Betwixt and Between

Right now, stunningly capable AI tools are cheap or free for ordinary people. That isn’t the end of the story. It’s a point in time. The direct result of unsustainable investor subsidy. VC capital is covering the gap between what these tools cost to develop and run, and what we pay to use them. This is the bet that subsidized scale today buys dominance tomorrow. It’s a reasonable bet, and a familiar one. Cheap doesn’t last.

“Betwixt and between” is an old phrase for being caught in the middle, neither here nor there. It’s where we are with AI: directly shouldering the costs of the buildout, with likely limited access the resulting benefit. As cliché as it is, we can’t seem to learn… eventually, the bill comes due.

We know this story

The early internet ran on venture money until the dot-com reckoning sorted the (few) real businesses from the rest – a perhaps necessary but nonetheless painful culling of the less capable described as the ‘cost’ of bringing in the information age. Ride-hailing operated below cost for a decade to kill the taxi and train us into the habit, then raised prices once we were hooked and the competition was gone. The pattern is consistent: subsidize adoption, capture the market, then charge what the thing actually costs (plus enough margin to earn an attractive return on all that invested capital) once leaving becomes painful. ‘Cheap’ may describe the customer-acquisition phase, but it won’t be the ongoing business model.

We are deep in the cheap AI phase now. The hyperscalers are spending on the order of hundreds of billions of dollars a year on data centers, funded by a mix of cash, debt, and equity prices that all assume boundless growth. That spending has to earn a return eventually. When it’s time to collect, the price of using AI goes up. The free tier will be the first to go.

Who gets squeezed

When capital starts asking for returns and compute capacity gets tight relative to demand, the subsidized users get squeezed first. Free accounts get throttled or metered. Monthly prices climb. What’s left for the consumer is the choked-down version that makes dancing yetis, while the genuinely powerful stuff stays with the enterprises that can pay for it.

There’s a deeper insult here, one I touched on last week in ‘Empty Calories.’ There is an economic case frequently made for subsidizing infrastructure investment with public resources to spur sustained growth. It would, however, be a stretch to suggest that civilization’s accumulated creative and intellectual output — absorbed wholesale to train these models — was fairly traded for what the courts have now settled at $3,000 a book. This wasn’t a few VHS tapes copied in someone’s basement, but effectively the talent and capability that made it, bought for a token settlement. The enterprises that now own that capability will put it to three uses: drive efficiency (read: fewer white-collar and creative-class jobs), aim their marketing with more precision (more empty calories to consume), and minimize the taxes that would otherwise help pay for the grid their data centers are straining. The gains consolidate to a narrow set of owners and investors that wagered ‘after-the-fact-licensing’ (aka, stealing) content to train the models would usher in the greatest land-grab of all time… $12 for Manhattan may have been expensive by comparison.

The public, on the other hand, gets the career disruption and the utility bills, and the privilege of paying to use what will become an increasingly stratified offering — further amplifying benefit based on ability to pay. Rather than ability.

Due and payable

I spent years developing energy projects: negotiating interconnection with utilities, working the queues, siting generation and storage. This isn’t abstract opinion about the buildout costs landing on the public. To date, grid infrastructure required to support data center capacity has been paid for by rate-payers as fees and taxes.

Start with the queue. More than 2.2 terawatts of generation and storage are stuck waiting to connect to the U.S. grid, nearly double the country’s entire installed capacity, and the median wait from request to operation now approaches five years. Into that gridlock, AI is dumping load at a pace utilities have never seen. In Texas, the large-load interconnection queue hit 226 gigawatts by late 2025, roughly four times a year earlier, and by mid-2026 had swelled toward 440 gigawatts, several times the all-time peak demand of the entire ERCOT system, with data centers making up nearly 90% of it. One Texas utility reported a 700% jump in large-load requests in a single year.

Then the bill comes due. In PJM, the grid serving 67 million people from the Mid-Atlantic to the Midwest, the capacity price utilities pay to guarantee supply went from about $29 per megawatt-day two years ago to the regulated ceiling of $329 this year, roughly a tenfold jump. PJM’s independent market monitor traced the majority of one year’s increase directly to data centers, around $9.3 billion, roughly 63% of the increase, that gets recovered from ordinary customers’ bills. Cumulative costs could reach $100 billion to $163 billion by the early 2030s, and the average family in that grid is looking at something like $70 more a month by 2028. All of us are already paying for the data centers, on our power bills, whether or not we use what runs inside them.

More power

The hunt for electricity to feed all this is its own remarkable story. Mothballed nuclear plants are being resurrected specifically to run data centers. Three Mile Island of all places, is being restarted under a 20-year deal with Microsoft, rebranded the Crane Clean Energy Center and now fast-tracked to 2027. Michigan’s Palisades plant has become the first fully shut-down U.S. reactor ever cleared to be brought back online. When crews inspected it they found stress-corrosion cracking on more than a thousand steam-generator tubes, far beyond what anyone expected, because it wasn’t maintained for an unforeseen restart. Utility-scale generation sites ignored for decades are now prized. Tech companies are placing bets on small modular reactors, a technology that has never operated commercially in this country, on the assumption it’ll be online to absorb some of this eventual load when it gets here. Water for cooling and land for the transmission footprint round out the rest of the tab.

Blowback

None of this has been unnoticed. A Gallup poll this spring found 71% of Americans oppose building data centers in their area; nearly half of them strongly. For perspective, that’s more opposition than Americans express toward a nuclear plant nearby, which sits at 53%. People would rather live next to a reactor than a server farm. That opposition is already hardening into local moratoriums, legal challenges, and a polarized issue in this year’s state and local elections.

So here’s the squeeze, with all the juice. The public pays for the buildout through its power bills, its watersheds, its career disruption, and its collective creative output built up until now. Irrespective of mounting backlash, as capacity tightens the subsidized consumer tier is the first thing rationed away. Corporations that can pay will keep access to the frontier of accelerating AI capabilities. The rest of us will land betwixt and between: we picked up the bill, and we don’t get to use the good stuff. That’s not the dramatic AI catastrophe people lose sleep over. But it is the likelier one, with cascading impacts that over time will become the worst realistic outcome conceivable.

Now what?

Three things.

First, don’t build your life or your livelihood on a subsidized tool. Use AI, by all means. But treat consumer access as a dependency to hedge, not a permanent utility, the same way you’d treat any input you don’t control. The capability you’ll actually have when the free version gets throttled is the capability you’ve built into yourself, your skills, and your relationships. Lean on the tool; don’t become its tenant.

Second, the public fight has to grow up. “No data centers anywhere, ever” is a losing and incoherent position, because the compute is going to get built somewhere. The useful fight is over who pays. The hyperscalers are flush and, right now, desperate for sites and power, which is exactly the moment to make them carry full cost: dedicated rate classes so households aren’t subsidizing them, full-cost interconnection, real terms on water and land. Oregon became the first state to create a separate data center rate class, and some two dozen states have since taken up similar reforms; it is precisely the right move. Get the beneficiaries to pay full freight now, while we have any leverage remaining.

Third, we need the honest middle. The conversation is stuck between “robots give us cancer” and “build everything, everywhere, no questions.” (The Atlantic’s recent piece on why everyone hates data centers gets at this well.) The grown-up position is that the compute is coming in some form, and the questions worth fighting over are where it goes, what it’s allowed to draw from the grid and the watershed, and who pays. Those can be addressed. ‘Full stop’ or ‘full go’ aren’t the only two settings.

There’s a physical dimension to this too, and it’s the thread I’ll pick up next week. The same load growth inflating your bill is starting to strain the grid itself: PJM’s last two capacity auctions cleared short of the reserves meant to guarantee reliability, the first shortfalls in the market’s history, even as substations and transmission labor to absorb loads that can swing by a gigawatt in seconds. As grid power grows more expensive and less certain, the ability to run the essential parts of your home on your own generation systems — to “island” — stops being an off-grid prepper indulgence and starts being sound economic arithmetic. That’s the subject of the next piece.

Either way, disruption will continue to occur. It always does. What you control is your exposure to it: how dependent you are on the subsidized version, how loud you are about who pays, and how much real capability you have built that doesn’t evaporate when the free tier does. Betwixt and between is a bad place to get stuck. But you can see it coming.

The dynamics described in these field notes are systemic; your household’s exposure to them is not. Crux Resilience provides full household resilience assessments, prioritized mitigation plans, and owner’s representation to carry the work through, from scoping and bid evaluation to contractor oversight.