Why is Everyone Building Compute?
- Prices are still climbing — more capacity is coming online, and rental rates are going up… significantly!
- Planning for 10 gigawatts — and are scaling to hyperscaler size, and the neoclouds are adding on top.
- This feels like the start of a new cycle — not the top the market keeps hunting for.
Every week the bear case gets louder: capex is out of control, compute is about to be commoditized, the buildout has to peak. And every week the data says the opposite.
This week the two biggest sellers of merchant AI compute, and , reported prices still rising even as they rush to add new capacity. So we are stepping back to answer the question underneath every capex headline: why is everyone still building compute?
1. Prices Are Skyrocketing
The clean bear thesis was that massive GPU deployments would commoditize compute and push rental rates down. This week ran the other way: even as capacity increased, pricing climbed, and contract terms improved alongside it.
- Nebius. Mid-term deals (1–3 years) are pricing at $20–$25M per megawatt (~$4.50/hr for GB300s), with prepayments covering 50–60% of capex; sub-six-month deals clear at $40–$50M per MW, near $9/hr. Its first capacity auction cleared 15% above the highest price it had ever charged for Blackwell, and it could sell all of its 2027 capacity today, but is holding some back for better economics.
- CoreWeave. Pushed through a ~25% price increase in July, with new contracts carrying contribution margins 5–10 points higher, even as active power grew 50% sequentially to 1.5GW and contracted power reached 4.2GW. Revenue doubled to $2.58B on a $104B backlog. CEO Michael Intrator’s read: customers “are increasingly making money from their AI products and are more willing to pay.”
This chart we posted yesterday: rates climbing even as supply floods in. The near-record GPU rental and memory spot prices suggest an across-the-board compute shortage. When price keeps rising while you build as fast as you physically can, that is not a glut. It is a shortage.
2. Everyone Is Planning for 10 Gigawatts
Here is what separates this cycle from the last one: the buyer list is getting longer, and the newest entrants are the size of the incumbents.
- The hyperscalers keep guiding up. Combined 2026 capex has reached ~$860B, up ~80%, on a path to ~$1.2 trillion in 2027. lifted its number to $220B and still warned it “will still not have enough capacity”; is at $205B and guiding “significantly” higher. Goldman now models global AI investment above $1 trillion in 2026 alone.
- SpaceX and Meta are becoming hyperscalers. SpaceX’s compute revenue hit $2.6B (+247%) and turned EBITDA-positive, with management targeting closer to 10GW than 5GW by end-2027 — a ~$200B business on their own math. Meta runs near 7GW and is doubling toward 14GW. The two names bears called “sellers who’ll flood the market” are two of the largest new buyers.
- The neoclouds are adding on top. CoreWeave and Nebius aren’t just soaking up hyperscaler overflow, they’re building their own, 1.5GW active and 4.2GW contracted at CoreWeave alone, and still selling out.
The bear case says “everyone building means oversupply.” The order books say everyone is building because no one can build fast enough. Even ’s $500B financing package isn’t enough. It’s going to put online about ten gigawatts, which is what we’re expected to need next year alone.
3. This Feels Like the Start of a New Cycle
Step back from any single print and the shape is hard to miss: this looks like the beginning of a build, not the end of one.
- The returns are accelerating. Amazon spelled out the math: a data-center shell is a two-year build with a ~30-year life, while the silicon inside turns over every five or six years. One building runs five or six chip generations. The first pays off the shell, the rest are mostly margin. That’s a ~2-3-year payback, not a sunk cost.
- Capital is being re-engineered around it. Nvidia has started backstopping GPU residual values on up to a quarter of a deployment, CME and ICE are launching compute futures, and GPUs are being securitized into a tradable asset class. You don’t build that machinery for the top of a cycle.
And the money is already rotating. This is the tell of a new cycle, not a dying one. The accelerator was the trade three or four years ago, memory took the baton this past year, and optics and semi-cap look next.
Follow the capex to the longest lead times: buildings take 24–36 months, power four-to-seven years, transformers past 160 weeks, switchgear through 2028. Of the ~$37.6B it takes to stand up one gigawatt, well over half never reaches a chip company, it goes to the physical plant.
Where We’re Finding Opportunities
Follow the cash, and here is where we are finding opportunities:
- The infrastructure layer beneath the chips. 57% of every capex dollar goes to land, power, transformers, cooling and fiber the longest lead times and the quietest pricing power.
- Optics and semi-cap as the 2027 rotation. The same setup that repriced memory, constrained supply, inelastic demand, pricing power on a lag is building in the next layer out.
- Sellers with a differentiated stack. Among the builders, those that own differentiated silicon and a resale route to monetize it stand out.
- Durable value over scarce cost. Memory is ~62% of a superchip’s build cost, but cost share isn’t value capture. Memory is cyclical, and today’s shortage funds tomorrow’s fabs.
The market keeps hunting for the peak in AI spending. We keep following the cash and every dollar of it points to a build still accelerating, not one rolling over. This doesn’t feel like the top. It feels like the start.