The difficulty is that doing this requires building a full vertical around the idea, and it's so hard to sell new big scary ideas to the business. So, the entire space is now fractured into 10000 vendors trying to sell you various nuanced aspects of the big picture. Combine with the cloud hysteria, and there's virtually zero chance you get to even experiment with this idea unless you take it up as a hobby and/or build a company around it.
To do this right for commodity x86 - that is, make it as or more reliable than a Z15 mainframe with equivalent performance - you have to do it on-prem and in a datacenter designed around the topology of the cluster. You cannot run something like this inside AWS/Azure/et.al. and expect it to perform well. You need direct fiber links between nodes without switches in the middle. Minimum of 5 nodes for N+2 redundancy. N+1 paths between every node. 2N+1 power/cooling. There is zero compromise you can take on the aspects of such a solution if you want to meet or beat IBM's insane numbers. On top of this, you would have to develop a software platform that is directly aware of your physical topology in order to properly leverage it. This implies writing new business software against your new platform. This is risky business. No reasonable shop would sign up for that ride unless you had 2-3 success stories and living, breathing clients to testify to the same effect.
For some carefully chosen, possibly hypothetical value of cluster,x86,server, reliable and performance.
To go from a hypothetical to a concrete would require an investment of time and money of sufficient magnitude to make the purchase/lease of an extant mainframe a compelling choice.
IOW, if my grandmother had wheels, she'd be a horse trailer.
As another commenter mentioned, Oxide Computer seems to be going for something not unlike this. People might compare their system-design target to the minicomputer (or the "supermini") as opposed to the mainframe, but that's just semantics at this point.
Low-latency networking is obviously relevant to anyone who might want to achieve similar levels of throughput+reliability by scaling out towards datacenter-scale computing. But modern mainframes are typically single systems - perhaps with a secondary system in an off-site data center for automated failover, but nothing much beyond that.