AMD has just announced an agreement with Meta Platforms that is strikingly similar to the agreement it signed last October with model-building company OpenAI, from a total data center capacity of 6 gigawatts (covering computing, storage, and networking) to 160 million AMD warrants offered to facilitate the deal.

Eight more similar deals would result in ten companies owning 100% of AMD within five years. (No, that won't happen, I'm just kidding. But we hope it does.) Seriously, OpenAI and Meta Platforms could sell their shares after the warrants convert (provided that technology installations reach certain milestones and AMD's stock price rises due to these successful installations)—giving OpenAI and Meta Platforms more funds to invest in GPUs!

Meta Platforms and OpenAI Trading

We live in a crazy new world, don't we?

In short, a significant difference between the Meta Platforms and OpenAI deals is that we're fairly certain the former (also aspiring to be one of the world's top AI model builders) has the financial resources to fulfill its hardware procurement commitments. OpenAI, on the other hand, has annual revenues of only tens of billions of dollars, and people are still speculating about how it will raise hundreds of billions of dollars in cash. We believe it will find a way, even at any cost, to eventually acquire massive amounts of data center capacity.

The agreement AMD signed with Meta Platforms is nearly identical in size and scope to the deal the social networking giant signed last week with AI hardware giant Nvidia. Nvidia hasn't disclosed the specific amount Meta Platforms will spend, only stating that the company will purchase the rights to millions of "Blackwell" and "Rubin" GPU accelerators from Nvidia. But as we've pointed out, many have overlooked the details: some of the funds will be allocated to cloud capacity, rather than on-premises deployments. We estimate that Nvidia's agreement involves 2 to 3 million GPUs, a large number of "Vera" Arm server CPUs for both internal and external AI systems, as well as the NVSwitch vertical scaling network for GPUs and the SpectrumX network for horizontal scaling clusters. The GB300 NVL72 equivalent chip alone could generate $110 billion to $167 billion in revenue. (The Vera CPU clusters will have superior performance, but this may not be enough to have a significant impact.)

Cloud computing and new cloud platforms

It remains unclear how much of Meta Platforms' investment has already been realized on paper over the next four to five years in cloud computing and new cloud platforms. Neither company has specified a timeframe for the deal, only stating that it is a "multi-year, generational strategic partnership."

AMD and Meta Platforms are more precise, stating that the agreement is for five years, beginning in the second half of 2026, when the first 1 gigawatt of systems will be launched. These systems will be based on custom MI450 GPU accelerators and the "Helios" open rack-wide v3 rack-class system co-designed with Meta Platforms.

Depending on the accelerators used, 1 gigawatt of capacity is estimated to be equivalent to 500,000 to 600,000 GPUs. Taking the middle value, the 6 gigawatts of capacity installed over five years would be equivalent to 3.3 million MI400 series GPUs. Based on an average unit price of $35,000, the cost of GPUs alone would reach $115.5 billion, averaging $23.1 billion annually. This aligns with AMD CEO Lisa Su's statement in a conference call with Wall Street that the cost would be "double-digit billions of dollars per gigawatt." Adding the costs of racks, networking (partly from AMD, such as Pensando DPUs), and storage, the hardware cost could reach $35 billion per gigawatt, with the remaining costs going towards infrastructure, power, and cooling.

AMD "Antares"

During the conference call, Lisa Su stated that Meta Platforms was an early adopter of AMD's "Antares" MI300X and MI350X GPUs, and she believes that the "Altair" MI450 series would have sold "very well" on Meta Platforms without such an agreement with the social networking company.

Su explained, "But what we're aiming for is transformative growth. Achieving gigawatt-level deployments, reaching six gigawatts within five years, is undoubtedly transformative for our future business development. Furthermore, they are leading the way in terms of models. They are optimizing for future workloads, and we are optimizing in tandem with them."

This brings us to another unique aspect of the deal. As part of the transaction, Meta Platforms will receive semi-custom MI450 series GPUs from AMD, with initial shipments starting in the second half of 2026 from Helios racks. This is the first custom product in the MI400 series; Lawrence Livermore National Laboratory received a hybrid CPU-GPU chip—the "Antares-A" MI300A—during the MI300 series, which contained six GPU chips and two CPU chips. Due to the chipplet design, other customers may also receive custom MI400 series products, provided their order volume is large enough to support the production of special products.

MI450X chip

The specific features of the custom MI450X chip, commissioned by Meta Platforms, are still unclear, but AMD CFO Jean Hsu stated that this custom chip requires no additional tape-out during the MI400 cycle. We also understand that this custom MI450X chip is optimized for Meta Platforms' own inference workloads.

Therefore, depending on Meta Platforms' optimization goals (e.g., optimizing cost per watt or pursuing extreme performance), its HBM stacked memory capacity may be more or less than the standard version, and the GPU clock frequency may be higher or lower. To achieve a better balance between HBM capacity and bandwidth and the inherent computing power of the socket, one could consider reducing the number of GPU chips, slightly increasing their operating frequencies, and increasing the capacity and speed of the HBM memory, thereby achieving higher capacity and bandwidth per unit of computing power while reducing heat generation during computation.

There are many different approaches to try—especially if AMD, as we would like, places the vector cores and tensor cores on separate chips. If this is the case, Meta could increase the number of tensor cores and decrease the number of vector cores, for example, if doing so helps improve its inference performance, while still continuing to use the MI450 socket.

Meta Platforms is also an early major adopter of the upcoming "Venice" Zen 6 Epyc 9006 CPU and will also procure a significant amount of the future "Verrano" Zen 7 Epyc 9007 CPU. These CPUs will certainly be used in the Helios AI rack, but will also be deployed within Meta Platforms to run more general-purpose applications to support Facebook, Instagram, and other apps.

The initial 1 GW of capacity has been secured, with the remaining 5 GW of capacity to be contracted out between now and 2030. The OpenAI and Meta Platforms businesses alone have brought 2 GW of capacity commitments to AMD, meaning AMD can sign agreements with suppliers to fulfill these orders with lower risk. An average of 1.25 GW of new capacity will be added annually from 2027 to 2030, and each new capacity contract strengthens AMD's confidence in fulfilling each batch of production capacity.

As we pointed out in our previous analysis of the OpenAI deal, assuming the issuance of these warrants grows roughly linearly, and AMD's stock price also grows roughly linearly to $600 by 2030, the total value of the 160 million shares of OpenAI and Meta Platforms stock would reach approximately $69 billion by the end of these contract periods. (Details of this stock transaction are disclosed in AMD's 8-K filing with the U.S. Securities and Exchange Commission.) If AMD prices its AI computing engines and racks slightly lower than Nvidia's (which we believe will happen), this equates to approximately 2 gigawatts of GPU system capacity. This is equivalent to buying hardware with stock instead of cash, thus enjoying a discount. Triple-ing AMD's stock price in five years may be a challenge, but it's not impossible.

The key point is that through this deal, AMD is expected to receive approximately 40% of the revenue from Meta Platforms' AI accelerator business, while Nvidia will receive around 50%. AMD's own MTIA devices and (if the rumored) Google TPU deal ultimately goes through, could add another 10%. These accelerators combined will account for more than half of the $600 billion ($327 billion) that Meta Platforms CEO Mark Zuckerberg pledged the company to invest in data centers by the end of the next decade. We know this is just a rough estimate on the back of a napkin, but the same applies to budget planning four years from now.

We'll wait and see. But one thing is certain, as Lisa Su said at the end of the conference call: "We've made a big investment in Meta, and Meta has made a big investment in AMD."

Source: Compiled from nextplatform

View more at EASELINK