My very first project at Google [1] drove over $100M in revenue. Team of 3, wasn't intended to be revenue-facing at all, but it turns out that when you give a tool to the AdSales folks that shows the customers all the other queries they could be advertising on, they buy a whole lot of additional ads.

It's very easy to underestimate how small changes can add up at scale. Like another commenter said, Google makes about $80B in revenue. It's not hard to make a change to UI or ranking that impacts conversion rates by 0.125%. One of the changes I did for a later project would've lost a billion dollars for Google had it not been caught by A/B testing. Heck, a non-visible change I made (moving the results above the ads in DOM order, without changing styling, so that blind people didn't need to sit through ads before getting their results) was blamed for $50M in lost revenue before we ran an experiment and confirmed there was no effect. Marissa Mayer had done the exact same experiment 8ish years before and measured a $50M loss; the difference was that everyone was on dialup in 2002 but broadband in 2010, and so what was an unacceptable latency penalty that caused everyone to not click on ads had become an unnoticeable change.

[1] http://searchengineland.com/google-wonder-wheel-17093

Not to say that individual engineers can make large impacts at that scale, cause hell a million for a do change seems large. That being said, you were on a team of 3 that drove 100 million in revenue, so you didn't have that large an impact as a single engineer, and your other example was something that was assumed based on 8 year old data, and the impact didn't exist when actual measurements were taken.

I'm still having a hard time believing that a single person created 100 million in value for Google for a product that isn't currently being sold. It seems more like a result of a price war for self driving car talent as companies are scrambling to try and take the whole self driving car market before anyone else

These things are related. You don't get a price war over someone that results in paying them $100m if you don't believe that person will create over $100m in value.

And since this person went off, started their own company, and sold it a couple years later for $1 Billion, it seems whoever was offering that paycheck was right to do so.

That's the thought process behind investing, but just because the investment risk was taken doesn't mean they actually created that value. Also with strategic investments sometimes it isn't that your investment generates that much revenue for you, but ithe does deny your competition access to that resource

Well, keep in mind that we also don't know the timescale this was over. If it was $120mm in a year, then that's perhaps crazy, but 120 million over 5 or ten years?

With 80 billion in revenue and ~75K employees according to a google search, that means that there's an expectation that an average hire will drive $1 million in revenue per year. To be paid 120M/year you'd need to drive significantly more than that, but to be paid 20M/year, that's now someone who is 50x the average employee, not 200x the average employee, and that's reasonable. I think we all know an engineer or two who is extraordinary.

I know engineers that did that.

so:

when you're earning $1 it's not hard to double your money when you're earning $1 billion it's hard to double your money but when you're earning $1 billion it's a lot easier to earn an extra million or two.

Google's quarterly revenue is roughly $20B[1] - that's $80B per year. I think that it is perfectly plausible that an engineer could achieve a 0.125% increase in yearly revenue. Now, disentangling that from other possible contributing factors, getting the right visibility from management, etc. is distinct from that.

I have no problem believing that MapReduce contributed $XB to Google's revenue over the course of the past 20 years.

[1] https://www.sec.gov/Archives/edgar/data/1652044/000165204416...

MapReduce was published in 2004. The paper was two people; the implementation was probably a lot more. Even then, it was just a re-engineering of an earlier code base (read that as multiple available cat skinning technologies). By 2014, it was no longer used, so that's 10 years. Lastly, it wasn't exactly novel.

https://en.wikipedia.org/wiki/MapReduce#Lack_of_novelty

FWIW, it appears they're giving away the TensorFlow stuff because they lost the mindshare battle between MapReduce+Hadoop and don't want that to happen again.

Again, the $100M figure is an easy (heroic) claim to make but hard to substantiate. Shit's hard. Big shit is really hard.

You basically don't know what you're talking about, factually speaking, along a wide range of fact dimensions.

Good to know that. I was worried.

> By 2014, it was no longer used

[Citation needed]

I think my point on complexity probably stands. At early days, people were creating products from ground up. Now it's more incremental. At early days 100M might have meant doubling the revenue, but now, it's a mere fraction.

You can shave off latency and 1% improvement could result in 0.1% revenue gain (made up numbers).

You can improve relevancy of the ads by 1% and contribute to revenue similar amounts.

It's not really super hard.

(also @google)

Scale makes it easy to have an outsized impact.

Consider that a single CPU core costs ~10$/yr in electricity, and that google has a really big computer[1]. The questions are then "how many CPUs does google have", "what percentage of those can any given change affect", and "how many such changes can one engineer make in a year?"

And that's just one type of potential change.

[1]: http://www.clickhole.com/article/future-now-google-has-confi...