YouTube Cobalt Evolution: From Legacy Engine to Chromium Crobalt

Name: Cobalt 27 - Delivering the Next Generation of Amazing YouTube Devices
Uploaded: 2026-04-29T14:28:49+00:00
Duration: 30 min 17 s
Channel: YouTube Device Partner Summit
Description: Summary and key takeaways on YouTube Cobalt Evolution: From Legacy Engine to Chromium Crobalt, covering The Cobalt Legacy YouTube launched the original web

YouTube Device Partner Summit

Apr 29, 2026

•

30 min video

•

2 min read

YouTube video ID: IDvzUp-0H3s

Source: YouTube video by YouTube Device Partner Summit — Watch original video

PDF

YouTube launched the original web engine in 2013 under the name “Steel” for PlayStation 3. The platform rebranded to Cobalt in 2015 when the Starboard porting layer arrived. Evergreen, introduced in 2019, lets YouTube push fixes to devices in hours rather than days or weeks. Since Evergreen’s debut, partners have received 394 million updates, and the service now supports over 1 billion hours of daily watch time.

The Crobalt Future

Crobalt replaces the bespoke web core with Chromium M138, creating a focused “Cobalt embed layer” that bridges Blink to living‑room requirements. Starboard 18 trims the abstraction layer from 273 functions to 117, relying on standard POSIX APIs and cutting partner integration time from months to days. V8 MagLev integration speeds boot sequences by up to 30 %, while new support for Web Transport, WebGL, and Chrome DevTools gives developers direct TV‑platform debugging.

“The philosophy here is minimize your work, maximize your flexibility.”
“What used to take months of engineering effort now takes a few weeks or even just a few days of integration.”

Accelerating Development with AI

Gemini writes roughly 65 % of POSIX and PLB tests and has performed more than 700 automated pull‑request reviews. AI‑driven compiler optimizations lower browser watch latency by 11 % and shrink binary size by 356 KB. In‑field health‑monitoring agents filter alert noise, pinpoint root causes of crash spikes, and keep web developers and QA teams from “working in the dark.”

“You can think of it [Cobalt Companion] as almost like your own personal cobalt pair programmer.”

Migration and Roadmap

The Cobalt 27 LTS branch lives on GitHub, with an official release candidate expected in the coming months. Early‑access reference ports for Linux and RDK are already available; an AOSP port follows soon. Samsung and LG lead early‑access testing for Crobalt, while Broadcom, Realtek, MediaTek, Novatech, and ML Logic evaluate performance on their SoCs. Cobalt Companion assists partners with debugging, log analysis, and feature implementation throughout the migration.

Mechanisms at Work

Evergreen’s updatability bypasses traditional OTA cycles, delivering fixes directly to in‑field devices. The Crobalt architecture swaps the custom engine for a Chromium stack, using a thin embed layer to satisfy TV‑specific constraints. Starboard 18’s reliance on POSIX APIs reduces custom code and accelerates integration. Cobalt Companion parses logs, detects missing asset directories, and suggests concrete code fixes, further easing partner engineering burdens.

Takeaways

Evergreen has delivered 394 million updates, allowing YouTube fixes to reach devices within hours instead of weeks.
Crobalt swaps the custom web core for Chromium M138, cutting Starboard functions by 56 % and achieving up to 30 % faster boot times.
Gemini now writes about 65 % of POSIX and PLB tests, performs over 700 automated PR reviews, and its compiler tweaks reduce watch latency by 11 % while shaving 356 KB from the binary.
Starboard 18 reduces the abstraction layer to 117 functions, leveraging standard POSIX APIs and shrinking partner integration time from months to days.
Cobalt Companion acts as an AI pair programmer, parsing logs, spotting configuration errors, and suggesting code fixes to lower engineering effort for partners.

Frequently Asked Questions

How does Evergreen updatability shorten the deployment cycle for YouTube fixes?

Evergreen pushes code changes directly to in‑field devices, bypassing traditional OTA update windows. This direct delivery lets fixes appear in hours rather than days or weeks, as demonstrated by the 394 million updates already deployed.

What performance gains does Crobalt provide over the legacy Cobalt engine?

Crobalt adopts Chromium M138 and integrates the V8 MagLev engine, delivering up to 30 % faster boot times and new support for standard web APIs like Web Transport and WebGL. The streamlined Starboard 18 layer also reduces integration overhead, accelerating overall performance.

Who is YouTube Device Partner Summit on YouTube?

YouTube Device Partner Summit is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.

Does this page include the full transcript of the video?

Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.

Helpful resources related to this video

If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.

Ergonomic Mechanical Keyboard For Software Developers Recommended

Provides physical comfort and efficiency for engineers spending long hours writing code and debugging complex Chromium-based architectures.

Amazon →

Ultrawide Monitor For Software Development

Offers increased screen real estate for side-by-side code editing, log analysis, and utilizing Chrome DevTools simultaneously.

Amazon →

Book On Chromium Browser Architecture

Deepens understanding of the underlying engine (Blink/V8) that powers the new Crobalt architecture.

Amazon →

High Speed Network Switch For Lab

Essential for hardware engineers and SoC vendors to manage high-bandwidth testing and data transfer for multiple reference port devices.

Amazon →

Links may be affiliate links. We only include resources that are genuinely relevant to the topic.

Summarize another video

Full Transcript YouTube

You're definitely my favorite host.
Um, so hello everyone. Um, and welcome
again to the 2026 YouTube Device Partner
Summit. We're excited to have you here
and we're delighted to share more about
what we've been working on with Cobalt.
So Cobalt's mission has always been to
bring the very best YouTube experience
to every screen, you know, and today
we're announcing the most significant
technological evolution of Cobalt since
its inception. Cobalt 27 and it's truly
delivering the the next generation of
amazing YouTube devices.
So I'm Lamb Buouie. Um it's nice to meet
you all. Um I'm one of your hosts during
our time here together. I lead the
Cobalt partner engineering team and um I
work to make sure that the best YouTube
experiences um work really well on your
platform. So a lot of you already work
with me and know me well.
We also have my colleague from the
Cobalt team, Andrew Savage, who is our
technical lead for our Cobalt 3P
ecosystem, who will join us a little bit
later.
With that, here's what we have in store
today. So, the Cobalt Legacy, we'll
first take a look back at Cobalt history
to see how far we've come together.
And then next, we'll talk about our
Crobalt future. So, Crobalt is here, and
we're excited to tell you what we've
been cooking up for the past year.
And next is accelerating Crobalt with
AI. Everyone is using AI these days and
Crobalt is no exception.
Then we'll talk a little bit about how
we've been using AI to boost our
velocity and later on how we can help
boost your velocity as well.
And then we'll give you a quick tour of
getting started and migrating to Cobalt
27. It's much easier than you think. And
finally, we'll give a quick sneak peek
at how we're looking to improve quality
of life for you partners.
Okay,
now let's take a step back and take a
look at how far we've come with legacy
Cobalt. This is stuff we've done
together with you partners and it's
something to celebrate our work together
for the last 13 years.
So in 2013, Cobalt was started
originally as support for PlayStation 3.
you're seeing an opportunity to bring
YouTube to the platform via Chromium.
Originally, this was actually called
Steel. Uh most people don't know this.
And then later in 2015, we started to
see some real value in having a web
platform on devices and we rebranded
Steel to Cobalt, which you all know.
This also added the starboard porting
layer. So, this evolved from a large
custom engine to a lean abstraction
layer, simplifying porting. Cobalt to
partner platforms that you all use today
and all familiar with.
So as our growth exploded, it became
increasingly clear that partners had to
bear huge development resource costs and
also infrastructure costs to keep Cobalt
up to date to our requirements in
particular our device maintenance
requirements.
So we developed an updatability feature
called Evergreen. I think you all know
Evergreen. So in 2019 that allowed us to
reduce partner effort and costs while
also enabling us to deploy fixes and in
on to infield issues from the YouTube
side within hours.
This previously might have taken
partners days to weeks via OTAA. So this
was a major win for ecosystem that still
delivers value to our users today and in
the future.
In 2021, we migrated development to
GitHub. So, we adopted the standard
open-source workflow and we really tried
to make it easier for to collaborate
with partners and also develop a much
more frictionless and visible
development model so you can keep track
of what we do and we can also work
better with with you partners.
So, thus far we've openly collaborated
with 45 active partner teams and merging
over 9,500 PRs since Cobalt. This is
such an amazing number to take a look
at.
And now in 2026 with Cobalt, we're
coming full circle and moving closer to
Chrome, transitioning to a modern
Chromium powered engine for the next
generation of modern devices.
So with Cobalt, we've Cobalt, we've had
this mentality of one Cobalt every
platform. And for over a decade,
Cobalt's been the heart of YouTube on
TV. We've grown to power hundreds of
millions of active devices,
various form factors, including TVs, OTT
devices, game consoles, monitors,
projectors, and now even in cars. So,
across thousands of models, Cobalt can
be found just about everywhere you look
in your home.
Again, with our updatability feature,
Evergreen, we've delivered 394 million
updates in to real infield devices.
149 million updates were since this time
in 2025. So, this is truly a staggering
number given that we started in 2019 for
updatability for us.
So again, Evergreen has been one of our
most impactful features for users and
partners, saving partners significant
development and deployment costs to the
infrastructure and also testing.
So all this while we can deploy fixes
and optimizations and new features in
the most efficient way possible to our
mutual users.
So finally looking at the big picture,
we've hit over 1 billion hours of daily
watch time. It's just unbelievable the
scale that we're at here with Cobalt. 1
billion is is insane number. I never
would have thought we'd we'd hit that at
least so soon.
And you know what? The message is clear.
The world loves YouTube and watching
YouTube on the amazing devices that
we've built together.
So, I also want to take a moment to
highlight all the amazing technology
we've introduced to the world with
Cobalt and with you partners. We've
introduced new codecs such as VP9, AV1,
uh, Eclipse audio, which is AMF, and of
course, our bread and butter
experiences, YouTube, YouTube TV,
YouTube Kids, Music, and our most recent
product, NFL Sunday Ticket, which has
been an incredible success, no doubt due
to our work together.
So, again, I'd like to take a moment to
honor the effort and the partnership
that we've had together. Your feedback,
deep collaboration, and commitment to
high quality are the foundation of
everything we've built so far together.
So, I just want to take a moment to
thank you all. So, appreciate it.
All right, with that, I'll hand it off
to Andrew to tell you a little bit more
about Crobalt.
Hi everyone. As Lamari mentioned, I'm
Andrew. Uh, I lead the Cobalt third
party team. Um, all right. Let's
transition into what's new. We just
spent some time reflecting on the
incredible legacy we've built together
over the last decade. But as we all
know, the living room is evolving faster
than ever. Users are demanding richer,
heavier, and more dynamic applications,
and standardizing our tech stack is the
only way we keep up without burning out
our shared engineering teams. So today,
we'll be talking about our latest
project
to bring Cobalt into the modern world.
We call this project Crobalt.
So, what is Crob anyways? It's a fun
name, but the implications behind it are
massive.
For years, we've operated somewhat
parallel to the broader web ecosystem,
maintaining our own bespoke web core to
meet the unique constraints of living
room devices.
But the web ecosystem has evolved
quickly, and there are many capabilities
of the web and your TVs that we're not
taking advantage of. It's time we do.
Crobalt represents our move to unify the
YouTube on TV experience with the
bleeding edge of the modern web.
Specifically, Crobalt is our new next
generation chromiumbased architecture.
And Cobalt 27 is the first broadly
available version of Cobalt to be be
fully powered by this new Crobalt
architecture right out of the box.
This means we are stepping away from a
heavily customized engine and adopting
the exact same Chromium foundation that
powers the majority of the global web.
To really appreciate the scale of this
upgrade, let's look at where we are
today with the Cobalt stack.
Currently, the YouTube web app, all of
the front-end code sits on top of our
custom Cobalt engine.
Underneath that bespoke web core is the
Starboard layer.
Over the years, Starboard has grown into
a very large complex abstraction layer.
Every time we needed a new feature, we
had to build custom pipes through the
cobalt core layer and starboard. It got
heavy. It took a lot of your time to
port and a lot of our time to maintain.
Now look at the nextG cobalt stack. We
are replacing that bespoke engine with
Chromium M138.
This instantly brings industry standard
performance and reliability directly to
your devices.
We bridge the powerful Blink engine to
YouTube's specific living room
requirements using a focused Cobalt
embed layer and add small customizations
for H5 VCC and media. And at the
foundation, a radically simplified PZIX
aligned starboard. We stripped out the
complexity.
This modernized stack unlocks massive
potential for the user experience.
We're talking about next level graphics,
raw computing power, and full support
for nextgen APIs.
By moving to the V8 Mag Lev JavaScript
engine, we are seeing up to a 30% faster
uh boot time that translates directly to
faster YouTube app boot times and
incredibly fluid UI navigation.
Plus, we are bringing in standardized
network APIs like web transport for low
latency, high performance streaming.
We are adopting standard web APIs which
means less bespoke maintenance for
everyone. Here
you can see our YouTube playable demo on
the TV.
Many of the playables use WebGL under
the hood and instead of needing to
heavily edit them to have them come to
TVs, we're now able to run many of them
out of the box. You can check these out
and play the games yourself at our demo
booth. Please give it a try.
Pivoting back to Starboard, let's talk
about the new foundation because this is
where your teams are going to feel the
most relief. We heard your feedback.
Integrations were taking too long and
the cost of maintenance was taxing your
resources and bottom line.
We've spent the last few years working
towards improving the Cobalt partner
experience. Starboard 18 is simpler, 56%
simpler to be exact. from 273 functions
in starboard 15 to 117 functions in
starboard 18. We've made significant
cuts and expect to keep cutting it down.
You now have to implement far fewer
custom starboard APIs because we are
leaning heavily on the standard PZIX
APIs already built into your systems.
This inherently stabilizes Starboard.
If you need custom integrations, you can
leverage our Starward extensions or
platform services, giving you more
flexibility than ever without bloating
the core. The philosophy here is
minimize your work, maximize your
flexibility.
Because of the drastically simplified
starboard layer and our reliance on pix,
your integration timelines are going to
shrink.
We wanted to make this the easiest
integration you've ever performed.
The results from our early access
program exceeded even our expectations.
Anecdotal evidence from our top partners
shows that one of the G once the GN
build environment is configured, the
rest is remarkably simple.
The work we've done to simplify
Starboard slashing the size of the API
is really paying off.
One partner took their legacy Star Wars
16 implementation from Cobalt 25 and
simply copied it into Crobalt to get a
functional port.
After removing a few deprecated
functions and minor bug fixing, Crobalt
more or less just worked.
What used to take months of engineering
effort now takes a few weeks or even
just a few days of integration.
This is an incredible outcome and is
part of what we have been working
together with all of you partners on.
Less time on integration infrastructure,
more on innovation and cobalt features
to ensure our mutual users around the
world get the absolute best YouTube
experience.
And to ensure both your and our
developers can develop faster every step
of the way, we are fundamentally
changing the de debugging experience.
With Crowba, we are bringing modern
tools directly to the TV platform. Our
web developers and QA teams will no
longer be working in the dark. They can
use the exact same standard Chrome dev
tools they use on desktop to inspect
network requests, analyze the DOM,
profile performance, and trace
JavaScript right on the TV. It's a night
and day difference for troubleshooting.
Our developers have already had a great
experience with this. Some of my
co-workers helping show off the demos
today and tomorrow shared their
experience using the new tools in two
completely different cases.
One was telling me about the memory
infra tool that allows us to get fine
grain insights into me memory usage and
to let us optimize performance and
memory usage in precise ways and quickly
validate those changes.
In a different vein, another co-orker
told me about her experience developing
the demo page for the YouTube playables
that we're demoing today.
She heavily utilized the tools to debug
the JavaScript code directly on the TVs
it ran on and was able to quickly
iterate by changing front-end code on
the TV itself instead of falling back to
a time-consuming dev environment
deployment workflow.
Another great tool highlighted in this
slide shows us where repaints are
happening.
This gives us immediate insight into
what pieces of the app are performing
well and what's taking expensive cycles
when it's not the focus of the page.
We're already using this tool to make
the app as efficient as possible and
expect there to be long and lasting
benefits from having these tools
available.
We haven't just added dev tools though.
With Cobalt 27 comes a much more
comprehensive usage of UMA metrics.
UMA reports data about how Cobalt is
running in field back to us, giving us
valuable infield metrics on real world
performance and stability.
This will help us catch and address
issues experienced by users more
effectively.
So, we have a unified Chromium
architecture, a drastically simplified
porting layer in Starboard 18, and
worldclass debugging tools. But what if
we could make the integration and
certification process even faster?
That brings us to chapter 3, Crobalt
Accelerated by AI.
I talked a lot about simplifying our
starboard porting layer and relying
heavily on standard pix functions
instead.
We want to make sure that these
functions are working well across all
our platforms, but there are tons of
these functions and no good open- source
test suite for us to use. So, we had to
write many ourselves.
Because the PZIX APIs are well
specified, we were able to use Gemini to
write around 65% of our PZIX and PLB
tests, it can read the spec, write the
test, build the code, run the test, and
continue to iterate until it's happy it
has met the requirements, leaving it to
us to give a final review before
submitting the code.
Another massive timesaver has been the
automated code reviews Gemini can
perform on our GitHub repository.
Gemini has performed over 700 of these
reviews on our PRs, catching anything
from minor style fixes to complex race
conditions that need to be addressed.
Using this as a way to ensure your code
is in the best state possible before
sending it to human reviewers has eased
a ton of burden for both the developers
and the reviewers.
Gemini also helped us evaluate compiler
and tool chain configurations that
optimized loop unrolling for TV cache
sizes.
This allowed us to reduce brows watch
latency by 11%. And binary size by 356
kilobytes on our reference platforms.
This was a great win that took hours to
verify versus days of or upwards of
weeks that it would have previously
taken for a human to identify and
verify.
Let's look at a specific situation in
which we use Gemini to great effect.
The new tools I talked about give us a
lot of helpful information, but I need
to emphasize that it is a lot of
information.
Parsing through it can be difficult,
especially when we're navigating a
codebase as large as Chromiums, and it
takes a human a ton of time.
This is where Gemini can help immensely.
It is able to quickly and precisely
evaluate massive amounts of information.
for example, reading through all of the
memory allocations happening in Cobalt
and produce meaningful insights into
that information.
What I'm showing here is Cobalt's memory
usage on Android TV while running
YouTube. On the left, we can see the
memory is slowly growing over time,
indicating a memory leak.
It's easy to identify that there is a
memory leak, but it can be very hard to
identify where that leak is coming from
when there are so many different threads
requesting resources.
We can look through the allocations
ourselves and use some tools to get
insights into where the leak may be
coming from, but even better than that
is feeding all that information into
Gemini and having it flag the
allocations that are potentially
problematic.
In this case, it helped us find and fix
some significant leaks. Sorry about
that.
Uh, in this case, it helped us find and
fix some significant leaks in just a
couple of days. And we were able to
stabilize memory usage.
Another place A has been able to help is
by giving insights into infield health
metrics.
Beyond just filtering alert noise, our
AI tools can help pinpoint potential
root causes for complex stability issues
much faster, sometimes surfacing subtle
issues unique to specific hardware
platforms and rollouts.
This means accelerated debugging and
resolution of infield crashes and
stability problems, leading to more
robust devices and reduced engineering
effort for partners.
In this example, there's a real alert
triggered on a crash spike at 9:00 a.m.
by our monitoring system.
The alert system dispatches the
investigation to our intelligent AI
health monitoring agent before sharing
databacked analysis with our on call
owner.
The AI agent made a quick determination
with a low volume of app starts, 131 in
this case, showing five crashes. the
results are likely skewed and a false
alert. The conclusion is to continue
monitoring as more data is available and
after 1 hour we see the baseline levels
return to the expected norm.
While this is what we expect from alert
monitoring, the analysis was largely a
human task in the past that takes time
to review and process. Adding AI into
our work streams has been meaningful in
focusing on what's important and not
worrying about what's not.
With all the talk about new features,
I'm sure you want to know our plan for
releasing Cobalt 27. I'll let Lamb
discuss the details of where Cobalt 27
is today and how you'll get started with
it. Thank you.
>> All right. Thanks, Andrew, for that tour
of Halt 27. Super excited to get this
out in the field. All right. So, now
that you know a little bit more about
Cobalt 27, I'd like to give you a quick
tour about getting started and migrating
your platform to Cobalt 27.
So, taking a look at Cobalt 27
branching, I just want to point out a
few changes here. So, there's two
branches that you need to consider.
There's the main branch and there's a 27
LTS branch. Um, these are already
available on GitHub.
U main is our development branch. 27 LTS
is stable and this is what you should
use for your development and
certification.
So we already have this 27 LTS working
branch development GitHub. So you can
start integrating this even today. So we
plan to announce an official 27 LTS RSC
tag in the coming months and this is
when you can largely consider the
starboard to be stable. Um but I would
even encourage you to start even earlier
and provide feedback for us.
Following RC 27 LTS stable will come
later this year which you can use for 27
uh certification.
We also have migration guides available
in GitHub under starboard doc which is
highlighted right here. They're super
short. It's super easy. Um
would love to see get your feedback on
how how your integration goes. But it's
been great seeing the simple
integrations from other partners.
And as Andrew mentioned, some partners
have already started integrating and
sharing feedback with us. They want we
want to resolve any issues before we
release later this year for stable.
So the sooner you let us know about
issues on your platform, the more likely
it'll get fixed and optimized before 27
LTS stable. So please get started early.
I think I've said that enough times.
Um, so in addition to the getting
started guide, I'm also excited to
announce that we also have working ports
already available for you to try and use
as a reference. So we have reference
ports available for Linux for our RDK
reference device and also AOSP is coming
soon as well. So stay tuned for that.
So again, Crobalt's definitely here and
we're delighted to be with the strong
interest that we've seen since last
year. With all these exciting changes
coming from Cobalt 27, many partners
have already been working very closely
with us as early adopters to ensure
Cobalt 27 is certification ready.
And so I just wanted to shout out a
little bit here. So shout out, you know,
to Samsung and LG for being the first
Crobalt early access partners. delighted
to be working with our SSC vendors as
well, Broadcom, Realtech, MediaTek,
Novatech and ML Logic to start testing
and validating uh Crobalt on their
platforms and also Calabra and Negalia
who've been working with to bring our
RDK and ASP reference ports to life.
All right, that's it for the migration
integration. So, um on to the next
thing. AI is everywhere. You've seen it
all day today.
There's these new features announced
today such as AI ask search and you also
saw YouTube AI solutions from other
teams that help partners with
certification and testing.
Andrew touched on how AI is transforming
how we do cobalt development and
debugging today as well. So what's next
for the future of cobalt development for
partners?
So with these incredible advancements
and impact on accelerating our
workflows, we also want partners to
leverage our Cobalt AI tools to help
make your integrations better optimize,
perform better, and also pass
certification faster.
So what I'd like to introduce to you
today is something we've been working on
to help you all Cobalt developers be
more productive with the help of AI.
There's something that we're referring
to as Cobalt Companion.
So, let's walk through a real Cobalt
companion scenario that we've actually
encountered and we've leveraged Cobalt
Companion to help us to fix a real
launch crash.
You'll see here the Cobalt developer
working on integrating Cobalt 27 on
their next generation device. However,
she's running into a crash issue. I'm
sure some of you have seen something
like this before.
So previously this could require a lot
of deep investigation and a lot of
background context to understand the
codebase. It may not be obvious how to
solve this especially if it's a new team
member as well. However, Cobalt
Companion can really help out here.
So the developer says Cobalt Companion
Cobalt 27 is crashing with the sigabort.
Can you read this Cobalt log and help me
fix this?
In that time about 10-15 seconds, Cobalt
Companion will analyze the Cobalt logs
and if necessary the total Cobalt uh the
local Cobalt integration as well.
And it came up with a thoughtful
recommendation for our Cobalt developer
here.
And Cobalt Companion responds, "No
problem. It looks like you didn't
configure your evergreen content
directory correctly, causing assets to
fail to load. Would you like me to fix
that for you?" Yeah, of course.
So this is really cool just to be able
to also identify the problem but also
offer to fix it. So not only can um can
do that but also saves a lot of precious
time by being able to do that. YouTube
has been heavily leveraging this ability
across all facets of of our cross
functional teams and we want to share it
with you to improve your integrations.
So another scenario here, Andrew touched
on a bit earlier was addressing a YTS
performance issue.
In this case, she asked Cobalt Companion
in plain language.
Can you help me run, debug, and optimize
my Cobalt 27 integration for the
automated YTS certification test 2?
Or can you also tell me what new YouTube
features I'll get by updating to Cobalt
27 versus Cobalt 25?
Of course, there's so much cobalt
companion can do and it'll only grow
more capable over time.
So, these are actually real example use
cases. And what I want to highlight here
are the skills and abilities of the
Cobalt Companion will actually grow over
time.
Cobalt companion will stay up to date
with the latest debugging, testing, and
integration techniques developed by our
teams. So, your teams will be
immediately be able to tap into them.
The companion can also mentor and guide
your teams on implementing new features
and to learn new techniques as well. So
you can think of it as almost like your
own personal cobalt pair programmer to
help you with integration.
And for partners, you can also
contribute to help improve its
capabilities.
cobalt's largely open source and some of
this much of this will be all in open
source and we even customize um and you
can even customize local functionality
to match your platform needs and
requirements so it can be truly
personalized for your team's needs.
So again there's so much opportunity and
potential in the age of AI and we really
want to work together with you all to
explore what we can do together. While
some of these cases may be simple, some
are quite complex, but pursuing
solutions have never been as simple as
it is today.
With that being said, so if Copa
Companion is something that you'd be
interested in piloting or sharing
feedback or even collaborating on, we'd
love to hear from you. Come talk to us,
share us your thoughts and ideas and how
else we can um support you.
And to close, I just want to say a
massive thank you to all you partners.
You are the reason YouTube has brought
so much joy to all these users around
the world. We're incredibly excited to
continue this work with you and we can't
wait to see the incredible experiences
we'll build together on the next
generation of Cobalt powered devices.
Thank you.

Help & FAQ

YouTube Living Room : 2027 SW and 2028 HW Technical Requirements Overview

YouTube Device Partner Summit

May 09, 2026

Watch Read Summary

2026 Product Innovations

YouTube Device Partner Summit

May 09, 2026

Watch Read Summary

Airport Karen Thinks She’s Too Important to Follow the Rules

ArrestFlix

May 15, 2026

Watch Read Summary

Late Passenger Tries to Force His Way Through TSA Checkpoint

NYDETECTIVE

May 14, 2026

Watch Read Summary

My phone stole my life. This is how I got it back.

Sean McRae

May 18, 2026

Watch Read Summary

Airport Passenger Arrested After Attempting to Board Flight Without Ticket | Police Bodycam Analysis

The RWP Dossier

May 13, 2026

Watch Read Summary

Liver Lobules (Portal Triad)

Dr Matt & Dr Mike

May 14, 2026

Watch Read Summary

Popular Summaries Today

负债1000万走投无路，赴美仅两年开工厂！他的逆袭到底靠什么？｜从睡机场到开工厂｜众叛亲离的绝望｜前“股神”的美国救赎｜

Jack的世界

May 19, 2026

Watch Read Summary

The price of iron ore - Relocating an entire city in Sweden | DW Documentary

DW Documentary

May 18, 2026

Watch Read Summary

Lecture 11: One-Shot Deviation Principle and Bargaining

MIT OpenCourseWare

May 18, 2026

Watch Read Summary

Lecture 13: Infinitely Repeated Games

MIT OpenCourseWare

May 18, 2026

Watch Read Summary

Lecture 6: Imperfect Competition

MIT OpenCourseWare

May 18, 2026

Watch Read Summary

PDF

The Crobalt Future

Accelerating Development with AI

Migration and Roadmap

Mechanisms at Work

Takeaways

Frequently Asked Questions

How does Evergreen updatability shorten the deployment cycle for YouTube fixes?

What performance gains does Crobalt provide over the legacy Cobalt engine?

Who is YouTube Device Partner Summit on YouTube?

Does this page include the full transcript of the video?

Helpful resources related to this video

Share This Summary

Embed This Summary