# AI's real constraint is the grid: what the gigawatt deals and capacity auctions show

- **Published:** 2025-10-15
- **Updated:** 2026-10-12
- **Category:** Hardware & Infrastructure
- **Author:** Sinthetix
- **URL:** https://sinthetix.com/blog/ai-power-demand-grid-limits

> Three chip deals in three weeks are priced in gigawatts. Grid forecasts and a record capacity auction show why power, not silicon, now sets the pace.

In the span of three weeks, one AI developer has announced chip agreements measured not in units or dollars but in gigawatts. The unit matters. It signals that the planning constraint for AI infrastructure has moved from how many accelerators can be manufactured to how much electricity can be delivered, and when. This post lays out what the announcements say, what the grid data says, and what it means for anyone buying AI capacity.

**Key takeaways**

- OpenAI announced at least 26 GW of chip deployments across three deals between September 22 and October 13, 2025.
- The IEA projects global data-centre electricity use more than doubling from about 415 TWh in 2024 to about 945 TWh by 2030.
- PJM's latest capacity auction cleared at its price cap, with forecast peak load up more than 5,400 MW year over year.
- Power, not chips, increasingly determines when new AI capacity actually comes online.

## Three deals, one unit

The announcements, in order [5][6][7]:

| Date | Partner | Scale | Stated start |
| --- | --- | --- | --- |
| Sep 22 | NVIDIA | at least 10 GW of NVIDIA systems | first gigawatt, second half of 2026 [5] |
| Oct 6 | AMD | 6 GW of AMD GPUs | first 1 GW of MI450 Series, second half of 2026 [6] |
| Oct 13 | Broadcom | 10 GW of OpenAI-designed accelerators | start in second half of 2026, completing by end of 2029; racks across OpenAI facilities and partner data centers [7] |

NVIDIA also said it intends to invest up to $100 billion in OpenAI, released gradually as each gigawatt goes live [5]. AMD issued OpenAI a warrant for up to 160 million AMD shares, with the first tranche vesting at the initial 1 GW deployment and later tranches as deployments scale toward 6 GW, subject also to AMD share-price targets and technical and commercial milestones [6].

The sum is at least 26 GW of announced capacity. These are agreements, not energized facilities: the NVIDIA release describes a letter of intent with details to be finalized [5], and the Broadcom announcement describes a signed term sheet [7]. Treat the headline figures as intentions that still have to clear construction, supply-chain and, above all, grid-connection hurdles.

To make the scale concrete, consider that a gigawatt is a power rating, not an energy figure. A single 1 GW site running continuously would use on the order of 8.76 TWh in a year. That is simple arithmetic from the rating, not a figure from any of the announcements.

## What the demand curves say

The International Energy Agency's April 2025 analysis put data centres at about 415 TWh of electricity in 2024, roughly 1.5% of global consumption, and projected that figure to more than double to about 945 TWh by 2030 in its base case [2]. In the US, data centres are on course to account for almost half of electricity demand growth through 2030 [1].

A second forecast points the same direction. A Lawrence Berkeley National Laboratory report, released by the Department of Energy in December 2024, estimated that data centres used 176 TWh in 2023, about 4.4% of US electricity, up from 58 TWh in 2014, and projected that data centres would consume about 6.7% to 12% of US electricity by 2028; DOE's summary describes the 2028 range as 325 to 580 TWh [3]. The share would rise from 4.4% to between roughly one and a half and nearly three times that level.

The two forecasts differ in method and geography but point the same way: rising data-centre demand against a grid that is slow to expand [2][3].

## Where the grid pushes back

Forecasts are one thing; the wholesale market is where scarcity becomes a price. In July 2025, PJM Interconnection, the operator of the grid across much of the eastern US, cleared its capacity auction for the 2026/2027 delivery year at the FERC-approved cap of $329.17 per MW-day across the whole footprint. The prior year cleared at $269.92 per MW-day for most of the region [4]. PJM's forecast peak load for 2026/2027 rose by more than 5,400 MW year over year, "driven largely by data center expansion, electrification and economic growth." PJM estimated the result could raise some customers' bills by 1.5% to 5%, depending on how wholesale costs pass through [4].

Two cautions on that data. PJM does not attribute a specific share of the load increase to data centres, so the cause is stated as a mix of drivers. And a capacity price at its cap tells us supply was tight, not why.

The IEA adds a structural point. A data centre can be operational in two to three years, while the wider energy system needs longer lead times for planning and building infrastructure [2]. That mismatch in timelines is the heart of the problem. A rack order can be placed this quarter; a substation or a transmission upgrade cannot.

## What this changes

**Capacity will arrive unevenly.** Announced gigawatts are staged, and the first tranche for NVIDIA and AMD is dated to the second half of 2026 [5][6]. Locations with available power will fill first.

**Costs may surface in surprising places.** Higher regional capacity prices flow into electricity bills [4], including the bills of the businesses that host their own compute.

**Efficiency becomes a procurement criterion.** When power is the binding input, tokens per watt matters as much as tokens per dollar.

## Practical implication

If your AI plans depend on large dedicated capacity, ask providers where it will physically sit and when its power is contracted, not just when the hardware ships. For most businesses buying inference through an API, the near-term effect is indirect: expect capacity and pricing to be shaped by the grid as much as by chip supply.

## Sources

1. [AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works](https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works), International Energy Agency, 2025-04-10
2. [Energy demand from AI (Energy and AI)](https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai), International Energy Agency, 2025-04-10
3. [DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers](https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers), U.S. Department of Energy, 2024-12-20
4. [PJM Auction Procures 134,311 MW of Generation Resources; Supply Responds to Price Signal](https://insidelines.pjm.com/pjm-auction-procures-134311-mw-of-generation-resources-supply-responds-to-price-signal/), PJM Interconnection (Inside Lines), 2025-07-22
5. [OpenAI and NVIDIA announce strategic partnership to deploy 10 gigawatts of NVIDIA systems](https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems), NVIDIA, 2025-09-22
6. [AMD and OpenAI announce strategic partnership to deploy 6 gigawatts of AMD GPUs](https://www.intelligentcio.com/north-america/2025/10/06/amd-and-openai-announce-strategic-partnership-to-deploy-6-gigawatts-of-amd-gpus/), AMD press release (reposted by Intelligent CIO), 2025-10-06
7. [OpenAI and Broadcom announce collaboration on 10 gigawatts of custom AI accelerators (Broadcom news page); AIwire coverage at dev.aiwire.net/2025/10/13/ was also checked](https://www.broadcom.com/company/news/product-releases/63691), Broadcom, 2025-10-13

## How Sinthetix helps

An engineering studio for AI agents, automation, and internal tooling. We build the systems that make operators scale.

[Our services](https://sinthetix.com/services) | [Talk to us](https://sinthetix.com/contact)

Sinthetix is an AI engineering studio. We ship agents, automation, and internal tools for operators who need results, not reports. [About Sinthetix](https://sinthetix.com/about)