The Local AI
Opportunity
What happens when the demand for a new good is unbounded, while its supply is tightly constrained?
Introduction
When discovered, every node in the tech tree produces a new marginal unit of account for its era. The mainframe had a “timeshare credit”. The internet had the megabit. Intelligence has tokens: the fundamental unit of account for the AI era. These units obey an old law, the one Jevons observed in coal:
“It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth.”W. Stanley Jevons, The Coal Question, 1865
Since ChatGPT’s launch in 2022, the tokenmaxxing argument has tracked very tightly in advance of greater capabilities. We’ve used more tokens as AI has gotten more capable, and it’s unlikely that a real demand cap exists.
The first chatbots answered simple questions:
They consumed 10s of tokens per interaction, and the daily interaction was <100 tokens.
Then came short-horizon agents that could do simple things — respond to emails, browse the web, do research for you:
At the end of a chat, you’d have probably consumed at least 5,000 tokens at a time.
Next came strong agents that could carry long-horizon tasks — building pieces of software end to end, or reverse-engineering an app — runs of tens of hours at a stretch:
At this point, it was pretty easy to consume 100k tokens or more, and you’d be happy to consume even more as memory, context, and tool calling kept improving.
Finally, we hit the 24/7 loop: agents that are never prompted, and never stop running.
New information is simply an injection into their stream of tokens, and they run all day on behalf of their human. The first of these are already here — Grok Bot, Instinct, Muse, Dot.
The supply of tokens, by contrast, is not growing nearly as fast — because supply hits the real world, and Silicon Valley no longer knows how to build there. And even if it did, supply is a function of energy, land, and regulation: three limiting factors, every one of which is set to worsen from here. The thesis of this essay is contained in the distance between those two curves, and it can be drawn in one figure.1
What this creates is a large unmet area under the curve that simply finds a new home:
Local inference.
Instead of covering the planet in datacenters, you can just sell one GPU per person to 1B humans! Apple and Nvidia both understand this strategy, while OpenAI and Anthropic don’t.
The same thing happened in the 1960s with IBM. They were unwilling to build a personal computer, and bet their future on the mainframe instead. It turned out to be easier to ship a million computers than to build a few giant ones. It sounds cliché, but the same thing is happening again — and everyone is asleep at the wheel.
The Math Behind Agents
A Billion Loops
Let’s start with something simple, and conservative: one billion humans using personal agents that run 24/7 at 50 tokens per second — a reasonable speed, a little slower than ChatGPT today.
The agent reads all of their context — their messages, their computer screen, their emails — and is connected to every app they use over the course of a day.
Here is what that would take with modern datacenters:
| Tok/s required | 5 × 1010 tokens per second |
| Tokens per day | ≈ 4.3 quadrillion |
| H200 equivalents | ≈ 80,000,000 GPUs |
| Power at roughly two joules per token | 100 GW, continuous |
| Energy per year | ≈ 880 TWh |
| Cost to build | ≈ $3.5 trillion |
One Hundred Reactors
100 GW of continuous draw is ~100 nuclear reactors running flat out, dedicated to inference. It is roughly twice the peak load of the entire Texas grid. 880 TWh a year is 20% of all the electricity generated in the United States — for personal agents alone.
This math makes no sense, but no one has yet questioned it. Their solution?
“Build more!”
There are serious headwinds that are about to turn into a storm in 2027, because the reality is much more unappealing than Silicon Valley would have you believe.
The Three Caps
The Grid
A datacenter is, before anything else, a connection to the electric grid. Google now calls transmission the single biggest threat to its datacenter expansion, with connection waits in parts of the country surging past a decade.2 A 100 MW connection averages seven years, and Berkeley Lab data show interconnection waits have doubled.3
These delays are not administrative accidents. They are the physical schedule of high-voltage construction: routing, permitting, transformer backlogs, eminent domain. Chip production is much simpler and faster relative to erecting transmission lines — which is how you end up with GPUs sitting in warehouses, waiting for electricity that does not exist yet.
The Neighbors
Let’s forget about the grid for a second. Datacenters are massive projects that continually generate noise, siphon resources from neighborhoods, and create an eyesore for communities. People are not thrilled to have them in their backyards. Data Center Watch counted $64B of American projects blocked or delayed by local opposition;4 by early 2026 the tally hit roughly $130B in a single quarter, with 833 opposition groups across 49 states.5
And the opposition is broad, not fringe: three in five Americans oppose a datacenter near their community, across Democrats, Republicans, and independents alike.6
The State
And where the neighbors organize, the state eventually follows. In August of 2026 Governor Greg Abbott of Texas halted the approval of datacenters statewide pending grid audits,7 an order that froze four hundred and seventy-four gigawatts of pending connection requests, five times the record peak demand of the entire ERCOT system.8 New York went further, signing a one-year ban on large AI datacenters; nearly two-thirds of New Yorkers polled would keep the ban after it expires.9
Nor is this a passing mood of two statehouses. Datacenters are a leading issue of the 2026 midterms, with candidates in both parties running against them — one side on moratoriums, the other on stricter regulation — and fifty-eight percent of Texas voters opposed to construction in their own communities.10 The political price of saying yes to a datacenter is rising in the only poll that matters to a politician, and nothing about the demand curve in Figure 1 will make that price fall.
Energy, land, regulation: three caps, each sufficient on its own, together binding. The supply curve of cloud intelligence is not going to follow demand. It is going to flatten, and the wedge in Figure 1 is going to open.
The Paper Pipeline
Breaking Ground
Against the hundred gigawatts of Chapter I, what is actually being built? Goldman Sachs projects American datacenter additions of 13.6 gigawatts in 2026 and 36.3 in 2027 — the most optimistic reputable schedule in print.11 But a schedule is not a building. Jefferies observes that construction has begun on only twelve of the twenty-four gigawatts planned for 2026, that as much as eighty percent of the 2027–28 pipeline has not broken ground, and that fifteen to twenty gigawatts a year is the realistic pace.12
The Hiking Cycle
Even the paper pipeline rests on an assumption that no longer holds: cheap money. The buildout has been financed with more than $300B of debt, much of it floating-rate private credit, and the Federal Reserve is once again raising rates into it.13 The strain is already visible at the margin. CoreWeave, the largest of the so-called neoclouds, has watched its bonds trade past sixteen percent yields while credit default swaps imply roughly sixty-three percent odds of default within five years;14 Gartner expects as many as eighty percent of neoclouds to cease independent operations by 2030. The Dallas Fed and the Bank for International Settlements have each warned that private credit is mispricing the sector’s risk.15
None of this requires a crash to matter. It requires only that the marginal datacenter — the one at the end of the queue, financed at the top of the rate cycle, opposed at the county meeting — not get built.
An exponential demand curve along with a supply curve flattened by energy, land, and politics is about to create a trillion-dollar opening for running AI on local computing devices.
The modeled version of that area amounts to $5T of displaced cloud demand by 2030.16 That’s 550M personal AI computers that have to step up and meet the demand for AI.
There are three candidates:
- Just build more, bro
- Put them into space
- A GPU per person
It is our belief that history is repeating itself, and the simplest solution is likely the one that catches this unmet demand.
- The interactive model, with its assumptions stated and a slider for the degree of datacenter backlash: local-ai-opportunity.deepshard.chatgpt.site. ↩
- “Google warns transmission delays are now the biggest threat to data center expansion,” Network World, reporting Reuters. networkworld.com. ↩
- MarketScale, on utility spending and interconnection waits, citing Berkeley Lab data (marketscale.com); The Load Growth, “It takes 7 years to plug in a power plant to the grid” (theloadgrowth.substack.com). ↩
- Data Center Watch, $64 billion of data center projects have been blocked or delayed amid local opposition. datacenterwatch.org. ↩
- The Decoder, on local resistance across eleven states; NBC News, on state and local opposition gaining steam, with figures for Q1 2026. the-decoder.com; nbcnews.com. ↩
- University of Pennsylvania Annenberg survey, reported in coverage of the 2026 midterms. batterytechonline.com. ↩
- “Data center approvals in Texas halted until audits completed, Gov. Greg Abbott says,” KSAT, August 3, 2026. ksat.com. ↩
- Tom’s Hardware, on the 474 GW of connection requests subject to the pause. tomshardware.com. ↩
- Siena Research Institute polling on New York’s one-year ban, September 2026. ↩
- Texas Politics Project polling; midterm coverage of datacenter politics. batterytechonline.com. ↩
- Goldman Sachs Research, US data center power demand projected to double by 2027. goldmansachs.com. ↩
- Jefferies, via Data Center Dynamics: half of US capacity planned for 2026 has started construction. via datacenterdynamics.com. ↩
- Crypto Briefing, on the Federal Reserve’s hiking cycle and the sector’s debt load. cryptobriefing.com. ↩
- Fierce Network, Wall Street’s anxiety catches up to neocloud CoreWeave, reporting Reuters bond and CDS figures and the Gartner estimate. fierce-network.com. ↩
- Invezz, on the FOMC minutes and the blind spot in AI financing; Dallas Fed research on AI debt and rates. invezz.com. ↩
- Under the model of note 1: one billion eventual agent users, one hundred tokens per second, three agent-hours per person per day, twelve dollars per million output tokens, and a degree of backlash that holds datacenter growth to forty percent of the unconstrained build rate. ↩