Abstract:
In order to develop AI, Musk
even made his own gas turbine blades
. We all know that computing power is tight, and no matter which major AI manufacturer it is, they are all trying to find ways to get stuck in their computer rooms. But I never expected that Lao Ma was not building his own data center this time, nor was he investing in a power plant to solve the power bottleneck, but he was working on a more basic link.
According to The Information, SpaceX is preparing a gas turbine blade casting plant in Texas.
Relevant recruitment positions have been posted on the SpaceX official website.
As soon as the news spread, Musk himself quickly claimed it and even threw out an even more exaggerated figure:
Of the AI computing power produced in 2027, approximately 15GW may not be able to be turned on that year.
And behind the insufficient power supply, transformers, wiring, liquid cooling systems... everything is a problem. The production bottleneck of gas turbines in gas-fired power generation “lies in the casting of blades and guide vanes.”
Shorten gas turbine production time by 18 months
So I’ll give you a detailed look at Musk’s logic this time:
AI lacks computing power → build data centers → data centers lack power → engage in solar power generation.
But for now, solar energy can only be said to be a longer-term solution:
Both SpaceX and Tesla are building 100GW of solar energy production capacity per year as quickly as possible, but in the next few years, natural gas will still be needed to supplement and start solar power generation.
The old horse can draw cakes, but the AI is waiting impatiently——
SpaceX predicts in the second quarter of this year that the company will have more than 2GW of computing power online by the end of 2026; by the end of 2027, the scale will be close to 10GW.
In order to run these GPUs, SpaceX even plans to prepare a total of 20GW of power supply, cooling and electrical infrastructure projects. Musk himself estimates that even if a quarter of it is delayed, the power generation capacity that will actually come online by the end of 2027 may still be close to 15GW.
The problem is that even if you deploy your own gas turbine, it is not easy to buy it! often! fee! strength!
Gas turbine giant GE Vernova disclosed in the second quarter of this year:
Gas turbine orders and capacity bookings have reached 116GW, and are expected to reach at least 125GW by the end of the year. The company has signed up for its production capacity position in 2031.
Its CEO has also publicly stated before that GE Vernova is expected to sell out all its 2030 production capacity by the end of 2026.
Continue to dismantle, as Musk himself said, it is the
blade
that blocks the production capacity of the gas turbine. ——Other components of the entire gas turbine can be obtained 12-18 months earlier than the blades and guide vanes.
△
AI generationIf you need something, just make it yourself. This is very Musk.
He himself also let go of his harsh words again:
By casting it in-house at SpaceX, we can bring the gas turbine online 18 months earlier, which will bring about profound changes.
AI infrastructure competition is no longer limited to GPUs
Let’s just say that I didn’t expect that AI would be involved in materials science + precision manufacturing by 2026.
But it cannot be said that Musk has zero basis for doing this.
The blades and guide vanes are at the first impact point of the ultra-high temperature airflow and need to withstand both high temperature and the mechanical stress caused by high-speed rotation.
This kind of casting is not simple. Fortunately, SpaceX itself has experience in manufacturing turbomachinery for the Raptor rocket engine that works under extreme working conditions.
In addition, some clever netizens have already thought of:
Computing power requires electricity, electricity requires gas turbines, gas turbines require casting, and casting yield is a problem that can be solved through machine learning.
Hey guys, another one is the rhythm of taking off with the left foot and the right foot spiraling (doge).
Judging from the current recruitment situation, Lao Ma's "Blades and Vanes Foundry" in Bastrop, Texas, covers positions in all aspects from aerospace and machinery, to automation and control, to operations, processing, and materials, and it seems to be moving in the direction of complete precision casting.
However, there is currently no evidence that the foundry has been put into operation.
△
AI generationIt can only be said that the vertical integration of AI companies is no longer limited to GPUs and data centers.
As Musk said, any slowdown in GPU, power generation, power grid equipment, transformers, cables, liquid cooling, chillers, and networks will affect the final token output.
Now, a gas turbine blade, like the latest AI chip, may determine how much computing power is available for AI in the future.
Comments