Specialist
Senior Product Manager at Microsoft Corp
Agenda
- Tier 1 and Tier 2 cloud purchasing momentum for key electronics components and semiconductors into Q1 and Q2 2022
- Positioning of key cloud data centre suppliers, particularly Intel (NASDAQ: INTC), AMD (NASDAQ: AMD) and Nvidia (NASDAQ: NVDA)
- Adoption of custom designed semiconductors, particularly AI inference accelerators and ARM (Advanced RISC Machines)-based server CPUs (central processing units), highlighting Google (NASDAQ: GOOGL) and Amazon (NASDAQ: AMZN)
- CXL (computer express link), DDR5 (double data rate 5) and cloud DRAM and NAND purchasing outlook
Questions
1.
Could you give an update on the current operating environment for hyperscalers and their procurement since our last Interview [see Forum Pulse: Hyperscaler Data Centre Procurement Trends & Chip Shortage Impacts – Q3 2021 – 26 August 2021]? What activity are you observing in the market?
2.
What might this wait-and-watch approach you mentioned mean for CAPEX for hyperscalers or the actual outlay from a dollars perspective? Could there still be growth but potentially decelerating, or a drop-off as hyperscalers work through a bit of their inventory?
3.
How much difference is there between the major tier 1 cloud hyperscalers and the buying activity occurring in tier 2 cloud or supermassive large enterprise data centres? Are the purchasing dynamics different and are they equally cautious there? What are your thoughts on activity from other players that are not as large as Google or Amazon?
4.
You referenced flat YoY server purchasing. To what extent do you expect purchasing volumes to be oriented more towards refreshing existing hardware within the data centre vs new build-outs? Will the priority for 2022 be more refresh-focused, given the strength of buying over the last few years?
5.
Could you discuss the strength of demand for specific categories? Could you outline the buying requirements for CPU [central processing unit] vs GPU [graphics processing unit] vs memory going into Q1 and H2 2022?
6.
What expectation do you have for the unit volume percentage decline for CPUs? Is it 1-2% or 5%?
7.
What pricing increases do you expect YoY? Are they in line with your expectations, or could you outline the CPU pricing environment?
8.
Could you discuss the attach rate for GPUs on data centre servers going forward? How do you expect that to grow, based on some of the requirements that you may observe in AI training and inference, or video processing?
9.
To what extent have supply chain issues for certain categories been a headwind around deploying servers for the last year? I realise a large portion of this is contract-based. How might this impact the market?
10.
What are the current lead times for CPU and GPU? Are they at 26 weeks today?
11.
How much do you expect hyperscalers to be over-forecasting? Over the last year, double-booking and panic buying has been a theme in other sectors, so to what extent is this playing out with hyperscalers? How much are they choosing to hedge those risks?
12.
What is the current state of inventory for CPU and GPU memory for leading hyperscalers? Has there been a structural shift in the amount of inventory bigger hyperscalers will hold because of the last two years? Are they holding or building inventory such that we can expect a bigger pull-back in H2 2022 or into 2023?
13.
Could you elaborate on your expectations for DRAM and NAND pricing going into Q1-Q2 2022?
14.
What is the cost of outfitting a single server with DRAM or NAND? How could this trend going into 2022 and 2023? Could it be structurally higher? Could you explain the proportionality to DRAM buying and server and the server builds?
15.
You referenced custom-design ASICs [application-specific integrated circuits] used for offloading specific functions from the CPU. What’s happening with the silicon within the data centres today and to what extent is the current work being done to pull different functions out of the CPU? For example, how strong is the push there to offload networking capabilities onto a SmartNIC or drive up GPU attach rates?
16.
You referenced some of the things that have been done already and some of the announcements from some players. How aggressive have Google and Amazon been in pushing custom chips into their data centres? How do the top 3-5 largest and tier 2 cloud players such as Alibaba approach that? How concerted are the efforts from these types of second-tier cloud providers to also follow through on custom chip design?
17.
What proportion of workloads do you expect to shift to custom processors for the biggest tier 1 cloud players? Over what timeline do you expect this to play out, and what are your 2-5-year expectations around the presence of these custom chips within hyperscaler data centres?
18.
Does the CPU share loss of x86 disproportionately impact Intel vs AMD? Do you expect this to impact Intel a bit more given the issues around manufacturing it’s had and the competitive dynamic between both companies?
19.
Could you discuss GPU for AI training and inference? What are your expectations for custom silicon adoption in AI training and inference? Nvidia has talked a lot about AI inference and the advantages it can bring there. How is Nvidia positioned there vs how do you expect hyperscalers to strategise about inserting their own custom silicon on AI inference and training?
20.
It would seem hyperscalers’ performance wouldn’t be as good as what Nvidia can offer for AI training or inference. Are there specific types of workloads for which the custom chips would offer better performance economics? For example, what is your assessment of Google’s TPU [Tensor Processing Unit] or Amazon’s Trainium and do they compare to Nvidia’s leading edge?
21.
What are the market share splits in AI inference? There’s Nvidia, custom design and a series of start-ups that have talked about getting involved here. You referenced 50/50 for a custom silicon vs GPU for training. What’s your take on AI inference? Do you expect hyperscalers to try this on their own when using custom ASICs? As CPU or traditional x86 GPU fall off as a proportion of the total AI inference, what replaces it?
22.
To what extent are smaller start-up players in consideration, such as Cerebras, SambaNova, Graphcore and Groq? Do you think of any as potentially leading in what they offer?
23.
Do you have any updated thoughts on the positioning of memory players such as Samsung, Hynix and Micron? How could these players be positioned going forward given some of the developments or announcements you’ve observed around their roadmaps? What are your thoughts on DDR5 adoption?
24.
Is there anything you would like to highlight, such as the companies you think are worthy of a specific shout-out? Do you have final thoughts?
