Nvidia's (NVDA) high-flying shares have come under pressure lately due to concerns about tougher competition, but those concerns aren't cause for panic at this point.
The concerns -- caused in part by remarks from VC Chamath Palihapitiya, and perhaps also by new details shared by Alphabet/Google (GOOGL) about a chip it unveiled earlier this year -- have stoked fears that Nvidia's booming server GPU business will face stiffer competition in the battle to power AI-related data center workloads.
And this debate is likely to keep raging in the weeks to come, as Intel (INTC) and various startups make fresh announcements about AI server co-processors that will soon ship in volume. But as it does, a few things are worth keeping in mind:
- A large and booming market that has been seeing limited competition will probably see new players join the fray.
- Nvidia doesn't have a monopoly on designing chips that are well-suited for AI-related work.
- Nvidia does have an unmatched developer ecosystem that serves as a big competitive advantage in one valuable portion of the AI co-processor market.
- Nvidia appears intent on doing whatever it takes -- via R&D investments, partnerships and other moves -- to hold onto its current leadership position.
- Enterprises and cloud providers thinking about adopting a new chip or hardware platform tend to take their time evaluating it before signing off on large orders.
The AI co-processor space certainly looks a lot more competitive than it did 12-to-18 months ago. In October, Intel unveiled its Nervana Neural Network Processor (NNP); it relies on technology obtained via last year's Nervana Systems acquisition, and is positioned as an alternative to Nvidia's flagship Tesla V100 GPU for the very demanding task of training deep learning algorithms to do things such as detect objects within images, analyze voice commands and make content recommendations.
Intel has made impressive claims about the NNP's memory bandwidth, I/O connectivity and (thanks to a unique architecture it calls Flexpoint) processing efficiency. It has also been with major Nvidia client Facebook (FB) to test and optimize the NNP.
Another big Nvidia client, Google, has begun deploying a second-gen version of its Tensor Processing Unit (TPU) within its data centers. Whereas the first-gen TPU (unveiled in 2016) could only handle AI inference work (e.g., running AI algorithms against real-world requests), a competitive space in which Intel, Nvidia, Xilinx (XLNX) and others participate, the second-gen version can also handle training, a field where Nvidia has dominated.
Google claims the chip, which is deployed via boards containing 4 chips and is believed to have been developed with Broadcom's (AVGO) help -- can deliver a whopping 45 teraflops (TFLOPs) of performance and supports 600GB per second of memory bandwidth. One qualifier: Unlike Nvidia and Intel's offerings, which support many AI software frameworks, the TPU is only meant to work with Google's (admittedly popular) TensorFlow framework.
And as noted in a recent column by The Information, quite a few chip startups are now also taking aim at Nvidia. These include Cerberas Systems, which says it's working on a training chip that provides unmatched memory and I/O bandwidth; Graphcore, which says it's developing a training chip featuring over 1,000 cores and promises a big performance edge relative to Nvidia; and KnuEdge, which is working on a unique chip architecture that both supports thousands of cores and can work in tandem with various third-party chips (CPUs, GPUs, FPGAs, etc.).
But as Nvidia faces all this competition, the fact that the company's CUDA GPU programming interfaces (APIs) effectively become an industry standard among AI researchers looking to train deep learning algorithms remains a giant advantage. In addition, the tools, software runtimes and APIs found in Nvidia's Deep Learning software development kit (SDK) have been widely adopted for both training and inference work.
Thanks to all of this, it probably won't be enough for a rival to develop a training solution that moderately outperforms Nvidia's. Major price/performance gains will be needed to make big Nvidia clients think about jumping ship.
And Nvidia is certainly doing whatever it can to keep itself from falling that far behind. The Tesla V100, which has seen brisk sales since launching this summer, is a monster of a chip that provides major performance gains relative to its predecessor (the Tesla P100) with the help of a new type of processor core (called tensor cores) specifically meant for machine learning work. And a successor to the V100 should arrive next year.
Nvidia's AI co-processor lineup also includes inference-focused products such as its Tesla P4 and P40 GPUs, and was recently expanded to include the Titan V, a $3,000 PC solution that helps AI researchers test machine learning models. Between its server and automotive efforts, AI-related investments appear to make up a sizable portion of Nvidia's R&D spend, which is expected to total $1.8 billion in fiscal 2018 (ends in Jan. 2018).
In addition, as a visit to the news section of its website makes clear, Nvidia has been working overtime to extend its AI ecosystem and partner base. Over the last couple of months alone, the company has unveiled AI-related alliances with everyone from GE Healthcare to 3D rendering software firms to construction giant Komatsu to Taiwan's Ministry of Science and Technology. It has also unveiled new partners for its Deep Learning Institute, which aims to train AI researchers in the use of its products, and announced investments in several startups whose machine learning-powered offerings rely on Nvidia GPUs.
Nvidia's share of the AI training co-processor space probably won't stay at its present sky-high levels over the long run. Intel might especially be well-positioned to gain some ground towards the back half of 2018, given the NNP's specs and the chip giant's huge enterprise and cloud reach. And there's a chance that at least one of the AI chip startups making big promises will eventually back up its talk.
But between the company's ecosystem, R&D investments and stellar execution, one shouldn't bet just yet on Nvidia's leadership position being seriously threatened. And it's worth keeping in mind that the pie that Nvidia, Intel and others are battling over has been growing pretty rapidly.
Jim Cramer and the AAP team hold positions in Nvidia, Alphabet, Broadcom and Facebook for their Action Alerts PLUS Charitable Trust Portfolio. Want to be alerted before Cramer buys or sells NVDA, GOOGL, AVGO or FB? Learn more now.
More of What's Trending on TheStreet: