Concrete is being poured for another data center in the desert outside of Phoenix. Despite the building’s unremarkable appearance—low, rectangular, and industrial—the power lines that supply it convey a more powerful message. There is no metaphor for six gigawatts. This type of power is typically only found in small cities.
With a multi-year commitment to deploy up to six gigawatts of AMD Instinct GPUs, supported by the unique MI450 architecture, that is the bet Meta Platforms has made with Advanced Micro Devices. The first shipments are expected to occur in the second half of 2026 and are linked to an initial gigawatt. The symbolism is hard to exaggerate.
| Category | Details |
|---|---|
| Buyer | Meta Platforms |
| Chip Partner | Advanced Micro Devices |
| Architecture | AMD Instinct MI450 |
| Total Commitment | Up to 6 gigawatts of AI infrastructure |
| Estimated Deal Value | ~$100 billion (multi-year) |
| Reference | https://www.amd.com |
Nvidia dominated the discussion of AI chips for years. Executives from data centers discussed CUDA lock-in, supply problems, and H100 allocations in whispered tones. AMD was positioned as the rival; it was believable, getting better, but it came in second. That story is complicated by this deal.
Just looking at the structure makes me think of corporate chess. AMD has awarded Meta performance-based warrants to buy up to 160 million shares, or around 10% of the company. These rights will vest in tranches as the stock price and shipment milestones are reached. When that first gigawatt ships, the first tranche will unlock. Future ones will depend on Meta reaching the entire six-gigawatt goal.
Perhaps no significant semiconductor cooperation in recent history has incentives that are so closely matched. The market responded immediately. Following the announcement, AMD’s stock surged 7%, challenging key levels such as the 20-day moving average. Zooming out, however, reveals a more restricted image: the stock is still down about 19% from its peak for the year. Despite their excitement, investors are apprehensive. This situation is tense. Is it time for a fundamental reassessment? Or just a stimulant for a stock that has been experiencing stress?
The impact is huge in terms of money. AMD anticipates increasing its income by tens of billions of dollars in the coming years. Last quarter, the data center sector, which houses EPYC CPUs and Instinct GPUs, reported $5.4 billion, a 39% increase from the previous year. Over the next three to five years, CEO Lisa Su has projected that segment will increase by more than 60% annually. The Meta collaboration seems designed to achieve that goal.
The atmosphere at AMD’s Santa Clara facility was different than it was two years ago. Engineers discussed scaling alongside hyperscalers as well as catching up. Meta is more than just a customer. Through the warrant, it gains some ownership and becomes a strategic ally whose performance is correlated with AMD’s.
However, the warrant creates a risk of dilution. AMD will issue a sizeable stock stake if all milestones are reached. The CFO maintains that the transaction will increase non-GAAP earnings per share on paper. In reality, stockholders will keep a careful eye on the number of shares.
The scope of this collaboration has an almost industrial feel to it. Six gigawatts of processing power suggests whole rack-scale systems, with Helios servers running in synchronized rows and being cooled by complex liquid cooling systems. It portends a future in which AI infrastructure is more about power grids than it is about experiments.
Through 2026, Meta intends to invest more than $135 billion in capital projects. In addition to using Google’s TPUs for specific workloads, it has already committed to millions of Nvidia processors. It’s possible that the AMD agreement is more about diversity than replacement.
Investors are wondering if this action merely increases Meta’s capacity or if it cannibalizes Nvidia orders. It’s still unclear. It is evident that AMD today has a large, assured, multigenerational consumer base.
This solidifies AMD’s position as the main substitute for high-performance AI CPUs from a strategic standpoint. These days, the AI conflicts are real. They have a contract.
But danger still exists. The pivotal moment occurs with the first shipment in late 2026. Delays in scaling up MI450 production could cause warrant vesting to halt and postpone revenue recognition. There is a risk of execution. When deadlines are missed, the chip business is harsh.
It’s difficult to ignore how drastically this agreement alters perception. AMD’s AI goals felt aspirational for years. They now have tangible power goals in gigawatts to support them. If—and this is a big if—delivery goes without a hitch, that scale de-risks the growth story.
The subject of valuation is still difficult. AMD is valued at a premium, about 50 times EBITDA’s trailing price. that a few of prices are experiencing exceptional growth. The argument for it is strengthened by the Meta deal. However, it also creates an unforgiving level of expectation.
This is a classic case of “sell the news.” Profits could be made by traders who made speculative purchases. If investors analyze the warrant structure and determine that the revenue ramp is slower than the headline implies, momentum may wane.
As this develops, it seems as though AI infrastructure has moved into a new stage. More ecosystems—CPUs, GPUs, networking, and software stacks—integrated across multi-gigawatt deployments are being discussed than single-chip benchmarks.
There is more to the 6-gigawatt gamble than just a purchase agreement. It is a statement that the focus of the AI warfare is changing from scarcity to scale. One factor will determine if it changes the competitive hierarchy in the end: execution. The story will never be the same if AMD delivers on schedule, at performance, and at volume. Otherwise, the agreement turns into a bold footnote.
For now, the cranes are still hoisting steel over data centers in the desert. The next phase of the AI arms race is being plugged into place somewhere within those incomplete barriers.