The point he's making is salient enough if you've worked with sensor networks, specifically if you're trying to track physical objects and predict what it will do next. You run up against a sort of computational corollary to Nyquist-Shannon (http://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_samplin...).
In real life, we just sample data as fast as the computation will allow and accept that a large fraction of the world's information will not be part of our computation and that the computation is imperfect. For example: I'm sensing some time-varying function, perhaps an unpredictable aircraft, and it takes 10 sensor sampling cycles to compute my approximation of speed/velocity/turn rate/etc; I'm not going to compute every sample and watch my picture of the world lag behind by 10n (for n samples) -- I'm only going to run the computation on every 10th sample. It doesn't make Turing wrong, it just makes [sequential] sampling and computation of multiple unpredictable inputs slower than the rate at which the unpredictable inputs appear.
But if you had a computer that had enough I/O bandwidth to be able to read all the sensors in one clock cycle, and enough CPU power to be able to run the computation on those readings in one clock cycle, you'd have a computer that was capable of performing the computation optimally. When you add more sensors to the configuration or you have a new computation that is too complex for this computer to solve in one clock cycle, then the computer can no longer be capable of solving that computational problem as stated. Obviously, you can change the problem to make it acceptable to sample instead as you mentioned.
If I'm understanding the principle in question and Mr. Akl's result, he states that the number of computations required to solve a problem is unbounded. If the definition of a universal computer is that the memory and run time is infinite, but the number of computations performed per time slice must be bounded, then there will always be a problem that a realized universal computer cannot solve.
A comment up this thread states that it is silly to have a restriction that all the data must be processable in one time slice because you can just buffer it. I'm not aware of any buffering solution that does not require additional computation to read the value and store it in the buffer.
In his short list of misconceptions and responses, I think that #3 is closest to this argument. I don't know if it would be an absolute requirement that the initial collection of the data must remain an intrinsic part of the universal computer as opposed to the simpler definition that the observed data required for solving the problem must simply be supplied to the computer.
I think I'd lean toward the latter definition because it doesn't seem right to me to expect the universal computer to be responsible for both observing and recording every event in the universe as opposed to just expecting it to be able to perform a calculation on any given set of observed events.
In real life, we just sample data as fast as the computation will allow and accept that a large fraction of the world's information will not be part of our computation and that the computation is imperfect. For example: I'm sensing some time-varying function, perhaps an unpredictable aircraft, and it takes 10 sensor sampling cycles to compute my approximation of speed/velocity/turn rate/etc; I'm not going to compute every sample and watch my picture of the world lag behind by 10n (for n samples) -- I'm only going to run the computation on every 10th sample. It doesn't make Turing wrong, it just makes [sequential] sampling and computation of multiple unpredictable inputs slower than the rate at which the unpredictable inputs appear.