The broad-brush shape of the server market was drawn up many years ago:
- x86 processors took over from vendor-designed (and sometimes manufactured) RISC ones, which themselves had overtaken CISC types;
- Microsoft’s Windows operating system took over from vendor specific Unix variants tied to their dedicated RISC processors;
- Dell EMC overtook HPE as the world’s largest server vendor in 2017; IBM’s revenues declined as a result of its decision to exit the x86 server market in 2014;
- The proportion of servers that are virtualized (i.e. able to run multiple operating sessions simultaneously) increased dramatically once Vmware had produced viable virtualization solutions for x86 servers from 2003 onwards.
The dominance of Windows is shown in the Figure above which shows the annual spending on servers by operating system. The traditional market is somewhat predictable, so those of us who live this market spend much of our time looking at the anomalies, including:
- How IBM‘s approach has managed to maintain its z System mainframe’s position as the world’s most important server despite pre-dating the rise of semi-commodity product types;
- Why Intel’s Itanium processor failed despite its enormous potential benefits;
- How new suppliers with service approaches managed to divert end-user spending from on premise to cloud solutions, in the process adding ‘operational’ to end-users’ ‘capital’ expenditure;
- How to define servers now the number of processors and cores each one has grown tremendously;
- How effectively Chinese vendors can compete effectively with US ones;
- How new trends in cybersecurity, GPU compute, Generative AI, edge computing, etc. will affect server adoption.
Throughout the last 40 years we’ve constantly looked at the advantages of an open technology approach over closed proprietary ones, while organizational computing has often moved from centralized, to client-server, to grid architectures. Over the last few years edge computing has become more important, while the current trend fo GPU-based compute for Generative AI applications is reinvigorating what we useed to think of as technical – or High Performance – computing.
It has become somewhat difficult to track server shipments as a single box can now encompass hundreds of cores; nevertheless the total number is relatively predictable. My figure above shows the unit shipment of servers each year by processor type. x86 types have been growing, as have CISC (which includes IBM’s z Systems); RISC and Itanium have been falling.
The same is true of the installed base (I show the annual totals by processor type in the Figure above). The order-of-magnitude difference in the number of x86 servers installed explains why many component suppliers, software companies and end-users have little experience of other types.
As for suppliers; what was a three-horse race between HP, Dell and IBM (see my Figure above for the annual revenues from servers for the leaders) has become a two-horse one between HPE and Dell. IBM’s total revenues have fallen since its withdrawal from the x86 market, leaving it more time to concentrate on its Power and z System servers. The success of Chinese vendors Lenovo and Huawei has been limited by the US government’s decision to restrict their access to certain technologies, as well as bans on their use in a number of countries. Nvidia has recently become a notable server vendor on the strength of the large machines it sells (largely to Cloud providers) housing multiple GPUs. The business of the Japanese server vendors Fujitsu, Hitachi and NEC have declined significantly to the point that they no longer feature as major worldwide server vendors.
2023 was a pretty dull year for the server market. These are the highlight by supplier:
- Dell was massively concerned with generative AI, which was the main topic of its press releases between May and December; topics include Project Helix with Nvidia and with AMD Instinct accelerators in December. It also announced that is was working with CoreWeave, Imbue, Hugging Face and Meta on Generative AI projects in the year. Other subjects included introducing new PowerEdge machines in January, Security offerings in March, NativeEdge software in Mayn and its expanded its AI offerings in July. Its corporate announcements included its intention to acquire Moogsoft in July, the addition of Steve Mollenkopf to its board of directors in September and an update to its long-term financial framework in October.
- HPE was most concerned to promote its Greenlake cloud services in 2023 – a subject more important than generative AI, at least in its press releases; these had few mentions of cybersecurity and even less of its Proliant servers and edge computing (both included in a single announcement).Its corporate announcements in the year included the acquisition of Pachyderm in January, as well as Axis Security and OpsRamp in March. It extended its private cloud portfolio running in Equinix data centers. It also talked up its Partner Ready Vantage programme and replaced its CFO.
- Supermicro has a different go-to-market strategy than other leading sever vendors. As mainly an OEM builder of kit for other suppliers and cloud services companies it’s press releases are more product focused than most. In the year 12 of its 26 press releases were of product launches based on Intel-, AMD- and Nvidia-based servers alongside a number of storage, high performance computers and telco edge products.
- IBM made a number of product launches in the year, all of which were covered by ITCandor. In descending order of importance its 115 press release in the year covered issues as follows: corporate (46), generative AI and watson/watsonx (36), quantum computing (16), cybersecurity (14) and product launches (3). In August it announced that it had sold on the assets of the Weather Company to Francisco Partners, having separated the company from IBM itself. It also completed the acquisition of Apptio in the same month.
- Huawei focuses on corporate issues in the 230 press releases it published in 2023. It doesn’t mention its PowerEdge servers once and OceanStor storage systems only four times. Of the major issues talked a lot about by the other major suppliers only generative AI (mentioned ten times) appeared. As for product launches its main emphasis is on its 5G telecom offerings
- Lenovo is a major Chinese supplier which has built its server business significantly since taking on IBM’s offloaded x Server portfolio in 2014. Like Huawei, its global position has been hampered by the restrictions and mistrust of the US government. In 2023 it’s press releases concentrated mainly on corporate issues (65 of the total of 110). Product launches were the next most important issues (covered in 22 releases); however servers were insignificant – there were no mentions of its ThinkServer brand, only 3 mentions of ‘infrastructure and 5 of HPC systems. Generative AI was a subject that has grown in its coverage, covered 16 times in the year). It opened an in-house workstation and server factory in Hungary 2022, from which it had already shipped 1 million machines by June 2023.
Of the other vendors Nvidia was perhaps the most important; it has expanded its business from being a GPU component supplier to server vendors to being a server vendor in its own right. It too has been affected by the US government legislation, preventing it from selling its most powerful GPUs in un-favoured countries.
Since 2003 Vmware’s introduction of virtualization into the x86 has stimulated a major change in the make up of the server market; nevertheless even in 2023 the majority of x86-based servers sold were still of non-virtualized (i.e. physical-only) machines. Nowadays x86 -based machines are the virtualization servers of choice, having overtaken the markets from supplier-specific (typically RISC-chip based and running Unix) machines.
In terms of shipments the virtual machines running on virtualized servers have grown exponentially since Vmware’s introduction of its hypervisors (see my Figure above) and, while physical-only servers still out-shipped virtualized ones in 2023 (15.2 million verses 12.0 million) there are now many more applications running as part of a suite of workloads on virtual machines than as single one dedicated to a single server.
In future I expect there to be major growth in GPU-enhanced (mainly Nvidia) products, US-government restrictions to continue to separate the types of machines sold by country and region and for the established list of major vendors to stay much the same, with the exception of the growth of Nvidia. I also hope to see a revival of the Japanese suppliers as global players (possible) and the establishment of at least one European supplier to compete strongly in the world-wide market (highly unlikely).
Let me know what you think, whether you agree with my analysis and, if so, whether I can help you with your server strategy.