News Elon Musk’s Nvidia-powered Colossus supercomputer faces pollution allegations from under‐reported power generators

[jp7189]

Part of the problem here is that this was not a new, greenfield datacenter build. He bought an old Electrolux appliance factory on the cheap that was not wired for the amount of power that was going to be needed, then moved in and started running before the transmission situation could be worked out. That’s how he supposedly “built” it so quick: they already had a structure and they didn’t wait for the planning/build-out of necessary utilities.

Then he doubled the amount of GPUs and the associated power requirements before the utilities could catch up with the power draw of his initial buildout and already plans to go bigger, seeming to guarantee they will continue to operate an on-site power generation facility indefinitely, in a residential area, that is made of many portable units that are less efficient, more polluting, and less regulated than large permanent installations.

Can we leave Elon off the discussion? It seems to be clouding it with emotion. Xai interviewed several cities for this project and Memphis won the project with promises of available power and monetary incentives to attract this. They failed to deliver and xai is trying to make the best of it. This is Memphis city's fault not bad planning on xai's part.

 

[jp7189]

I'm peripherally involved in one of the largest datcenter buildouts on the planet (1GW initial, 2GW future expansion). Politicians made lots of promises to attract the buildout because the tax revenue per large area is orders of magnitude beyond any other use. Plus everyone wants more high tech jobs in the area. The problem is no one wants it in their neighborhood, like a landfill or a nuclear plant. Strong public opposition caused the politicians to change stance, and the project has become much harder, and of course investors feel betrayed. In other words, the Memphis thing seems quite common from my perspective.

 

[bit_user]

I'm peripherally involved in one of the largest datcenter buildouts on the planet (1GW initial, 2GW future expansion). Politicians made lots of promises to attract the buildout because the tax revenue per large area is orders of magnitude beyond any other use. Plus everyone wants more high tech jobs in the area. The problem is no one wants it in their neighborhood, like a landfill or a nuclear plant. Strong public opposition caused the politicians to change stance, and the project has become much harder,

Why does it need to be so large? This sounds like the "factory farm" problem, where most of the downsides are a result of its scale. Perhaps just having more, smaller datacenters is the way to go.

 

[HansSchulze]

Do tell, what are the emissions products and numbers from these natural gas generators vs, say, a natural gas power plant, like the 1100MW natural gas power plant in Memphis?

And how long would it take to build a 1GW NGPP next to this data center?

 

[HansSchulze]

Why does it need to be so large? This sounds like the "factory farm" problem, where most of the downsides are a result of its scale. Perhaps just having more, smaller datacenters is the way to go.

Sorry, the cluster needs all gpus to be within about 100m for speed. If you added distance, they would need 5x more GPUs to get the same performance.

 

[bit_user]

Sorry, the cluster needs all gpus to be within about 100m for speed. If you added distance, they would need 5x more GPUs to get the same performance.

I don't believe that. I think a popular misconception is that these firms doing AI training have all of their GPUs simultaneously working on the same model, at any given point in time. Actually, they have multiple models concurrently in development and multiple experiments being conducted, at any given point in time. That's not to mention the fact that any AI service provider will have a significant portion of their fleet doing inference workloads, which are very granular. So, the maximum number of GPUs you need in a single place is probably quite a lot fewer than the total number. That's as far as xAI goes.

As for what @jp7189 is talking about, we know even less about that datacenter and who or what it's for. If it's for a multi-tenant cloud provider, like Amazon or Google, then the need to pool all of the resources together in a single location is way less, still.

 

[USAFRet]

I don't believe that. I think a popular misconception is that these firms doing AI training have all of their GPUs simultaneously working on the same model, at any given point in time. Actually, they have multiple models concurrently in development and multiple experiments being conducted, at any given point in time. So, the maximum number of GPUs you need in a single place is probably quite a lot fewer than the total number. That's as far as xAI goes.

As for what @jp7189 is talking about, we know even less about that datacenter and who or what it's for. If it's for a multi-tenant cloud provider, like Amazon or Google, then the need to pool all of the resources together in a single location is way less, still.

Not so sure I disagree with that closeness aspect.

Stock traders on Wall Street get their servers as physically close to the trading servers as possible.
Reduce the transmission time. Milliseconds count.

 

[bit_user]

Not so sure I disagree with that closeness aspect.

Stock traders on Wall Street get their servers as physically close to the trading servers as possible.
Reduce the transmission time. Milliseconds count.

If they're working on the same thing or interfacing with the same exchange, then latency does indeed matter. As I said, whether they're all being used in a single collaboration is a whopper of an assumption.

 

[jp7189]

Why does it need to be so large? This sounds like the "factory farm" problem, where most of the downsides are a result of its scale. Perhaps just having more, smaller datacenters is the way to go.

Well, when it has a master plan all the big guys can share common infrastructure and gain efficency both during the buildout and operationally. In this case it's brown field - permanently contaminated by previous industrial waste and unsuitable for farming or residential. When individual operators build alone there's tons of depluication and waste.

 

[bit_user]

Well, when it has a master plan all the big guys can share common infrastructure and gain efficency both during the buildout and operationally. In this case it's brown field - permanently contaminated by previous industrial waste and unsuitable for farming or residential. When individual operators build alone there's tons of depluication and waste.

Regarding the brownfield aspect, maybe they should put a smaller datacenter there and set aside the rest of the area for other industrial purposes. That obviously leaves the problem of having to build the remaining capacity elsewhere, but sometimes compromises are necessary to unblock progress.

 

[jp7189]

Regarding the brownfield aspect, maybe they should put a smaller datacenter there and set aside the rest of the area for other industrial purposes. That obviously leaves the problem of having to build the remaining capacity elsewhere, but sometimes compromises are necessary to unblock progress.

It's also worth noting that AI training datacenters can be built almost anywhere because their bandwidth requirements are modest. Full service cloud datacenters can only be deployed nearby major exchanges. For example the datacenter I've referenced has 200,000 fibers coming directly off a continental exchange. Typically what you see is datacenter sprawl that expands in rings around the exchange, and that ultimately results in a mess of kludey and inefficient infrastructure and tons of duplication. One large site (that is thoughtfully planned) is better in almost every way.

Regarding this particular brown field, it's been vacant for 35 years. There's no industry that cares to use it. They all left for other countries long ago.

 

[jp7189]

I don't believe that. I think a popular misconception is that these firms doing AI training have all of their GPUs simultaneously working on the same model, at any given point in time. Actually, they have multiple models concurrently in development and multiple experiments being conducted, at any given point in time. That's not to mention the fact that any AI service provider will have a significant portion of their fleet doing inference workloads, which are very granular. So, the maximum number of GPUs you need in a single place is probably quite a lot fewer than the total number. That's as far as xAI goes.

As for what @jp7189 is talking about, we know even less about that datacenter and who or what it's for. If it's for a multi-tenant cloud provider, like Amazon or Google, then the need to pool all of the resources together in a single location is way less, still.

Too true re: AI datacenters there are tons of different task (training and other) that are distributed to many clusters via a scheduler. I'm sure they would like to have all gpus at 100% all the time, but i doubt it ever happens. Though I've never seen a statistic about it.

Re the site I've referenced its multi operator. Google, Amazon, Microsoft, and others will all have a footprint onsite. They are no strict AI providers that I'm aware of yet, but it wouldn't be surprising to see Oracle/OpenAI put something there too.

 

[thestryker]

They failed to deliver and xai is trying to make the best of it. This is Memphis city's fault not bad planning on xai's part.

I'm curious where you saw this because they're contracted for 150MW and I've not seen anything indicating they're not getting it. This might have been enough power to run the initial installation, but the hardware has been doubled so it's nowhere near enough. That would imply the city is actually not to blame for the situation.

 

[jp7189]

I'm curious where you saw this because they're contracted for 150MW and I've not seen anything indicating they're not getting it. This might have been enough power to run the initial installation, but the hardware has been doubled so it's nowhere near enough. That would imply the city is actually not to blame for the situation.

I believe there's an article her on Tom's at the initial deployment of generators that talks about, but my first Google result was from fortune . Com:

"xAI’s first facility isn’t even fully online yet. xAI has fired up temporary generators as MLGW works to accommodate the company’s initial power requests. MLGW has had to schedule construction for a new power substation, paid for by xAI, just to get the first 150 megawatts of power xAI has requested."

Edit to add:

In my mind, the power company ceo saw a single customer asking to buy up a ton of unused capacity and his eyes turned in to dollar signs and happily agreed to whatever the requirements were. I bet he assumed they could buildout faster than a datacenter could buildout.

 

[jp7189]

While I was Googling the other info, I saw that Memphis just greenlighted the purchase of another 3 parcels of land by xai to support their expansion to 1M GPUs. It doesnt sound like the city is mad at xai. Here's a quote about the proposed expansion:

Ted Townsend, president and CEO of the Greater Memphis Chamber, said: “This significant expansion by xAI reinforces Memphis’s position as a premier destination for technological innovation.

“Their investment in this million-square-foot facility, along with xAI’s collaborative partnership, demonstrates the tremendous momentum we’re building in the Digital Delta. Memphis continues to prove itself as the ideal location for companies leading the future of technology.”

 

[bit_user]

Ted Townsend, president and CEO of the Greater Memphis Chamber, said: “This significant expansion by xAI reinforces Memphis’s position as a premier destination for technological innovation.

I wonder how much it's costing local tax payers. Given that datacenters tend not to create many local, long-term jobs, I doubt they're going to recoup the tax subsidies via income tax revenue. If the property values of nearby homes go down as well, it could compound the losses to the city.

 

[jp7189]

I wonder how much it's costing local tax payers. Given that datacenters tend not to create many local, long-term jobs, I doubt they're going to recoup the tax subsidies via income tax revenue. If the property values of nearby homes go down as well, it could compound the losses to the city.

Tennessee has no income tax, so number of jobs created doesn't matter either way.

However they do charge a 30% tax each year on the value of personal property (different from real property) which is generally considered every asset inside a building.

I don't know, but have to assume xai was given a discount on that as that would be a crippling amount.. approx $1B per year.

 

[thestryker]

I believe there's an article her on Tom's at the initial deployment of generators that talks about, but my first Google result was from fortune . Com:

"xAI’s first facility isn’t even fully online yet. xAI has fired up temporary generators as MLGW works to accommodate the company’s initial power requests. MLGW has had to schedule construction for a new power substation, paid for by xAI, just to get the first 150 megawatts of power xAI has requested."

Had more time to look and based on this given the past tense it sounds like that substation is actually done:

XAl has spent $35 million to construct a substation to support MLGW. XAI didn’t just contribute money, xAl rolled up its sleeves and managed the construction project to ensure that it met all of MLGWs requirements and went as expeditiously as possible. xAl is not even close to being done on this front, as xAl is committing an additional $20 million to build another substation.

https://www.actionnews5.com/2025/04/26/emotions-run-high-schd-town-hall-xai-environmental-concerns/

In my mind, the power company ceo saw a single customer asking to buy up a ton of unused capacity and his eyes turned in to dollar signs and happily agreed to whatever the requirements were. I bet he assumed they could buildout faster than a datacenter could buildout.

That's not quite how it works, but I'm 100% certain you're right about the mindset. MLGW doesn't actually generate power so they have to get approval from the TVA before making such promises.

 

[thestryker]

While I was Googling the other info, I saw that Memphis just greenlighted the purchase of another 3 parcels of land by xai to support their expansion to 1M GPUs. It doesnt sound like the city is mad at xai. Here's a quote about the proposed expansion:

Ted Townsend, president and CEO of the Greater Memphis Chamber, said: “This significant expansion by xAI reinforces Memphis’s position as a premier destination for technological innovation.

“Their investment in this million-square-foot facility, along with xAI’s collaborative partnership, demonstrates the tremendous momentum we’re building in the Digital Delta. Memphis continues to prove itself as the ideal location for companies leading the future of technology.”

Crooked city officials shouldn't really be of any surprise though. I have no doubt those who stand to gain don't particularly care about the downsides of these actions.

The last couple of years here The Oregonian has done several in depth pieces about data center (Amazon is the biggest) expansion in The Dalles. There were lawsuits from the city the to try to stop them from uncovering information about data center resource consumption. They uncovered a bunch of shady activity by city officials even the people living in the city didn't know about.