HFT infrastructure has become affordable relative to the past, definitely not the numbers being quoted here. For example if you want microwave market data at all 3 US equity datacenters all you have to do is pay up to nasdaq, no need to build out your own infrastructure anymore.
Nearly every bank now has FPGA order entry with under 1us risk checks and most well under that. For some can use a shared instance at no additional cost or dedicated instance for 50k per unit or a monthly cost of like 2k . The jitter at the exchange gateway is now an order of magnitude greater than the risk check latency. I guess theoretically that could eventually be made available to retail although not currently.
So a rough estimate to get something up and running.. let's say you want 20 boxes at 20k each in each colo, rack space would be ~30k per month, some fast switches 20k each, cross connects ~60k/month, 20 dedicated ports at each colo for order entry to round robin across ~30k/month, wireless market data with itch FPGA ~150k/month (rough estimate)
So 3.2 M per year in recurring costs, 1.2 M upfront for servers, and say 120k for switches. and maybe outsource FPGA market data feed handlers (under 1 us latency to build books) so another 500k per year depending on how many you need.
With that one would effectively be as fast as anyone on the street from a pure infrastructure perspective with 20 boxes at each colo. Granted now you need to compete with good code or put some trading logic on FPGA and actually have some alpha (highly unlikely...)
So still rather out of reach for retail but well within reach for institutional money. Granted broker dealers (HRT/IMC/Tower) still have an advantage since they run their own risk checks and can put their strategies on the same box whereas none of the banks will let me do that yet unfortunately. Also they can send day ISO orders whereas none of the banks will let me do that directly either without going through their software SOR. So definitely a subset of trades that can't be made without being a broker dealer but still enough left over that it isn't necessary to be one. And some HFT broker dealers do actually route flow through the banks in order to get the rebate tiers and take the latency hit in order to do so.
But the other aspect that I think makes it unreachable is that it's pretty much necessary to have order book data these days from most if not every exchange. I spend 40k a year to store it all of it on aws and approach 10k a month in compute on aws. I do a pcap from each colo so don't have to pay for historical data going forward but if you don't have that looking at a 400k-1M to get historical book data depending on how much history one needs.
As far as "AI" or let's say non-linear models.. that is currently accessible to retail. I have a strategy that runs eod on the close using out of the box models from sci-kit learn and completely written in python that I could execute on IB for example. Granted the data I use isn't available at a reasonable cost to retail.
Nearly every bank now has FPGA order entry with under 1us risk checks and most well under that. For some can use a shared instance at no additional cost or dedicated instance for 50k per unit or a monthly cost of like 2k . The jitter at the exchange gateway is now an order of magnitude greater than the risk check latency. I guess theoretically that could eventually be made available to retail although not currently.
So a rough estimate to get something up and running.. let's say you want 20 boxes at 20k each in each colo, rack space would be ~30k per month, some fast switches 20k each, cross connects ~60k/month, 20 dedicated ports at each colo for order entry to round robin across ~30k/month, wireless market data with itch FPGA ~150k/month (rough estimate)
So 3.2 M per year in recurring costs, 1.2 M upfront for servers, and say 120k for switches. and maybe outsource FPGA market data feed handlers (under 1 us latency to build books) so another 500k per year depending on how many you need.
With that one would effectively be as fast as anyone on the street from a pure infrastructure perspective with 20 boxes at each colo. Granted now you need to compete with good code or put some trading logic on FPGA and actually have some alpha (highly unlikely...)
So still rather out of reach for retail but well within reach for institutional money. Granted broker dealers (HRT/IMC/Tower) still have an advantage since they run their own risk checks and can put their strategies on the same box whereas none of the banks will let me do that yet unfortunately. Also they can send day ISO orders whereas none of the banks will let me do that directly either without going through their software SOR. So definitely a subset of trades that can't be made without being a broker dealer but still enough left over that it isn't necessary to be one. And some HFT broker dealers do actually route flow through the banks in order to get the rebate tiers and take the latency hit in order to do so.
But the other aspect that I think makes it unreachable is that it's pretty much necessary to have order book data these days from most if not every exchange. I spend 40k a year to store it all of it on aws and approach 10k a month in compute on aws. I do a pcap from each colo so don't have to pay for historical data going forward but if you don't have that looking at a 400k-1M to get historical book data depending on how much history one needs.
As far as "AI" or let's say non-linear models.. that is currently accessible to retail. I have a strategy that runs eod on the close using out of the box models from sci-kit learn and completely written in python that I could execute on IB for example. Granted the data I use isn't available at a reasonable cost to retail.
