/pages/posts/anti-scaling-pro-ai
The Pro/Anti AI Discourse Is Rotten, The Real Fight Is Elsewhere

The Discourse Is Rotten.

The Real Fight Is Elsewhere.

If you’ve spent any time online in the past three years, you’ve seen the script. AI will either save humanity or destroy it. It’s either the greatest creative tool ever built or the death of art. The rhetoric has calcified into two warring camps, each more interested in moral posturing than in understanding what’s actually happening. Meanwhile, the people building the infrastructure are laughing all the way to the bank, because while you’re fighting about whether a diffusion model can make “real” art, they’re securing permits to pump billions of gallons of water, burn fossil fuels in residential neighborhoods, and build the surveillance backbone that will outlast every argument you’ve ever had about neural networks.

This essay isn’t about whether AI is “good” or “bad.” That framing is broken beyond repair. I think AI is an extraordinary medium, and I think artists who use it are making art, full stop. I also think the environmental destruction happening right now, in service of compute demands that have almost nothing to do with the public-facing AI tools people actually use, is one of the most underreported stories in tech. These two beliefs don’t contradict each other. The contradiction is in a discourse that forces you to pick a side between “AI is magic” and “AI is evil,” while the actual power brokers build unpermitted power plants and buy $2.8 billion worth of new gas turbines.

What Actually Eats the Compute

Here’s a question almost nobody in the pro/anti AI debate asks: what are all these data centers actually for? The public conversation assumes they’re training the next frontier model, pushing toward some inevitable AGI horizon. And some of them are, sure. But the overwhelming majority of global compute capacity isn’t going to open-ended research. It’s going to surveillance, ad targeting, behavioral tracking, and the perpetual motion machine of personal data extraction that has defined the internet economy since before LLMs existed.

The numbers are staggering, and they predate the current AI boom by more than a decade. A 2007 EPA report to Congress found that US data centers already consumed 61 billion kilowatt-hours annually — 1.5% of total US electricity consumption — and projected that figure to double by 2011.¹ By 2016, a Lawrence Berkeley National Laboratory study put the number at 70 billion kWh, and that was before the modern transformer era.² These facilities weren’t running diffusion models. They were tracking your clicks, building ad profiles, storing your location history, and processing the endless telemetry that surveillance capitalism demands.

Today, the surveillance infrastructure has only metastasized. Flock Safety operates over 80,000 AI-powered cameras across 5,000+ municipalities in 49 states, performing more than 20 billion vehicle scans every month.³ The company invests over 25% of its revenue into AI and machine learning R&D, not for artistic purposes, but for license plate recognition, gunshot detection, predictive policing, and automated evidence linkage.⁴ Flock’s network integrates directly with platforms like Palantir, creating a seamless pipeline from street-level surveillance to federal intelligence analysis. This is where your compute is going. Not to curing disease. Not to solving climate change. To tracking whether your car visited a particular neighborhood and flagging you for follow-up.

The “AI data center” boom isn’t primarily a story about training large language models. It’s a story about scaling up the infrastructure of control — and wrapping it in language so vague and future-oriented that communities don’t realize what’s being built next door until the generators are already running.

Environmental Destruction in Real Time

The Potomac Conservancy published an open letter in April 2024 documenting how a data center project in Frederick County, Maryland — the Quantum Loophole development — had racked up “numerous ongoing violations of environmental regulations,” including a stop-work order from county inspectors “due to environmental hazards.”⁵ The violations included sediment and pollution discharge into waterways, the kind of damage that doesn’t make headlines but permanently alters local ecosystems.

This is not an isolated case. It’s a template.

In Memphis, Tennessee, xAI’s Colossus data center has become the defining case study in corporate impunity. The company set up dozens of unpermitted gas turbines just south of the Tennessee-Mississippi border, creating what the Southern Environmental Law Center describes as a “de facto power plant” that releases “smog-forming pollution, soot, and hazardous chemicals like formaldehyde” with no permits, no public input, and no notice to nearby communities.⁶ When the NAACP filed a lawsuit last month seeking an injunction, xAI had been granted permits for 15 turbines but was actively operating 46.⁷

The company’s response? According to SpaceX’s IPO filing, xAI will buy $2.8 billion worth of natural gas turbines over the next three years, including $2 billion specifically for “mobile gas turbines” — the exact type it’s being sued over.⁷ This isn’t a company scrambling to comply. This is a company that has calculated the cost of lawsuits and environmental penalties and decided they’re cheaper than doing it right.

The localized health impacts are only beginning to be documented. A 2026 study in Frontiers in Climate — the first of its kind — examined Virginia’s “Data Center Alley” and found evidence linking data center operations to respiratory illness, cardiovascular disease, waterborne illness risk, and mental health impacts from chronic noise exposure.⁸ An MIT study published the previous year documented how communities near Northern Virginia’s more than 450 data center facilities experience “increased electricity costs, degraded power quality, and the infrastructural strain of hosting the world’s largest concentration of digital infrastructure.”⁹ The researchers note that the benefits — tax revenue for states, massive demand for utility companies — “often do not extend to surrounding communities and certainly do not outweigh negative local impacts.”¹⁰

The environmental justice dimension is hard to ignore. Data centers are disproportionately sited in low-income communities and communities of color, places where residents lack the political capital to fight permitting decisions. The 2020 paper Recalibrating Global Data Center Energy-Use Estimates found that the industry’s carbon footprint had been systematically underestimated, in part because the localized emissions from backup diesel generators — which run during peak demand and outages — were poorly tracked.¹¹ When your business model depends on running diesel engines in residential neighborhoods, you don’t want that data to be easy to find.

The China Counter-Narrative

The hyperscaler narrative goes like this: the United States must build as fast as possible, environmental concerns be damned, because China is coming for AI dominance and whoever scales compute first wins. This is convenient messaging if you’re trying to secure tax breaks, fast-track permits, and billions in infrastructure investment. It is also, by almost every measurable metric, false.

China is not winning the AI race by building more data centers. It’s winning by building better ones, powered by cheaper and cleaner electricity, and by developing models that achieve comparable performance at a fraction of the compute cost.

DeepSeek’s V4-Pro, released in April 2026, is a 1.6 trillion parameter Mixture-of-Experts model — the largest open-weight model ever shipped. But only 49 billion parameters are active per token, meaning the vast majority of the model’s capacity sits idle during any given inference pass.¹² The result is frontier-level performance at inference costs that are difficult to fully appreciate: in May 2026, DeepSeek made a 75% API price cut permanent, bringing the cost to approximately $0.435 per million input tokens.¹³ For comparison, GPT-5.5 costs roughly $15 per million input tokens — DeepSeek is approximately 34 times cheaper for comparable tasks.¹⁴

The model is released under an MIT license. Anyone can download, modify, and deploy it. There are no API gates, no usage tiers, no corporate permission structure. The weights are on Hugging Face. The training methodology is documented. This isn’t a corporate product — it’s a public resource that happens to outperform models costing orders of magnitude more to run.

On the infrastructure side, China’s advantage is structural and growing. Chinese data centers pay less than half the electricity rates of their American counterparts, with state-subsidized wind and solar feeding hyperscale facilities via dedicated transmission lines.¹⁵ The country’s “Eastern Data, Western Computing” initiative relocates energy-intensive compute to western provinces with surplus renewable generation, while the US grid struggles with interconnection queues that can stretch for years.¹⁶ By 2030, China is projected to have roughly 400 gigawatts of spare power capacity — triple the expected global data center demand.¹⁷ The US, meanwhile, faces a projected 44-gigawatt electricity shortfall for data centers in the next three years.¹⁸

The efficiency gains aren’t limited to DeepSeek. Chinese labs across the board have embraced sparse architectures, aggressive quantization, and context caching that reduces compute by 75-90% on repeated queries.¹⁹ The result is a national AI ecosystem that processes more queries, trains more capable models, and does so with a smaller environmental footprint per unit of intelligence delivered. The hyperscaler argument that the US must sacrifice its environment to stay competitive collapses the moment you look at how the actual competition is operating.

The Actual Doomsday Scenario

If you want something to be genuinely afraid of, forget the paperclip maximizer.

In March 2026, Sam Altman stood before an audience of infrastructure investors at BlackRock’s summit 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.”²⁰ The backlash was immediate — commenters pointed out that utilities are regulated, low-margin businesses, which isn’t exactly the business model that supports a $300 billion valuation.²¹ But the backlash missed the deeper implication. Altman wasn’t accidentally revealing a business model. He was describing a worldview in which knowledge itself becomes a metered resource, distributed through proprietary channels controlled by a handful of corporations.

This is the doomsday scenario that could actually happen. Not AGI killing humanity — which requires a chain of speculative assumptions about emergence, agency, and physical capability that don’t hold up to scrutiny. But the gradual, piecemeal enclosure of human knowledge behind API gates and proprietary abstraction layers? That’s already underway, and it accelerates with every data center that comes online.

Consider the trajectory. CUDA and ROCm are already de facto gatekeepers — if you want to do high-performance computing, you run Nvidia’s stack or you run into walls. The major cloud providers increasingly offer AI not as software you can inspect and modify, but as a remote service you rent by the token. The models are black boxes. The training data is proprietary. The weights are “open” only when it serves a marketing narrative. And every year, the stack gets taller and the base gets further from anything an individual can understand, repair, or replicate.

The Altman vision takes this to its logical conclusion: intelligence as a utility bill. You don’t own the model. You don’t understand the model. You can’t run it without permission. You pay for access, and if the price goes up or your account gets suspended or the provider changes terms, your access to what has become essential cognitive infrastructure disappears. The historical parallel isn’t Skynet. It’s the enclosure of the commons — the transformation of shared resources into private property, mediated by technology just complex enough that fighting back requires expertise most people don’t have time to develop.

This is why the “pro-AI” position isn’t about cheering for corporations. It’s about recognizing that AI as a technology is not the threat. The threat is the concentration of AI infrastructure in unaccountable hands, the environmental devastation required to sustain that concentration, and the deliberate construction of technical and economic barriers that prevent individuals and communities from owning their own tools. A teenager can download DeepSeek V4-Pro and run it on consumer hardware. The same teenager cannot build a data center, negotiate a power purchase agreement with a utility, or get permits for industrial cooling infrastructure. The gap between those two capabilities is where the real power lies.

Where to Actually Direct Your Energy

The discourse has trained people to treat AI as a monolith — you’re either for it or against it. This is by design. It collapses a vast landscape of technical, economic, and political questions into a single binary, which makes it easy to mobilize outrage and very difficult to build coalitions around specific, winnable fights.

The fights that matter are specific and winnable. Oppose data center expansion in residential areas. Show up to zoning hearings. Support the NAACP’s lawsuit against xAI. Demand that environmental impact assessments be required before permits are granted. Support right-to-repair legislation that keeps hardware accessible and understandable. Pressure municipalities to reject Flock Safety contracts. These are concrete actions with concrete outcomes, and they matter far more than any position you take on whether AI-generated images count as art.

If you’re worried about surveillance, fight the cameras and the data brokers, not the diffusion model. If you’re worried about environmental destruction, fight the unpermitted generators and the water withdrawals, not the local artist using Stable Diffusion. If you’re worried about corporate concentration, fight the proprietary API gates and the infrastructure monopolies, not the open-weight model anyone can download and run at home.

The astroturfed discourse wants you exhausted and distracted, fighting proxy battles while the actual infrastructure gets built. Don’t give it what it wants. The most radical thing you can do is refuse the frame — and fight the things that are actually happening.

EPA Report to Congress on Server and Data Center Energy Efficiency (2007) ↩
United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (2016) ↩
Flock Safety Solutions Overview — 80,000+ cameras across 5,000+ communities, 20 billion+ monthly scans ↩
Flock Safety Inc. Form S-1, SEC Filing (2025) — 25%+ revenue into AI/ML R&D ↩
Potomac Conservancy: A Maryland data center project has violated environmental protections (2024) ↩
Southern Environmental Law Center: xAI built an illegal power plant to power its data center (2026) ↩
TechCrunch: Musk’s xAI is being sued over its data center generators — now it’s buying $2.8B more (2026) ↩ ↩²
Frontiers in Climate: Health implications of the rapid rise of data centers in Virginia (2026) ↩
MIT Climate & Sustainability Consortium: Investigating the Ecological Impacts of Data Centers (2025) ↩
Ngata et al., The Cloud Next Door: Investigating the Environmental and Socioeconomic Strain of Datacenters on Local Communities, MIT/UMass (2025) ↩
Masanet et al., Recalibrating global data center energy-use estimates, Science (2020) ↩
DeepSeek-V4 Technical Report (April 2026) — 1.6T MoE, 49B active parameters, MIT license ↩
DeepSeek API Pricing — $0.435/million input tokens after 75% permanent price cut ↩
DeepSeek-V4: Capabilities, Pricing, and How It Compares to GPT-5.5 (Intuition Labs, 2026) — 34x cheaper than GPT-5.5 ↩
Al Jazeera: Can China lead the world in AI while going green? (2025) ↩
Brookings Institution: China’s new energy superpower path (2025) ↩
Fortune: China races to power AI data centers with renewables (2025) ↩
Morgan Stanley Research: US faces 44-gigawatt data center power shortfall (2025) ↩
Intuition Labs: The DeepSeek Effect — How Chinese AI Labs Are Out-Innovating the West (2025) ↩
Quartz: Sam Altman says OpenAI’s future is selling intelligence by the meter (2026) ↩
Business Insider: Sam Altman wants OpenAI to be a utility (2026) ↩

/pages/posts/anti-scaling-pro-ai The Pro/Anti AI Discourse Is Rotten, The Real Fight Is Elsewhere