0:00
2027年，Crusoe将推出Spark Liquid，这是我们的第六代模块化数据中心，为客户提供NVIDIA最先进的IT硬件和架构，包括NVL72液冷GPU的接入。
0:16
- 模块化数据中心的优势在于帮助我们提升获取令牌的速度。
0:20
它让我们能将大量现场作业转移至工厂的受控环境中完成。
0:26
- 你可以把它看作一个集装箱，内装GPU，由专门从事AI的人设计，仅专注于AI。
0:33
没有中间商参与。
0:35
端到端全流程由Crusoe管理，我们称之为垂直整合。
0:41
我们获得了市场推出时间和合理的成本结构，这正是支撑我们一些最先进云服务的基石。
0:50
- 这些产品全部由Crusoe制造，我们提供的零部件基本在装配线上组装完成。
1:01
这就像是在编排一支舞蹈。
1:03
一名工人在这里完成特定工作。
1:06
另一名工人准备下一块面板，我们的工程师团队则测量周期时间，并向设计团队反馈。
1:16
- 最终呈现出真正神奇的成果，这背后是数十年的经验积累，我们从大师级电工到学徒项目均在招募。
1:28
- 想想设计一个托管AI服务时，需求能被负责场地准备和模块数据中心建设的人完美理解，这真是令人惊叹。
1:41
- 挑战之一是如此小空间内的巨大功率。
1:45
Spark Liquid帮助我们排出热量，将Spark模块内GPU密度提升约5倍，朝着可持续方向发展，不需使用水冷，极大减少了环境足迹。
2:00
- Crystal Spark Liquid冷却系统专为迎接NVIDIA Verra Ruby和NVL 72设计，带来2027年你能想到的最高效设计。
2:11
- 所以我们已经为下一代芯片设计和工程做好准备。
2:15
离岸制造是做不到的。
2:17
这需要灵活应变的能力。
2:20
- 当我们在操作中节省几秒，周期时间表现就会提升，因为我们的目标是每年从工厂生产出数千套设备。
2:33
- 这种控制力和可预测性是我们成为AI行业可靠合作伙伴的关键。
2:39
- 我们在模块化计算领域领先市场，这代表着一场前所未有的工业革命。
2:45
- 有巨大机遇，让美国重新成为曾经的工业强国。
2:53
- 速度优先，质量保证。
2:55
你知道，很多团队会告诉你两者可以兼得，而我们确实做到了。
0:00
In 2027, Crusoe is launching Spark Liquid,
0:04
our sixth generation of modular data centers,
0:06
to give our customers access to the cutting edge
0:09
of NVIDIA's IT hardware and architectures,
0:12
including the NVL72 Liquid Cool GPUs.
0:16
- So the advantage of modular data centers
0:18
is it helps us improve speed to token.
0:20
So it allows us to bring a lot of the field labor
0:23
and into a controlled environment in the factory.
0:26
- Think of this as a container, packed with GPUs,
0:30
designed by people who do AI and AI only.
0:33
There are no middlemen involved.
0:35
The whole end-to-end equation is managed by Crusoe,
0:38
which we refer to as vertical integration.
0:41
We get the time to market, we get the right cost structure,
0:45
and that's actually the building block we're putting
0:47
under some of our most sophisticated cloud services.
0:50
- Every single one of these products
0:52
is manufactured by Crusoe,
0:55
and we're bringing kit of parts
0:57
that basically comes together in an assembly line.
1:01
This is like choreographing a dance.
1:03
One worker is doing a certain activity here.
1:06
What another one is getting their next panel ready to come,
1:09
and our team of engineers are basically
1:12
measuring our cycle time,
1:13
and providing feedback to our design engineering teams.
1:16
- You end up with something truly magical,
1:18
and it takes just literally decades of experience,
1:21
and we're hiring anywhere from our master electricians
1:25
all the way through apprenticeship programs.
1:28
- It's amazing when you think of designing
1:30
a managed AI service to be able to have your requirements
1:35
perfectly understood by people who prepare the site
1:38
and build the module data center.
1:41
- One of the challenges is the amount of power
1:43
that's in such a small space.
1:45
Spark Liquid allows us to extract that heat
1:48
and increase the GPU density inside the Spark modules
1:51
by roughly five X to bend the arc towards sustainability,
1:54
and so we don't consume water in the cooling process,
1:57
and it's leaving a much, much smaller footprint.
2:00
- Crystal Spark Liquid Cooled is designed specifically
2:03
to welcome the NVIDIA Verra Ruby and NVL 72,
2:06
and you get the most efficient design
2:09
you could think of in 2027.
2:11
- So we're already designed and engineered
2:13
for the next generation of chip.
2:15
You can't do that if you're off-shore.
2:17
It requires the ability to be flexible.
2:20
- So when we shave seconds in an operation,
2:24
our cycle time performance improves,
2:26
because ultimately what we want to do here
2:28
is we want to have thousands of these units
2:31
coming out of this facility every year.
2:33
- That control and predictability is the key
2:36
of what makes us reliable as a partner in the AI industry.
2:39
- We are leading the market in modular compute,
2:42
and this isn't part of an industrial revolution
2:44
that has not happened before.
2:45
- There's an enormous amount of opportunity
2:47
to ultimately turn America back
2:50
into the kind of industrial powerhouse that we used to be.
2:53
- Build it fast, build it right.
2:55
You know, a lot of teams will tell you
2:57
that you can do both, but we're doing both.
回覆