Antipasto: il cloud sovrano che non c’è

• Media Stream AI (MSAI) says it bought 10,000 GPUs (graphics processing units), worth hundreds of millions of pounds, from Lenovo – but Lenovo says it has not sold the company a single GPU
• MSAI allegedly built the UK’s first sovereign large language model (LLM), Mother, “built, trained, and hosted entirely within British infrastructure” […] he outlined the company’s plans to develop an industry-leading £50m AI data centre in Salford
• But an investigation by TBIJ and the Manchester Mill has found that this data centre is currently an empty retail unit
• The data centre expert said construction costs for a data centre of this scale would likely be around £185-300m for the build, and another £450m for the GPUs and IT.

Portata principale

https://www.wheresyoured.at/the-ai-industry-is-lying-to-you

Sentiamo parlare dell’esplosione che viene del settore dei data center.
L’economista Paul Kedrosky ha analizzato la capacità di aggiungere potenza di calcolo ai datacenter, attenzione qui si parla di progetti previsti, non di effettiva entrata in funzione:

Le aggiunte di capacità dei data center statunitensi si sono dimezzate dal terzo al
quarto trimestre 2025, poiché le sfide della coda di carico persistevano. Il declino
sottolinea le difficoltà dell'attuale contesto di sviluppo e segnala una conseguente
attenzione ai progetti di pipeline esistenti.
[...]
that of the 241GW of disclosed data center capacity,
only 33% of it is actually under active development

Il rapporto aggiunge anche che la maggior parte dell’energia impegnata (58%) è per “utilità solo fili”, il che significa che il fornitore di servizi è responsabile solo di ottenere energia alla struttura, non generando l’energia stessa, il che è un grosso problema quando stai costruendo interi campus costituiti da server AI affamati di energia:

The total announced pipeline looks huge at 241 GW —
about twice US peak electricity demand — but most of it is not real. 
Only a third is under construction, with the rest a mix of hopeful permits,
speculative land deals, and projects that assume power sources nobody has
actually uilt yet.
In particular, much of it assumes on-site gas plants, a fraught assumption
given current geopolitics.

Provando a ricostruire i numeri

• Only 33% of announced US data centers are actually being built, with the rest in vague levels of “planning.” That’s about 79.53GW of power, or 61GW of IT load.
  • “Active development” also refers to anything that is (and I quote) “…under development or construction,” meaning “we’ve got the land and we’re still working out what to do with it.
  • This is pretty obvious when you do the maths. 61GW of IT load would be hundreds of thousands of NVIDIA GB200 NVL72 racks — over a trillion dollars of GPUs at $3 million per 72-GPU rack — and based on the fact there were only $178.5 billion in data center debt deals last year, I don’t think many of these are actually being built right now.
• Even if they were, there’s not enough power for them to turn on.
• NVIDIA claims it will sell $1 trillion of GPUs between 2025 and 2027, and as I calculated previously, it sells about 1.6GW (in IT load terms, as in how much power just the GPUs draw) of GPUs every quarter, which would require at least 1.95GW of power just to run, when you include all the associated gear and the challenges of physically getting power.
• None of this data talks about data centers actually coming online.

Fonte: mappa di fractracker: https://experience.arcgis.com/experience/5a4d072ad01449bba5698a80103fb909

Installare le GPU

With the acceleration of NVIDIA’s GPU sales, it now takes about 6 months to install and operationalize a single quarter’s worth of sales […]
It’s becoming very obvious that data center construction is dramatically slower than NVIDIA’s GPU sales, which continue to accelerate dramatically every single quarter.

Even if you think AI is the biggest most hugest and most special boy: what’s the fucking point of buying these things two to four years in advance? Jensen Huang is announcing a new GPU every year!

La reale portata del mercato

There Is Only 5GW of Global Data Center Capacity Actually Under Construction, And Every Huge, Multi-Gigawatt Project You Read Is Going To Take 2 to 4 Years Or More To Complete — And Wood Mackenzie Believes Capex Growth Will Slow In 2026

Before we go any further, I want to be clear how difficult it is to answer the question “how long does a data center take to build?”. You can’t really say “[time] per megawatt” because things become ever-more complicated with every 100MW or so. As I’ll get into, it’s taken Stargate Abilene two years to hit 200MW of power.
These numbers also are based on 5GW of capacity, meaning about 3.84GW of IT load, or about $111.5 billion in GPUs and associated gear, or roughly 57.5% of NVIDIA’s FY2026 revenue that’s actually getting built.

Not to sound unprofessional or nothing, but what the fuck is going on? We have 241GW of “planned” capacity in America, of which only 79.5GW of which is “under active development,” but when you dig deeper, only 5GW of capacity is actually under construction?

Casi specifici negli USA

• Building A Data Center Is Difficult, And Nobody Has Built A 1GW Data Center Yet
• There are campuses — such as Stargate Abilene — which promise to reach 1.2GW, but nearly two years in sit at two buildings at around 103MW of critical IT load each
• Microsoft claims to have brought around 4GW of data center capacity online in the last two years, but it’s unclear how much actually got built.
• In an analysis of all announced groundbreakings and land acquisitions, it appears that Microsoft has only finished the first phase of its Atlanta and Wisconsin data centers.
• It is unclear where this capacity could be.

https://www.wheresyoured.at/where-are-all-the-data-centers/#no-really-where-are-all-the-data-centers

E se avessimo detto un sacco di cazzate?

Poniamo anche il caso che siano tutte minchiate, probabile.
Però una notizie recente non risuona con l’entusiasmo di Anthropic, che ha aumentato i limiti di utilizzo settimanale di Claude code del 50% dopo aver ottenuto l’esclusiva su Colossus, il Datacenter di xAI costruito in Tennesee da 300 MW.
Questo
La domanda è: se ci fossero tutti questi GW di potenza, perché la disponibilità di claude, dopo essere stata limitata per mesi, viene aumentata del 50% con 300 MW da xAI?

https://www.cnbc.com/2026/05/06/anthropic-spacex-data-center-capacity.html