$IREN (+1,55%)
$META (+0,14%)
$MSFT (-0,24%)
$NBIS (+2,34%)
$EQIX (+0,52%)
$DLR (+1,18%)
ein toller Beitrag von $IREN (+1,55%) auf x (Jim Liu) weshalb ich ihn hier teilen möchte, wirklich sehr gut geschrieben.
A fair skeptical question I’ve been asked on $IREN (+1,55%) is: If $IREN (+1,55%) overplays their hand, could hyperscalers can simply say, “Forget it, we’ll build ourselves.”? The answer is that hyperscalers are definitely acquiring power and building datacenters themselves but that’s still not enough; I will go into detail.
Projections for power shortfalls by 2030, 2035 are huge but seem far into the future, so let’s use 2028: by 2028 there will be a 36GW shortfall (1) (sources provided by number in comments). A general shortfall number seems hand wavy so let’s look at what this shortfall actually means.
1. The first misunderstanding is the assumption that if the all the tech CEOs and US government is aware about this problem then US will fix the shortfall within 1-2 years, after all the US is still the most powerful nation on earth. Well the upper bound for quickly building infrastructure is China. China is adding 290 GW to their grid in 2025 (2). 290 GW is a massive number right? Who needs 3GW from $IREN (+1,55%) when China can add 290 GWs in one year? Well 290 GWs is peak generation - sources like window, solar are intermittent and cannot sustain peak capacity and batteries only smooth out the power curve, not generate more power. With batteries and base generation to smooth it out and accounting for transmission losses, the effective continuous load supplied will be 22% of the peak generation or to 65GW (3). Now 65GW is still alot, so what’s the deal with 3GW? Well you see, the 65GW that came out in 2025 was planned in China’s 14th 5-year plan (4). Although China’s 14th 5-year plan was announced in 2021, the work needed to meet deadlines started in 2017/2018. Now this 65GW was due to their high projections in heavy industry and manufacturing, EVs, air conditioning, etc. China is the country to over builds infrastructure to maintain employment numbers, this is where high speed rail operates at a loss and prime real estate in Tain Fu District Guangzhou to fall 40% due to overbuild. However, in 2018, China never accounted for AI HPC. Even with manufacturing demand and EV adoption slower lower than projections and extra GW allocated to AI HPC, China still faces a shortfall of power for AI HPC (5). Let’s face it, the US builds infrastructure much slower than China, this is not a 2-3 year solvable problem. Even if small nuclear reactors ($OKLO ) first come out in 2028, it will take at least 1-2 years for them to ramp up to fully supply the shortfall.
2. Hyperscalers will never put their GB300s on the Chinese grid for obvious reasons. The 36GW shortfall is for the American grid. So what’s happening in North America? Hyperscalers are trying some uncertain projects like $MSFT (-0,24%) Three Mile Island restart scheduled for 2028 (6) and small modular reactors. Now geeze, $MSFT (-0,24%) has plenty of time until 2028, if one of it’s best plans to get more power in 2028 is to restart Three Mile Island, what kind of plans does it have for 2026 or 2027? $NBIS (+2,34%) 300MW of IaaS is not going to be enough to fulfill demand. You hear that Meta is getting very serious about its AI ambitions. It announce it’s 1GW Ohio DC called Prometheus. Closer inspections shows that the 1GW DC will be powered by Two 200MW natural gas facilities coming up in Q3 2026 and 440MW of solar (7). Solar has a capacity factor of 18%. Meta’s 1GW Hyperion DC has ~480MW of power at this time. You see the short fall? $META (+0,14%) announces it’s 2027 Hyperion DC to have 1.5GW IT Load (2.26GW of generation to get 1.5GW of IT Load), but upon closer inspection the first two gas turbines are expected to come online in 2028 and the third is expected to come online 2029 (8). Now 2028 is actually pretty fast for 2.26GW, how did Meta even secure the deal? It’s paying 3.2B extra for the gas turbines that will be to be owned by Entergy (9) on top of paying all the regular costs for the DC power. If $META (+0,14%) is paying 3.2B of extra cost for 2.26GW in 2028/2029, how much would 2.2GW in 2026 (Childress + SW1) and 3GW by 2027 for $IREN (+1,55%) be? With With xAI’s 1GW Colossus 2 is coming out in 2026, is Meta going pray that it can fill in the power gap for it’s Prometheus, or chill out until late 2028 when Hyperion is ready?
3. In NA, there were $64 billion of DC blocked or delayed in 2024/2025 (10) because the grid simply doesn’t support that much and people don’t want their electricity bills going up. So that’s pushing GPUs DC’s to two hot areas: energy abundant areas like West Texas and behind the grid gas turbines. West Texas is under Ercot and BTC miners taken all the early power capacity area since 2017 and have queue up the power pipeline to 2030. Crusoe and Lancium contracted out its power to Oracle and OpenAI. Everyone else is bragging about their power pipeline in West Texas but there’s not enough power as Ercot will deny over half of power pipelines between now and 2030 (11). IREN does not disclose its power pipeline rumored to be 5-6GW (12) and only public announces the 3GW with interconnect agreements.
4. Now what about gas turbines behind the meter? Isn’t that what xAI did for Colossus 1/2 and what DataOne did for $NBIS (+2,34%) ? Well gas turbines have a wait time of 5-7 years (13). Colossus took their gas turbine from abroad, disassembled it and then is rebuilding it in Mississippi (14). I love $TSLA (+0,81%) mission but let’s be real here: Europe/Japan/China aren’t in excess of power plants and Elon conveniently omitted where the gas turbines came from: it probably came from a 2nd/3rd world country who sold out their infrastructure with the byproduct being their people’s suffering. This isn’t really scalable. DataOne is a spin out as BSO’s GPU data center division who have been in the datacenter business since 2024. DataOne/BSO like $IREN (+1,55%) had the foresight to procure long lead-time items.
5. So now the question is: what does it mean for $IREN (+1,55%) to over play their hand? Has CPU datacenter operators like $EQIX (+0,52%) ($75B), $DLR (+1,18%) ($55B), and many similar sized privately owned operators like BSO overplayed their hand? Why doesn’t AWS, GCP, Azure build more of it’s own datacenters instead of let $75B and $55B market cap “middle-men” develop? The GPU datacenter operator space will undoubtedly be larger than the CPU datacenter operator space and $IREN (+1,55%) will be 2-3x larger than $EQIX (+0,52%) and $DLR (+1,18%) once it’s said and done. And that’s if $IREN (+1,55%) doesn’t move up the software stack.
Sources:
(1)https://x.com/sectorsignals/status/1873257677595484382?s=46&t=5M46IuHFFx0VtfxNNuG8NA
(3) https://makayda.com/blog/what-is-capacity-factor
(4)https://cset.georgetown.edu/publication/china-14th-five-year-plan/
(9)https://lailluminator.com/2025/08/20/entergy-meta/
(10) https://www.datacenterwatch.org/report
(11)https://x.com/umbisam/status/1968272986584351104?s=46&t=5M46IuHFFx0VtfxNNuG8NA
(12)https://x.com/fransbakker9812/status/1964911703059238927?s=46&t=5M46IuHFFx0VtfxNNuG8NA
