Deep learning is having a profound effect on the microprocessor and server markets. In the face of a declining server market during the last six months, NVIDIANVDA tripled its data center revenues. The reason? Deep learning applications like voice search, image recognition, and translation are taking off, and NVIDIA’s graphics processing units (GPUs)1 are the fastest processors for training deep learning applications.
NVIDIA’s rapid penetration of the data center has pushed its market capitalization to $90 billion. It has eclipsed Qualcomm’sQCOM market cap and is more than half that of IntelINTL. Now the market is grappling with whether or not its growth is sustainable. Important to ask, how large is the data center market for accelerator chips like the GPU?
At its 2017 Investor Day, NVIDIA provided an estimate of the total available market (TAM) for the data center—$30 billion by 2020. This is a startling number—it’s 19 times NVIDIA’s data center revenue run rate of $1.6 billion and larger than Intel’s $17 billion in data center revenue last year.
To do a reality check on these estimates and dimension the addressable market for GPUs and other accelerators, we analyze how server revenue flows into silicon revenue, shown in the chart below. According to IDC, global server revenues were $53 billion in 2016, of which $46 billion were x86 servers and $7 billion, non-x86 servers like mainframes and IBMIBM Power systems.
The bulk of the $46 billion in x86 server revenue pay for components such as CPUs from Intel, GPUs from NVIDIA, and memory from Samsung. The rest is kept by server manufacturers such as HPHPQ and Dell as gross margin. According to our estimates, x86 manufacturers margin is roughly 10%, leaving 90% or $41 billion to the component manufacturers.
Intel’s Xeon server CPUs generated ~$13.6 billion2 in revenue, accounting for roughly a third of the $41 billion server component market. Despite rapid growth, NVIDIA’s GPU revenues accounted for $0.8 billion, less than 3% of component spend. The remaining $26.6 billion of component sales went toward memory, storage, and motherboards.
Ultimately, NVIDIA’s revenue potential will be determined by the number and intensity of GPU accelerated data center applications. Basic web applications require no accelerators. Video streaming and analytics applications benefit from a single GPU per server. At the most extreme, training deep learning applications require up to 8 GPUs per server, allowing NVIDIA to capture 75% of the component TAM.
So, how large will the market be for accelerators in the data center? As shown below, in five years, if accelerators capture 10% of server component spend, we estimate that would generate $4.5 billion in silicon revenue. If adoption is particularly strong, say 20%, this doubles to $9 billion. In our view, $4.5–$9 billion is a reasonable range for potential data center accelerator revenues. To reach NVIDIA’s $30 billion scenario, accelerators would have to capture 65% of the component market within three years, which seems unlikely.3
A key assumption in this analysis is that total server spend will remain flat over the next five years. While this may seem overly pessimistic,4 server revenue from 1996 to 2016 has not grown5 despite a massive internet boom spanning PC, mobile, and cloud. Moore’s Law along with virtualization software made it possible for servers to do much more at a lower cost. We expect GPUs and other accelerators will continue this trend.
- We use GPUs and accelerators interchangeably in this article. The latter includes FPGAs and ASICs. These chips are also used for non-AI workloads, which for simplicity, we include in our deep learning TAM estimates. ↩
- We assume 80% of Intel’s $17.2B data center revenue was for CPUs with the balance going to storage, networking, and other components. ↩
- This implies 87% of all processor spend will go toward GPUs. If all software became deep learning programs, this might happen. ↩
- This time may be different and we could see sustained server growth over the next few years. We modeled this scenario but found that it did not materially affect the accelerator TAM. Component share remains the dominant variable. ↩
- IDC Server Tracker, 1996–2016 ↩
ARK's statements are not an endorsement of any company or a recommendation to buy, sell or hold any security. For a list of all purchases and sales made by ARK for client accounts during the past year that could be considered by the SEC as recommendations, click here. It should not be assumed that recommendations made in the future will be profitable or will equal the performance of the securities in this list. For full disclosures, click here.