A quick note: I’m trying out the scrollytelling style The New York Times Magazine uses, because I love how they tell stories. If you’d rather just read, there’s a text-only version at the top.
In my policy class, we learned about utility regulation: the rules for who gets to build power infrastructure and what they can charge for it. Around the same time, I kept seeing headlines about insane AI valuations and new apps launching every week. All of that has to come from somewhere, and a lot of it comes back to the stuff we were studying.
It reminded me of high school, when I was the youth representative on Stafford County’s utilities commission. It was about as boring as it sounds, but I wanted to learn how the stuff underneath everything actually works. I didn’t expect utilities to end up at the center of AI.
In early 2024, Meta described its AI plans by counting chips: 350,000 Nvidia H100s by the end of the year.1 Nobody talks like that anymore. AMD’s deal with Anthropic in July was for two gigawatts.2 OpenAI’s deal with Broadcom is for ten.3 Jensen Huang says a one-gigawatt AI data center costs $50 to $60 billion all in.4
The next step is already happening. CME and ICE both want to list futures on GPU rental prices. CME was set to launch H100 and B200 contracts on October 5, until the CFTC extended its review to November 9, citing “novel or complex issues.”5 Once compute has a futures market, it has a public price for next year. So I went through the deals, the rental data and the first forward curves to figure out what that price says, and whether I agree with it.
Part I
Counting in Gigawatts
Zooming out from one chip to everything the labs have signed
- Custom chips
- AMD
- Nvidia
- Cerebras
- Potential (2028 TPUs)
- Actually running
This is one GPU,Start with one GPU, an Nvidia H100. It rents for about $2.85 an hour right now.6
Ten racks is about a megawatt, roughly 720 GPUs.
A thousand megawatts is a gigawatt. That’s about what one nuclear reactor puts out.
Here’s what’s actually running. OpenAI had about 1.9 gigawatts at the end of 2025.11
Data for this graphic
| Buyer | Counterparty | Chip | GW | Announced |
|---|---|---|---|---|
| OpenAI | AMD | AMD MI450 | 6 | Oct 2025 |
| OpenAI | Broadcom | Custom | 10 | Oct 2025 |
| OpenAI | Cerebras | Cerebras | 0.75 | Jan 2026 |
| OpenAI | Amazon | Custom (Trainium) | 2 | Feb 2026 |
| Anthropic | Microsoft and Nvidia | Nvidia | 1 | Nov 2025 |
| Anthropic | Google and Broadcom | Custom (TPU), 2026 | 1 | Oct 2025 |
| Anthropic | Google and Broadcom | Custom (TPU), 2027 | 5 | Apr–Sep 2026 |
| Anthropic | Google and Broadcom | Custom (TPU), 2028, potential | 10 | Sep 2026 |
| Anthropic | Amazon | Custom (Trainium) | 5 | Apr 2026 |
| Anthropic | AMD | AMD MI450 | 2 | Jul 2026 |
| OpenAI | Nvidia | Nvidia (letter of intent, shelved; not drawn) | 10 | Sep 2025 |
Two things stood out to me.
First, Nvidia barely shows up. Some of that is just how Nvidia sells. Its chips mostly reach the labs through clouds like Oracle and Microsoft, under contracts written in dollars, so they don’t appear here. Nvidia is fine. It booked $89 billion of data-center revenue last quarter.12 But when a lab signs with a chipmaker directly and picks the chip itself, it keeps picking the cheaper gigawatt. That makes sense once you remember the labs are the most cash-strapped buyers in this market. OpenAI expects to burn close to $280 billion through 2030, according to the Financial Times.13
Second, signed and running are very different numbers. A gigawatt deal is basically an option on power that might take years to show up. The typical power project that came online in the U.S. in 2025 waited more than five years to connect to the grid.14 The chip isn’t the bottleneck anymore. The watt is.
Part II
The Curve
What the first GPU futures are pricing in
Here’s what each GPU rents for this week.6 Nvidia’s newest chip, the B300, costs 4.4 times as much per hour as an A100 from 2020.
Now divide by how much power each chip draws. The spread drops to about 1.4 to 1.7 times. Per watt, the market is already pricing these chips almost the same.
The H100 had a big year. Silicon Data’s index went from $2.00 in December to $2.85 now.6 A four-year-old chip went up more than 40%.
My call is lower than that. I think the H100 is under $2 an hour by the end of 2027.
Data for this graphic
| GPU (Oct. 12, 2026) | $/GPU-hour | Rated power | $/kW-hour |
|---|
| Date | Series | $/GPU-hour |
|---|
The per-watt result is what got me interested in futures in the first place. If rent mostly tracks the power a chip uses, then compute is starting to trade like an energy commodity. The exchanges seem to agree. ICE’s second compute contract, with a startup called NATIVX, settles on an “energy-normalized” compute index and is meant to trade next to ICE’s power and gas futures.17
So why do I think the H100 curve is still too high? It comes down to what happens inside a data center when power is the limit.
Picture a data center with a fixed amount of power. Every square here is one megawatt. You can’t get more power, so the question is what you put on each megawatt.
Here’s one megawatt filled with H100s, about 1,400 of them. At today’s index they earn about $4,100 an hour together.
Now swap in B300s. Each one draws more power, so only about 700 to 900 fit. But they earn $4,750 to $6,050 an hour, depending on how you rate their power.
But you have to pay for them. At about $69,000 per GPU slot,7 spread over five years, that’s $1,100 to $1,400 an hour. After that, swapping ranges from 11% worse to 13% better. Basically a wash, which is why people still rent H100s.
That changes when newer chips earn about a third more per watt. Then pulling the H100s out pays for itself. Vera Rubin is already in full production.12 I don’t think the H100 dies because demand goes away. I think it gets kicked out of its power slot.
That’s also why I’d read the H100 and B200 curves together, as a spread. Both slope down, but the H100 slopes down faster. If I’m right about eviction, that gap gets wider.
There’s already a floor under part of this market. It’s just private. Nvidia now guarantees minimum revenue to some neoclouds that buy its chips and will rent back their idle capacity.18 In one SemiAnalysis example, the guarantee works out to an average of about $2.36 an hour over six years.19 That’s a put option, written by the company selling the chips. Lenders want that protection because, as SemiAnalysis puts it, there’s no active derivatives market to tell them what a GPU will be worth later. That’s the real case for compute futures. Lenders need a public price for what a GPU is worth in three years, and right now they’re guessing.
Part III
Watts or Dollars
Which chip you should buy depends on what you’re short on
So who should buy which chip? The Next Platform puts Nvidia’s Vera Rubin systems at about $40 billion per gigawatt and custom accelerator systems at about $15 billion.20 Nvidia’s pitch is that it gets more work out of every watt.
I think the answer depends on what you run out of first. If you have money but not power, every watt matters, so you buy whatever does the most with each one. If you have power but not money, every dollar matters. I built a small model to see where that line actually falls. Drag the dot.
How much better does Nvidia need to be?
The lines show how much more output per watt Nvidia needs to beat a custom chip. Below the green line, custom chips win either way. Above the blue line, Nvidia wins either way. In between, it depends on whether watts or dollars run out first.
If watts run out first
If dollars run out first
My model. Systems are written off over five years and a $15 billion facility over 20; electricity is $0.7 billion per gigawatt-year. Financing and taxes are left out. Method in the notes.
With my defaults, Nvidia needs about 1.25 times the output per watt of a custom chip to win when power is the constraint, but about 1.9 times when money is. My guess is the real number is somewhere in between. If that’s right, it explains the ledger. Hyperscalers and neoclouds are short on power, so they keep buying Nvidia. The labs are short on cash, so they sign for custom chips.
What I’m watching
H100 under $2 an hour by the end of 2027.
Silicon Data’s illustrative strip has December 2027 around $2.16.15 I’m taking the under.
I’m wrong if it holds above $2.50.
The H100–B200 spread gets wider.
Newer chips hold their value better than the curve implies while power stays tight.
I’m wrong if the B200 falls as fast as the H100.
2028 is when money gets tighter than power.
Nvidia is already backstopping its customers’ lenders,18 and CoreWeave’s interest expense is a quarter of its revenue.21 When that flips, custom chips win more of the spending. I expect Broadcom’s AI revenue to reach 40% of Nvidia’s data-center revenue in 2028, up from about 19% last quarter.22
I’m wrong if it’s still under 25%.
Lenders, not labs, become the first big users of GPU futures.
They’re the ones holding six-year risk on chips that reprice every year.
I’m wrong if the first big hedger named publicly is an AI lab.
The standard
Compute used to be counted in chips. Now it’s bought in gigawatts, and soon it’ll have a futures curve anyone can look up. If you want to know where AI infrastructure is heading, I’d watch that curve more closely than any product launch.
Looking forward to writing more. Stay tuned.