The Bill Nobody Sees: On AI, Classism, and Who Really Pays

I noticed it first on a Tuesday, doing nothing important.

I was scrolling through Instagram — half-awake, half-present, the way most of us exist online — when a feature I'd used without thinking suddenly sat behind a prompt. Upgrade to Instagram+. A small thing. A minor inconvenience. The kind of thing you're supposed to shrug off and move on from.

I didn't move on.

Something about it nagged at me the way things do when they're pointing at something larger than themselves. I kept scrolling. I kept thinking. And somewhere between that paywall and a reel about heat maps over India, something in my brain started pulling threads.

I should tell you upfront: I use AI. Every day, probably. I find it useful, sometimes remarkable, occasionally genuinely moving in what it can do. I'm not writing this from the outside, from some position of clean-handed critique. I'm writing this from the inside — as someone who benefits from these tools and still can't shake the feeling that something is deeply, structurally wrong with how this is all unfolding.

If it stayed with me — just a person paying attention — imagine what it means for the people it's actually happening to.

The Meter Is Running

In March 2026, Sam Altman stood at the BlackRock Infrastructure Summit in Washington DC and said the quiet part out loud.

"We see a future where intelligence is a utility like electricity or water, and people buy it from us on a meter and use it for whatever they want to use it for."

He wasn't being sinister. He was being honest about the direction things are heading — a world where cognition itself is commodified, where the ability to think faster, search better, write cleaner, decide smarter is something you pay for by the unit. Like electricity. Like water.

The comparison is meant to sound democratic. Utilities are things everyone has access to, right? Everyone gets water. Everyone gets power.

Except — do they?

If you grew up in India, you already know how that story actually goes. Power cuts that hit some neighbourhoods and not others. Water tankers that arrive in some colonies and not others. The meter model has never been neutral. It reflects existing inequalities rather than flattening them. The person who uses more and can afford more gets more. The person at the margin either adapts or goes without.

Now apply that to intelligence. To the tools that help you write a better CV, understand a legal document, navigate a system that wasn't designed with you in mind. If those tools move entirely to a pay-per-use model, priced by consumption, accessible to whoever can afford the tokens — we haven't democratised cognition. We've just found a new axis to sort people by.

This is not a hypothetical. It is already happening, one paywalled feature at a time.

The Backyard Deal

Around the same time Altman was speaking in Washington, a California startup called Span was announcing a partnership with Nvidia that I found quietly astonishing.

The idea: install a compact AI compute node — called an XFRA unit, roughly the size of an HVAC box — outside residential homes. The node uses your home's excess electrical capacity to run AI workloads for cloud providers. In exchange, homeowners receive discounted electricity and internet bills, or potentially a flat arrangement of around $150 a month.

The framing is neighbourly. Helpful, even. You have spare electricity. We need compute. Let's work something out.

But sit with it for a moment. The company deploys this infrastructure six times faster and at one-fifth the cost of building a conventional data centre. The efficiency gains are enormous — and they accrue entirely to Span and its clients. The homeowner gets a discounted utility bill. The AI economy gets distributed, cheap, scalable compute embedded into residential neighbourhoods.

Who captures the value? Who absorbs the infrastructure? Who is being asked to host — physically, electrically — the backbone of a cognitive economy they may or may not be able to fully participate in?

This is being piloted in suburban America. But the logic — of offloading cost and infrastructure onto the people least positioned to negotiate — is not unique to America. It is the oldest logic in the book.

What My Sustainability Class Knew

There's a term for this at the macro level. Carbon leakage.

It describes what happens when developed nations, facing strict environmental regulation and growing public pressure to go green, shift their emission-intensive activities to countries where regulation is lighter and labour is cheaper. Their sustainability metrics look clean. The pollution just moved. The only winner, researchers note, is the public image of the institution that outsourced the damage.

The United States, Japan, and many Western European nations have managed to outsource significant portions of their carbon emissions this way — the actual production happening somewhere else, the consumption and the credit remaining at home.

I learned this in a sustainability elective in my MBA. It stayed with me because it named something I’d felt but couldn’t articulate: the way "progress" in one part of the world can be precisely contingent on extraction in another. The way clean hands in one place require dirty hands somewhere else.

Now look at data centres. India's data centre capacity has already expanded past 1,500 MW and is projected to reach 6.5 GW by 2030, fuelled largely by the global demand for AI compute. More than half of India's existing data centres already experience temperatures above 35°C for over 90 days a year. A conventional 100 MW data centre consumes nearly 2 million litres of water per day — in a country already dealing with severe water scarcity and accelerating heat stress.

Meanwhile, in 2026, India's heatwaves are arriving earlier and running hotter than any recorded season before them. Roughly 380 million people — three-quarters of India's workforce — labour in heat-exposed sectors. Those who can afford air-conditioned homes, cars, and offices move through the crisis insulated. Those who cannot bear it on their bodies.

The AI economy needs cooling. It is consuming water and power in one of the hottest, most water-stressed countries in the world. The global tech industry calls this "digital infrastructure development." The people living next to these facilities — and the farmers whose groundwater is shrinking — might use different words.

This is carbon leakage's next chapter. Call it compute leakage. The benefits of AI concentrate in the Global North. The heat, the water draw, the grid strain — that gets distributed to places that were not consulted.

The Invisible Workforce

Here is the part that should be impossible to ignore, and somehow largely is.

AI systems do not become "safe" through code alone. They become safe because thousands of human beings sit in front of screens for hours every day and label what is harmful, what is violent, what is too dangerous for a model to reproduce. They are shown the worst of what humans produce — child abuse, graphic torture, sexual violence, beheadings — and asked to categorise it, so that the model learns what not to say.

Researchers who interviewed 113 data labelers and content moderators across Kenya, Ghana, Colombia, and the Philippines documented over sixty cases of serious mental health harm — PTSD, depression, insomnia, anxiety, suicidal ideation. Workers reported panic attacks and symptoms of sexual trauma, without access to mental health support, under pressure to meet punishing productivity targets.

In Jharkhand and Uttar Pradesh in India, female workers viewed up to 800 violent clips each shift. They earned between £260 and £330 a month. They signed NDAs that prevented them from discussing what they saw — even in therapy.

One worker described what happened over time: "By the end, you don't feel disturbed — you feel blank." The trauma doesn't announce itself immediately. It surfaces in the quiet hours, in dreams, in a flatness that wasn't there before.

The outsourcing model exports the psychological damage of AI development to populations with the least access to mental healthcare. When a company announces its model is "safe" — when it runs ads about responsible AI, when its CEO speaks at summits about the future being bright — it rarely mentions the people who made that safety possible, or what it cost them.

These workers are not edge cases. They are load-bearing. The entire project depends on them. And they are, by design, invisible.

The Boomerang

Here is a different kind of cost, one that landed closer to home for a lot of people in 2025 and early 2026.

Companies, chasing efficiency and spooked by the pace of AI development, made sweeping decisions to reduce their workforces. The logic was swift and confident: AI can do this now. We don't need as many people.

Gartner now projects that 50% of companies that attributed headcount reductions to AI will rehire for similar functions by 2027 — often under different job titles, as if the rebranding might soften the embarrassment. Forrester found that 55% of employers regret those cuts. 73% of organisations that executed AI-driven staff reductions failed to come out financially ahead. One analysis found a clear and consistent pattern: AI successfully manages about 60% of repetitive workflows. It catastrophically fails at the remaining 40% — the part that requires institutional knowledge, human judgment, contextual nuance, the ability to hold a difficult conversation with a real person having a hard day.

So what actually happened? People lost jobs — real jobs, with real consequences for real families — because of a corporate experiment that didn't account for what it couldn't measure. And now those same companies are quietly rehiring, under new titles, having learned an expensive lesson.

The companies will recover. They have capital. They have the ability to recalibrate.

The workers who were let go in the interim — who missed rent, who took worse jobs, who watched their industries restructure around them — they paid the tuition for that lesson. Nobody's reimbursing them for it.

This is what reckless adoption looks like at human scale.

The Saturation Myth (And Why It Matters Here)

There's a version of this essay that gets dismissed with a familiar move: everyone has access to AI now. This is all available for free. You're overstating the divide.

I want to push back on that, because it’s the same logic applied to creator economy debates. You’ll hear: "don’t become a creator, it’s too saturated." But statistically, only about 1% of the global population creates content. 1%. That’s not saturation. That’s barely a beginning. The saturation is a perception problem, not a statistical one — created by the fact that the people who are creating content are all creating content about content creation, which makes it feel ubiquitous.

Same with AI. It feels like everyone is using it, fluently, confidently, productively — because the discourse about AI is dominated by the people who are already inside it. Tech Twitter. LinkedIn thought leaders. People with reliable broadband, expensive devices, time to experiment, English as a working language, and enough economic security to treat failure as a learning opportunity.

If you zoom out: who has consistent electricity? Who can afford the subscription tiers that unlock the actually powerful features? Who has the foundational literacy to use these tools meaningfully, rather than just prompting badly and getting confused output? Who has the psychological safety to experiment with something new when job security is already precarious?

The gap between "AI is everywhere" and "AI is accessible" is vast. It is not closing as fast as the headlines suggest. And in the meantime, the people bearing the cost of building it are — almost universally — not the people who will most benefit from it.

What This Actually Is

I want to name it.

This is not just subscription fatigue. It is not just a labour market disruption or an environmental externality or a policy failure. All of those things are real, but taken together, they describe something in shape.

The AI economy, as it is currently structured, is a system in which:

The infrastructure — compute, cooling, data centres, power draw — is being built on and in the Global South, consuming resources that are already scarce and worsening conditions that are already dangerous.

The invisible labour — the human annotation, moderation, and psychological exposure that makes AI "safe" — is being performed by workers in developing countries, under NDA, without adequate mental health support, for wages that reflect how little the industry values what they’re doing.

The workforce disruption — the hasty layoffs, the experiments in replacement, the collateral damage — lands hardest on workers with the least cushion to absorb it.

And the benefits — the productivity gains, the cognitive augmentation, the tools that genuinely do make life easier and work better — concentrate among those already positioned to access and leverage them.

When intelligence becomes a utility priced on a meter, who will be able to afford to think?

That's not a rhetorical question. It has an answer. And the answer is who it always is.

The Reckoning I'm Still In

I said at the start that I use AI. I want to come back to that.

I'm not interested in a purity argument. I'm not suggesting we abandon these tools or that anyone who uses them is complicit in something unforgivable. That kind of thinking — all or nothing, clean hands or dirty — is how we avoid actually grappling with complicated things.

What I am saying is that the people making decisions about AI adoption, deployment, pricing, and labour — they have a responsibility that isn't being fully met. And the rest of us, who use these tools and benefit from them and will continue to do so, have a responsibility to keep asking who isn't in this conversation and why.

The reel about India’s heat maps stayed with me. The content moderation research stayed with me. The Gartner report about the rehiring wave — that stayed with me too. None of this is classified information. It’s not buried in academic journals. It’s out there, findable, readable.

It just requires deciding that the people it's happening to are worth paying attention to.

I'm not an expert in AI. I'm not an economist or a climate scientist or a labour rights researcher. I'm someone who reads, and notices things, and couldn't let this go.

I think that's enough of a reason to say it out loud.

If it stayed with me — just a person paying attention — imagine what it means for the people it's actually happening to.