It's not a problem with infiniband but a problem of sw mutating needs vs actual hw assets: matching is hard. Hw cost money so anyone try to be moderately tight, looking at life expectancy and supposed sw changes in that timeframe, but hw and sw have different lifecycle. At a certain point in time sw infra change while hw is still the same, at another point in time available/current iron change but software on it remain the same.
In classic times that's just a matter of balancing both, carefully designing both, but in modern time developers just "produce code", operation is pushed aside, communications are constrained often by the policies and the result it's a mess => companies choose "the cloud" witch means outsource the issue to third parties. Those third parties count on the fact that even if anything change in the end a computer is always a computer so they can live on a changing world because if single companies have irregularly changing needs on average the rate of change is similar. Also on scale redundancy needs lower the costs and that's why we switch from classic big iron to desktop crap packed in racks. Now we start seeing more and more the limit of such model, and no solution is here.
Big iron is long lost, now way to expensive not much in absolute terms but in demand terms: try to sell a Power system, the customer look at specs and say "hu?! a couple of PCs from nearby grocery store are equally powerful!". Such iron modularity coming from classic mainframes (infiniband in the end is crylink changed a bit by SGI and others) is considered too expensive, on the other side of the spectrum small-board ARM super-blades tentative still results in fails.
Until we accept that hw and sw can't really be two separate world and both must be open to being able to interoperate on scale we will not find any really scalable solution.
Sw side it's not different: these days is the containers mania days because after many have suffered burn with full-stack virtualization on x86 they look for something with less overhead and equal design: hw-sw decoupled, possibility of run an infra you do not know buy pre-made images/containers from third parties who maintain them without needing to know your infra. The result again is an enormous amount of layers and giant attack surface no one really know how extended it is. Since coms are needed we punch holes here and there, and on the other side "safety features" arrive to compensate, making more complicatedness layers on complicatedness instead of admitting the basic issue: in sw terms our OSes are badly designed for current needs, they are a relic form a long lost past where they aren't even good but just cheaper alternatives to something better.
If we ever admit those errors probably we will be able to design modular iron and modern OSes marginalizing those issue, but so far no one want such a nightmarish adventure...