Today, AMD announced it is acquiring ZT Systems, a leading hyperscale systems design and ODM house for GPU rack systems. I want to thank AMD CEO Lisa Su for her time yesterday to break the deal down for me.
Details Of AMD’s Acquisition Of ZT Systems
- AMD acquires ZT’s rack-scale systems design and manufacturing assets for $4.9 billion in cash (75%) and stock (25%). This is additive to the $1 billion AMD already invested in ZT over the past year.
- The design focus will be on Instinct AI GPUs, Epyc CPUs and AMD/partner networking.
- AMD will spin out and sell ZT’s manufacturing assets, keeping the systems design capabilities.
- ZT CEO and founder Frank Zhang leads the manufacturing business (to be sold), while ZT President Doug Huang leads design and customer enablement, reporting into Forrest Norrod, the head of AMD’s Data Center Solutions Business Group.
- The deal has been approved by the AMD board and is expected to close in the first half of 2025, subject to regulatory approvals. It is expected to be accretive for AMD on a non-GAAP basis by the end of 2025.
ZT Systems Background
- Designs, integrates, manufactures, and deploys rack-level, hyperscale AI systems
- Rumored to drive $10 billion in annual revenue, with AWS and Azure as its largest customers
- Approximately 1,000 design and customer headcount and 1,000 manufacturing headcount
- Private company founded in 1994 and headquartered in Secaucus, New Jersey
- History: 1994 desktop PCs and pedestal servers; 2004 focus shifts to datacenter servers; 2010 rack-scale design and integration; 2013 hyperscaler play; 2024 ships “hundreds of thousands of servers annually”
How Acquiring ZT Positions AMD For Big Growth In Datacenter AI
The growth in AMD’s Instinct GPUs has been impressive—from $0 in the first half of 2023 to a $4.5 billion 2024 forecast—driven by big investments and effort in hardware and software. However, compared to AMD’s own market forecast for AI accelerators and GPUs of $400 billion by 2027, the company needs some accelerants to help it grow parabolically and take what I call its “unfair” share.
Although this has been improving, AMD currently has two main competitive disadvantages in AI infrastructure: its software and its system scale and maturity. AMD already addresses this well for non-AI EPYC servers and PCs, but not so much for AI racks. AMD could build the capabilities on its own organically, but how much time would that take?
AMD has already executed three small software tuck-in acquisitions (Silo AI, Nod.ai and Mipsology) to help it with mid- and high-level software abstractions and to help customers to customize LLMs. It has also made major strides with ROCm AI optimizations and PyTorch and Hugging Face support for both Instinct and EPYC. I expect the company will make more software acquisitions in the future.
While AMD could not have a $4.5 billion annual forecast for Instinct if it was lacking in systems capabilities, what it already has is not enough to get its unfair share of the $400 billion market that’s coming. The AI infrastructure game is not just a chip game; it has become a more vertically oriented system and software game. “Chip” vendors are expected to deliver the full rack and software stacks to achieve year-on-year improvements to performance, efficiency, quality and time-to-market. And the ZT acquisition is targeted at accelerating AMD’s “above the chip” and “below software” capabilities for AI servers.
I believe this acquisition, if executed with Lisa Su’s typical precision, will be an accelerant the company needs to drive parabolic revenue growth for both Instinct and head-node EPYC with the hyperscalers, tier-two CSPs and even some of the largest on-prem facilities for government agencies and financial, energy and pharma enterprises.
A Short Path To Engineering Integration
I’m also bullish about the culture fit between the two organizations. When I talked with her about the deal, Su emphasized the long relationship between AMD and ZT. “Our team has been working with them for many years,” she said. “They did some of our first EPYC designs and MI 250 designs with us, and they’ve been fully engaged on MI 300 designs. And so we’ve gotten to know them very well.”
That alignment extends to customers, too. Su talked about how focused Frank Zhang has been on the datacenter and cloud market over the past 15-plus years. This means that, rather than going broad, ZT has targeted relatively few—very important—companies to sell to. And while Su couldn’t share any customer names, given that ZT is a private company, she did point out that “Every one of their customers is our customer.” So, although it always takes work to integrate engineering teams from one company into another, I think it bodes well that everyone involved is going to be serving the same customers they have been all along.
Staying Out Of The Systems Manufacturing Business
Meanwhile, I like the decision to (eventually) jettison ZT’s manufacturing, sales and support, as these functions would be highly dilutive. For comparison, Supermicro’s net margins are in the mid-single digits while AMD’s are 25%. Related to this, Su told me that AMD would not enter the systems business as Nvidia has with DGX. I am mixed on this as DGX provides Nvidia with significant revenue and margin and gives it a vehicle to sell the total solution. For sure, hyperscalers and tier-one OEMs would not want AMD to get into the systems business, but AMD needs to get something for not competing with its customers. Doing so does not seem to be hurting Nvidia much.
Su believes that customers value choice and bespoke solutions, instead of an approach that dictates building data centers around a predetermined combination of CPU, GPU and networking in a certain form factor. Su says that, with this deal, AMD is going to flip that thinking around. “We’re going to say, ‘You know what, I’d love for you to take my CPU and GPU and our open networking standard, but, actually, I’m going to design the system around you. So tell me what you want your data center to look like.’”
There is another competitive wrinkle here. ZT Systems designs, manufactures and deploys Nvidia systems—including for AWS and Azure, if the rumors are true. In that context, it is unclear exactly what the ZT acquisition means from a design standpoint for AMD’s prime datacenter AI merchant chip competitors, Nvidia and Intel. Once the deal is closed, I would expect all design activities for Nvidia and Intel to cease. AMD says that manufacturing for the competitive systems will continue, which makes sense assuming that the manufacturing group is indeed spun out and sold.
While some may understandably criticize AMD in some areas, pinpoint execution has become the hallmark of Su’s tenure. Precision execution is exactly what is needed to turn this investment into gold for the company by increasing revenue and taking market share. Compared to AMD’s Xilinx acquisition of a few years ago, this one looks simple. This deal further reinforces the lead that AMD and Nvidia have built on the rest of the pack for AI chips. I believe this acquisition will be positive for the company and enable it to take much more of that projected $400 billion datacenter AI market in 2027.
