Google's Gemini is SOTA and Google is valued at 4 trillion. Now it's ironic of course: Google became a 4 trillion thanks to ads. Which now allows Google to serve no ads in Gemini. It doesn't matter to Google that they're operating at a loss with Gemini: they can throw hundreds of billions at the problem over a few years, serving zero ads, while people make fun of the AI firms showing ads, like ChadGPT.
OpenAI is not just facing Anthropic and its "no ads" ad, it's also facing "we're a 4 trillion company and can run Gemini at a loss for decades" Google, "we pretend we'll put datacenters in space so that xAI can tap into the SpaceX warchest", openweight models which anybody can run on rented hardware for next to nothing compared to ChadGPT's price, etc.
It doesn't matter if others can survive without ads: all that matters is that they can survive long enough without ads so that OpenAI doesn't become another trillion dollar company.
Electricity generation is the constraining factor, but the sun does not turn off in space. xAI data centers in space drives cost to zero, even with inferior models.
I see no other future than SpaceXai winning.
Solar in space produces 30% more power, and doesn't turn off at night, meaning you don't need batteries. That means power costs, say, 25% of what it currently does measured against terrestrial solar and batteries.
The 75% electricity discount needs to pay for launch vehicles, specially designed satellites, and the inability to service the hardware or resell it when it's EOL for the data center.
It's a gamble. Maybe it'll turn out to be a slight edge, maybe it'll turn out to fail, but it's not a sure thing and it certainly isn't going to hugely decrease the cost.
Especially since they're competing against Google and their custom designed hardware that's far more power efficient for AI. It's not clear that NVIDIA running at a 75% dollar discount beats Google's best TPU in compute per dollar.
We can make chips faster than we can build power plants.
There are far too many variables still unknown to all parties. Anyone trying to say with certainty "X will lose", whether X is terrestrial or space based DCs, is lying and probably trying to sell you something.
90% of his predictions didn't materialize, he's full of shit, how do people keep falling for it over and over again?
Camera based fully self driving cars
Robotaxi fleets by 2020
Whatever the fuck were these tesla tunnels in LA
Humans on Mars by 2024
The sub 25k tesla
etc.
But watching him battling basic physics is very funny, not gonna lie.
Marginal cost of launches keep coming down for SpaceX with reusable rockets and lifetime of satellites is long.
Before downvoting, would you mind quoting the relative cost of batteries vs. solar panels for a 150kW solar-powered satellite?
OK.
At a good location (~25% capacity factor), you need about 600 kW of panels to average 150 kW. Utility-scale solar runs roughly $0.50–$1.00/W installed, so call it ~$450K–$600K. Overnight storage (say ~16 hours) requires ~2,400 kWh. Adding a buffer for cloudy days, say 4,000–7,000 kWh total. At roughly $200–$350/kWh (utility-scale Li-ion), that's ~$1M–$2M.
In a favorable orbit, capacity factor is ~90–100% (GEO or sun-synchronous), so you need roughly 160–170 kW of panels. Space-qualified solar panels historically cost $100–$300/W. Even optimistically at $50–$100/W with newer manufacturing, that's 167 kW * $100/W = ~$17M optimistically, or 167 kW * $200/W = ~$33M realistically. You also need space-rated power management, thermal systems, and radiation-hardened electronics.
Even ignoring launch costs entirely, space solar is roughly 10–20x more expensive than ground solar + batteries, driven almost entirely by the enormous cost premium of space-qualified solar panels. Ground-based solar is extraordinarily cheap now (~$0.50–1/W), while space-grade panels remain orders of magnitude more expensive per watt.
The ground option wins overwhelmingly. The space option would only start to make sense if space-grade panel costs dropped to near terrestrial levels, which would require a revolution in space manufacturing.
FYI you'd need 2x the solar panels of the ISS to run a single rack of NVIDIA GB300, and microsoft just built a datacenter with 4600 of these racks.
[1] https://sunwatts.com/150-kw-solar-kits/
[2] https://ourworldindata.org/grapher/cost-space-launches-low-e...
[3] https://everydayastronaut.com/starlink-group-4-5-falcon-9-bl...
150kw is just enough to power a single gb300 rack, the rack alone weights 1500kg+
Guess he is not as bright as he thinks he is.