To The Moon

Summary

• Advanced Micro Devices called its AI aspirations "engagements" on its earnings call, but engagements don't bring in revenue.

• Engagements are merely proof-of-concepts and may not lead to sales.

• Nvidia Corporation continues to bring in revenue invoices with its complete H100 AI package in hardware and software.

• Until AMD gets through its "long journey" in software, Nvidia's products should remain on top, especially with Hopper-Next releasing while AMD just begins ramping MI300.

AI,Circuit board

Andy

There's a pretty intense debate around Advanced Micro Devices, Inc.'s (NASDAQ:AMD $Advanced Micro Devices(AMD)$ ) ability to compete with Nvidia Corporation (NASDAQ:NVDA $NVIDIA Corp(NVDA)$ ) in the artificial intelligence, or AI, space, especially in the data center sector. And the former is chasing the latter because it doesn't have a competitive product being bought up by customers. Not yet, at least, says one side of the debate. And just last week, with AMD's earnings report, an interesting choice of words was used by CEO Lisa Su to describe the chase: engagements. Engagements is a word with a very particular meaning in the business world. But Nvidia doesn't describe its H100 AI accelerator business using this word, yet it'll report a massive increase in its data center revenue in a few weeks with receipts, the kind that comes from invoices. So what gives with the wordplay?

The culmination of AMD's AI ambitions rests on its MI300 accelerator product. It's marketed to be AMD's direct challenge to the fast-adopted H100 accelerator from Nvidia, the latter being in the field for about nine months at this point. The MI300 is supposed to be AMD's answer to take AI data center market share away from Nvidia. The problem is it's still in pre-production, launching and ramping production in Q4 of this year. That means the direct competitive product is, at best, a year behind Nvidia's.

Now, just because Nvidia has a head start with its product doesn't mean AMD can't come on the scene, rip some wins out of the hands of Nvidia, and steal the show. It can certainly happen - though, there's a software aspect to it that I'll discuss later.

The problem is there's no real way to know at this point, and "engagements" are about as cheap as water from the ocean.

But what are engagements, exactly?

Defining The Situation

According to Lisa Su, they're customer "programs" with the intention of deploying its MI line of AI accelerators (emphasis added):

AI cluster engagements grew by more than seven times sequentially as multiple customers initiated or expanded programs supporting future deployments of Instinct MI250 and MI300 hardware and software at scale.

- Lisa Su, AMD's Q2 '23 Earnings Call.

In the business world, this is the very first level of working with a potential vendor or customer (depending on which perspective you're coming from). To "engage" with a vendor means to get more information about what it can provide and go as far as getting its hands on the product to mock a proof of concept. Proof of concepts can take anywhere from a month to six months or longer before any business decisions are made; it all depends on the complexity of what's being proven.

These engagements are what Nvidia had to do over a year ago when it began sampling its H100 accelerator. Those engagements have turned into invoices. But before I get too far down the road, I'll map out where AMD is right now on the timeline.

With MI300 production not starting till Q4, the volume ramp will trail into 2024. So while the mass production of the product begins, it won't have sufficient inventory to sell until a few months later once it ramps. Then there's the delivery to the customer and the installation and setup thereafter. If AMD is just now having proof of concepts shipped, there's a minimum of three-to-six months before these companies decide to purchase. Moreover, AMD is only providing "early system access" to its lead customers - not all of them.

Engagements with top-tier cloud providers, large enterprises, and numerous leading AI companies significantly expanded in the quarter. We are providing early system access and sampling both products with our lead AI, HPC, and cloud customers now...

- Lisa Su, AMD's Q2 '23 Earnings Call (emphasis added).

So even the engagements are only turning into proof of concepts with its lead customers. Those lead customers will likely apply the most scrutiny to their proof of concept. However, if they approve, the remaining customers may be more comfortable with their proof of concept. But with AI, the product must work for each of them specifically, so there are very few shortcuts to the evaluation.

But more to the point, if only initial lead customers are evaluating, the number of customers lining up to order will be thin come Q4. Furthermore, this puts the purchase decision late in the fourth quarter. Delivery and installation will not likely occur until Q1 '24 once volume production allows for shipment. From there, widespread availability may not come until Q2 '24.

This is corroborated by other analysts in the tech space (emphasis added):

...the chip will only begin sampling to major cloud and data center providers in the third quarter, with production to ramp up in the fourth. Some analysts believe that means the MI300 won't be widely available until mid-2024; that would put it about 18 months behind Nvidia's H100, which began shipping in volume at the end of 2022 and beginning of 2023.

With engagements only happening now, it means no revenue for the next two quarters, at the very least, with the third quarter beginning to show some receipts. So the process for AMD to turn engagements into invoices is nowhere near where the market would like it to be. Incremental revenue gains in the data center segment for MI300 won't happen until Q1 '24 at best.

Engagements Don't Necessarily Lead To Invoices, Software Kind Of Does

To move into the more speculative side for a moment, these engagements may also be a red herring altogether. In my assessment above, I'm giving the benefit of the doubt AMD will see these engagements turn into invoices. They instead may be cloud and AI customers trying to find other solutions to Nvidia's H100 package because they can't order enough H100s to fulfill their AI needs. After all, Nvidia admitted to ramping volume up more than it was to satisfy demand in 2H '23. So these customers kick the tires with the only other vendor in the field to see if they have something to offer to supplement their AI needs.

Ultimately, these customers may find the alternate solution in their proof of concept unworkable or too costly to implement. You might be thinking, "How can it be too costly? Nvidia's H100 setup is the most expensive thing a cloud provider can buy right now." While this is true, this is where software comes in. Another solution requiring more software integration and several new software components may require more labor (development time) to get to a workable solution. This is TCO - total cost of ownership.

With Nvidia's accelerator, customers are getting its proprietary package. Some may say this is the exact problem. However, it's also a supported solution where Nvidia is incentivized to work with its customers to tweak the software because it can provide the most expertise, and its customers become entrenched. AMD is working toward an open-source software solution. While this sounds free and broad-community supported, it also requires a lot of customization to tailor it to the customer's needs. And much of that lands on the customer to figure out.

Even beyond my skepticism, AMD admits its software path will be long. Cloud customers don't have "long" (emphasis added):

AMD added that it had developed its own open-source software stack called ROCm, but also acknowledged it had a long "journey" ahead on the software front.

And AMD is open-sourcing because it cannot work on the software solely internally, either because of a lack of talent or dedicated resources, or both.

Nvidia has been working on its software for over 16 years (with CUDA), while its AI package has been around arguably since 2017. This has made Nvidia a software powerhouse and not just a hardware engineering company. AMD's reliance on open-source software, which is all-new, requires it to make the same inroads while also dealing with the open-source community to advance the software. There are certainly advantages to open-source software; I use them every day. But for the biggest customers in the world, a proven and minimal configurable-intensive solution is what moves business forward. AMD has always been a follower in software, while Nvidia has been a leader, and this is still the difference between the two.

Engagements Are Not Investment Worthy

Engagements are a far cry from revenue, especially without a proven product. Sure, benchmarks here and there may show AMD's MI300 beating out the H100 in certain aspects and certain tests, especially on the memory side. However, pure hardware performance is nothing more than a paperweight (a very expensive one) without software. And Nvidia maintains a significant lead in software for GPUs and AI compute.

Also, none of this addresses Nvidia readying its next-generation product just as AMD is at full production levels. It will put AMD behind again in hardware before it ever gets going.

In March of next year, Hopper-Next (believed to be called Blackwell) will be launched. This would put Nvidia, without question, back on top of the hardware side. I expect Nvidia will up its memory game and procure Micron's (MU) HBM3 Gen2, which will put to rest any questions about Nvidia's accelerator's ability to keep up in the memory bandwidth bottleneck game. The timing aligns for Micron's product release of early CY24 with a likely Hopper-Next Q4 '24/Q1 '25 release.

"At the core of generative AI is accelerated computing, which benefits from HBM high bandwidth with energy efficiency," said Ian Buck, vice president of Hyperscale and HPC Computing at NVIDIA. "We have a long history of collaborating with Micron across a wide range of products and are eager to be working with them on HBM3 Gen2 to supercharge AI innovation."

Overall, I wouldn't be too excited about AMD's AI data center prospects. Its MI300 is only in the engagement phase of the sales cycle, while Nvidia is finalizing its Hopper-Next accelerator architecture. Once AMD overcomes the "long journey" in software, I'll be all ears. Until then, AMD's revenue may be incremental to the recovery in data center as a whole - if it ever recovers to the same levels, which I doubt - as GPUs begin to take over the data center architecture, pulling CPU demand down for AMD and Intel (INTC) over the next few years.

Source: seeking alpha