The Data Center Dine and Dash

Welcome to Product Cocktail, where the takes are as polarizing as a shot of Fernet—but the insights come together like a perfectly crafted daiquiri.

I met my now-wife just as COVID was starting to percolate in Seattle in early 2020 (#memz). By happenstance (a cautionary quarantine en route to visit her parents), in December of that same year, I visited her childhood summer escape—a family vacation home in Niantic, CT.

Niantic is a beautiful, quiet beach community roughly equidistant from Boston and New York, without the insanity of the Hamptons or Nantucket, but with one unique aspect: a stunning view of the Dominion Millstone Power Station across Niantic Bay in Waterford, CT (not to be confused with the glass Christmas ornaments of Irish origin).

The nuclear power plant sitting on the horizon is somewhat of a punchline, and largely fades into the background amidst the chilly-to-downright-balmy swims in the ocean, visits to the neighborhood market for Italian ice, and the kitsch of neighborhood Fourth of July parades.

Little did I know that during the last few years, a conflict had been brewing involving that very reactor.

In 2023, Developer NE Edge and Dominion Energy proposed a 300 megawatt hyperscale data center sitting on 5% of the Millstone property, plugged directly into the existing nuclear reactors. They positioned this as an opportunity to create "a ready to market, non-carbon solution to the pressing needs of the regional Cloud infrastructure." In exchange for the privilege, they would commit to a $231M host fee over 30 years.

So began a three year fight by the Concerned Citizens of Waterford and East Lyme (where Niantic sits).

These aforementioned concerned citizens told NE Edge and Dominion to GFY, and they won: the project died in March 2026, when NE Edge missed its deadline to secure permits and defaulted on the host agreement.

They're not outliers. In a Gallup poll taken in March 2026, 71% of Americans opposed an AI data center in their area, compared with 53% who opposed a... nuclear power plant.

In May 2026, Anthropic took over SpaceX's Colossus 1 site, using the additional inference to bump Claude Code usage limits in response to customer complaints. How much capacity? More than 300 megawatts.

Generally speaking, the foundational model companies make huge margins on inference (70%+) and lose money building infrastructure.

The buildout is increasingly being passed on to the customers at the frontier side of the equation. Cheap AI is still cheap. In fact, it's technically cheaper... in a "better bang for your buck" sense.

There are two curves running in opposite directions: fixed capability and frontier.

When measuring a fixed capability level, the prices have collapsed: "the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year."

Epoch estimated that OpenAI's o3 achieved 75% on the GPQA Diamond benchmark (PhD-level science questions) at $0.30 per question in January 2025. Less than 18 months later, GPT-5.6 Luna (one of their budget models) matched it for $0.0004—"a 725-fold drop." Last year's frontier work is essentially free 99.

Interestingly, the fastest drop is just behind the frontier. Epoch calculates a 66% per quarter drop in the cost of performance near the state of the art, slowing to 32% per quarter (4.7x/yr) two years after that level is first reached.

OpenAI and Anthropic are still cutting the prices on their cheaper models. Haiku 5.5 is 10x cheaper than Haiku 4.5 on paper (closer to 7–8x, since its new tokenizer counts the same text as more tokens); gpt-6-sol costs less than half what gpt-5.5 did; and gpt-6-luna matches Haiku 5.5 to the penny, at $0.10 / $0.50.

At the frontier, the labs have cranked it up to 11 in 2026.

OpenAI: 5–8x
Aug 2025 — GPT-5: $1.25 / $10
Sep 2026 — GPT-6-Astra: $10 / $50

Anthropic: 2x
Nov 2025 — Claude Opus 4.5: $5 / $25
Jun 2026 — Claude Fable 5: $10 / $50

Google: 2x, same model
At launch — Gemini 4 Argon: $2 / $10
Dec 1 — Gemini 4 Argon: $4 / $20

Price per million tokens, input / output.

Google provides an interesting counterpoint. Flash (their mid-range model) prices have gone up in every generation:

Google Flash: up every generation
Gemini 2.5 Flash: $0.30 / $2.50
Gemini 3 Flash: $0.50 / $3
Gemini 3.6 / 3.8 Flash: $0.75 / $3.75
Jan 1 — Gemini 3.6 / 3.8 Flash: $1.50 / $7.50

Price per million tokens, input / output.

What's the implication?

The spread has exploded between Corolla and Cayenne quality inference. Now, Fable costs 100x as much as Haiku 5.5. In November 2025, Opus 4.5 cost 5x as much as Haiku 4.5. Astra: 100x gpt-6-luna. Before: GPT-5 cost 25x gpt-5-nano.

Each task has turned a casual summer hot dog into a Joey Chestnut token-eating contest. Agents use four times as many tokens as chat interactions. Multi-agent systems use 15x as many tokens. At I/O in May, Google announced they were processing 3.2 quadrillion tokens per month, up 7x year over year. (It raises the question: how many tokens would it take to run a Delorean-based time machine?)

This is a fairly opaque process for the user. Unlike buying 66 hot dogs (approximately $395.34 + tax at Coney Island), users aren't shown the price upfront. The token count and cost show up after the fact. With reasoning models, you don't even see everything you pay for: per Anthropic's docs, "You are billed for the full thinking process, not the thinking content visible in the response."

In spite of this, customers can't get enough tokens.

Amy Hood (Microsoft): "Customer demand continues to exceed available capacity."

Anat Ashkenazi (Alphabet): "We're still in a supply-constrained environment." (Welp, guess this didn't age well: "Google is the only player that isn't compute-constrained: they make the silicon.")

Andy Jassy (Amazon): "We will still not have enough capacity to meet all the demand we have in 2026... this dynamic will also be true in 2027."

Demand outrunning supply is the textbook setup for a price increase, and the frontier is where it's showing up.

It's not quite as dramatic as the "end of the Millennial Lifestyle Subsidy" that we saw in the early 2020s with Uber, DoorDash, WeWork, et al., but the fully loaded frontier model cost is starting to leak through to users. Buyer, beware.

Even if you haven't felt the cost pressure yet as a developer (do you even tokenmaxx, bro?), the cost has shown up to at least one group: gamers.

If you thought the rent was too damn high, try paying five hundred twenty five thousand six hundred dollars for a stick of DDR4.

That's MBA-speak for "if you were planning on buying a PS5 or building a PC anytime soon, you are COOKED, son."

To wit, in 2025, Micron quit selling consumer RAM and SSDs, citing that "the AI-driven growth in the data center has led to a surge in demand for memory and storage," and that it was exiting "to improve supply and support for our larger, strategic customers." (Shout out to one of my nicest clients back at Uncle D. Sorry that my direct-to-consumer strategy didn't work out for you 😅.)

In April, PlayStation raised the price of their consoles $100-$150. Nintendo kicked up the cost of the Switch 2 in May by fifty bucks. The last domino fell when Xbox followed suit in the summer, citing "console storage and memory prices [increasing] by more than 2.5x" to explain a $100-$150 increase. Back in my day, late generation consoles got cheaper. (None of this was attributed directly to AI, but... 🙃)

Spoiler alert: gamers aren't the only AI-outsiders feeling the squeeze.

Data centers have raised wholesale prices about 7% nationwide (Dallas Fed), but so far the only place they clearly show up on household bills is the PJM grid (13 states + D.C.). Everywhere else, the cost is spread too thin to see... yet.

And most importantly, the communities sitting next to data centers are particularly impacted. The ~4M people who live within a mile of the 550 EPA-regulated data centers are disproportionately non-white, with above-median air pollution. In the Memphis area, xAI installed 27 unpermitted gas turbines powering Colossus 2, triggering an NAACP suit. Unfortunately, we've seen this pattern of environmental racism before.

Once hidden costs show up, voters, politicians, and regulators push them back onto whoever is writing the checks, and often there's a quiet retreat.

Towns shitting on data centers is the new black. Fewer people support building AI data centers in their area than approve of President Trump.

Data Center Watch identified 75 projects worth $130B "blocked or delayed" in a single quarter (Q1 2026), while opposition groups grew from 396 to 833.

Even the AI industry's home turf—SF and Oakland—passed 45-day moratoriums on new data centers. Nationally, there are 35 local ballot measures concerning data center development in the November 3rd midterm elections.

Erin Brockovich (yes, that one) recently cited transparency as the number one complaint she's heard about data centers.

Bessemer, Alabama's mayor signed an NDA with data center developers requiring the city to destroy "confidential" information about the project—including notes city officials made—raising serious public-records concerns. Voters ousted him after a 16-year-tenure in a September 22, 2026 runoff election.

Eleven days after, Amazon dropped NDAs. (Keeping up with the Joneses is a key part of their strategy, apparently.) AWS CEO Matt Garman, in full retreat: "We no longer use nondisclosure agreements with the government agencies we work with on our projects." Past tense, quietly.

In fairness, industry research firm SemiAnalysis claims this hasn't moved the needle in measurably delaying already-funded hyperscaler capacity; they only moved their year-end forecast ~1% in response to the first two quarters of opposition efforts.

Cheap AI isn't over, but other stakeholders picking up the tab might be.

Costs have started making their way back to the people who decide to spend: AI customers through frontier price hikes, and data center developers through "no" votes, moratoriums, and dropped NDAs. Meanwhile, gamers, ratepayers, and the communities next door have been paying all along.

The cost was there all along, it just became more visible.

Issue 14 asked "What's on the tab?" This one is about who loses credit card roulette. For years, it was investors, gamers, ratepayers, and towns. Increasingly, it's whoever buys the inference, and that comes as an increased cost or a "no."

Send me questions, feedback, and cocktail recipes:
[email protected].

_{Icons made by Icongeek26 from www.flaticon.com.}