I want to start off by saying that I'm not rightly sure why you have such statistically-significant discrepancies in your curves. I performed a quick, little, ad-hoc analysis using your provided data almost certain that it would support Irlanur's assertion that your results could be explained by instrumental drift, but I don't believe that to be the case after looking over my ad-hoc analysis.
So, I thought I'd just try to add to the conversation in what way I can and see if it helps to elucidate things.
I took the liberty of assuming that you had an unknown that fell right smack dab in the middle of your calibration curve for each replicate measurement (signals of 8.0241E6 and 7.0058E6, respectively). The results were as follows:
Replicate 1: Unknown = 2.09 +/- 0.04 ppm
Replicate 2: Unknown = 1.82 +/- 0.03 ppm
Admittedly, I didn't double-check my work, but I do believe it is correct. Also, I corrected your y-values for the blank. I also did NOT include the blank as a data point (0,0) on your curves. Some analysts do and some don't. Personally, I choose not to. To me, the above results show that a theoretical unknown would yield VERY different values if analyzed by each curve, statistically-speaking. Assuming that your Excel sheet presents things in a chronological fashion, how far apart were these replicate measurements taken? Days, weeks? Minutes? Hours?
This is the part where I start asking you leading questions, MrHappy, and can't, unfortunately, share with you my interpretations. After all, our goal on these forums is to teach moreso than it is to just divulge the answers. But, I will say that I don't think there are absolute answers to all of these questions, only that some answers are more likely than others.
So, these results (again, assuming your Excel sheet is chronological in moving from left to right) suggest that your sensitivity (your slope) lessened with time. Let's assume any intercept problems are insignificant. Does this loss of sensitivity suggest that the concentration of your standards reduced with time (hence yielding lesser instrumental signal) or does it suggest that your instrument was no longer measuring the analytes in your standards as efficiently? Or is a mixture of both problems more likely? How stable were your solutions? What's the "Achilles heel" for instruments like Flame AAs and ICPs? For instance, for gas chromatography work, sample introduction (reproducibility of injection and inlet split) is the Achilles heel. Can you draw any comparisons for an ICP? Do you know the parameters for your ICP's torch, and was it the same across both analyses?
Also, this one detail confuses me deeply: did you really re-measure the DI water blank each time? Or only the first time? I ask because most of your standards had appreciable changes in signal from the first replicate to the second. But your blank, which should arguably have the worst precision (as it contains the least amount of analyte), has no change in signal. SEVEN sig figs? I mean that's better than an analytical balance. That's hard to believe. That would really only make sense if you didn't really re-measure the blank or if your standards were all outside of the working range of the instrument. I will say that some of your standards are a bit concentrated for ICP work. I like to keep things sub-ppm, but your linearity is fine, even wonderful, in fact, so it seems unlikely that that's the problem. Plus, Mg tends to have a much higher concentration working range than other metals, so your range seems fine.
If you REALLY DID re-measure the blank each time and had no change in signal, what does that tell you about the stability of your solutions with time? After all, that blank implies that you should be getting (roughly) 7 sig figs of reproducibility accross each replicate, and you're no where near that. That should REALLY tell you something about your solutions . . .
This should help get you started. Again, I can't rightly say that I know exactly what's going on. I do have some educated thoughts on the matter, but, also, a lot of my conclusions would rest on information only you have (sample prep procedure, sample matrix, primary standard used, instrumental parameters, etc., just to name a few).
I hope this helps at least some.