James Gardner, director for retail banking in the global financial services group at Getronics, looks at the economics of BNP Paribas' recent on-demand computing deal with IBM and asks whether the bank might have got a better price from Amazon.
Recently , I saw this
little gem of news. Apparently BNP Paribas is using an on-demand computing service from IBM to run its risk application. According to the story, they rent the capacity they need, and can increase or decrease it depending on what they're doing.
Now, these risk application are fairly compute hungry. I actually did some work with some of these once-upon-a-time, and frankly, you start running into some upper limits of the hardware quite quickly. The problem is that you have to perform millions and millions of little calculations in what can be quite a small time window. And the smaller the time window you can run in, the better the information you can give to your people doing transactions. On the systems I worked on, we were running up against hardware limits (mostly as a result of an expanding portfolio) just doing an inter-day run. But the business wanted intraday, and it just didn't seem that there was any real way to deliver that within the constraints of the cycles we had available.
That's why the BNP Paribas deal is rather interesting to me. What is the result when you can buy any number of cycles you need just when you need them?
When you start adding hardware, it is rather accepted that you get decreasing returns (in terms of work completed) for each additional cycle you shove at the problem. So when you add hardware, you get more done, but not as much as gets done as the last time you added hardware. The economics of this means that at a certain point, the value of the additional work done by a cycle is less than the acquisition cost of that cycle. So you stop acquiring.
So the benefit of the IBM deal to BNP Paribas is that it lets them go way, way down the curve for cycle acquisition, probably to the point where they have some real competitive advantage. When you are sharing the cost of cycles, you definitely get more bang for the buck. Its no different when you buy electricity: perhaps the power station is capable of running every air conditioner in the land a few degrees above freezing throughout the during summer, but I don't have to pay for that capacity.
BNP Paribas is no small time bank. Neither are any of the large financial services players who are likely to make the news in deals of this magnitude with IBM or, any other vendor.
But have you been watching Amazon and their compute on demand service
, EC2? Its amazing. With a web service request, you can order, load, and execute a Linux image that does anything you know how to code. Need more capacity?Send off a web service call.
You pay for the instances based on the amount of cycles, bandwidth and storage you use. It means that anyone can offload capacity to Amazon at any time, without a long term, IBM style contract.
So if you're a bank, running the risk application- essentially lots of machines doing Monte Carlo - in the cloud means you can go down the economic curve just as much as BNP Paribas and IBM. In fact, it would seem to me that BNP Paribas could have gone to Amazon and gotten as much or more for their money.
Let's work this out using information gleaned from various Google searches.
We know the BNP deal is for 2500 machines (with an option for up to 2500 more), and that the current capacity is being shared by one other customer at present (Aura, a telco, I believe). The deal is a "multimillion dollar commitment", over three years, but there is also talk of it being some kind of JV.
I've not been able to find the dollar value of the BNP Paribas contract, but at the IBM site, I found out that the lowest blade server they make - the HS20, can be bought for $1609. That's a server similar in configuration to the virtual one that Amazon provides. Let's be conservative and say that the bulk discount shaves off 30% of the price of this server when you buy 2500 at once. Upfront, that's $2.8 million dollars in infrastructure. Add to that a data centre, management staff (2500 expensive machines? They are hardly just going to leave them in the racks when they die, unlike Google), and it would be reasonable to add a few more million. Let's say 2 million per year just to keep the calculation going (there is talk in the articles I found of 450 staff).
They are sharing the capacity with one other customer at present, so we can reasonably say only 50% of the cost goes to BNP. With these numbers, my back of the envelope calculation is $4.4 million over three years.
Now let's calculate the same thing for Amazon EC2. Firstly, it costs ten cents per compute hour. For 2500 machines, that's $250 an hour for all those cycles. Storage is extra, and so is bandwidth that is used outside Amazon. For this application, these are likely to be not that significant, so we can ignore them.
Let's assume that BNP does inter-day, processing only, and that the run takes all night. Therefore they use that capacity for 12 hours. In other words, the facility costs them $3000 a day. There are 52 weeks in a year, but markets are open for only 5 days of each week. Ignoring holidays, we have to do 260 runs a year, for an annual cost of $780,000. Over three years, that would amount to the grand sum of $2.34 million dollars. For 2500 machines available on demand.
Now I know that some of this is drawing a long bow in the absence of real data, but it seems to me that BNP could have bought the cycles for less cash by going to Amazon than IBM. And it would have been easier to get capacity at Amazon- and for a shorter period of time. IBM don't have a web service call to bring up new machines images (they sell hardware!), and I bet they can't bill in increments of ten cents like Amazon.
If I am close at all, then this suggests to me that compute in the cloud is something that bankers should be thinking about, and not just for risk calculations. Any time there is a huge server farm that sits idle most of the time - Internet banking is another example - there are powerful economics available to offload those servers.
I guess the next question is whether any bank would dare to offload servers to Amazon the way they would to IBM? I'd not be surprised to see innovative banks doing it, frankly, once Amazon have proved themselves a bit more.