zlacker

>>g-mork+(OP)
> it is possible to put 500 to 1000 TW/year of AI satellites into deep space, meaningfully ascend the Kardashev scale and harness a non-trivial percentage of the Sun’s power

We currently make around 1 TW of photovoltaic cells per year, globally. The proposal here is to launch that much to space every 9 hours, complete with attached computers, continuously, from the moon.

edit: Also, this would capture a very trivial percentage of the Sun's power. A few trillionths per year.

>>gok+h4
We also shouldn't overlook the fact that the proposal entirely glosses over the implication of the alternative benefits we might realize if humanity achieved the incredible engineering and technical capacity necessary to make this version of space AI happen.

Think about it. Elon conjures up a vision of the future where we've managed to increase our solar cell manufacturing capacity by two whole orders of magnitude and have the space launch capability for all of it along with tons and tons of other stuff and the best he comes up with is...GPUs in orbit?

This is essentially the superhero gadget technology problem, where comic books and movies gloss over the the civilization changing implications of some technology the hero invents to punch bad guys harder. Don't get me wrong, the idea of orbiting data centers is kind of cool if we can pull it off. But being able to pull if off implies an ability to do a lot more interesting things. The problem is that this is both wildly overambitious and somehow incredibly myopic at the same time.

>>rainsf+RA
Yeah it does not make a whole lot of sense as the useful lifespan of the gpus in 4-6 years. Sooo what happens when you need to upgrade or repair?

>>nwelli+R11
This is a question that analysts don't even ask on earnings calls for companies with lowly earthbound datacenters full of the same GPUs.

The stock moves based on the same promise that's already unchecked without this new "in space" suffix:

We'll build datacenters using money we don't have yet, fill them with GPUs we haven't secured or even sourced, power them with infrastructure that can't be built in the promised time, and profit on their inference time over an ever-increasing (on paper) lifespan.

>>Lalaba+Q41
> This is a question that analysts don't even ask

On the contrary, data centers continue to pop up deploying thousands of GPUs specifically because the numbers work out.

The H100 launched at $30k GPU and rented for $2.50/hr. It's been 3 years since launch, the rent price is still around $2.50.

During these 3 years, it has brought in $65k in revenue.

>>acchow+S51
They can run these things at 100% utilization for 3 years straight? And not burn them out? That's impressive.

>>hdjrud+K71
Not really. GPUs are stateless so your bounded lifetime regardless of how much you use them is the lifetime of the shitties capacitor on there (essentially). Modulo a design defect or manufacturing defect, I’d expect a usable lifetime of at least 10 years, well beyond the manufacturer’s desire to support the drivers for it (ie the sw should “fail” first).

>>vlovic+yf1
The silicon itself does wear out. Dopant migration or something, I'm not an expert. Three years is probably too low but they do die. GPUs dying during training runs was a major engineering problem that had to be tackled to build LLMs.

>>mike_h+YE1
> GPUs dying during training runs was a major engineering problem that had to be tackled to build LLMs.

The scale there is a little bit different. If you're training an LLM with 10,000 tightly-coupled GPUs where one failure could kill the entire job, then your mean time to failure drops by that factor of 10,000. What is a trivial risk in a single-GPU home setup would become a daily occurrence at that scale.