There's peril and promise in generative AI

What makes a good technology stock? Dominic Rizzo, Portfolio Manager of Global Technology Equity Strategy at T. Rowe Price can answer that question, having invested in technology stocks since 2015 across three regions (the Americas, Europe and Asia), across all different market capitalisations from small, medium and large through to mega cap, and across all three subsectors of hardware, software and translational technologies. 

His answer is to adopt a back-to-basics approach, which looks hard to do when you are considering something as new as generative AI. In this wire, I take a look at his approach, before looking at the generative AI ecosystem and Rizzo's way to understand the opportunities across the three subsectors. 

Note: This wire is based on the webinar, Unveiling the Promise and Peril of AI: Insights for Investing in the Global Technology Sector, on Thursday 19 October 2023. 

Rizzo’s investment framework

Rizzo's investment framework for technology stocks is based on four factors: 

1. Lynchpin technologies: Rizzo starts by identifying companies that sell mission critical lynchpin technologies. “I want to find those technology companies that, if you pulled them out of the ecosystem, their customers would be dramatically worse off," he said.
2. Innovating in a secular growth market: Rizzo suggests that we think of companies that are taking market share and are fast while also growing it. 
3. Improving fundamentals: Rizzo looks for accelerating revenue (growing faster in the future than in the past), operation margins that are expanding, and proven free cash flow conversion that is improving.
4. Reasonable valuations: According to Rizzo, "Where I think you get burned with technology stocks is when you buy the incredibly expensive stocks because those are too expensive and also the really cheap stocks. They’re really cheap for a reason, they’re often selling lifeboats to sinking ships."

Technology innovation cycles

Rizzo believes that to understand generative AI and the investment opportunities, you also need to understand how computer technology goes through a cycle of centralising computing to decentralised computing to centralised and decentralised again, and that cycle repeats. Take hosted main frames (centralised) to client servers (decentralised) to the world-wide web and sites like Amazon's AWS (centralised). 

Machine learning, Rizzo reminds us, goes back decades. What's new is generative AI. Rizzo believes we are at the centralising stage. As he says,

"AI, particularly generative AI and the training of the large language models right now means we’re in the centralised compute phase building out these huge GPU clusters and hyperscalers in order to train these models. Eventually we’ll shift over to decentralised stage, putting that AI at the edge: AI on your smart phone and in the car."

Generative AI uses complex, highly capable analytical components: hardware in the form of graphic processing units to undertake parallel processing; and large language models to supply the context. As he notes, "Large language models can be applied to chat, ideas generation, cybersecurity, theft detection, software writing, so many different areas, ship design, semiconductor design."

If generative AI can be put into effectively every sector of the economy, what does this mean for the end markets? For the technology sector? According to Rizzo, 

"This is an incredibly silicon-intensive process, lots and lots of GPUs from the likes of Nvidia (NASDAQ: NVDA), and potentially from the likes of AMD (Advanced Micro Devices (NASDAQ: AMD)). Hyperscalers are going to try and design their own chips," he said. 

Rizzo notes that AMD thinks the AI chip markets is worth around US$30bn in 2023 but will scale to US$150bn by 2027. Placing this in context, he reminds us that the entire current semiconductor market today is US$500bn. AI chip manufacture is adding an incrementally large cap on top of that existing market.

As for software, "We’re going to put AI into every software," says Rizzo. He cites two examples illustrate the potential: 

• Synopsys (NYSE: SNPS) sells software that chips are designed on. Put AI into the software, they get a 20% increase in contract earnings - because AI adds so much productivity to their customers.
• Microsoft (NASDAQ: MSFT) Co-pilot launches in November, with Rizzo noting it will cost roughly US$30 per user per month, "but will add huge amounts of productivity to their customers."

The ecosystem for generative AI investment opportunities

When Rizzo thinks about industries using generative AI and their different stages of growth, his considers who helps those companies to adopt generative AI by enabling it. 

The first of these is the chip ecosystem, what Rizzo calls, "the foundation of artificial intelligence." This includes "companies you know and have heard of, like NVIDIA (NASDAQ: NVDA), but there are many different lynchpin companies within that ecosystem."

Nvidia makes their chips at Taiwan Manufacturing Sector Company (NYSE: TSM). They make it on equipment from ASML Research (NASDAQ: ASML), design their chips on software from Synopsys and Cadence Design Systems Inc (NASDAQ: CDNS). The servers consume a ton of memory from the likes of Samsung Electronics (KRX: 005930). AI servers can consumer anywhere from four to eight times the amount of memory than a traditional server does." he said.

Secondly, there are the infrastructure enablers, who are actually going to help implement this incredibly complex technology. "The most obvious ones are the large cloud companies in the US like Amazon AWS and Microsoft Azure," he said. 

"You will also need IT service companies to help you implement this technology if you're an enterprise."

"Over time, one of the most obvious use cases for artificial intelligence is writing new software. Which means you'll need a lot more databases," Rizzo said. 

Next comes the foundational models. 

Rizzo notes there are many of these in the marketplace today. The most important large language model is GPT-4 by OpenAI. He also noted Microsoft has Turing-NLG. He called out how Meta (NASDAQ: META) has also done some really great work on open-source language models with its LLaMA technology. There is also Google with PaLM2 and Amazon with its Titan text. 

Then, there are the open-source models. Rizzo thinks the open-source community will be really important for artificial intelligence. 

"I think the open-source community will be one of the ways we will get to see the AI engine over time," he said. 

Finally, there is the application layer at the top. Rizzo thinks this is the hardest part to understand.

"We don't know who is going to win at the application layer. Yet at the beginning, right now, we are beginning to see things like Chat GPT, Google Bard, Amazon Code Whisperer, beginning to take off. Again, it's very early," he said. 

Productivity enhancers and speculative bubbles

Rizzo's view is that Moore's law - the number of transistors in a given area would double every two years and this would historically happen at roughly the same cost - doesn't readily apply to AI chip. He explains further:

"It's becoming really hard to make these incremental chips. Moore's law is slowing down. And you have to trip the chips into doing different things, better things, in different ways. 

This means using bigger pieces of silicon: the GPUs are enormous. Another way is via chiplets. There are many different things you have to do to trip the chip to get more performance for roughly the same cost. Yet it's no longer roughly the same cost. The capital intensity of manufacturing has on up a lot...In 2023, we're on a path to north of US$100bn if this capital intensity keeps up."

But the potential for productivity enhancers is huge.  And while that means opportunity, Rizzo cautions that in history, other productivity enhancers have cause speculative bubbles: the railroads, telecommunications, the internet. 

"I'm not saying that artificial intelligence is going to definitely result in a speculative bubble. All that I'm saying is that, historically, productivity enhancers for the global economy can result in speculative bubbles," he says. 

Successful navigation 

Rizzo has two tools he uses to navigate this field responsibly - his investment framework (see above) and a framework for AI, to identify the AI winners:

• Compute resources and capital: AI is very expensive. Look to see if the company's capital expenditure budget is forecast to increase each quarter. 
• Data sets: So companies can rip apart the information and get the insights.
• Distribution: Rizzo notes that a company needs to be able to take AI, push it into the customer base, in order to monetise it. "Because if you don't monetise it, how are you going to pay for it?" he asked.
• Talent: Rizzo notes there are only a few people in the world who can do this, and the major companies in the US, the hyperscalers, have cornered the market on a lot of this scarce talent. Over time, with AI software being able to help generate its own code, talent will be more diffuse.

And while there are disruptive innovations which create new winners, Rizzo doesn't think AI is a disruptive innovation. It's a sustaining innovation, where the old winners before the innovation become even more powerful, as they adopt and adapt the innovation into their products and businesses. 

Invest with confidence