Uh, is it just me or is this chart having a hard time with comma versus period for "thousands separator"? Broadcast.com is listed as having $10,961 cost per user, but comes up as $10.961 on the chart. [Sale price of 5.7B for 520K active users]
Another interesting metric would be cost per user minute (would probably need to be normalized over a time period like day, week, month, or year, or quarter).
450000000 is a lot of users. Metcalfe's law says that the value of a communication network is proportional to the square of the connected users. Then add to this the amazing growth they are seeing. I don't think people quite realize what this means.
EDIT: In general, it baffles me how many armchair opinionators there are. Isn't it more interesting to try to understand why FB paid $16B, rather than saying why they shouldn't have?
Metcalfe's law is clearly wrong after a network get's overly large. I could in theory call a random person in China but if I have no way of communicating with them it's practically worthless to do so.
But seriously: I've had people on conference calls when they couldn't understand the language, but someone wanted them to see slides. The fact they could join the call (in that case on Skype) definitely made the network more valuable.
You might argue that the fact you can't speak to some of them drops the value per user, and that is true to some extent. But the fact they are on there and attract other people (some of whom you can speak to) adds value for you, too.
Orders of magnitude. Sure, adding someone else in china is in theory worth something. However adding one more friend that I regularly communicate with on a network is worth more to me than every person in China.
PS: For large networks X Log X is probably much closer to reality than X ^2. Just compare the amount of internet bandwidth between NY to California vs the bandwidth between the US and China.
Metcalf's law is not suggesting "The value of the network to any random node within the network." Because that measure is very subjective in the ways you say.
Instead, it's "The total value of the network to all nodes". Because the more people in the network, the higher the probability that the nodes you DO care about are also connected.
I'd extend the rule further and suggest that the value of a network scales as some Odlyzko and Tilly suggest, but with an additional negative function subtracting value (v) from the network:
v = n(log(n) - f(n)
That is: as the network grows, the added value of each additional member is reduced (log(n)). Further, each additional member of the network exacts a cost to the network as a whole as well. Doing some simple modeling, I suspect that this isn't a strictly linear factor, but itself grows with n, quite possibly as the original Metcalfe's law suggestion. That is: any given member is increasingly less likely to be a positive contribution to the network, but might well present an equal opportunity to be a net negative to the group as a whole:
v = n(log(n)) - kn^2
Where 0 < k < 1 and n > 1.
Moreover, let's look at some group sizes which might allow us to estimate for k in various contexts.
For software team size, it's very typical that a core engineering group has a size of 5-10 members, more or less. This suggests that k is about 0.2 for software development: every added team member exacts a cost, within a single group, of about 20%. This sees value grow for 1 < n < 8, then fall with larger n, hitting negative values at about n=14.
For an elementary school classroom where the ideal class size seems to be around 22-25 students, k would be around 0.08.
For Dunbar's Number, the number of relationships people can manage (100 - 300, typically set at 150), k is between 0.028 (100), 0.2 (150), and 0.113 (300).
For city sizes, it's likely that different cities offer different matches of positive and negative factors. k of 0.0005 gives value max at n ~ 10,000, k of 0.0006 is ~ 100,000, and 0.000007 is around 1 million.
To scale to 1 billion users with net positive value means you have to keep k to less than 0.00000001. That is: any one member can have only a 1 in 10 million chance of being annoying to other members.
> To scale to 1 billion users with net positive value means you have to keep k to less than 0.00000001. That is: any one member can have only a 1 in 10 million chance of being annoying to other members.
Or you could just design a network where new users don't annoy existing users, and reduce it down all the way to zero.
In the case of cities, physical distribution means that even within a given city, the interactions of its citizens, while potentially very high, are generally reasonably low. It's less my direct contacts (likely within a fair approximation of Dunbar's Number) that are high, than my 2nd and 3rd order possible contacts which are high.
In a small town, those 2nd order connections are inherently constrained to the size of the population: my 300 direct contacts may expand to the 3,000 or 30,000 of a small town, but not the million or more of a large metropolis.
Similarly, for more complex organisms, you also have more complex immune systems. An interesting (and staggering) recent fact I ran across is that the individual cell mortality rate among ocean lifeforms is about 20%. Per day. If you're a cell in the ocean, you've got 1:5 odds of not being here tomorrow, because of the viral load:
> The value to huge communications networks isn't in making connections but in avoiding making them at all.
The social networking startup I'm at is focused entirely on that problem. It's taken a lot of work to have the scope of possible interactions be the entire network, while at the same time, limiting the interactions to any given user to just those with a positive value.
Since the algorithms are not omniscient, we also have a ruthlessly efficient feedback mechanism whereby users can indicate when they have received a negative value communication with just a single swipe.
There's tons of interesting problems in this space. :)
we also have a ruthlessly efficient feedback mechanism whereby users can indicate when they have received a negative value communication with just a single swipe.
That's helpful, though you've really got to recognize that a huge part of the problem is assessing indirect feedback.
If you've seen Derek "Veritasium" Muller's "The Problem with Facebook" videos, one of the challenges is that interactions with FB content are hard to gauge. Dating sites have a similar challenge, in that feedback on interactions ("how did the date go") are rarely collected. As opposed to YouTube where a huge signal is "did the user stay on the page for the duration of the entire video". If I watch 5 seconds of a 3 minute video (or 30 seconds of a 60 minute vid), odds are I wasn't very impressed.
There's also the challenge of sorting out abuse of moderation systems, particularly those trying to get legitimate (but unpopular) voices banned or restricted.
The good news is that there are some people whose interactions are so widely negative (spammers and trolls) that you can attack them head on and reduce the cost coefficient significantly (recognizing that the cost constant is constructed of both a specific value and the number of connections). Spam is as annoying as it is because a single spammer affects so many other users.
The flipside is that you can increase the network value by finding people others really want to connect with. Here I see G+ as being horribly naive (also YouTube) in repeatedly making recommendations that I'm absolutely not interested in, without offering me an opportunity to say "don't show me this person" or "don't show me this product / video / category" ever, ever again. One of my long-standing challenges to the "deep data" (snooping) perspective is: rather than compile a massive dossier on my and attempt to bother me by way of it, when you do find me in an intentional mood to find something, get really good at figuring out whether you're offering me what I want or not. Why Google should know my location to within 2 feet every 60 seconds of the past five years ... but not be able to tell me the dot pitch of the monitor I'm shopping for, strikes me as a stunningly obtuse mismatch of data focus.
That said: yes, the ability to dismiss stuff I don't want to see and be bothered with is absolutely useful. As I'd repeatedly said at G+: let me say "not now", "not this hour", "not today / this week / month" (essentially: timeouts). And of course "not ever". G+'s blocking feature is also grossly inadequate. Some people, yes, I simply don't want to deal with. For others, I just don't want to see their insipid posts, or deal with their insipid comments on my own posts.
450 million is a number that immediately brings up the question, how many are real? This is almost as if the site is trying to make up an excuse for the purchase. However to me, having watched different entities out there I am always curious as to what is real and what isn't
Interesting dataset, though I do have some quibbles. For example, registered users aren't the same as monthly unique visitors, which is how CNET is best measured. And ecommerce like Zappos have completely different metrics. One thing I would like to see is the user numbers normalized against web users at the time, which helps put late 90s acquisitions in better perspective.
I think the more fascinating metric here is cost per employee which gives us an idea of economic value created per employee. Whatsapp destroys the other acquisitions on this metric.
it's odd that we focus so much on just user numbers yet talk so much about aqui-hiring.
In most cases it's not just the users they're trying to get, but also the talent (with the idea that most of them aren't one hit wonders).
Acting like startup acquisition is a single-metric game seems like a gross oversimplification. This chart only proves it by showing there is little correlation between user count and buy price.
Some of these acquisitions weren't done for the users.
Aardvark, Jaika, Dodgeball, Picasa, Broadcast.com, FriendFeed (missing), maybe Flickr as well as more recent acquisitions like DeepMind ($900M to Google, zero users) were bought for some combination of the team and the technology.
All this shows to me is how useless it is to compare by "cost per user". The note at the bottom says "Facebook acquired WhatsApp for 16 billion dollars, the largest startup acquisition to date. Cost per user was comparable to Google's YouTube acquisition.".
It's sort of close, but we're still talking about a swing of >30% there -- $5 billion dollars in WhatsApp's case.
There are many things missing from "cost per user" but I think the first thing to ask is "user doing what"?
Agreed - how do we know the value difference for a "user" of Zappos and a user of whatsapp (among other dimensions)? This graph is not very insightful, it's just visualizing the easy way to look at and quantify the purchase.
How many were they at YouTube when they got bought?
I tend to think more about acquisitions as capital per employee, as in, that's how efficient they are at creating value. It's more interesting that comparing users because of the "users doing what" question. I'm disappointed i can't find graphs about those. Last companies I checked (but I can't remember which ones), the capital was around $1m per employee. Except for WhatsApp.
Best acquisition? When was the last time you gave Youtube money (or Google money to watch videos)? The fact that their videos are no unwatchable because of quality stutters and the inability to rewind/jump, it's only a matter of time for a new video site to take over.
It's your right to complain about Youtube quality these days (from what I understand ISPs tend to throttle them mercilessly), but it was clearly an extremely good buy for Google, and Google makes tons of money off of Youtube ads.
When was the last time you paid Google for search results? Meh, can't be making that much money.
Are you kidding? After becoming a household brand by trojan tactics (starting with no ads), and then slowly boiling its users until almost every click leads to an ad with delayed skip, and the videos are covered by six layers of distracting banners and useless information?
If our patience is worth any money, Youtube is extracting it in spades now.
Yup. "Cost per user" strikes me as a very '90s metric, for lack of a better way to put it. It is an unfocused average. It presumes that the singular purpose of any acquisition is user count. This makes especially little sense in certain cases, i.e., the Facebook/WhatsApp acquisition, wherein a large proportion of the WA users are likely to be duplicative with FB users.
"User doing what?" is a decent start. But of course, we have a host of other considerations to ponder: technology, platform, integrations, competitive blocking and tackling, etc.
FWIW, I still can't wholly grok the WA purchase price. The price seems largely to have been driven by competitive bidding between FB and Google. Also a healthy sprinkling of defensiveness and fear, mixed with a dollop of "because we can." Factor in some "we believe our stock is overvalued, and as such, we will realize a discount on price correction," and I suppose you arrive at a figure that will end up south of FB's nominal purchase price. The one thing I can say is that Instagram looks likes a ridiculous bargain in retrospect, and especially in comparison.
That said, to whatever extent we do want to talk about user counts, it's worth noting some history. Around the time Facebook had an estimated $15 billion valuation, on the heels of Microsoft's advertising partnership/investment in 2007, Facebook had about 50 million active users. WhatsApp has 465M+. Apples and oranges, to some extent. But still. Worth considering.
Kind of dumb to compare U.S./Western Europe users to primarily non-U.S. users as is the case for Whatsapp. I'd bet a 19 year old in India is an order of magnitude less valuable from an advertising standpoint than a 19 year old in America. Not only because people have less disposable income in those countries, but because kids have less control over their parents' disposable income.
How are we defining a user here? For example in the case of youtube, do we count people who watch videos? People who sign up and comment? People who upload videos?
99 comments
[ 2.4 ms ] story [ 187 ms ] threadOn a related note, 11K per user is insane. What did Yahoo see in Broadcast.com?
http://en.wikipedia.org/wiki/Decimal_mark#Examples_of_use
450000000 is a lot of users. Metcalfe's law says that the value of a communication network is proportional to the square of the connected users. Then add to this the amazing growth they are seeing. I don't think people quite realize what this means.
EDIT: In general, it baffles me how many armchair opinionators there are. Isn't it more interesting to try to understand why FB paid $16B, rather than saying why they shouldn't have?
...but did they really put a LLS line through that dataset?
But seriously: I've had people on conference calls when they couldn't understand the language, but someone wanted them to see slides. The fact they could join the call (in that case on Skype) definitely made the network more valuable.
You might argue that the fact you can't speak to some of them drops the value per user, and that is true to some extent. But the fact they are on there and attract other people (some of whom you can speak to) adds value for you, too.
PS: For large networks X Log X is probably much closer to reality than X ^2. Just compare the amount of internet bandwidth between NY to California vs the bandwidth between the US and China.
Metcalf's law is not suggesting "The value of the network to any random node within the network." Because that measure is very subjective in the ways you say.
Instead, it's "The total value of the network to all nodes". Because the more people in the network, the higher the probability that the nodes you DO care about are also connected.
http://www.dtc.umn.edu/~odlyzko/doc/metcalfe.pdf
I'd extend the rule further and suggest that the value of a network scales as some Odlyzko and Tilly suggest, but with an additional negative function subtracting value (v) from the network:
That is: as the network grows, the added value of each additional member is reduced (log(n)). Further, each additional member of the network exacts a cost to the network as a whole as well. Doing some simple modeling, I suspect that this isn't a strictly linear factor, but itself grows with n, quite possibly as the original Metcalfe's law suggestion. That is: any given member is increasingly less likely to be a positive contribution to the network, but might well present an equal opportunity to be a net negative to the group as a whole: Where 0 < k < 1 and n > 1.Moreover, let's look at some group sizes which might allow us to estimate for k in various contexts.
For software team size, it's very typical that a core engineering group has a size of 5-10 members, more or less. This suggests that k is about 0.2 for software development: every added team member exacts a cost, within a single group, of about 20%. This sees value grow for 1 < n < 8, then fall with larger n, hitting negative values at about n=14.
For an elementary school classroom where the ideal class size seems to be around 22-25 students, k would be around 0.08.
For Dunbar's Number, the number of relationships people can manage (100 - 300, typically set at 150), k is between 0.028 (100), 0.2 (150), and 0.113 (300).
For city sizes, it's likely that different cities offer different matches of positive and negative factors. k of 0.0005 gives value max at n ~ 10,000, k of 0.0006 is ~ 100,000, and 0.000007 is around 1 million.
To scale to 1 billion users with net positive value means you have to keep k to less than 0.00000001. That is: any one member can have only a 1 in 10 million chance of being annoying to other members.
Or you could just design a network where new users don't annoy existing users, and reduce it down all the way to zero.
Sure. But that is just one of many possible cost-reducing mechanisms. I've expanded on this idea at more length here: http://www.reddit.com/r/dredmorbius/comments/1yzvh3/refutati...
In the case of cities, physical distribution means that even within a given city, the interactions of its citizens, while potentially very high, are generally reasonably low. It's less my direct contacts (likely within a fair approximation of Dunbar's Number) that are high, than my 2nd and 3rd order possible contacts which are high.
In a small town, those 2nd order connections are inherently constrained to the size of the population: my 300 direct contacts may expand to the 3,000 or 30,000 of a small town, but not the million or more of a large metropolis.
Similarly, for more complex organisms, you also have more complex immune systems. An interesting (and staggering) recent fact I ran across is that the individual cell mortality rate among ocean lifeforms is about 20%. Per day. If you're a cell in the ocean, you've got 1:5 odds of not being here tomorrow, because of the viral load:
http://www.reddit.com/r/dredmorbius/comments/1vyanj/viral_so...
So it is with other complex networks: higher levels of complexity require far more aggressive immune systems.
> The value to huge communications networks isn't in making connections but in avoiding making them at all.
The social networking startup I'm at is focused entirely on that problem. It's taken a lot of work to have the scope of possible interactions be the entire network, while at the same time, limiting the interactions to any given user to just those with a positive value.
Since the algorithms are not omniscient, we also have a ruthlessly efficient feedback mechanism whereby users can indicate when they have received a negative value communication with just a single swipe.
There's tons of interesting problems in this space. :)
we also have a ruthlessly efficient feedback mechanism whereby users can indicate when they have received a negative value communication with just a single swipe.
That's helpful, though you've really got to recognize that a huge part of the problem is assessing indirect feedback.
If you've seen Derek "Veritasium" Muller's "The Problem with Facebook" videos, one of the challenges is that interactions with FB content are hard to gauge. Dating sites have a similar challenge, in that feedback on interactions ("how did the date go") are rarely collected. As opposed to YouTube where a huge signal is "did the user stay on the page for the duration of the entire video". If I watch 5 seconds of a 3 minute video (or 30 seconds of a 60 minute vid), odds are I wasn't very impressed.
There's also the challenge of sorting out abuse of moderation systems, particularly those trying to get legitimate (but unpopular) voices banned or restricted.
The good news is that there are some people whose interactions are so widely negative (spammers and trolls) that you can attack them head on and reduce the cost coefficient significantly (recognizing that the cost constant is constructed of both a specific value and the number of connections). Spam is as annoying as it is because a single spammer affects so many other users.
The flipside is that you can increase the network value by finding people others really want to connect with. Here I see G+ as being horribly naive (also YouTube) in repeatedly making recommendations that I'm absolutely not interested in, without offering me an opportunity to say "don't show me this person" or "don't show me this product / video / category" ever, ever again. One of my long-standing challenges to the "deep data" (snooping) perspective is: rather than compile a massive dossier on my and attempt to bother me by way of it, when you do find me in an intentional mood to find something, get really good at figuring out whether you're offering me what I want or not. Why Google should know my location to within 2 feet every 60 seconds of the past five years ... but not be able to tell me the dot pitch of the monitor I'm shopping for, strikes me as a stunningly obtuse mismatch of data focus.
That said: yes, the ability to dismiss stuff I don't want to see and be bothered with is absolutely useful. As I'd repeatedly said at G+: let me say "not now", "not this hour", "not today / this week / month" (essentially: timeouts). And of course "not ever". G+'s blocking feature is also grossly inadequate. Some people, yes, I simply don't want to deal with. For others, I just don't want to see their insipid posts, or deal with their insipid comments on my own posts.
After all, that's the main way that we can consider a per-user cost to be "high".
In most cases it's not just the users they're trying to get, but also the talent (with the idea that most of them aren't one hit wonders).
Acting like startup acquisition is a single-metric game seems like a gross oversimplification. This chart only proves it by showing there is little correlation between user count and buy price.
and does it include the youtube legal deals and fines? probably not
There's probably a bug in the visualization library getting the thousands coma as a decimal separator, too bad :)
Aardvark, Jaika, Dodgeball, Picasa, Broadcast.com, FriendFeed (missing), maybe Flickr as well as more recent acquisitions like DeepMind ($900M to Google, zero users) were bought for some combination of the team and the technology.
It's sort of close, but we're still talking about a swing of >30% there -- $5 billion dollars in WhatsApp's case.
There are many things missing from "cost per user" but I think the first thing to ask is "user doing what"?
I'd forgotten completely about YouTube. At under $2B, that must have been one of the best acquisitions of the whole bunch.
YouTube for 1/10 the cost of WhatsApp.
I tend to think more about acquisitions as capital per employee, as in, that's how efficient they are at creating value. It's more interesting that comparing users because of the "users doing what" question. I'm disappointed i can't find graphs about those. Last companies I checked (but I can't remember which ones), the capital was around $1m per employee. Except for WhatsApp.
When was the last time you paid Google for search results? Meh, can't be making that much money.
Nice way to show I was wrong :)
If our patience is worth any money, Youtube is extracting it in spades now.
"User doing what?" is a decent start. But of course, we have a host of other considerations to ponder: technology, platform, integrations, competitive blocking and tackling, etc.
FWIW, I still can't wholly grok the WA purchase price. The price seems largely to have been driven by competitive bidding between FB and Google. Also a healthy sprinkling of defensiveness and fear, mixed with a dollop of "because we can." Factor in some "we believe our stock is overvalued, and as such, we will realize a discount on price correction," and I suppose you arrive at a figure that will end up south of FB's nominal purchase price. The one thing I can say is that Instagram looks likes a ridiculous bargain in retrospect, and especially in comparison.
That said, to whatever extent we do want to talk about user counts, it's worth noting some history. Around the time Facebook had an estimated $15 billion valuation, on the heels of Microsoft's advertising partnership/investment in 2007, Facebook had about 50 million active users. WhatsApp has 465M+. Apples and oranges, to some extent. But still. Worth considering.
Now is a good time to be in the yacht business.