Capacity Planning Used to Be Easier

Yesterday I wrote about what an internal cloud provider exposes to its customers (Who Gets a Choice at Scale). This post is the ground floor of the same problem. Capacity planning used to be easier. You projected growth, you ordered machines, you racked them, and the hard part was the arithmetic. The arithmetic is still the easy part. What changed is everything the arithmetic used to sit on.

Energy became the constraint

The first factor is one nobody used to write down, because electricity was never the scarce thing. Now it is. The EIA’s Short-Term Energy Outlook has US electricity generation hitting a record 4,367 billion kilowatthours this year, growing into 2027, with sales growth driven by data center development and manufacturing activity. 1

The lead times are the part that lands in a capacity plan. JLL’s 2026 Global Data Center Outlook says power, not location or cost, is now the primary site selection criterion because of multiyear grid connection waits, and that the average wait for a grid connection in primary data center markets exceeds four years. 2

When you are planning your own capacity, that shows up as: the building you want is three years out, the substation behind it is four, and neither wait is something you can spend your way out of. You used to plan compute. Now you plan power, and power is booked years ahead.

Memory became a resource like power

The second factor is that memory went from a component on a bill of materials to a resource you have to secure. NVIDIA’s CFO commentary for its second fiscal quarter of 2027 reports supply and capacity commitments jumping from $119 billion to $279 billion in one quarter, primarily related to the procurement of memory. 3 That is the procurement arm of the most valuable chip company on earth pre-committing a quarter-trillion dollars, because if they do not sign early, they do not get the parts.

Micron’s CEO said in June 2026 that his customers had signed long-term supply agreements worth about $100 billion, and that the company has no line of sight to when memory supply catches up with demand. 4 JLL has the average equipment lead time at 33 weeks globally and 42 weeks in the US, which is 83 percent above 2019 levels. 2

If you are building your own fleet, that shows up as: you cannot get the memory, the lead time blew out, and the price moved between the quote and the purchase order. If you are renting, your provider is in that queue and passes the constraint to you as allocation, and your capacity plan becomes a conversation about whose allocation you are in.

Import and export controls are in the plan now

The third factor used to be called trade policy, and most capacity planners could go years without thinking about it. Now it belongs in the plan. The Council on Foreign Relations has documented Beijing imposing export controls on gallium, rare earths, and at least seven other materials between 2023 and 2026, including components produced with them, which is the part that widens the net. 5

On this side of the ledger, the White House issued a proclamation in January 2026 under Section 232 imposing a 25 percent tariff on certain advanced computing chips, with a broader second phase planned. 6 The Bureau of Industry and Security’s Under Secretary told the House Foreign Affairs Committee in July 2026 that further regulatory action on chips and AI is coming. 7

If you buy your own hardware, the book makes the point directly: buying hardware exposes you to “market forces, supply-chain risk, and import and export controls that a provider otherwise absorbs or handles for you” (chapter 13). Chapter 3 makes the same point from the escalation side: no amount of escalating to an executive changes “import and export controls, supply-chain difficulties, or power limits” (chapter 3). Those passages used to read as hypotheticals. They are now a normal part of the job.

What used to be the whole job

Here is the older contrast, from the book: requirements can change faster than supply (chapter 2). That was always true. But it used to be that supply itself was the steady part. Demand moved, and suppliers kept up, because capacity could be added in quarters and paid for with a purchase order.

That is what broke. Supply is now the slow, gated, politically exposed part. Demand moves in months. Grid connections take four years. Memory is pre-committed a year out. Export controls redraw the vendor list overnight. The arithmetic was never the hard part. The assumption behind the arithmetic, that the market will still be there when the order lands, is what stopped holding.

Sources

1. U.S. Energy Information Administration, “EIA expects record electricity generation in 2026 and 2027,” Short-Term Energy Outlook press release, September 9, 2026: eia.gov — record US electricity generation of 4,367 billion kilowatthours in 2026, growth driven by data center development and manufacturing activity.

2. JLL, “2026 Global Data Center Outlook,” 2026: jll.com — “Power, not location or cost, will be the primary site selection criteria due to multiyear wait times for a grid connection”; average grid connection lead time for new 50 MW data centers exceeds four years in primary markets; global equipment lead time 33 weeks, US 42 weeks, 83 percent above 2019.

3. NVIDIA, “CFO Commentary on Financial Results for the Second Quarter of Fiscal 2027,” SEC filing, August 26, 2026: sec.gov — supply and capacity commitments increased from $119 billion to $279 billion, “primarily related to the procurement of memory.”

4. Anton Shilov, “Micron inks long-term supply agreements worth $100 billion — says it has no idea when RAM crisis will end,” Tom’s Hardware, June 25, 2026: tomshardware.com — long-term agreements covering 2026–2030; CEO Sanjay Mehrotra: “Even as we expect industry supply to improve gradually in 2028, we currently do not have line of sight as to when memory supply will be able to catch up with increasing demand.”

5. Council on Foreign Relations, “Dismantling China’s Supply-Chain Weapon,” 2026: cfr.org — Beijing imposed export controls on gallium, rare earths, and at least seven other materials between 2023 and 2026, including components produced with them.

6. The White House, Proclamation “Adjusting Imports of Semiconductors, Semiconductor Manufacturing Equipment, and Their Derivative Products into the United States,” January 14, 2026: whitehouse.gov — Section 232 tariffs on certain advanced computing chips and semiconductor manufacturing equipment, with a broader second phase planned.

7. Under Secretary of Commerce for Industry and Security Kessler, testimony before the House Foreign Affairs Committee, July 14, 2026 — “there will be future regulatory action in the area of chips and AI.”