The Smoke Blowing Up Telus’s Ass

Telus has a very lofty claim: that it plans to build out a pair of AI data centres here in Vancouver BC, specifically in the downtown core and in the Mount Pleasant neighbourhood. Thankfully the Mount Pleasant site is not zoned yet, and has been referred back to city planners, so we live in hope that it will be properly staked through the heart and turned to dust before its launch date of some time in 2029. The other one, which will occupy a site close to Rogers Arena on West Georgia St, is the greater concern, in our opinion.

In any case, Telus will be contributing “roughly CAD$1 billion” over the next five years, with the Canadian federal government contributing CAD$2 billion over the same time period. From this, they expect to build 150MW of capacity between the two Vancouver facilities and a third one in Kamloops. We wish to find out, through any and all analysis and information, whether this is feasible, and Telus (so far) have not given us quite enough information. We must resort to one of our favourite pastimes, mathematics, to make sense of the corporate buzzspeak infesting every news article on this subject.

Let’s begin.

Telus intends to acquire 60,000 GPUs for this cluster of three facilities. Due to their statements that the facilities will be state-of-the-art, we may assume that the chips they intend to install are the liquid-cooled Nvidia Vera Rubin racks, specifically the VR200 NVL72 configuration which Jensen Huang (CEO of Nvidia) announced as being in full production in January 2026. (The bigger Kyber racks have been delayed to 2028). This rack has a power draw of 190–230 kW and, obviously, contains 72 Vera Rubin GPUs.

Telus would need to buy 833 racks in order have 60,000 GPUs. At the lowest estimate of the power draw per rack, that would add up to about 158MW of capacity (which does not include the power draw required for running everything else in the facility that isn’t a GPU). We may infer that, if this is what Telus is working with, they are rounding up or down to conveniently big and easily marketable numbers.

Before we dig further into the murky, mud-filled cave ecosystem of data centre costs, a small note on the power draw question. For it is indeed a question, dear readers, and one we feel has not been adequately answered by the Telus minions who push press releases to all and sundry, yet refuse to answer our emails. We have found numbers ranging from 120kW to 350kW per Vera Rubin NVL72 rack from various sources online, with some hints from those who build and deploy such systems that the real-world thermal profile may be far more unpredictable than the manufacturer specs. We have taken the lower end of the estimate above on the assumption that Telus would do the same in order to present the best case scenario.

We found a recent quote for the Vera Rubin NVL72 rack which amounts to USD$7.8 million. For 833 racks, that presents a bill to Telus for USD$6.5 billion for the compute hardware alone. We must, of course, convert that to Canadian dollars and hope that the exchange rate does not weaken too much in the future, which results in a bill of CAD$9.12 billion. The paltry CAD$3 billion provided by Telus and the Canadian taxpayer “over the next five years” will not suffice to acquire the compute hardware alone.

But there is a further issue. According to CNBC, “Nvidia’s current-generation Rubin systems are in full production and begin shipping this fall to eight cloud partners, including Amazon Web Services, Microsoft Azure and Google Cloud.” Telus is not one of these providers. According to HashrateIndex, the following cloud providers are the ones offering Vera Rubin instances: AWS, Google Cloud, Microsoft (Fairwater AI superfactories), OCI, CoreWeave, Lambda, Nebius, and Nscale. Despite Telus’ claims that the Kamloops and downtown Vancouver data centres will come online “later this year”, it is flat-out impossible that that can happen with Vera Rubin racks because they cannot get the chips.

So, let us look at the math again. We will endeavour to be generous to Telus, which they most assuredly do not deserve, in the hope of landing on a reasonable scenario. Perhaps they are engaging in a little marketing silliness, and by “state-of-the-art” they actually mean “chips we can buy right now”, meaning last-generation Blackwell NVL72 racks instead of Vera Rubin racks.

Again, they require 833 racks in order to have 60,000 GPUs. The power draw of a Blackwell NVL72 rack is 120kW to 130kW, which puts the total capacity at about 100MW, giving Telus some overhead to actually power the facilities surrounding the compute hardware and suggesting that this is plausible, if nothing else. (Blackwell NVL72 racks are also liquid-cooled and ostensibly efficient with water usage.) The cost of a Blackwell rack is a more reasonable USD$3.5 million, which means USD$2.9 billion for the compute hardware. How nice, a more than 50% discount over Vera Rubin racks! Unfortunately, this still amounts to CAD$4.06 billion, and once again the money earmarked for this silly project is not enough.

At least the chips are available, however. If Telus can throw even more money into this slop-lined pit, they may indeed get something to put into their new data centres.

Oh yes, we must discuss the actual data centres. Can’t forget those. Very important topic.

It’s important to note that the large, power-hungry, liquid-cooled racks used for AI are very different from the more usual server racks in a regular data centre, such that buildings need to be retrofitted at great cost (or simply demolished and rebuilt) in order to accommodate them. Axis Intelligence has produced some interesting analysis on the raw cost, on average, to build and kit out an AI data centre. Their findings show that liquid-cooled racks incur quite a premium in the costs of everything surrounding the computing hardware, amounting to a total of USD$12 to 15 million per MW. Canada itself has an average cost of USD$7 to 12 million per MW and we find it reasonable to assume that the construction costs in Vancouver, a city notorious for the price of its real estate, will lean towards the high end of that. Taking a reasonable USD$12 million per MW cost for the non-compute portion of the project, Telus has a further bill of USD$1.8 billion dollars (CAD$2.52 billion) landing on its proverbial desk in order to fulfill its demented AI chatbot dreams.

If costs were the only issue with data centre builds, then we suspect there would be more data centres in the world. But nothing, of course, is that easy, and there are some things that an infinite pile of money can’t fix or hurry along. (Not many things, mind you, but some.) The lead time on actually breaking ground and then completing an AI data centre is, in a word, godawful. Although there are plenty of construction firms that can build the shell, the power equipment (and the expertise to actually install and connect it all) is in short supply; it is, in fact, being sucked up by US-based hyperscalers like Microsoft, EV infrastructure, and replacements for the aging power equipment used to do useful things, like keeping the lights on in Minneapolis. It takes years, in fact, to get the essential equipment needed to make sure that the compute hardware can always run, like backup generators, high voltage circuit breakers, and so on. We may also see that the costs of such will likely explode due to demand, as this is the trajectory we are witnessing for the high bandwidth memory chips built into the GPU modules.

Telus may wish that their data centres can deliver more than CAD$2 billion in AI revenue by 2028 (and how are we accounting for that particular figure, pray tell?) but we suspect that they will fall far short simply because time, money, resources, common sense, and the laws of physics are all against them. They are a comparatively small player in a field that contains major US firms with more cash than an entire flock of gold-hungry dragons, and those firms will out-spend and out-last them with ease in order to build their own AI projects. For comparison: Telus intends to spend CAD$70 billion on its entire infrastructure over the next five years. We suspect it will have to earmark quite a bit more of that amount to its AI ambitions, because Microsoft alone spent USD$80 billion on AI-related capital expenditures in 2025.

In short: the odds of Telus actually building, stocking, and powering an AI data centre anywhere in Vancouver in the next year is zero. The odds of it happening by the end of 2027 are close to zero unless they decide that a Rimouski-style project, with a few hundred GPUs, is an acceptable compromise. The odds of a larger data centre with 60,000 GPUs being completed, powered, and selling compute within five years are just plain bad. And, dear readers, if Telus manages to legitimately increase their AI revenue to CAD$2 billion in 2028, we promise most faithfully to grill and consume our favourite hat.

We shall watch with interest, and possibly popcorn, to see if or how Telus overcomes the basic logistics obstacles in their way. We shall also watch and support the efforts of No AI Vancouver, a local activist group who has rightly determined that a demented circus of a project run by rich, AI-bedazzled clowns can do nothing good for our city and province. And, of course, we will consider an update if said clowns ever actually respond to our emails.