds of patients most exposed aren’t abstract. They are often people in immediate discomfort, trying to get relief.

But the detail in Ars that should reframe how we think about “AI in healthcare” is the incentive design.

In a Senate hearing, Sen. Patty Murray pressed the administration’s HHS deputy secretary nominee on whether contractors make more money when they deny care; Ars reports CMS planning documents indicate that WISeR participants are compensated through a share of “averted expenditures”—with guidance describing CMS paying vendors 25 percent of a benchmarked cost when a request is denied. Ars also reports the CMS Office of the Actuary warning that participants will have an incentive to deny as many claims as possible.

You can believe—sincerely—that you’re fighting waste and fraud, and still create a machine that is rationally paid to say “no.” In other words: even if the model is “accurate,” the surrounding system is designed to prefer denial, then dare patients and clinicians to fight their way back to yes.

And that fight happens in the most modern and least accountable arena we have: the interface. A denial is not just a decision; it’s a compression. It takes a patient’s messy reality—pain, history, risk tolerance, the physician’s judgment, the calendar—and collapses it into codes, rules, queues, and deadlines. Ars reports WISeR aims for decisions within 72 hours, but providers reported waits stretching to weeks and even months, including at least one request pending after 83 days. When time is part of the clinical outcome, delay is not a neutral inconvenience. Delay is a form of denial that doesn’t have to call itself that.

The vendors’ struggles matter less to me than what they reveal about how fragile this kind of governance is. Ars reports one vendor asked to delay rollout and temporarily “auto-approved” everything to avoid backlogs; another had months of discrepancies tied to misunderstanding Medicare Part A vs. Part B; and one vendor’s weekly report showed more denials than approvals during a period Ars describes. If you’re an elderly patient trying to understand why your care is stalled, none of that reads as “innovation.” It reads as being trapped inside someone else’s half-built system.

We talk about AI like it’s mainly a question of capability: Can the model classify? Can it detect fraud? Can it match rules? In WISeR, the more dangerous question is constitutional: who holds the handle on refusal?

Ars reports providers describing “radio silence” when they sought help, and describes feedback complaining there was no way to get a human on the line. That’s the thing no one should accept as a growing pain. If you cannot reach a responsible human quickly, you do not have a healthcare process—you have a labyrinth.

The program’s defenders, as Ars reports, point to checks: corrective action plans for missed deadlines; quality scoring meant to discourage inappropriate denials; the claim that vendors use established CMS coverage determinations rather than inventing their own rules. But Ars also reports that the payment adjustment for poor quality scores appears relatively modest compared to the core incentive to deny, and that few denials are appealed in similar contexts even though many appeals succeed. That combination—strong incentive to deny, weak incentive to be right, and low appeal throughput—creates a predictable outcome: the system can be “compliant” while patients suffer.

If this is an experiment, then stop asking the public to evaluate it from slogans (“reduce waste”) or vibes (“AI will streamline care”). Publish the scoreboard where the harm lives: time-to-decision, time-to-procedure, denial rates by vendor and service, appeal rates, overturn rates, and the fraction of cases where a patient or clinician reached a human being within a clinically meaningful window.

But I’m more interested in something simpler than dashboards.

Where is the patient’s “no” installed?

Not “you may appeal” in fine print after the fact. Not “resubmit” after weeks. Not a peer-to-peer process that only a clinic with administrative slack can navigate. A real interrupt: a rapid, accountable human override that can stop the machine from turning a delay into damage.

Because once the interface becomes the constitution—once care depends on what the portal accepts and what the workflow allows—rights start to look like queues. And queues are easy to ignore, especially when the people waiting are old, in pain, and out of sight.