00:00:00
So fulfillment centers ship into this building.
00:00:03
This building sorts those packages.
00:00:05
Then they flow out to our delivery stations.
00:00:07
Inside this 1,000,000 square foot warehouse in Northern
00:00:10
California, amazon packages are handled by hundreds of
00:00:13
people and hundreds of robots, all increasingly driven
00:00:16
by tech's biggest craze.
00:00:17
Generative AI underpins everything we're doing here
00:00:20
with perceiving, grasping and moving products with the
00:00:23
Robin arms to managing fleet congestion with our
00:00:26
Pegasus robotic drives.
00:00:28
For years, Amazon has sped up package delivery two-day,
00:00:31
one-day, and now, more and more, same-day made
00:00:35
possible by more workers but also by rapidly improving
00:00:38
algorithms and AI-enabled robots.
00:00:41
What you see here is I like to call it our dance floor.
00:00:44
And thanks to being an online retailer from the start,
00:00:47
Amazon has a big advantage when it comes to the
00:00:49
massive data needed for generative AI.
00:00:51
Absolutely. Amazon has been better at it than probably
00:00:56
every other retailer out there.
00:00:58
Think better predictions of exactly what you're going
00:01:00
to order from where and when.
00:01:01
Hundreds of thousands of robots and more efficient
00:01:04
delivery routes.
00:01:05
But not all the change that could come from generative
00:01:08
AI is positive.
00:01:09
In the event that they were able to leverage generative
00:01:14
AI and fully automate the fulfillment center, I think
00:01:18
that would be problematic.
00:01:20
The other downside that we don't talk about enough is
00:01:22
the negative impact on the environment: the costs of
00:01:25
running data centers, the use of electricity, the use
00:01:29
of water for cooling.
00:01:30
Still, Amazon says AI helps cut costs and its carbon
00:01:34
footprint thanks to more efficient planning.
00:01:36
It seems subtle, but at this scale, getting like just
00:01:41
one more product in the right spot means that it's
00:01:43
shipping less distance when you order it.
00:01:45
Better speed, lower distance traveled, better
00:01:48
sustainability.
00:01:49
CNBC visited Amazon's largest California sort center
00:01:52
and a same-day warehouse nearby to see firsthand how
00:01:55
it's putting AI to work at every step of operations,
00:01:58
and sat down with Amazon's head of transportation
00:02:01
technology to find out just how far the e-commerce
00:02:04
giant plans to take
00:02:05
it.
00:02:14
When Amazon Prime launched in 2005, two-day shipping
00:02:17
was virtually unheard of.
00:02:19
And although it's now standard and free for millions
00:02:21
of items if you're a Prime member, it's a grueling
00:02:23
logistical lift.
00:02:25
Along here, you see Robin arms, which are robotic arms.
00:02:29
They're loading packages, with employees, onto Pegasus
00:02:33
drive units. Those Pegasus drive units are then
00:02:35
sorting packages by neighborhood.
00:02:38
Steve Armato started at Amazon as a software engineer
00:02:40
before Prime, in 2001.
00:02:42
A lot of the things you see today, those weren't there
00:02:45
in 2001.
00:02:47
We had five fulfillment centers back then.
00:02:49
Now we have hundreds.
00:02:51
The delivery vans that you see in the neighborhood,
00:02:53
none of that was there.
00:02:54
Traditional retailers like Walmart and Target were
00:02:57
selling online, but they hadn't started making
00:02:59
promises of speed.
00:03:00
Back then, you know, mail order when Amazon started,
00:03:03
you'd be lucky if you could get something in 2 to 3
00:03:05
weeks. And, you know, Amazon would still promise like
00:03:08
a week and they would yet get it to you in a few days.
00:03:10
So that was amazing.
00:03:12
So how did Amazon pull it off?
00:03:14
The short answer is data.
00:03:16
Long before generative AI became all the rage with the
00:03:18
release of ChatGPT in 2022, general AI was a huge
00:03:22
differentiator for Amazon.
00:03:24
As an early online-only retailer, Amazon had a unique
00:03:27
ability to collect mass aggregate data on shopping
00:03:29
behavior and use it to create algorithms to maximize
00:03:32
sales and speedy logistics.
00:03:34
We've been working on AI over 25 years.For employees,
00:03:38
a lot of it is around ergonomics and safety.
00:03:41
For customers, it's around vast selection, great
00:03:45
speeds.
00:03:46
Exploiting technology to drive e-commerce sales.
00:03:49
That's essentially what Amazon has done since '97.
00:03:52
Since the beginning.
00:03:53
They are, I would say, hands down the most data heavy
00:03:58
and data savvy company.
00:04:01
It's not that Walmart and Target and Costco and others
00:04:04
don't have their own reams of data, but they're
00:04:07
looking at things a little bit differently, and they
00:04:09
have much older systems.
00:04:12
Amazon is decades younger than its major retail
00:04:15
competitors, but its stock value and footprint have
00:04:17
grown incredibly fast.
00:04:18
Hundreds of warehouses, more than 1.5 million U.S.
00:04:22
employees and more speed.
00:04:24
In 2014, amazon launched Prime Now with some
00:04:27
deliveries arriving in an hour or less.
00:04:29
Then in 2018, Amazon vastly increased its driver
00:04:32
network with the launch of its Delivery Service
00:04:34
Partner program, where it contracts driving out to
00:04:36
some 4,400 small delivery businesses that employ
00:04:39
390,000 drivers.
00:04:42
By 2019, one-day shipping was the norm.
00:04:45
Then in 2020, Amazon began using transformer
00:04:47
architecture, the backbones of what we know of today
00:04:50
as generative AI, to develop models for demand
00:04:53
forecasting and supply chain optimization.
00:04:55
By 2022, it was rolling AI transformer models into its
00:04:59
robotics. All that made shipping times even faster.
00:05:02
Today, drivers are delivering 20 million packages per
00:05:05
day across 20 countries, and in the first quarter of
00:05:08
2024, more than 2 billion items arrived the same or
00:05:12
next day.
00:05:13
60% of our deliveries for Prime customers in March were
00:05:18
same-day or next-day, so there are a lot of those fast
00:05:20
orders for our top 60 metropolitan areas.
00:05:23
Could you have gotten to that 60% number without
00:05:25
generative AI?
00:05:26
Well, I think we've been working on this for decades to
00:05:30
get to this speed.
00:05:31
And it's a combination of engineering, people,
00:05:34
processes, and technology.
00:05:36
Generative AI is a big unlock for us, particularly for
00:05:39
new products where we have sparse or no history for
00:05:43
that sales history for that product.
00:05:45
It's going to come faster because of generative AI.
00:05:51
But all this speed comes at a huge cost, in actual cap
00:05:55
ex, but also human labor, a burden that can be
00:05:57
reduced, Amazon says, with the use of robots.
00:06:00
Before robotics, pickers would need to walk distances
00:06:05
between aisles to pick products, kind of like a
00:06:07
library. And now it's being brought to you more like
00:06:11
self-service. And so that's that's great for
00:06:13
ergonomics. It's great for less walking.
00:06:15
Amazon has faced scrutiny in recent years over its
00:06:18
workplace injury record, with federal citations for
00:06:21
safety violations and a year-long Senate probe that
00:06:24
found Prime Day was a major cause of worker injuries.
00:06:27
Amazon has appealed the citations and said the report
00:06:30
ignores progress it's made, and it says AI can help.
00:06:33
One algorithmic improvement is to take our faster
00:06:36
selling products and place those on the shelves at
00:06:39
waist height. That's your ergonomic power zone.
00:06:42
So less reaching, less bending.
00:06:44
Amazon's big shift to automation started in 2012 with
00:06:47
the purchase of Kiva Systems for $775 million.
00:06:51
Now, Amazon has deployed at least 750,000 robots, more
00:06:55
than double the number it had in 2021.
00:06:57
So generative AI helps with prioritization.
00:07:00
So some of the two-day deliveries might stand aside
00:07:04
and let the robot with a next-day delivery go on its
00:07:07
mission first and take a straight line to its
00:07:09
destination.
00:07:10
Amazon's next generation of drive units, called
00:07:12
Proteus, are fully autonomous.
00:07:14
They're actually outside of this fenced area, moving
00:07:17
things around. They're using generative AI and
00:07:20
computer vision to avoid obstacles and find the right
00:07:23
place to stop, the right place to park.
00:07:25
And then there's Robin arms, which Amazon says have
00:07:28
handled some 2 billion packages so far.
00:07:31
Would these 20 Robin arms be able to do what they're
00:07:33
doing without generative AI?
00:07:35
I think that we would see that they would require a lot
00:07:37
more training. And so generative AI has been really a
00:07:40
step function improvement in being able to infer from
00:07:44
our vast product catalogue about how to handle a given
00:07:47
product, even if I haven't seen that product before.
00:07:51
And although it's only in a couple of warehouses for
00:07:54
now, there's a humanoid robot called Digit that can
00:07:56
grasp and handle items in a similar way to how people
00:07:58
can. And Amazon has a new deal with AI startup
00:08:01
Covariant, hiring its founders and licensing its
00:08:04
models that help robots handle a wider range of
00:08:06
physical objects.
00:08:08
Of course, this brings up the big question, could
00:08:10
Amazon one day replace all warehouse workers if AI
00:08:13
helps robots get too capable?
00:08:15
There's kind of a balancing act for Amazon.
00:08:17
How can they implement automation to improve
00:08:21
efficiency and manage labor expenses?
00:08:24
But how can they do it in a way that complements their
00:08:27
use of humans and doesn't replace them?
00:08:29
Amazon says the robots work with people and they're
00:08:32
creating new roles.
00:08:33
We're investing over $1.2 billion to upskill more than
00:08:37
300,000 employees by end of next year.
00:08:40
One study found that each robot adopted in
00:08:42
manufacturing replaced about three workers.
00:08:45
Other research shows that companies that deploy more
00:08:47
robots add more jobs overall.
00:08:49
So someone needs to maintain this if it breaks down.
00:08:53
Or if something does get dropped on the dance floor,
00:08:56
we have a process and special training to go clean that
00:09:00
up. And so each of those creates new categories of
00:09:03
jobs, some of which have higher earnings potential as
00:09:05
well.
00:09:06
The truth is, at the end of the day, Amazon's
00:09:09
responsibility is not to employ, you know, kind of a
00:09:12
million Americans, even though it does.
00:09:14
Their responsibility is to their shareholders.
00:09:19
Another big way to please shareholders is to cut down
00:09:22
on the huge amount of time and money it takes to get
00:09:24
inventory from sellers to customers.
00:09:26
To do this, amazon has always used algorithms to
00:09:29
predict how much of what inventory is needed, when
00:09:32
and where.
00:09:33
Every product has like a regional nuance.
00:09:35
We recently regionalized our entire national network.
00:09:39
And by doing a regional network, that means that
00:09:42
products are more likely to ship from fulfillment
00:09:45
centers close to you.
00:09:46
What's new with generative AI is the ability to predict
00:09:49
where to place brand new items.
00:09:51
So we're able to use generative AI to create a link
00:09:54
between products we have seen before, where we do have
00:09:56
a sales history, and a new product we haven't seen
00:09:59
before yet, and get it in the right place the first
00:10:02
time. So when we place a product in the right place
00:10:05
ahead of time before you click buy, it's traveling
00:10:08
less distance, which is a win for speed and
00:10:11
sustainability.
00:10:12
Amazon also says AI is helping sustainability with a
00:10:15
specific model that makes better choices about which
00:10:17
packaging to use, and by reducing the number of
00:10:20
damaged items that get sent to and returned by
00:10:23
customers. Amazon says its AI is three times better
00:10:26
than humans at identifying damaged products.
00:10:29
We ship billions of packages.
00:10:31
We have the data about those packages.
00:10:32
So we're able to use computer vision together with
00:10:35
generative AI and that vast product and package data,
00:10:39
to then detect damages and being able to sideline a
00:10:43
package if we think it might be damaged before we ship
00:10:45
it to a customer.
00:10:47
But training and running AI is itself a carbon
00:10:50
intensive process, a fact that could make it hard for
00:10:53
Amazon to achieve its 2019 climate pledge to reach net
00:10:57
zero carbon by 2040.
00:10:58
By 2027, AI servers are projected to use up as much
00:11:02
power every year as a small country.
00:11:04
And Amazon Web Services has data centers filled with
00:11:07
servers running AI workloads, although this also gives
00:11:10
it an edge over other e-commerce players because it
00:11:12
can train its AI in-house.
00:11:14
Amazon has also invested $4 billion in AI startup
00:11:17
anthropic, which makes chatbot Claude a competitor to
00:11:20
OpenAI's ChatGPT.
00:11:22
And Amazon makes its own AI-focused microchips and its
00:11:25
own generative AI tools for developers, which are used
00:11:28
in operations.
00:11:29
We use tools like Amazon Q, Amazon Bedrock.
00:11:32
That allows us to evaluate different models against
00:11:36
the, you know, what does good look like?
00:11:38
So what are the metrics for success of this business
00:11:40
application?
00:11:41
One big metric for success that shareholders are
00:11:44
watching: if Amazon's huge investment in AI will
00:11:47
translate to profits.
00:11:48
I have yet to see huge lift in anybody's retail
00:11:52
business due to generative AI, including Amazon.
00:11:55
I think that a lot of their biggest impact has
00:11:58
happened because of the earlier investments, not
00:12:02
necessarily some of these more recent investments.
00:12:04
I've seen a lot of hype, but no actual numbers.
00:12:13
One area Amazon is hoping AI will translate to true
00:12:16
savings is the most expensive part of the delivery
00:12:18
process. Getting that package the last mile to your
00:12:21
door. Amazon is now using more than 20 machine
00:12:24
learning models to figure out the most efficient
00:12:26
routes for its delivery drivers.
00:12:27
If there is more congestion on a road or if a road is
00:12:31
closed, AI is able to help us determine whether that
00:12:35
diversion is still there or take that different route.
00:12:38
In 2021, CNBC talked to Amazon drivers about the
00:12:41
pressures and pitfalls of the job, from dog bites to
00:12:44
urinating in bottles to save time.
00:12:46
People are running through stop signs, running through
00:12:49
yellow lights. Everybody I knew was buckling their
00:12:52
seat belt behind their backs, because the time it took
00:12:54
just to buckle your seat belt, unbuckle your seat belt
00:12:56
every time, was enough time to get you behind
00:13:00
schedule.
00:13:01
The hope is, with better routes and vehicle
00:13:03
coordination, drivers will feel less pressure to cut
00:13:05
corners or skip breaks.
00:13:07
It should enhance the driver experience, but it will
00:13:10
still be challenging.
00:13:12
In 2022, Amazon also rolled out new fully electric
00:13:15
Rivian vans and now has 15,000 of them across the U.S.
00:13:19
They're equipped with large screens where the new
00:13:21
mapping and routing is displayed, as well as
00:13:23
AI-enabled cameras that watch the road, sides of the
00:13:26
vehicle and the driver.
00:13:28
There's no camera recording if the driver is not
00:13:31
driving and there's a privacy mode.
00:13:33
Another area where privacy often comes up is around the
00:13:36
huge amount of data it collects on shopping behavior.
00:13:38
Now, with generative AI, that data can be used to
00:13:40
generate better hyper-personalized product
00:13:43
recommendations to a shopper, thanks to new developer
00:13:45
tools like Amazon Personalize.
00:13:48
What do you say to people when they're concerned about
00:13:50
their privacy? Or that it feels creepy that Amazon
00:13:52
predicts their shopping behavior so closely?
00:13:55
Well, we're looking at aggregate sales history.
00:13:57
We're looking at geographic regions and their
00:14:01
behavior. If a particular customer were in San
00:14:04
Francisco, it's not about that one customer.
00:14:06
It's about what the aggregate behavior in that area
00:14:09
is.
00:14:10
Sellers can also use gen AI to write more targeted
00:14:13
product listings, or to generate images of their
00:14:15
products in different seasonal and lifestyle settings.
00:14:18
For shoppers, last year Amazon.com started populating
00:14:21
with AI generated review highlights.
00:14:23
Any product that you look at is going to have reviews,
00:14:26
hundreds or even thousands.
00:14:28
What I love about AI review summarization is it gives
00:14:31
you just a couple of quick bullet points, and so
00:14:33
that's helping me make a more informed buying
00:14:36
decision.
00:14:36
And in February, Amazon launched a new gen AI powered
00:14:39
conversational shopping assistant called Rufus, to
00:14:42
further streamline those product recommendations.
00:14:44
Some consumers may be put off by product
00:14:47
recommendations that are based on their purchase
00:14:50
history. Now Amazon may have to, if they don't already
00:14:54
implement, an opt out, some sort of feature where a
00:14:58
consumer could say, you know, please don't look at my
00:15:02
purchase history when giving me recommendations.
00:15:05
Despite concerns, Amazon is committed to injecting
00:15:08
generative AI into every possible step of operations
00:15:10
to boost speed, efficiency and eventually, it hopes,
00:15:13
its bottom line.
00:15:15
If you can reduce packaging, is that a good thing?
00:15:18
Yes. If you can reduce the amount of time and the
00:15:21
amount of miles you have to travel to get from point A
00:15:24
to point B, is that a good thing?
00:15:25
Yes. These are all good ideas, you know, kind of
00:15:28
definitely feel free to show us the receipts at any
00:15:31
time.
