Thanks for asking - this project excites our team on a regular basis!
NVIDIA's own SATURNV is the prototype for us: https://www.nvidia.com/en-us/data-center/dgx-saturnv/
NVIDIA uses it for in-house product development, and it is built entirely of NVIDIA DGX-1 modules. It was also world no. 1 in 2016: https://www.youtube.com/watch?v=q2yxcnXsCOA
It is also surprisingly physically small. Let me find you a video: https://www.youtube.com/watch?v=mszXq8ZhKHk
Skip to 2:34 where Jensen Huang says that 1 DGX-1 replaced 400 servers. And our project will combine 72 DGX-1s...replacing 400x72=? servers. Let's do the math  
Also this: https://www.youtube.com/watch?v=nVk416fVd3s Skip to ~14:15 on discussion on energy efficiency and space-saving.

As for mining, NVIDIA is a leading maker of crypto mining hardware: https://www.cryptocoinsnews.com/nvidias-shares-ride-on-ethereums-gains/
But AI supercomputers use chips/technologies different from hardware made specifically for mining.
However, we will be able to cost-effectively mine using a supercomputer as long as mining is not its primary occupation - in its idle time. We plan to do so.
Focusing on hardware and idle-time mining will enable us to be accountable to our (HPT) token holders in 2 key ways:

1. Our operating performance will be straightforward to track compared to other ICOs. The price of DGX-1 is public, thus token holders will be able to know how much money we reasonably need and what  for. Our ICO investors will find it easy to check whether the supercomputer has been bought, installed, and running. So half the operating risk disappears. The part of risk that remains is only whether we are able to build innovative services on top of amazing hardware. We hope that NVIDIA Inception incubating us can help answer this question.
2. We will be able to start paying a share of mining income to HPT token holders (hopefully) starting from month 1 that the supercomputer has been switched on.

Note: cryptocurrency mining is not what we specialize in. Its profitability depends on fluctuating crypto (such as ETH) prices and mining economics. Please check our Medium for a longer discussion on potential profitability of crypto mining income: https://medium.com/@housepanda/hpt-token-share-in-the-worlds-no-17-supercomputer-5752d53d3325

We are convinced that it will be far more profitable to provide unique, world-first prediction services built on the supercomputer's AI capabilities than to place primary focus on crypto mining.
For example, let's look at what Bloomberg Terminal is doing. Bloomberg Terminal is the classic example of a subscription service: US$24,000/year, hundreds of thousands of subscribers: https://en.wikipedia.org/wiki/Bloomberg_Terminal
House Panda is looking to build a subscription service starting at 99HPT/month providing future price calculations of various assets: cryptocurrencies, real estate and stocks. In our view, a realistic monetization scenario after 24-36 months of operation: 100,000 paying customers x 1,200HPT/year = ?
1HPT=$1 at time of ICO.

That is not to say that we only care about profit - ability to operate profitably enables everything else. In effect, we are looking to open access to a supercomputer to everyone who can pay a small amount - for the first time ever. Others might be able to build services on top of this hardware - not only us.

We are also planning a range of additional services and token uses, of which later.