The last time I did the "back of the envelope" calculations (about two weeks ago), it would cost about $12,000,000 to reach 50%. That's the cost to design, contract, receive, and hookup fully-custom ASICs to build a hashmonster. I believe it was 200 blades, each with 24 ASICs. Each ASIC would have 32 2xSHA256 cores running at 350Mhz, each with the hashes fully-implemented as custom gates, pipelined. (So while we're doing round 3 of the 3rd nonce, we're doing round 2 of the 2nd, and so on.) Each core, once its pipeline was full, could churn out one doublehash every 8 clock cycles. The total hashing power: 6.7THash/s. (That's from memory.)
However, this would only give you half if it was already half. Which it might be, for all we know.
However, this would only give you half if it was already half. Which it might be, for all we know.
But for the situation to be realistic we have to assume that the honest miners would already be using the same technology. Otherwise your scenario is only valid in the concrete moment when the technology is evolving.
And, by the way, we have no way to know whether or not dishonest miners already have this technology and are using it. They could be feeding into any number of pools and running as independents as we speak -- making us think the network is much more secure than it really is. (I don't think this is actually happening, but it can't be ruled out.)
You couldn't even start on something like this with less than $2,000,000.