There are six companies remaining in the mainframe industry - Hitachi, NEC, Fujitsu, Atos (formerly Bull), Unisys, and IBM. Three of those - Fujitsu, NEC, and IBM - still develop their own CPUs, though Fujitsu is likely to do so for only one more generation. If I had to speculate about where each of them is used, this is more or less what I'd end up with.
IBM is global and has, as best I can tell, between 3000 and 7000 Z customers. Many of these are on very large systems; IBM scales to far higher performance levels and core counts than other vendors.
Unisys is global, with higher concentrations in Latin America (especially for MCP) and East Asia (especially for OS 2200.) MCP is more common in banking and telecom, OS 2200 in airline and government. Both exited custom CPUs in the early 2010s but have fast emulators. I'd guess there's 800-1200 MCP sites and a much smaller number of OS 2200 sites. MCP uses an impressive high-level-language-oriented descriptor ISA, while OS 2200 uses a 36-bit ones' complement ISA with some unusual characteristics.
Fujitsu is mostly global, with the notable exception of North America, and has several distinct mainframe families. The ex-Siemens BS2000 lineup and the semi-IBM-compatible GS21 family use different software but on the same 390-based custom CPUs. The former ICL family, the 29-series descriptor systems, lives on in Britain in finance and government but is slowly declining and has been primarily run emulated for over 20 years. BS2000 is concentrated in Germany, the GS21 userbase in Japan. I'd guess there's a total of 1000-1500 Fujitsu mainframe customers, with a majority being in Japan, but I'm not super confident in that.
Hitachi, though they once had a thriving global business including North America, markets their systems exclusively in Japan. Until approximately 2020, they built custom CPUs, but in the latest generation - the AP10000 - they rebadge IBM Z running their own MVS-derived VOS3 operating system. I'd guess they have 200-300 sites, almost all in Japan.
Atos, which recently finished devouring the French computer company Groupe Bull, owns Bull's GCOS 7 and GCOS 8 operating systems. Both are unique ISAs, incompatible with each other and running completely different software; GCOS 7 is 32-bit, EBCDIC, vaguely MVS-like, and even includes a POSIX subsystem (which it uses for TCP/IP support.) GCOS 8 is older, ASCII, and 36-bit word-oriented. GCOS 7 is emulated on x86, GCOS 8 (as far as I'm aware) is still emulated on Itanium; I would guess the combined total of customer sites for G7 and G8 is less than 100, mostly in Western Europe but with a handful in North America and probably Africa.
NEC's ACOS-4 operating system is a distant cousin of Bull's GCOS 7, and NEC still designs its own processors to run it. These are fairly big systems by the standards of non-IBM mainframes - up to 48 custom-designed cores and 256GB RAM - and while the customer base is almost exclusively Japanese, they've historically sold a few ACOS systems elsewhere in Asia and even North America. I would guess that at this point there are 200-400 sites remaining, almost all in Japan.
[1] Bull is the exception; GCOS 7 and GCOS 8 are at least as incompatible as MCP and OS 2200 are. One could probably make a case for Fujitsu, but generally the markets for GS, BS, and VME have little overlap (except in Britain.)
Unlike in the West, where custom mainframe hardware (other than IBM) had largely died out by the early 2000s in favor of emulation, the Japanese mainframe vendors have generally been reluctant to abandon custom CPU development for their mainframes. NEC pushed emulation for several years for their ACOS-4 operating system, but ultimately, performance wasn't good enough and they returned to the embrace of custom hardware - including their shiny new NOAH-7 mainframe CPU, released last year and probably built on TSMC 28nm lithography.
Things are changing, though. NEC is the last one standing, and even they may not last much longer.
Fujitsu announced early last year that they would end mainframe sales by 2030. This surprised me. Fujitsu's GS21 and BusinessServer families - running on the same hardware, but with different operating systems for different regions - have been generally successful, and Fujitsu is the only mainframe hardware vendor other than IBM that retains a true global presence - with customers in Australia, Western Europe, and Japan itself. Even sales numbers have been very decent historically; the most recent generation, 2018's GS21 3600, was projected to ship 800 units in its first three years of availability. Those are respectable numbers. Why would Fujitsu break a working formula?
The actual numbers, though, are telling. In 2021, Japanese domestic mainframe shipments looked like this, according to IDC:
Vendor Units Shipped --------- ------------- Fujitsu 84 NEC 48 IBM 29 Hitachi* 19
These numbers are absolutely dismal, and raise serious questions about whether mainframes as a whole are dying. There are a few angles to interpret these numbers.
It could be all three.
Where does this leave NEC? Their user base is smaller than Fujitsu's, and they have no global base to draw on. Until recently, NEC had a public roadmap showing new systems in 2022, 2027, and 2032. The 2022 machine shipped as the i-PX AKATSUKI, a major generational upgrade, but the public roadmap disappeared after the AKATSUKI released, and it's unclear if the other two systems are still committed. Unlike Fujitsu, though, no end of life has been announced.
NEC's last public ACOS roadmap
