Inherent Limitations with Respect to HFT Hardware

Question

I am attempting to compile some aspects of high frequency trading that pose issues to all players in the field i.e., things that cause problems regardless of how much money the firm has to spend on hardware, co-location, getting the most pristine software developers, etc.

Some things that have come to mind initially:

1. Size: Generally it seems that most exchanges charge for co-location based on square-footage / some other similar metric for accounting for the space their system takes. But obviously this does not scale... even if a company has all the money to spend on co-location a large amount of space is not plausible nor even would it be advantageous necessarily.
2. Fees: This affects all HFTs, regardless of how much money they have.
3. CPUs: For the most part, they are all on the same CPUs (if they are in fact using that). However the difference lies in how they are able to manage threads and respective tasks being done on the CPU for HFT. But, something else I have been thinking is an inherent limitation (and perhaps my confusion is unfounded--and someone can point out why), is that even if you have numerous tasks running in several different threads, there still must exist a "critical processing thread", where stuff has to be done sequentially, no? The main reason I think this is the case is that given the result of whatever analysis is being done on that thread, that has to be passed to the next logic branch, right? Perhaps that sounds confusing and was phrased poorly, but problems of this nature is what I am getting at with this question.

This is what I have so far, I would appreciate any insight about perhaps my statements above being incorrect / misguided, and some more aspects of HFT that pose issues to all players in the field.

Answer 1

Size. You're mostly correct. It's market standard among data centers, not just those run by exchanges or hosting exchanges, to offer preferential rates only when you get to the scale that you have cages full of cabinets. This eliminates almost all HFT firms from having preferential rates because a large number of servers is unnecessary to trade in large scale on any given market.

Fees. True. Most venues charge fees that are uniformly applied to any participant. Whether you are a retail trader that does \$10B in monthly volume or a HFT firm that does \$10B in monthly volume, you should expect the same rates on a venue. The water gets murky here however because some firms have grandfathered rates, exchange memberships or enrollment in market making incentive programs which may give them preferential rates.

CPUs (or FPGAs). Yes. To the extent that a firm uses CPUs or FPGAs, they are mostly limited to commercial vendors because the cost of fabricating these competitively is outside the budget of the largest HFT firms.

Other areas which may be considered level-playing fields include:

Colocation. The top tier of exchanges generally have a third party auditor ensure that all the fiber lengths to the matching engine are identical.

Long-haul connections on common paths. Go West, Toronto-NJ, DC, Aurora-NJ, Hibernia etc. are all examples of the fastest paths possible today that are also owned by vendors rather than trading firms. This makes it a level playing field.