future transit prices

someone wrote, in response to my piece this morning...

Can you explain more about why you think transit prices will return to
the $200-$300/mbps. I've been quoted $40/mbps on a 50mbps commit
(95th%) ... which I think is pretty much as low as it's going to get.
I can understand prices going back up near $100/mbps over time, but
>$200 is much more than I'm expecting.

the way i think about this is that somebody has to carry the traffic to
wherever it's got to go. with a "top tier" of huge networks, the pricing
model gets smoother in two ways: (1) the distance insensitivity in sales
has a larger set of costs to average against; and (2) cost per bit-kilom
goes down as pipe size goes up. however, the cost per bit per second of
switching these is relatively constant over time (people, rent, depreciation
or lease of equipment).

a non-top tier provider who wants to get into the game will not be able
to make money at market prices until they fill their network to a
certain crossover point. (and if you buy your pipes too small you can't
get there at all, and if you buy them too large then you can never fill
the whole thing.)

a lot of networks, both top-tier and non-top-tier, have been selling
transit without being able to determine their costs other than at a very
gross level. the thought seems to have been, we have to charge what the
market will bear, and hope we're the last ones standing. but i think
we, as an industry, have pretty much burned all the cash we'll be able
to burn in that way.

when i look at the ingredients:

  worldwide presence (peering points, pops, whatever)
  worldwide L1/L2 costs between pops
  staff (engineering, operations, management, sales, marketing, etc)
  capital (for all those pops)
  rent (of things that aren't pops, like HQ offices)
  marketing, legal, travel, other goo
  and so on

it looks to me like you could run an OC48 backbone at 60% capacity and
make a sustained single digit NPM selling at $250/Mbit average, or you
could do an OC192 backbone at 60% capacity and single digit margins at
maybe $175/Mbit. perhaps an OC768 backbone running at 60% will be able
to make single digit NPM at $100/Mbit, but i'm really reaching on that one.

doing it for less involves either (a) not knowing your costs yet, or (b)
buying market share, or (c) cost containment strategies like using
assets that have been recently through the cleansing ritual of
bankruptcy, or (d) selling ahead of usage like getting 100Mbit/sec
commits from a lot of 20Mbit/sec customers. none of those things lasts

Regardless of which of us is right, I guess I'm still pretty safe if I
lock in todays rates for multiple years.

oh yeah, oh yeah, oh yeah.

How do you compute CGS on a network that is 25% utilized? Is it
expenses/current utilization or expenses/maximum capacity?

I think a lot of the low-ball pricing that is in the market is the result of
networks selling off underutilized capacity at discounted pricing just to
get some additional cash flow. This pricing probably doesn't take into
account the necessary capex that will be required to upgrade the network
when it approaches saturation.

There are a couple of other points to consider when talking about 60%
utilization of the backbone.

I can sell 20 T3's into my OC48 pop and still not ever see 900mb/s average
usage. The statmux effect varies depending on the nature of your traffic.
Spikes/peaks disappear in the core, and the larger your core, the better
able your average will multiply for you.

Second, 60% utilization -- is that assuming symmetric usage or not? Often
the same T3 of capacity can be sold to complimentary customers/usage
(access/web), etc.

And while staff/facilities/legal/etc may stay relatively high irrespective
of usage, I would be very surprised to see them scale at anywhere near the
same velocity as backbone capacity. A n OC48 backbone and an OC192 backbone
don't require different amounts of operations staff, maybe a variable (<<4x)
amount of front-line customer-service-staff, and like wise the equipment
isn't 4x more expensive than the OC48, etc. So I would say overhead per bit
per second is a high constant with a small variable (capacity dependent)
piece and when dividing into that capacity the numbers drop drastically. In
an access-network point-of-view. Obviously running datacenters changes the
model slightly.

Deepak Jain