The parent mentioned the cost of "fabricating that technology", not of processors themselves; this is vague, but AFAIK, fabs are getting more expensive quickly.
"The increase in cost and complexity is encapsulated in Rock's Law, which dictates that fab construction costs double every four years. That law is itself tightly connected to the better-known Moore's Law, which dictates that the number of transistors on a chip doubles every two years." (http://news.cnet.com/A-fab-construction-job/2100-1001_3-9810...)
"By 2007, the price of building a fab is expected to reach $6 billion."
$6 billion * 2 ^ 10 = 6 trillion in 20 years.
B) Not really it's a well known problem with AI. Over train a learning system and you get worse behavior.
C) Nope. Information theory puts strict limits on how much you can zoom in with a web cam. If an AI want's to earn about sub atomic physics it needs to look at the output of a particle accelerator. Can it make a better theory with existing data? Well see B.
D) http://en.wikipedia.org/wiki/Rock%27s_law The semiconductor industry has always been extremely capital-intensive, with very low unit manufacturing costs. Thus, the ultimate limits to growth of the industry will constrain the maximum amount of capital that can be invested in new products; at some point, Rock's Law will collide with Moore's Law.
B) This is nonsensical ...
C) You can deduce a ludicrous amount of information from relatively little sensor input. Humans don't extract a fraction of a percent of the maximum (in the information theoretic sense) information from our environment. See http://lesswrong.com/lw/qk/that_alien_message/
D) This is trivially, factually incorrect. A 3GHz P4 is cheaper today (in absolute or relative terms) than it was when it was brand new. A new Core 2 Duo will give you better performance/dollar than a P4 when it was new (which was better than a P3, which was better than a PII, etc).