AMD Locks in Up to 2.5 GW of Data-Center Capacity — and Tells You Where the Real Bottleneck Is

A Chipmaker Signs for Power, Not Just Chips

For most of the AI boom, the scarce resource everyone talked about was the GPU. The July 28, 2026 agreement between Core Scientific and AMD is a marker that the conversation has moved. According to Core Scientific’s own announcement, the partnership secures AMD the ability to deploy its AI solutions across up to 2.5 gigawatts of data-center capacity — a figure measured in electrical power, not silicon.

The concrete, contracted core of the deal is smaller and more precise than the 2.5 GW headline. As cloud-industry coverage detailed, the initial agreements cover 529 MW of critical IT capacity across five Core Scientific facilities in Texas, Oklahoma, Alabama, and Georgia, on initial 15-year lease terms with options for three five-year extensions. Of that, AMD directly enters leases for 377 MW at sites in Pecos and Hunt County, Texas, and Muskogee, Oklahoma, while a separate, unnamed neocloud leases 152 MW at Auburn, Alabama, and Dalton, Georgia. The path to 2.5 GW runs through an option: AMD can reserve an additional 1,925 MW through late December 2028.

The money frames the ambition. Core Scientific estimates the initial agreements represent more than $14 billion in potential base contracted revenue over the 15-year term. The deal also has an equity dimension: AMD received warrants to purchase up to 30 million Core Scientific shares at $23.47 each, with a portion vesting on signing — aligning the chipmaker’s interest with the data-center operator’s stock, not merely its floor space.

What “2.5 Gigawatts” Actually Means — and Doesn’t

The number invites misreading, so it is worth being exact. 2.5 GW is a ceiling, not a commitment: it is what the relationship could reach if AMD exercises its reservation option in full by December 2028. The firm, signed capacity today is 529 MW, and even that does not come online immediately.

1. Separate the contracted 529 MW from the optional 1,925 MW

The honest way to describe this deal is “a firm 15-year lease for 529 MW, with an option to scale toward 2.5 GW.” The 1,925 MW of additional capacity is a reservation AMD can choose to take through late 2028 — real optionality, but not yet demand converted into steel and transformers. Reporting that flattens the two into a single “2.5 GW deal” overstates what is contractually locked. For anyone modeling AI infrastructure supply, the number to track is how much of that option AMD actually converts, and on what timeline.

2. Note the timeline: capacity in 2027, not now

Even the firm capacity is a future good. Industry coverage puts the initial deployment — more than 500 MW of U.S. infrastructure — as beginning in 2027. That lag is the whole point. A gigawatt-scale data center is gated by power delivery, substation and grid interconnection, cooling, and construction, none of which move at the speed of a chip order. The deal is a bet placed today on power that will be usable a year or more out.

3. Read the neocloud slice as a structural signal

The 152 MW leased by an unnamed neocloud alongside AMD’s own 377 MW is not a footnote. It shows the emerging market structure: chipmakers, “neocloud” GPU-rental specialists, and colocation operators like Core Scientific are assembling into stacked supply chains where each layer contracts for power years ahead. AMD entering leases directly — rather than only selling chips to whoever builds the data center — is the notable move: the chipmaker is reaching down the stack to guarantee its hardware has somewhere to run.

Why Power — Not GPUs — Is the 2026 Bottleneck

This deal is one data point in a broader shift that ran through the industry’s 2026 earnings season. In their mid-2026 results, the largest hyperscalers reframed the discussion away from aggregate capital spending and toward time-to-energy, power procurement, large-scale networking, and the speed of converting infrastructure into revenue-generating compute. When the operators building the biggest clusters start talking about megawatts and interconnection queues instead of chip counts, the binding constraint has moved.

The physics is unforgiving. A single frontier-scale training or inference cluster can draw hundreds of megawatts continuously; assembling 2.5 GW of capacity is closer to planning a fleet of power stations than a server room. That is why a chipmaker signs a 15-year lease and why the useful capacity lands in 2027 rather than this quarter — grid interconnection, on-site power, and cooling are the long-lead items now, and they define who can actually deploy AI at scale.

For markets outside the U.S. hyperscaler core — most of the emerging world included — the lesson is transferable even though the deal is not. The constraint on running large AI workloads locally is not primarily access to chips; it is reliable, affordable, high-density power and the grid to deliver it. A country that wants domestic AI infrastructure has to solve energy and interconnection first, on multi-year timelines, exactly as AMD and Core Scientific are doing — which is why so much emerging-market AI compute will, for now, keep being rented from data centers built where the power already exists.

Where This Fits in the 2026 Build-Out

Strip away the headline number and the deal is a clear read on the state of AI infrastructure. AMD, a chipmaker, is contracting directly for gigawatt-scale power on a 15-year horizon because chips without power are inventory. The firm 529 MW, scalable toward 2.5 GW, is a serious commitment; the more-than-$14-billion revenue estimate and the warrant package show how deeply chipmaker and operator interests are now intertwined.

But the disciplined way to hold it is as a power-and-time story wearing a gigawatt headline. The contracted capacity is 529 MW, not 2.5 GW; the useful power arrives in 2027, not now; and the 2.5 GW ceiling depends on an option AMD has until late 2028 to exercise. For anyone tracking where AI can actually run — or planning to build capacity somewhere new — the transferable insight is the one the hyperscalers spent 2026 repeating: secure the energy and the grid interconnection first, because that, not the GPU, is what gates the build.

Frequently Asked Questions

Sources & Further Reading

• Core Scientific and AMD Announce Infrastructure Partnership — Core Scientific (investor relations)
• AMD secures up to 2.5 GW of Core Scientific data centre capacity — CloudComputing-News
• Core Scientific and AMD Partner To Deploy Up To 2.5 Gigawatts Of AI Data Center Capacity — Pulse 2.0
• AMD gets up to 2.5GW of compute for AI from Core Scientific under new agreement — Neowin