Nvidia’s Open‑Source LTX2 Text‑to‑Video Breakthrough, New Supercomputer, AI Health Tools, and Industry Shifts

Name: AI News: NVIDIA Rubin, LTX-2 Open Source, ChatGPT Health, GPU Prices and More!
Uploaded: 2026-01-09T05:34:08.143668+00:00
Channel: Matthew Berman
Description: Summary and key takeaways on Nvidia’s Open‑Source LTX2 Text‑to‑Video Breakthrough, New Supercomputer, AI Health Tools, and Industry Shifts, covering

Matthew Berman

Jan 09, 2026

•

3 min read

YouTube video ID: HLhfepgHLcg

Source: YouTube video by Matthew Berman — Watch original video

PDF

Open‑Source LTX2 Text‑to‑Video Model

Full release: Model weights (full and distilled), Laura adapters, and a modular training framework called Trainer are now publicly available on GitHub and HuggingFace.
Hardware focus: Optimized for Nvidia GPUs, especially the RTX 5090 (though prices may rise soon).
Capabilities:
Generate up to 20‑second clips with control over the first and last frames, enabling storyboard‑style stitching.
4K resolution at up to 50 fps, high‑quality lip‑sync, multi‑keyframe conditioning, 3D camera logic, and multimodal inputs (text, video, audio, etc.).
Use cases: Local inference, fine‑tuning, production deployment, or cloud‑based execution.
Recognition: Holds the #1 spot on the Artificial Analysis Open‑Weights leaderboard.

Nvidia’s "Reuben" Supercomputer (CES Announcement)

Next‑gen architecture designed for hyperscalers and AI model providers (OpenAI, Anthropic, AWS).
Performance & cost: Claims up to 10× inference cost reduction and 4× fewer GPUs needed for mixture‑of‑experts training compared to the previous Blackwell platform.
Cooling innovation: Uses 45 °C hot water for ambient cooling—no chillers required, dramatically improving power efficiency.
Scale: Each rack houses 72 Reubin units; a full system contains 1,152 GPUs across 16 racks.
Timeline: Expected to be operational at scale in about nine months, with production already underway for Nvidia partners.

Sponsor Spotlight: Grapile AI Code Reviewer

Purpose: Automatic AI‑driven code review, catching bugs and issues that humans miss.
Availability: Free for open‑source projects; offers a 14‑day free trial for other users.
Adoption: Used by major repos such as Nvidia, Posthog, and Storybook.

OpenAI’s ChatGPT Health Feature

Function: Allows users to feed personal health data (Apple Health, Whoop, Oura, medical records) into ChatGPT for proactive health recommendations.
Privacy: Health data is kept separate from model training; a dedicated UI area reduces accidental sharing.
Rollout: Available to all subscription tiers except Europe (due to stricter regulations). A wait‑list is open.

GPU Market Alert: Potential RTX 5090 Price Surge

Rumors: Reports suggest RTX 5090 prices could jump from $2,000 to $5,000 because of memory shortages.
Official stance: Neither Nvidia nor AMD have confirmed the hike, but industry chatter and a CES Q&A with Jensen Huang indicate they are exploring production tweaks (e.g., using older‑gen GPUs or lower‑DRAM variants) to mitigate the issue.
Advice: If you need a high‑end GPU now, consider purchasing before any price increase takes effect.

Nvidia Alpameo: Open‑Source Autonomous‑Vehicle Stack

What it is: A full‑stack, vision‑language‑action (VLA) model for autonomous driving, trained on only 1,700 hours of synthetic video data.
Open source: Models, simulation tools (AlpaSIM), and datasets are freely downloadable.
Key features:
End‑to‑end video‑in, actuation‑out pipeline (steering, acceleration, braking).
Vision‑only approach similar to Tesla, but can also incorporate LIDAR/radar.
Reasoning capabilities beyond pure perception.
Partnerships: Demonstrated with a Mercedes‑equipped camera‑only vehicle navigating San Francisco.
Impact: Provides a low‑cost entry point for car manufacturers to develop autonomous tech, though real‑world adoption may be limited by legacy industry inertia.

Anthropic Funding Round

Scale: $10 billion raised at a $350 billion valuation, positioning Anthropic at roughly half the valuation of OpenAI.
Financial health: Near profitability with strong revenue streams.
Backers: Continued investment from Nvidia and Microsoft, reflecting a "bet on everything" strategy across competing AI firms.

Other Highlights

The creator hints at future tutorials (e.g., integrating LTX2 with Comfy UI).
A reminder to like and subscribe if the content was helpful.

Nvidia’s aggressive open‑source releases—from the powerful LTX2 text‑to‑video model to the full autonomous‑driving stack—combined with groundbreaking hardware like the Reuben supercomputer, signal a rapid democratization of AI capabilities, even as GPU pricing pressures and data‑privacy concerns shape how developers and enterprises will adopt these tools.

Frequently Asked Questions

Who is Matthew Berman on YouTube?

Matthew Berman 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.

Summarize another video

Full Transcript YouTube

We now have a fully open-source open
weights texttovideo model. This is LTX2
and now it is completely open. You can
download and run it on your local
machine. Here's the announcement. The
release makes a full LTX2 development
stack publicly available. You get the
model weights, both the main model and
the distilled version, enabling local
inference, fine-tuning, and production
use across a range of hardware, which
means you can download it and you can
customize it for your own needs. They
are also offering Laura adapters, so you
get to see how they've customized the
model for specific styles and tasks, and
that might give you inspiration for how
you might do it. They're also providing
something called trainer, a modular
training framework that includes
multimodal data pipelines, reference
configurations, multi-GPU support and
evaluation tools. They are basically
giving you the full stack to do anything
you want with this texttovideo model.
Really the first time we're ever getting
this much support from a company. LTX2
is optimized for Nvidia machines and
specifically it can run well on an
RTX5090,
but those chips might be getting pretty
expensive very soon. More on that later.
A few more things about this model. You
can get up to 20 second generations, but
you have control over the first and last
frame. So, you can tie together these 20
second clips into however long you want.
The special thing about LTX in
particular is that it gives you kind of
this storyboarding control as if you
were sitting inside of a movie studio.
With these 22nd generations, they are 4K
and up to 50 frames per second. So very
high quality. It also has very accurate
lip sync. So obviously it comes with
audio and you get incredibly fine grain
control over everything including multi
keyframe conditioning, Laura based
customization, 3D camera logic and
multimodal inputs including text, video,
audio and more. So download the model,
download the code, download the pipeline
all from GitHub, from HuggingFace, run
it locally. Of course, if you don't want
to run it locally, you can use a cloud
provider. And all of this is listed in
this announcement. I'll link that down
below. And LTX2 has already taken the
number one spot on the artificial
analysis open weights leaderboard. This
is the number one video model that is
also open source open weights. Go check
it out. It's super cool. I'm thinking
about making more videos, maybe a
tutorial about how to plug it into Comfy
UI. Let me know if you want to see that
in the comments. And they even released
a research paper detailing all of the
techniques that they use to create this
model. Here's a few demos of what you
can actually get out of LTX. Look at the
quality. Look at the detail. Look at the
realism of the physics, the different
styles that you can use. All of this is
now available to you for free. And next,
I just got back from CES. I attended
Jensen's keynote where he announced
multiple new products including this
absolute beast of a supercomput called
Reuben. This is their next generation
architecture. This is not consumer grade
hardware. This is what the hyperscalers
and model providers like OpenAI and
Enthropic and AWS. This is what they're
going to be buying. Rubin delivers up to
a 10x cost reduction in inference and a
4x reduction in the number of GPUs
required to train a mixture of experts
model. This is all compared to the
Blackwell platform, their current
Frontier supercomputer. It's also much
more power efficient and something that
kind of blew my mind, it doesn't even
require water coolers. It's basically
cooling the system with hot water. And
if that sounds weird, that is literally
what Jensen said. Look at this.
>> Very importantly, the water that goes
into it is the same temperature, 45° C.
With 45° C, no water chillers are
necessary for data centers. We're
basically cooling this supercomput with
hot water. Is so incredibly efficient.
Each of these has 1,152
GPUs, which is just insane, and 16 racks
to power it all. Each rack has 72
Reubins. Elon Musk commented on this
expost. It will take another 9 months or
so before the hardware is operational at
scale, and the software works well. That
might be true, but it's already in
production. And they've greatly
simplified this machine. No cables, no
hoses, no fans. It is as close to being
cooled ambiently as you can be. And as I
mentioned, it's already in full
production and should be available later
this year to Nvidia's partners. And of
course, one of the biggest use cases for
all of this AI power is coding, which is
why I'm excited to tell you about
today's sponsor, Grapile. AI has gotten
pretty good at writing code, but what
you may not have realized is that with
the right tools, it can be really good
at reviewing code, too. That's where
Grapile comes in. Gretile is AI that
reviews all of your code automatically.
And the GPile team has spent a lot of
time on this reviewing code problem. And
they showed me some incredible bugs and
errors and issues that Greile was able
to find that humans overlooked. In fact,
they have an entire page dedicated to
showing off some of those crazy
findings. And Gretile is completely free
for open source and has some massive
repos using it today, including Nvidia,
Postthog, Storybook, and more. So check
out Graptile. They're offering a 14-day
free trial right now. I'll drop a link
down below. Check them out to review all
of your code automatically, adding
another layer of confidence to your
deployment cycle. Now, back to the
video. And next, OpenAI is launching
ChatGpt Health, a major feature that, at
least for me, is very welcome. ChachiPT
health is a dedicated experience that
brings your health information into
CHACPT so that it could proactively give
you health recommendations. So you can
plug in things like your Whoop device if
you have that Oro ring Apple health and
all of this. You plug it into Chad GPT
and it will have all of that information
for whenever you ask it health related
questions but also if you want it to
proactively give you recommendations
about your health. You can connect your
medical records, wellness apps,
anything. I already use ChachiPT to help
me understand test results from my
doctor, scans, different things about my
health. I'm asking ChachiPT first. And
of course, I'm going to a human doctor,
but I want to arrive at the doctor's
office better informed. And so now
giving it access to my health data seems
like a natural next step. Now, I know a
lot of you are probably not so willing
or excited to share all of your health
data with OpenAI. And I kind of
understand that. And this is actually
where I think Apple might be just once
again getting in their own way. For
folks who use the Apple ecosystem, they
already have all of your health data
built into their Apple Health app, which
people seem to be very comfortable
sharing with Apple. And they are very
big on local artificial intelligence. So
AI running on your phone, on your
laptop. And so you can imagine a world
in which you don't actually have to
share your health data with a third
party like Chad GPT and Apple could use
its AI to give you those proactive
recommendations to answer your questions
about your health data without it ever
leaving your phone. But of course
they're not doing that at least not yet.
And so that's where chatbt comes in. So
there is a wait list for this feature.
It is available to essentially all the
plans. the free, the go, the plus, the
pro plans, all of them have it as soon
as it goes public with one big exception
and that's Europe. Of course, Europe has
much higher regulation and thus they're
not going to have this available, at
least not on day one. So, this is going
to operate as a completely separate area
in Chat GPT. And the thinking is to
increase privacy and security so that if
you happen to share a chat or you're
showing off your screen like I do all
the time, you won't accidentally share
your health information. And of course,
your health information is not used to
train the chat GBT models. So go sign up
for the weight list and hopefully we get
it soon. I haven't actually played
around with it. I just pinged my contact
at OpenAI. Hopefully I get access and
I'll show it off to you when I can. All
right. I mentioned that GPU prices might
be going up and yeah, they actually
might be. And it's all because of the
lack of inventory of memory, one of the
core components of GPUs. We might be
paying up to $5,000 for the latest
RTX5090.
According to Newswire, Nvidia and AMD to
significantly increase GPU prices
starting next month. RTX 5090 increasing
from 2,000 to 5,000. However, there is a
community note on this post. Reports
suggest GPU prices might rise in early
2026 due to memory shortages and some
speculate high-end cards could fetch
$5,000, but neither Nvidia nor AMD has
officially confirmed these price hikes.
But the rumors are happening and
somebody just asked Jensen about it
during the live Q&A at CES. Tom's
Hardware in response to these potential
price increases of the latest generation
GPUs asked Jensen specifically, do you
think that maybe spinning up production
on some of the older generation GPUs on
older process nodes where there might be
more available production capacity would
help with that? Or maybe also increasing
the supply of GPUs with lower amounts of
DRAM. Are there steps that could be
taken or any specific color you can give
on that? And Jensen said, "Yeah,
possibly. And we could possibly,
depending on which generation, we could
also bring the latest generation AI
technology to the previous generation
GPUs, and that will require a fair
amount of engineering, but it's also
within the realm of possibility. I'll go
back and take a look at this. It's a
good idea." So, price increases
seemingly are coming. There's a few
different solutions that Nvidia is
exploring, but the entire industry is
feeling the pinch of memory shortages.
So, if you don't already have a 5090 and
you want one, now is probably a good
time to get it. And next, in what I
thought was probably the coolest
announcement of all the announcements
that Nvidia made, Alpa Mayo, their model
and really full stack architecture for
autonomous vehicles. Now, why is this so
special? Well, first of all, it's just
really cool. Autonomous vehicles is so
futuristic. It is so cool. And Tesla and
Whimo work incredibly well. In fact, my
friend drove me in a Tesla from my house
to the airport and with one button drove
us the entire way. No disengagements. It
was truly impressive. Now, back to
Alpameo. And so, how did Tesla and how
did Whimo get so good? Well, they had
millions and millions of miles of
training data to train their models on.
But that's where Alpameo really shines.
Nvidia is not a car company. They don't
have millions of miles of data to train
their models on. So what did they do?
They used synthetic data. They generated
synthetic video clips of a car driving
around different cities and used that
synthetic data to train the models. So
they didn't need millions and millions
of real world miles. They simply use
synthetic data. Now what also makes this
special? It is completely open source.
The models you can download. You can
play around with them. the entire stack
you can download. And basically, if you
own a car company, you can implement it
yourself. And it's also a reasoning
model. So, not only is it a vision
model, but it is a vision reasoning
model, which seems to be different from
Whimo and Tesla, but nothing confirmed
because those are completely closed
systems. According to the announcement,
Nvidia is the first to release an open
reasoning VLA, I think vision, language
action model designed to tackle longtail
autonomous driving challenges. This
Nvidia Alpameo family also includes
simulation tools and data sets for AV
development. It includes Alpameo 1, Alpa
SIM and physical AI open data sets
enable the development of vehicles that
perceive, reason, and act with
human-like judgment. And they actually
showed off their partnership with
Mercedes. And a Mercedes vehicle with
cameras only driving around San
Francisco quite successfully. And so
what Jensen said is Alpameo is an end
toend AI architecture. Video in,
actuation out, meaning it looks at the
world and then controls the steering
wheel and controls the acceleration and
the braking. It is so impressive and it
was trained using only 1,700 hours of
driving data, which is really not a lot,
but apparently it's really good. So, any
car company now can catch up to Tesla's
drastic lead in autonomous driving, but
I don't actually think they're going to.
Some of these legacy auto manufacturers
just cannot get out of their own way.
And so, even though all of these tools
are essentially free, obviously Nvidia
wants you to buy their GPUs to power it
all in the vehicle, I just don't think
they're going to be able to implement it
successfully. And it's also interesting
when you look at Tesla versus Whimo.
Whimo uses LAR, radar, all of these
different sensors which are quite
expensive. And Tesla took a completely
stripped down approach. It is vision
only. It is cameras on the road making
decisions. And it seems like the Alpameo
system is more akin to that approach,
although I've been hearing it also can
use LAR. And next, breaking. At the time
of recording this video, Anthropic has
raised a new massive funding round.
According to the Wall Street Journal,
Anthropic is raising $10 billion at a
$350 billion valuation. Now, for
context, Open AAI is at I think about a
$750 billion valuation. So, Anthropic is
coming in at just about half, but
Anthropic is nearly profitable. They are
generating a great amount of revenue.
And so, this is a continuation of
Anthropic success. This is in addition
to the 15 billion that Nvidia and
Microsoft are investing in Anthropic.
And if that sounds weird because
Microsoft is also invested in OpenAI,
well, so is Nvidia. They are just on a
bet on everything strategy. So that's it
for today. If you enjoyed this video,
please consider giving a like and
subscribe.