If I were to start a new isolated firewalled bitcoin network, working from the same base block as the existing network, with only a single slow host working on it, the chain would still get longer by 1 block every 10 minutes, since the difficulty would adjust after 2 weeks to ensure this.
Also difficulty changes only every 2016 blocks - not every 2 weeks. It is expected that hash rates grow/decrease within limits (= not moving from a few Terahashes/s to a few Megahashes/s in your example). also the new difficulty takes into consideration the average time it took to calculate the past 2016 blocks. For the difficulty to go down by 50%, the last 2016 blocks would need to be solved within 4 weeks, instead of 2.
The Bitcoin blockchain is essentially a global clock, where it is ensured more or less that it's hard enough for members to trigger a "tick" at the same time - and if they do, the network will solve it on it's own.
To be able to build a longer Blockchain than the official one, you need more hashing power than the rest of the network. This is also known as the "51% attack" and which is why people are currently worried that deepbit might become (again) big enough to do this in theory. In practice, deepbit will most likely NOT pull off a stunt like that, but it still poses some dangers and insecurities.