Core Scientific and AMD: in the 2.5 GW deal, the word that matters is “potential”

On 28 July 2026 Core Scientific (NASDAQ: CORZ) — the data-centre operator born out of bitcoin mining and repurposed for AI hosting — announced an infrastructure partnership with AMD, alongside its second-quarter results. AMD, here, is not selling chips: it is securing power and buildings where its own customers will install the chips AMD sells elsewhere. In the filings lodged with the SEC, the deal is written down twice, at two very different levels of detail. Separating them is how to work out what was actually signed on 27 July, and what remains an option.

What is signed: 530 MW, five sites, an option running to 2028

A day before the public announcement, on 27 July 2026, Core Scientific filed an 8-K with the SEC under Item 3.02 (Unregistered Sales of Equity Securities) and Item 7.01. The text, checked line by line, is precise: lease agreements (“AMD Leases”) with AMD for 377 MW at the Pecos and Hunt County, Texas, and Muskogee, Oklahoma sites; separate agreements (“Neocloud Leases”) with a third counterparty the filing calls “Neocloud”, for 152 MW at Auburn, Alabama, and Dalton Phase 3, Georgia. Total: 529 MW across five sites (the “approximately 530 MW”, rounded, of the quarterly release), on fifteen-year terms with three five-year renewal options. On the Neocloud leases, all three parties also signed credit support agreements protecting AMD’s equipment installed there and giving AMD the right, not the obligation, to cure a Neocloud default: on two of the five sites, AMD is not the direct tenant but the guarantor of a third party’s risk.

The same filing sets out a second figure: the AMD leases reserve a further 1,925 MW, exercisable under certain conditions by 28 December 2028. Added to the 530 MW already leased, that produces the roughly 2.5 GW seen in the headlines — the difference between a contract and an option with a deadline.

The consideration: AMD received a warrant to purchase up to 30 million shares of Core Scientific, at an exercise price of $23.47 (the volume-weighted average price over the preceding five trading days: market price, not a discount), expiring 27 July 2031. Shares vest at 12,222 per MW leased: roughly 6.5 million have already vested from the 27 July leases, consistent with the 530 MW figure. This is not an agreement paid for in power alone: it is power plus an equity kicker that grows with the relationship.

What is announced: up to 2.5 GW, over $14 billion

On 28 July came the joint press release, via Business Wire: “Core Scientific (NASDAQ: CORZ) and AMD (NASDAQ: AMD) today announced a partnership to shape the future of AI infrastructure, with AMD to secure up to 2.5 gigawatts of data center capacity to support end customer deployments of AMD AI solutions.” Mathew Hein, chief strategy officer of corporate development at AMD: “AI deployments are accelerating rapidly, and bringing that compute online requires trusted infrastructure partners with the scale and power to support the next era of AI.” Adam Sullivan, Core Scientific’s CEO, spoke of a “strategic relationship” meant to “grow our relationship meaningfully over time.”

The same release, attached to the subsequent 8-K on quarterly results, lines up three figures under one heading — “Recent Business Developments” — separated by a single word, “potential”, repeated three times: “potential to support up to 2.5 GW of leasable capacity”, “more than $14 billion of potential base contracted revenue”, “more than $24 billion of potential contracted revenue” (across Core Scientific’s whole customer book, 1.1 GW leased). The one figure measured in the present tense reads differently: “Billing for 437 MW of capacity as of mid-July, representing approximately $635 million in average annualized colocation GAAP revenue.” Between the 437 MW already billed and the 2.5 GW potential ceiling lies a gap of nearly sixfold.

Furnished, not filed: what that means for the reader

Neither 8-K contains Item 1.01, “Entry into a Material Definitive Agreement” — the box the SEC reserves for a material contract just signed, normally attached as an exhibit. In the 28 July filing, AMD is never named in the body text: the name appears exactly once, inside the release attached as Exhibit 99.1. Both filings close with the same clause: the information “is furnished and shall not be deemed ‘filed’ for purposes of Section 18 of the Securities Exchange Act of 1934”. A “filed” document exposes the company to full statutory liability for its content; a “furnished” one conveys information under a lighter obligation — the route Regulation FD provides so a company can inform the market without pre-disclosing a deal to a select few.

The exception is the 27 July Item 3.02: that part is filed, because the law requires it when unregistered securities such as a warrant are issued — but it is narrow disclosure, calibrated to the warrant, not to the commercial contract. The actual lease agreements are not attached as exhibits to either filing: what is public, we know from a press release and a summary description, not from the text of the contract. That is not an irregularity — it is a common, legitimate disclosure choice, and a measure of the weight the issuer places on it.

The asymmetry that is itself news

Whoever needs to announce, as a rule, is not whoever is buying. That week AMD published a dense run of its own releases — Anthropic (22 July), the AAI 2026 product launch (23 July), a partnership with South Korea’s Ministry of Science and ICT (27 July, the same day it signed the leases with Core Scientific, per the SEC record). On ir.amd.com and newsroom.amd.com there is no trace of the Core Scientific partnership, in a list that is anything but empty in those days. At the SEC, AMD’s CIK shows no 8-K filed between 15 July and 4 August — only a Form 4 and a Form 144 — and the next 8-K, on 4 August, is AMD’s own quarterly results. Before drawing a conclusion, we tested the method on AMD releases known to exist — Anthropic, Microsoft on 20 July — and found them without difficulty, correct title and date. The absence is not a limit of the tool: it is a fact.

What this means for AI-capacity buyers in Europe

A chipmaker that secures power, buildings and an option on a further 1,925 MW changes the question to ask a compute supplier. Availability of AI capacity is increasingly a matter of energy and real estate, not of processor price lists: announced capacity is not available capacity, as we have written before. Counterparty risk shifts too: what matters is no longer only who sells you the cloud, but who physically runs the site — and sometimes, as above, not even the party that anchored the capacity is the direct tenant, but the guarantor of a third party.

If your capacity sits inside a campus anchored to a large supplier, who do you actually depend on? What happens if the anchor exercises its option before the deadline? Put the same distinction in writing to your supplier: how much of the capacity sold to you is billed today, how much is under a fifteen-year contract, how much is only an option with an expiry date? Paying in cash rather than equity does not by itself resolve dependence on whoever owns the site, just as a proprietary network fabric remains a constraint regardless of how you paid for it.

The method, applied here

Checking what a supplier has actually signed — how much is contract, how much is option, who guarantees whom at which site — is not an analysis done once and filed away. It is a check that runs against the client’s contracts and systems, with a register of critical dependencies: who supplies what, under which contractual commitment, with which deadline, with which alternative — and an audit trail ready to show an inspector or a board. The same system holds together an organisation’s contracts, suppliers, archives, management software and documents inside a single operating model, on which AI agents execute decisions with a human operator in command: for large enterprises, defence, government and healthcare. Always in two modes — on-premise, on autonomous machines that do not require deep integration into the client’s network, or dedicated cloud, with a dedicated VPN and a data centre in Italy — and always with shared management: the client should not need to already have someone in-house administering AI systems.

Do you need to work out how much of the AI capacity you have bought is billed, how much is under contract, and how much is only your supplier’s option? Let’s talk it through in thirty minutes.