Generative AI, how many chips are needed?

Since AMD launched its "Antares" MI300X and MI300A compute engines in early December, we have been carefully considering overall AI spending forecasts, and more specifically infrastructure and accelerator spending forecasts. With generative AI marketing soaring at what seems like escape velocity, it’s important to temper lofty expectations.

At the Antares conference, AMD CEO Lisa Su said the company is revising its forecast for the data center artificial intelligence accelerator market to include GPUs and other devices, and will likely allocate some of its CPU spending to artificial intelligence Work smart. Actually happens in the real world. The revision was so eye-popping that we had to seek some sort of validation or alternative opinion from market researchers around the world due to AMD's stated expectation of generous funding.

Specifically, a year ago, AMD expected the total addressable market for data center AI accelerators to reach $30 billion by 2023, growing to more than $150 billion by the end of 2027 at a compound annual growth rate of about 50%. That sounds like a lot of money, but with revisions a year and a half later, Su said AMD's data center AI accelerator market will reach $45 billion this year and will grow at a compound annual growth rate of more than 70% increase. It will reach over US$400 billion by 2027.

Now, keep in mind that this prediction only applies to AI accelerators. Not the server host, not the interconnect, not the main memory and flash on the host, not any software or services for the AI ​​platform being built. We've been thinking about this, but it still sounds too big to be true. (Of course, the IT world could get even crazier than it already is, and this model seems to suggest that's exactly what's going to happen.)

This week we were pleased to see a report from IDC discussing spending on generative AI hardware, software and services, which also included some commentary on the broader AI market, allowing us to build our own The model looks like it might not be available until 2027. We openly admit that this model is a bit like knitting a six-foot-long scarf out of a piece of dental floss. However, in the absence of a complete IDC report, we have no choice but to take what makes sense we can find and build on it.

By 2023, companies will invest more than $19.4 billion in generative AI. When we were confused about how CAGR is calculated, we spoke with IDC's press relations staff and learned that IDC believes that $6.8 billion will be spent on generative AI hardware, software, and IDC stated in the promotion of its report that by The global generative artificial intelligence market will more than double to US$40.1 billion in 2024, with a compound annual growth rate of 86.1%, reaching US$151.1 billion by 2027.

This paragraph from IDC’s public statement on the global generative artificial intelligence market is critical:

“GenAI infrastructure, including hardware, infrastructure as a service (IaaS), and system infrastructure software (SIS), will be the largest area of ​​investment during the build-out phase. But by the end of the forecast, GenAI platforms and application software will gradually surpass infrastructure, The five-year CAGR is 99.6%. Similarly, GenAI services (including IT and business services) will be almost equal to infrastructure spending, with a five-year CAGR of 94.2% by the end of the forecast. It will account for 29.0% of overall AI spending, a significant increase from 10.8% in 2023.”

We plugged these data points into the spreadsheet as usual and tried to fill in the gaps as usual. Of course, we don't guarantee that these numbers are IDC numbers, other than the black ink numbers, and again we point out that any numbers in bold red italics are our estimates based on our assumptions about growth rates and the difference between endpoints and years. How generative AI spending was divided between hardware, software and services last year.

We do this just to try to get a first-order approximation of the total addressable market for generative AI and overall AI. We think we can do a much better job than any GenAI chatbot given the same information.

Anyway, take a look at the table we generated, then we'll browse it:

First, if we know the overall spending on GenAI in 2023 and 2027 and its share of overall AI spending, we can calculate the overall level of AI spending in these two years. It's just simple math, which is why these numbers appear black. If we assume that AI spending will grow at roughly the same revenue growth rate in 2022 and 2023, we can estimate that GenAi will account for approximately 6.2% of all AI spending in 2022. Once we have that data point, we can not only calculate the results that work Spending on all AI hardware, software and services will reach $109.2 billion by 2022, and this also represents a CAGR of 36.7% over the forecast period, with total AI spending reaching $521 billion by 2027 Dollar.

Now, let's talk about GenAI's division between hardware, software, and services, as hinted at in that big announcement above. Our best guess is that spending on GenAI infrastructure will be approximately $3.2 billion in 2022, including $1.9 billion on GenAI platform and application software and $1.7 billion on GenAI services. Using the CAGR figures for these two categories (generally speaking, GenAI software and services), we can calculate the revenue levels in 2027, keeping in mind that we have to have GenAI software be larger than GenAI hardware by 2027, and by 2027 By 2020, GenAI hardware will exceed GenAI services.

The number we show in our GenAI hardware, software, and services spending estimates is obviously not the only way to bring this data together—it's just one possible solution to the simple equation behind it all. But we think the numbers pass the smell test. If IDC's vision of the GenAI world and our interpolation and interpretation of the data disclosed to the public are "true," generative AI hardware sales will exceed $50 billion in 2027, and global generative AI hardware sales will Total sales were $151.1 billion. AI spending.

If the split between hardware, software, and services were similar for all types of AI, roughly $125 billion of the $521 billion in total AI spending would be spent on hardware. This applies to all hardware, not just the accelerators in AMD's predictions we cited above, which makes a lot more sense to us in a market that's going to see not only intense acquisitions, but also increasing competition. number.

This all assumes GenAI doesn't dominate all AI spending. Even IDC believes GenAI accounts for less than one-third of total AI spending.

*Disclaimer: This article is original by the author. The content of the article is the personal opinion of the author. The reprinting by Semiconductor Industry Watch is only to convey a different point of view. It does not mean that Semiconductor Industry Watch agrees or supports the view. If you have any objections, please contact Semiconductor Industry Watch.

Today is the 3624th issue of "Semiconductor Industry Observation" shared with you. Welcome to pay attention.