Convopage
See the entire conversation
Benedict Evans
@BenedictEvans
You spend 50-100x more on your smartphone than Google or FB make from you in ad revenue. They pay for their clouds out of that ad revenue
68 replies and sub-replies as of Oct 10 2016
Benedict Evans
@BenedictEvans
That is, Google and FB’s users spend far more on computing power than they can themselves. Edge > cloud, in aggregate
Benedict Evans
@BenedictEvans
It makes a lot of sense to try to move as much computing as possible to the devices that your users already paid for.
Benedict Evans
@BenedictEvans
Computing power in a smarpthone is often a LOT more than the amount in the cloud that can economically be allocated to that user.
Benedict Evans
@BenedictEvans
Computing on the device uses power, but then so does turning on the radios and transmitting it to the cloud.
Benedict Evans
@BenedictEvans
And the great thing about ML is that training the model & running it are two different things. Trend: train in the cloud, run on device.
Benedict Evans
@BenedictEvans
(TIL)
Jon
@athst
if you add up all iOS devices, do you think Apple could control more aggregate computing power than Google/FB?
Adriano Carmezim
@AdrianoCarmezim
even running the model can be computationally quite expensive depending on the model application
Benedict Evans
@BenedictEvans
Sure. But turning on the radios might use a lot more battery
ruizhang
@silverhawk_ny
not quite , report about the breakdown of smartphone power, screen and CPU way more than wifi
Benedict Evans
@BenedictEvans
processing image recognition locally uses less power than transmitting the image
ruizhang
@silverhawk_ny
could be, if you refer to image filter but for object recognize, must relay to cloud to train algorithm
Peter MacRobert
@PeteM
calculating fitness of a sample is usually (ideally) a linear function. Dot product of a vector.
Brendon J. Wilson
@brendonjwilson
Will we see competitors stealing machine learning parameters / models from each other's apps?
Nathan Taylor
@ntaylor963
The parallel was the hype/fashion in 1999 for thin client. Then runtime in browser of course grew. Moore's Law on both sides.
Guillaume Mispelbaum
@gmispelbaum
what about train and run on the device? Allows business models with lower margins
Guesswork
@guessworkco
this will work for some algorithms like decision tree. But will fail for most of the clustering algorithms.
Eva Aymn
@evaAymn
would u tell me what you mean by "computing" !
Jeff McLeman
@jeffmc
But RF uses more, because RF is unpredictable. You have 4 bars, but that means you hear the tower, not vice versa, so boost
Benedict Evans
@BenedictEvans
yep
Thomas McMahon
@thmcmahon
if you think of the combined computing power of all our phones as one computer - it's a big one.
ruizhang
@silverhawk_ny
maybe we should use p2p more for smart phone
raoulduck
@ChatonDeGuerre
But compute in the cloud can accommodate oversubscription by factor of 50.
#Erlangs
Ted Herman
@tedherman
If that were the only cost involved.
Benedict Evans
@BenedictEvans
transmission uses power.
Alec Lazarescu
@alec1a
Desktop maybe, mobile it's too much battery drain potential.
Benedict Evans
@BenedictEvans
transmission uses power. LOTS of power
Adam D'Angelo
@adamdangelo
I think the real conclusion is that for data-heavy work, you want to move the computation to the data
Alec Lazarescu
@alec1a
Yes but anything run on a CLIENT can be manipulated there to cause you grief so careful what you calculate there
Alec Lazarescu
@alec1a
Fair point. Depends on specifics. Burning straw man sending vs mining a Bitcoin. 😀
Ruben Martinez Jr.
@rubencodes
Long term though the opposite makes sense. Decrease upfront cost to bring the next half of the world online.
Sachin Tandon
@cugwmui
Logical, but why does it seem that apps are moving away from edge computing and into the cloud? Even image processing etc..
Judaaaahhhh
@jdrch
Depends on the application. Games & storage, sure. But something like FB & Google Search can't be run locally. At all
Benedict Evans
@BenedictEvans
‘as possible’
Judaaaahhhh
@jdrch
ah good point
Dan
@danielkein
it makes even more sense to tell people it's for privacy reasons.
Nicholas Perry
@ultimape
Yes! So much this. Big reason I dug into
@monaxHQ
's stuff was this potential of what they call "Participatory Architecture"
Leon Overweel
@LayOn_OverWhale
spend on buying my phone or spend on buying things on my phone?
Adam D'Angelo
@adamdangelo
where do you get 50-100x? I get more like 2-5x
Benedict Evans
@BenedictEvans
does Facebook make $200-300/user?
Adam D'Angelo
@adamdangelo
Google makes $50B/year from say 1B users. So $50/user/year.
Benedict Evans
@BenedictEvans
$85m run rate, but 3bn users. ARPUs of $13 (FB) and $28. 25-50x, say.
Adam D'Angelo
@adamdangelo
what are you using for the cost of the device?
Benedict Evans
@BenedictEvans
now? $250-600. Android ASP ~$225
Adam D'Angelo
@adamdangelo
the device lasts more than 1 year. I also think the average is below $200 if you look globally at all 3B users.
Adam D'Angelo
@adamdangelo
I think if you go out to 3B users the average is more like $100/year
Adam D'Angelo
@adamdangelo
so that gets you to 4-8x
Benedict Evans
@BenedictEvans
$200 ASP /$13 ARPU is 15. ASP of FB users rather higher - $300+
Benedict Evans
@BenedictEvans
‘I bought this phone last year for $200. You have a global capex of x% of $13/user/year’.
Adam D'Angelo
@adamdangelo
1. most of the phone cost isn't going to computation; 2. the phone is idle most of the time, but cloud servers are working
Adam D'Angelo
@adamdangelo
3. compute per dollar is way more efficient in server settings
Benedict Evans
@BenedictEvans
A truck can carry more than a car, but your customers drive themselves to wal-mart for free
Adam D'Angelo
@adamdangelo
the vast majority of shipping happens between factories and the local walmart, not in cars driving things home from the store
Benedict Evans
@BenedictEvans
One can swing the precise number around - point is that the idea there is ‘infinite compute’ in the cloud v device is wrong
Adam D'Angelo
@adamdangelo
yeah, I disagree with the overall conclusion, not just the details
StJohnDeakins
@StJohnDeakins
more like "there is (almost) infinite compute in 5Bn devices"-and many more advantages to distribution besides
@adamdangelo
Pimentón
@BrettS69
So why don't Google/FB sell a home cloud personal server?
Benedict Evans
@BenedictEvans
because ‘where is the cheapest computation?’ is not the most relevant question for that product. For others it is.
Jon Duffy
@jon_duffy
and that's why it's good to sell hardware (if you can)
Tajar Varghese
@tajarvarghese
wow. This is a dumb comment. You don't take into account parallel processing and utilization.
Tajar Varghese
@tajarvarghese
and you mistake asp for cost. Margins for fb and google are MUCH higher than aapl.
Benedict Evans
@BenedictEvans
do you think I don’t know that? 😁
Tajar Varghese
@tajarvarghese
seems like you don't know a lot of things.
flavio 웃 graf
@youflavio
while you know a lot. Right? Discuss idea not the person.
@BenedictEvans
Tajar Varghese
@tajarvarghese
fair enough!
Tajar Varghese
@tajarvarghese
look at the gpu chip stack layer. There's a reason it's used in cloud rather than a normal CPU.
Edwin Zuidema
@EdwinZuidema
how do you see this balance for wearables and smart home? Cheaper and smaller devices, more data/sensor oriented
Alexandre Ferrari
@aferrari
So, whatever they can put to run in device-level, they will. Forcing you to pay more for the device.