The good news is that legacy apps are not precluded from cloud computing. The easy part is dealing with the legacy apps of tomorrow. It's becoming a defacto standard that all new development leverage a cloud architecture. Cloud architecture is a refinement of internet architecture which is a refinement of n-tier, client/server, and mainframe architectures. It's all about evolving as the technologies mature.
The hard part is what to do about the legacy apps of today. There are at least four options available: leave alone, enable with services, migrate key algorithms, and finally migrate. The leave alone option does not necessarily mean nothing can be done. For some mainframe apps there are other options. For example in my past we moved a large number of CICS applications from a mainframe to an HP9000 server with the benefit of higher throughput. In fact the Oracle Exalogic platform framework specifically identifies its ability to be used for CICS applications in the literature. And some applications can be moved with few changes such as a Smalltalk application ported from a mainframe to blades for a large brokerage to speed up execution. Never say never.
Option two, enable with services, is a pretty popular approach with those who will not wait. A large insurance company is working hard to enable its mainframe applications representing hundreds of millions of dollars in value to talk on the enterprise service bus (ESB) via services (I know because I designed the architecture). It is possible to keep that large investment largely intact and just change how it interacts with other applications. Through the ESB, a SOA concept, the mainframe apps have to respond in seconds in order to deliver web page results to prospective customers in a reasonable time, defined as no greater than eight seconds. Not only does this approach work, it can work well extending the life of applications to interact with the cloud at the software level while not participating at the hardware level (although I'll argue that with virtualization on the mainframe, they're just as much cloud as Zynga's games).
Migrating algorithms that are within applications rather than entire applications can deliver a huge benefit. Think of a pricing, scoring or inventory management application. The algorithm that helps management figure out what to do (the price, the score, or the shortages) can be move to, or replicated in, a cloud model enabling cloud applications to interact with the algorithm. To a cloud application that may be all that is needed, especially if that algorithm is exposed as a service. In good SOA design services should give answers, not expose raw data, and often all that's needed is the answer.
Finally migrating applications wholesale to cloud makes sense for some mainframe apps. However there is no requirement it happen today. The big issue has always been the business case; justifying the investment. CIO's have pushed application managers to put so much lipstick on these apps that they now resemble pigs. Bolt ons for integration, data management, identity management, internet access, storage tiering, systems management, and on and on combined with years of change requests have created systems so complex they carry a legacy tax. I have worked with several companies where just touching a mainframe application carries a minimum $1,000,000 fee. That's a pretty high buy-in, yet CIO's wonder why they're spending 70% of their budgets on maintenance. That first $1M is non-value added! CIO's need to expedite the sun-setting of these applications by building the case for change based on both the cost to change and the cost of not changing and accruing these legacy taxes. Sometimes you have to do what's right instead of what's easy and this is it.
The challenge to building the business case is the expectation that because computer code isn't physical it's easy to change. I balk at that argument with facts; if that were so why do software changes cost so much? I had a CIO challenge me to get the CFO to understand so I asked the CFO how much money had been spent to update/modify his highest producing factory while the CIO was to determine how much money has been spent to update/modify the most used application. The difference was staggering; the application costs were well over an order of magnitude higher although the plant was twice as old. So I asked the CFO if he were to invest that amount of money into operations, how much of the plant would have to change. He said he'd never do it, it wouldn't make sense. Instead he would build and outfit a new plant. Ah ha! So why do we use a different approach with software?
I'll admit that was a somewhat special case, but in the end it's all about foundation. Mainframe apps were built with a different architecture, intent, vision, everything. It's like investing in swords and armor and then being surprised when your enemy shows up in a kevlar body suit with his M4 and Seal Team Six tatoo. Ask the French how well their Maginot line worked.