Elena VossMore than 500 megawatts. That is the US data-center capacity AMD agreed to secure from Core Scientific beginning in 2027, under a deal that could scale to 2.5 gigawatts. A chip supplier has taken on the job of securing a place for its chips to run.
Key takeaways
- AMD said its first Helios rack-scale systems were installed at Meta in October 2026 and were running production workloads.
- AMD agreed to secure more than 500 MW of Core Scientific data-center capacity beginning in 2027; the arrangement could scale to 2.5 GW.
- AMD reportedly agreed to backstop a $300 million Goldman Sachs loan for Crusoe to purchase AMD AI chips.
- AMD committed $1 billion to Meta’s planned public-equity offering for AI infrastructure.
- AMD’s MI300X, previewed in 2023, has 192GB of HBM3 memory.
Meta’s first racks do not measure a six-gigawatt deal
Reuters identified Meta as an AMD data-center-chip customer in November 2021. AMD reported $1.5 billion in Data Center revenue in Q2 2022. By February 2026, The Wall Street Journal reported that Meta had agreed to acquire up to 6 gigawatts of AMD Instinct GPUs in a deal valued above $100 billion, with 1 gigawatt planned for deployment in 2026. Meta’s reported commitment gives AMD an identifiable buyer, but the ceiling and deployment plan do not measure installed GPUs.
AMD said in October 2026 that its first Helios rack-scale systems were installed at Meta and that Instinct GPUs were running production workloads there. AMD did not disclose the scale, utilization or economics of the broader agreement.
In Q3 2026, AMD reported $8.5 billion in Data Center revenue, up 97% from a year earlier and accounting for much of its $13.38 billion in total quarterly revenue. The comparison with Q2 2022 spans different quarters, and the segment includes CPUs and other products. The segment figure leaves Instinct GPU revenue unknown.
Nvidia’s packaging reservations leave AMD’s allotment unknown
AMD needs packaged accelerators to deliver more Helios racks. TSMC said its CoWoS advanced-packaging technology was growing at an 80% compound annual rate, while Nvidia had reserved most of its capacity. TSMC’s growth rate gives no figure for AMD’s allocation. Neither the first installed racks nor TSMC’s expansion establishes how many more AMD can deliver.
Software work follows the rack into production
AMD previewed the MI300X in 2023 with 192GB of HBM3 memory for large-model workloads. It later claimed up to 1.6 times the inference performance of Nvidia’s H100 HGX in selected workloads. An independent 2024 comparison found theoretical specification and total-cost-of-ownership advantages for MI300X, but said software bugs held the system back.
That comparison predates Meta’s 2026 Helios installation. A second-source buyer still has to pay engineers to migrate, optimize and support workloads after the hardware arrives. Intel, Google, Microsoft and Meta formed the Ultra Accelerator Link group in 2024 to advance a standard for connecting accelerator server chips. That work addresses the hardware interface; buyers must assess the software work separately.
AMD is helping finance the demand it wants to serve
AMD reportedly agreed to backstop a $300 million Goldman Sachs loan for Crusoe to buy AMD AI chips. AMD’s credit would support the customer’s purchase, making financing terms part of the chip sale.
Meta planned to raise $95 billion for AI infrastructure, including $35 billion through public equity offerings with AMD, Nvidia and Broadcom as cornerstone investors. AMD said it would invest $1 billion in Meta’s offering. The $95 billion remains a plan; final allocation and pricing have not been established. AMD’s investment commitment is distinct from accelerator revenue. Nvidia’s own infrastructure financing puts AMD’s move in context: suppliers are competing over which customers can afford to bring capacity online, as well as which chips they choose.
Crusoe’s loan supports an identified chip purchase, and Meta’s financing plan belongs to a customer already running AMD systems. Core Scientific’s agreement goes a step further: AMD agreed to secure data-center capacity itself. Meta’s reported six gigawatts is a ceiling for GPU purchases; Core Scientific’s more than 500 megawatts is capacity for deployment. Neither figure establishes how much of that space will carry paid workloads.
Frequently asked questions
What are the terms of AMD’s planned Taalas acquisition?
AMD announced plans on November 5, 2026, to acquire AI-chip startup Taalas for about $2.9 billion in cash and stock. Of that amount, $1.4 billion is contingent on conditions AMD did not specify.
Is AMD involved in AI financing beyond Meta and Crusoe?
Yes. Anthropic plans to raise $26 billion from cornerstone investors including AMD, Google and Nvidia, according to a November 10, 2026 confirmed relationship record. The available evidence does not state AMD’s individual investment amount.
What is AMD forecasting for Q4 2026 revenue?
AMD projects approximately $17 billion in Q4 2026 revenue. That is a companywide forecast, not a disclosed forecast for Instinct GPU revenue or data-center revenue alone.
Disclosed scale of AMD-linked AI infrastructure commitments
| Arrangement | Disclosed scale | Timing | What the figure measures |
|---|---|---|---|
| Meta’s reported AMD GPU agreement | Up to 6 GW; 1 GW planned for deployment | 2026 deployment plan | Potential GPU purchases, not installed or utilized capacity |
| Core Scientific capacity agreement | More than 500 MW; could scale to 2.5 GW | Beginning in 2027 | Data-center capacity AMD agreed to secure, not disclosed paid-workload utilization |
| Crusoe chip-purchase financing | $300 million | No deployment date disclosed | Goldman Sachs loan that AMD reportedly agreed to backstop for an AMD-chip purchase |
| Meta AI-infrastructure financing plan | $95 billion total; $35 billion through public equity; AMD committed $1 billion | Planned | Customer infrastructure financing, distinct from accelerator revenue |
Meta has running Helios racks and a reported agreement for up to 6 gigawatts. For the more than 500 megawatts AMD agreed to secure from Core Scientific, the unanswered question is how much capacity will hold operating, paid-for systems.