R retweeted
How big are the data centers today?
Today, there are only 14 data centers on Earth that are bigger than 250MW. For context, 250MW is a single Siemens SST-800 steam turbine.
So whilst the press talk this capex cycle up a lot, so far this is not much energy that is flowing to compute. Not yet.
There are only a handful of big data centers operating today, but there are a lot more big data centers being built right now.
The wave of big new data centers coming online still hasn’t hit. Frontier models are still training on sub-GW machines.
If you build cooling towers, it’s quite possible to build a single terrestrial facility that can emit 5GW of heat. Nobody has done that yet, but it’s theoretically possible (although maybe not politically).
The largest data center is more valuable than 2x data centers that are half its size, because a larger machine can more easily train and output the frontier.
AI is heavy industry.
Whoever can build the largest coherent synchronous machine, should be able to train the most powerful model.
Building such a facility, now the supply chain is tight, takes 5-7 years. There are a limited number of sites capable of hosting a GW scale data center… all of those sites are now bought up.
Many countries do not have a single site capable of hosting a GW scale machine.
Now the shape of AI can vary wildly, you can have many small models, a sort of swarm of AI, all running on lots of little compute machines. Or you could have one giant coherent machine that hosts a single giant super intelligence.
Inference can be a distributed swarm or a consolidated deity, or something in between.
But these things are not the same, they have strengths and weaknesses, they serve different applications better or worse.
Humanity is going to build it all.
But different organisations already have distinct trajectories and strategies for what they want to do, there is much more diversity in the market and the economy than people think. There are also myriad unserved gaps.
Humanity is going to build it all.
R retweeted
Elon Musk, Pete Hegseth, and anyone else who played a role in cutting off Ukraine's ability to defend itself must answer for it. Congress must get to the bottom of this.
R retweeted
We wanted to do more for the people who’ve been with us the longest. So with Fable 5.2, the longer you’ve been subscribed, the more Fable resets you get.
R retweeted
You can tell Gunn earnestly believed this bullshit was going to change the world. Hilarious.
R retweeted
What if 25%+ of the front is sustained through robotics by 2029? China will be feeding Russia technological innovations that tilt things in Russia’s favour over time if it’s not treated high priority enough.
R retweeted
thinking about the chinese no bitches dance
R retweeted
🚨🗣️ Antonio Cassano on Cristiano Ronaldo:
“Cristiano attacked me on whatsapp for saying that the real ronaldo is Brazilian” 🫣🇵🇹
R retweeted
Yusufu Ssuuna Kiweewa was the eldest son of Kabaka Mwanga II, and a soldier who fought in both World Wars. He was educated at Mengo in Kyaddondo and at Buddo Kings College in Busiro, before joining the King's African Rifles, which was then the most famous army on the African continent. He rose steadily through the ranks and by 1940 he had reached the rank of Lieutenant, a position he earned through service and not through birth alone.
Unlike his younger brother, Kabaka Daudi Chwa II, who was given military honours without training, Ssuuna took part in the real fighting of the First World War from 1915 to 1919. The war on the African continent was fiercest in the East and the North, and Lieutenant Yusufu Ssuuna Kiweewa was there throughout. Yet when he was on leave from the army in 1945 and unrest broke out in Buganda, the colonial rulers blamed him for it and exiled him to the Ssese Islands.
He died in 1949 at the age of 51 and was buried in the tombs at Nnabulagala in Kasubi. The 1900 Agreement mentions Prince Ssuuna and grants him land measuring eight square miles. His father, Kabaka Kirevuntende Basammulekkere Mukasa Mwanga II, had also given him responsibility for the palace on Lubowa hill, which those who understand it call Ejjembe Lubowa of Kabombo.
R retweeted
Ukraine could have WON the war 4 years ago.
In the Fall of 2022, Ukrainian forces liberated Kherson and a thousands of square kilometers of the Kharkiv region.
Thousands of russian prisoners were being captured and russian defenses were collapsing all along the line.
Then the supplies from the Biden administration mysteriously slowed down leaving Ukrainian forces without artillery.
Later, it emerged that the "as long as it takes" crowd feared a russian collapse and deliberately restrained Ukraine.
It's good that Trump's gang lost now but Ukraine has few real friends in either party.
A lot of death and suffering could have been spared...
The US is at best a friend of convenience, at worst an adversary.
I only trust regular Americans who genuinely support Ukraine, never their leaders.
R retweeted
Elon Musk has Ukrainian blood on his hands. He needs to come in and explain what the hell happened when he altered a global conflict with the flick of a switch.
See you in January, @elonmusk.
Ruben Gallego reveals to TMZ DC's @jacob_wass how Senators score better offices, takes aim at Elon Musk 👀
R retweeted
$MSFT $NBIS $CRWV
microsoft borrows more cheaply than anyone in this chain and still pays nebius about $3.97 an hour per GB300 rather than owning the racks, and a big part of that looks to me like a payment to keep the hardware risk off its own books.
over five years the rent per GB300 comes to about $173,900. owning the same GPU for five years would cost about $87,900 to $131,100:
-- the chip is $51,400 to $90,300.
-- its share of the data centre adds $24,600 to $28,900.
-- power and staff add about $11,900.
so microsoft pays 33% to 98% more than owning would cost, $42,800 to $86,000 per GPU, or $4.3bn to $8.6bn on 100,000 GPUs. some of that buys speed, since nebius had power and buildings microsoft didn't, and nothing disclosed says how the premium splits.
the same contract also looks very different year by year. B300 spot is about $9.06, and if it falls 42.9% a year, the flat $3.97 is cheap for two years and expensive for three. on 100,000 GPUs:
-- staying to the end comes out even.
-- leaving after year two saves about $5.5bn.
-- leaving after year three keeps about $4.6bn and avoids another $4.6bn of overpayment.
i doubt microsoft breaks a contract for $4.6bn. a lab funded round by round, paying 2026 prices in 2029 for chips three generations old, has every reason to renegotiate, and for coreweave, who its biggest customers are and how they're funded is really the same question as what its chips are worth.
ALT Graph comparing Microsoft’s rental rates for GPU usage against market prices over five years, highlighting costs and savings.
