Lead Tech Analyst and CEO Beth Kindig recently joined Real Vision’s Nico Brugge to discuss her AI outlook on leading AI stock Nvidia, while sharing which AI stock she believes may outpace Nvidia’s returns through 2030.
This AI stock’s opportunity is in the AI inference market, which will begin to take shape when large language models (LLMs) migrate and operate locally on AI-capable client devices, such as PCs and smartphones. Kindig has boldly stated in Forbes, and on CNBC, and Bloomberg that Nvidia will reach a $10 trillion valuation by 2030. Yet, she believes this AI stock may outpace Nvidia’s stock and provide investors with a larger percentage return.
Click here to watch the full interview on RealVision.
We built a leading AI portfolio beginning with Nvidia’s AI thesis in 2018, with our AI allocation of 45% in 2023 helping push us to a 131% cumulative return since inception. Now, we’re closely tracking what we believe is one of the next explosive growth waves in AI – and it’s not the cloud. Learn more here.
Training Versus Inference
Nvidia had surged to briefly become the world’s most valuable company due to its impenetrable moat in the data center GPU market, which was built upon the CUDA software platform for the purposes of training AI models. Eventually, we will see a shift from AI training to AI inference, which leaves the market open for competitors.
Kindig explained in the interview with Brugge that Nvidia’s H100 transformer engine was the impetus for Chat-GPT’s moment. Chat-GPT, and its competitors, are essentially large R&D departments for training models. We are in the midst of AI training, and what follows will be the AI inference market. As Kindig explains, “when you take the models and you bring them to the edge, and you run those models and have it make predictions based on live data for actionable results, that’s inference.” She pointed out that “for the most part, it’s agreed that inference will be a larger market than training once the ecosystem is mature.”
Currently, there’s one primary headwind to the inference market; devices are not powerful enough to handle the requirements to run AI at the edge. Kindig says that “one of the things holding back inference is our client devices, so our PCs and mobile. Inference runs best close to the data, and we don’t have powerful enough devices for inference, for where AI needs to go.”
AI PCs are currently working on solving this critical bottleneck, with NPU, GPU and CPU equipped devices packing the necessary power and efficiency to operate AI models locally, on-device and without relying on data being sent to and from the cloud.
Kindig told Brugge on Real Vision that she believes one AI stock is well positioned to capitalize on the long-term opportunity arising in AI inference — that stock is AMD.
Why AMD Can Outpace Nvidia Through 2030
Nvidia will need to rise nearly 250% by 2030 to reach Kindig’s $10 trillion target, yet she thinks AMD has the potential to provide a larger return over that time frame.
She told Brugge that her “time horizon would be that we see really nice movement by 2027, but we really need this 2030 time period to play out, and there’s a few reasons. Number one, Nvidia has the training market cornered right now. Training requires a lot of compute power, and they’ve gone through architectural changes that have defied Moore’s Law. This is things like Tensor Cores, which do matrix computations; floating-point precision, moving from 16 point [FP16] to 8 point [FP8], [the transformer engine switches back and forth which] increases accuracy while also increasing speed [depending on the workload]. So, all of those things, Nvidia has 98% of the GPU market and is crushing it, but a lot of that is training.”
Core to this thesis on AMD is giving time for the budding inference market to take off and mature – Kindig explains that “where AMD is going to compete with Nvidia is a market that is very early, so we need time for that to mature, which is inference. Many people may get that confused, because we are fully in the AI market today because Nvidia is putting up those huge data center numbers. We are in the data center training market today; one day, we will be an AI market led by inference.”
Kindig told Brugge that there are a “few reasons” that AMD could do better than Nvidia in inference and etch a niche, with the primary reason being that inference is “one way to circumvent CUDA.” CUDA is Nvidia’s proprietary software stack that has essentially locked developers into its GPU ecosystem, and what has driven its ~98% market share in AI GPUs.
For a deep dive on CUDA and how it’s Nvidia’s moat and first line of defense in the AI accelerator market, read more here and here.
How AMD Can Fend Off Nvidia
AMD is equal to Nvidia on hardware in many regards, but CUDA has locked in Nvidia’s monopoly; however, it’s likely that Big Tech and developers will seek alternatives to CUDA to limit reliance on Nvidia for the entirety of the hardware stack for AI development.
Kindig notes that CUDA will be the “biggest hurdle for sure” for AMD to compete against, “but after that, it’s probably product roadmap versus product roadmap, meaning that for everything AMD does, can Nvidia do better, by 6 months.” Put differently, Nvidia took the industry by storm with its transformer engine-equipped H100s, which saw extreme demand outstrip supply for multiple quarters. No company could compete at the time with a similarly spec’d GPU that could provide the same level of AI computing performance.
