Uh oh… Sid is about to rant about politics…
This morning, Axios published a leaked memo from the National Republican Senatorial Committee — the Senate GOP’s campaign arm — privately warning AI companies that data centers have become, and I’m quoting their own internal polling characterization here, as popular as spent nuclear waste. The memo frets that Sherrod Brown has made data centers his “de facto opponent” in Ohio, that Jon Husted is down eight points wearing the “face of data centers” like a millstone, and that if voters’ perceptions aren’t fixed quickly, “the campaign against them will expand far beyond Ohio.”
Meanwhile, Governor Shapiro — a man with a 2028 calendar on his desk and no intention of ignoring it — signed an executive order slapping “strict guardrails” on data center development in Pennsylvania. There are over 4,000 data centers online in this country and more than 3,000 proposed. And the entire national conversation about them has collapsed into two positions, both of which are wrong, and neither of which would survive ten minutes in a room with an actual practitioner.
Position one: AI requires lots of data centers. Data centers are bad. Ergo, AI is bad. This is the environmental-backlash syllogism, and it has the elegant simplicity of all arguments that skip the hard part.
Position two: AI needs GPUs and data centers. No data centers means China wins. AI good. Go USA. This is the national-security syllogism, and it has the elegant simplicity of a bumper sticker written by a lobbyist.
Here’s the thing about zero-sum framings: they’re seductive precisely because they relieve you of the obligation to understand the system. And I say this with love, because I’ve now watched a party campaign committee, a governor’s office, a Senate race, and roughly four hundred LinkedIn thought leaders argue about megawatts without once asking the question that actually determines the megawatts.
Which is: what are the models doing with all that power?
The game already changed. Nobody sent the memo to the memo-writers.
Let me be clear about where we actually are. The model capability plateau is real. Not “the models stopped improving” — they haven’t — but the marginal returns on raw intelligence have flattened to the point where the competitive game has moved. The question is no longer “how smart will the model be.” The question is how efficiently and accurately a model can address large volumes of enterprise data. That’s the game. That’s been the game for a while now. The frontier labs know it, which is why the interesting releases lately are about context, tools, and agentic reliability — not IQ points.
And here’s the part that should reframe the entire data center debate: efficiency per task is improving at a pace the energy sector has never seen anywhere, for anything. The IEA — not a booster organization, not a lobbyist shop — reported this year that power consumption per AI task is declining at a rate unprecedented in the history of energy. Google published a full-stack measurement showing the median Gemini text prompt consumes 0.24 watt-hours — about nine seconds of television — after a 33x reduction in energy per prompt and a 44x reduction in carbon in a single year.
Thirty-three times. In twelve months.
Name me another energy-consuming technology in human history that did that. The steam engine took a century to triple its efficiency. Aviation took fifty years to halve fuel per passenger-mile. I’ll wait.
So why is total data center electricity demand still climbing — up 17% last year, with AI-focused facilities up 50%, on a path from roughly 415 TWh in 2024 toward 945 TWh by 2030? Because we are burning the efficiency gains on waste. Not on ambition. On waste. And the waste has a specific, identifiable, fixable engineering signature that nobody in Ohio, Harrisburg, or the NRSC’s memo shop is talking about.
The most expensive way to read an accounting ledger
I’d bet that upwards of 90% of the agentic prompts running in enterprises today start the model off with limited or no tools — so the model does the only thing it can do. It codes its way to success. It writes Python to connect to the database. It writes more Python to parse what came back. It writes a third round of Python to fix the second round. Billions of tokens, spent on boilerplate connection logic, to answer a question a well-built tool could have answered in one call.
That is the most expensive possible way to get a simple result out of an accounting ledger. It is also — and I want to be fair here — a spectacular source of technical debt, because every one of those ad-hoc scripts is now load-bearing in someone’s workflow and documented nowhere.
The other day I listened to CIOs jubilantly describing how their users were “just pointing the model at the database” and getting amazing results. Amazing! And I sat there doing the thing where you smile politely while your soul files a formal objection. Because we know — we have the receipts — what happens when you point a raw model at a raw enterprise schema. Spider 2.0, the enterprise-grade text-to-SQL benchmark, showed models that score above 90% on academic toy databases collapsing to success rates in the single digits and low twenties on real enterprise schemas — GPT-4o managed roughly 6%, and even the best reasoning models landed around 17–21%. The dominant failure mode? Schema linking — the model simply picking the wrong tables and columns — accounts for the largest share of errors, because a thousand-column warehouse without an ontology is not a data source, it’s a Rorschach test.
So when I say results from untooled database access run wildly inaccurate — the figure I use is up to 70%, and the benchmarks suggest I’m being generous — I’m not being a curmudgeon. I’m reading the literature. The CIOs celebrating those “amazing” results are celebrating answers nobody has checked, produced by a system that fails most of the time on exactly this task, at maximum token cost. That’s not a productivity revolution. That’s a very enthusiastic random number generator with a power bill.
This, incidentally, is what I work on every day on the AI Data Platform: vending the right data to the model at the right time, with the ontology and the tooling that turn a guess into an answer. It is unglamorous. It is also where the entire environmental and economic case for AI now lives.
Yes, about China
The efficiency-model argument deserves an honest treatment, because the “China wins” crowd waves it around and the “AI bad” crowd ignores it. China’s labs — DeepSeek foremost among them — have done genuinely impressive work squeezing near-frontier capability out of constrained hardware. Trained for a fraction of the cost, on export-compliant chips, with architectural cleverness born of necessity. Respect where it’s due.
But low-energy models come at a cost, and the cost is ambition and creativity. Combine that with what I’ve written about at length — the average problem, agentic systems defaulting to the statistically average path because that’s what the training data rewards — and you get models exquisitely suited to high-volume tasks where average is genuinely good enough. Summarization at scale. Classification. Routine transformation. That’s a real market and China may well own it.
The real productivity boost — the part that shows up in GDP rather than in press releases — comes from the larger models. And those are the ones “claiming” to need the computing power. But interrogate that claim! The primary reason the big models burn so much compute isn’t that reasoning is inherently ruinous. It’s that we hand them no tools and force them to code their way to every answer, then we act shocked at the electric bill. We built a Formula 1 car, gave it no map, and are now debating whether racing itself is immoral because the driver keeps circling the parking lot.
The tool is a form of prompt engineering
If the industry wants to actually solve this — the cost problem, the accuracy problem, and yes, the data center problem the NRSC is losing sleep over — the answer lies in tools, and in improving how models reach for tools.
And here I have to sigh at my own industry, because our first attempts were weak. MCP and A2A arrived feeling like thin wrappers around RESTful endpoints — interfaces that are semantically miserable for a model to consume. And the results were predictable: tool definitions bloating context windows before a single user message arrives. Anthropic’s own testing found a modest five-server, 58-tool setup consuming ~55,000 tokens of pure overhead — a quarter of a context window, gone, on descriptions of tools the model mostly won’t use. Practitioners have measured setups where tool schemas eat 30–50% of available context. The ecosystem is now busily inventing tool search, progressive disclosure, and code-execution modes to claw back what a better-designed interface never would have surrendered.
Remember: the tool is a form of prompt engineering.
Every tool description is instructions to the model. Every schema is a semantic contract. When you expose an endpoint whose interface was designed for a REST client rather than for a mind that reasons in language, you are writing a bad prompt and shipping it as infrastructure. Semantically legible tools — tools shaped around what the model needs to understand, backed by an ontology that tells it what the data means — are the difference between one accurate call and forty rounds of speculative Python.
Do that, and the token math changes. The energy math changes. The data center math changes. Not by decree, not by executive order, not by a media campaign explaining “who benefits, who pays, and why a community should want one” — by engineering.
The actual economic lever
So I sigh, because the argument playing out in Ohio has an undertone of politics and valuation, and almost no undertone of engineering. One side needs data centers to be a monster; the other needs them to be a patriotic necessity; both need them to be simple. And the actual economic lever of this era is no longer mythical AGI, which was always a valuation story wearing a lab coat. The lever is tooling. Boring, unglamorous, ontology-having, semantically-designed tooling that lets models do more with radically less — less compute, less energy, less hallucination, fewer data centers per unit of actual productivity.
The IEA has already shown per-task efficiency improving faster than anything in the history of energy. The benchmarks have already shown where the waste lives.
The path to environmentally defensible AI isn’t fewer models or smaller ambitions. It’s high-efficiency tools that vend the right data to the right model at the right time.