AMD plans to spend $4.9 billion to acquire hardware company ZT Systems. It should get some of its money back when it spins off ZT’s manufacturing operations, leaving AMD with design talent and server-deployment capabilities. Already serving about a quarter of the server-processor market with Epyc, AMD’s aim with the acquisition is to sell more AI accelerators (GPUs/NPUs). The stock market and commentators reacted favorably to the news.
AI Market Potential Justifies Acquisition Cost
The acquisition amount is a huge sum, but data-center NPUs are a huge business. Simplifying the analysis, assume ZT ultimately costs AMD $4.5 billion, assume AMD’s Instinct NPUs have a 50% gross margin, ignore the cost of capital, and assume a target payback period of three years. The chipmaker must sell $9 billion more Instincts, or $3 billion per year on average, than it would’ve otherwise. That’s a big hurdle, but AMD MI300 sales are already at a $4 billion annual run rate.
Too Little or Too Much?
The optimal forward integration for a chipmaker is unclear. In AI, market leader Nvidia has gone as far as to offer systems and whole racks. Without analogs to Nvidia’s networking technologies, targeting hyperscalers that have proprietary hardware, and needing to preserve PC and server OEM relationships, AMD has taken a half-step forward. This could be a metastable degree of integration if customers really fall into one of only two camps: those requiring only chips and those requiring complete systems.
However, semiconductor companies have become risk averse over the past dozen years. Gone are the days when they committed to an investment without tangible customer support. We suspect that at least one hyperscaler all but required AMD to bulk up its design capabilities before it would increase Instinct purchases.
Skepticism is in Order
Nonetheless, skepticism is in order. ZT isn’t AMD’s first system-company acquisition under Lisa Su’s leadership. In 2012, when she was AMD’s general manager, the company acquired SeaMicro for $300 million. Three years later, it shut down the business. Those were dark days for AMD. It lacked a competitive CPU, a problem that a company building microservers from Intel Atom processors wasn’t going to solve. To justify the positive reviews earned upon announcing the deal, AMD must show within a few quarters that ZT is improving its top line.
Bottom Line
AMD is likely trying to meet customers’ needs by taking a half-step toward matching Nvidia’s forward integration but without the full step that would land on their toes. The acquisition also advances AMD’s capabilities beyond those of smaller AI competitors, including Intel, which in the past offered reference architectures for its data-center products. For AMD to succeed as an NPU supplier, it must rapidly grow sales. System expertise should help the company win business and accelerate deployments. Within a year, we should know if the ZT acquisition is paying off.