Elon Muskwrote on X on 31 October that there will be “a total of around 1.2M Blackwells at Colossus 2 by end of year”. If all of them sat in GB300 NVL72 racks at HPE’s published spec, they would draw about 2.2 GW of rack power, by our arithmetic, before cooling.
That is about 1.8 times the 1.2 GW of self-generating capacity that Mississippi permitted xAI in March, according to DatacenterDynamics. The post had about 1.19 million views by Sunday evening.
Business Pill · Lead Time Explained
A short explainer on lead times: why big infrastructure takes years from plan to operation. It teaches the general idea only and says nothing about this story.
The key insight: By our arithmetic on HPE’s published rack figures, the 1.2M Blackwells Musk posted would draw about 2.2 GW of rack power, about 1.8 times the 1.2 GW of on-site gas generation Mississippi permitted xAI.
Musk’s post, a reply to the account @XFreeze, reads: “There will be a total of around 1.2M Blackwells at Colossus 2 by end of year”. On 31 October he also named four Colossus 2 buildings holding about 860,000 GB300s between them, as we reported in Musk Says Colossus 2 Will Hold 1.2M Blackwells by Year-End.
Power implied by 1.2M Blackwells in GB300 NVL72 racks at HPE’s published rack figures (our arithmetic), against the 1.2 GW of on-site gas generation Mississippi permitted xAI in March 2026 (DatacenterDynamics) and the 495 MW of at least 27 turbines on the site as of 7 April, our earlier calculation from a 30 October court order’s figures. Rack power only, before cooling.
What Was Posted on X
@XFreeze There will be a total of around 1.2M Blackwells at Colossus 2 by end of year
Elon Musk posted on 31 October, replying to the account @XFreeze: “There will be a total of around 1.2M Blackwells at Colossus 2 by end of year”. The post had about 1.19 million views by 1 November.
He who shapes power, data centers, chips and models into an economically feasible endpoint, disproportionately shapes AI/SI right now.
Why
Power is THE binding constraint in AI/SI. Few really have it.
Data centers donโt work without power. This is not about promises of power tomorrow. Or forecasted power in 2028. Power TODAY.
Chips canโt operate profitably unless installed in a functional data center. If you can install chips, itโs because youโve solved the power problem.
Models arenโt economically viable unless run on cost effective infrastructure which requires a functional data center with low input costs and reliable management. This canโt happen without power.
AI/SI ROI doesnโt pencil unless models generate profitable outputs which necessarily needs functional data centers as the critical input. Again, power.
Ergo, he who can shape data centers, chips and models together into a tightly coordinated stack is someone who has solved power.
Solving power allows you to disproportionately shape AI/SI right now.
Chamath Palihapitiya quoted Musk’s post on 1 November and wrote: “Power is THE binding constraint in AI/SI. Few really have it.” He added: “Solving power allows you to disproportionately shape AI/SI right now.”
The Rack Arithmetic
HPE’s QuickSpecs for its version of the rack, dated 8 September 2026, say the system holds “72 NVIDIA Blackwell Ultra GPUs” and that “The rack’s TDP (thermal design power) is 132 kW nominal with an EDPp (electrical design power โ peak) of approximately 155 kW”.
Dividing 1.2 million GPUs by 72 gives about 16,700 racks, by our arithmetic. At 132 kW each, that is about 2.2 GW. At the 155 kW peak it is about 2.58 GW.
HPE also writes that it “suggests that data centers should provision their bus ways to be able to support 192 kW EDPp”. Provisioned at that level, the same racks would need about 3.2 GW of electrical capacity, by our arithmetic.
What xAI Has Permitted
DatacenterDynamics reported on 11 March that “xAI has received permission to install 41 natural gas turbines at a site in Mississippi that will generate 1.2GW to power its data centers in the area”.
The report says the turbines, at a former Duke Energy plant site in Southaven, “will be used to power xAI’s Colossus 2 data center” and “will also supply power to the firm’s upcoming Colossus 3 data center in Mississippi”. xAI said on X it was “thrilled that MDEQ approved our permanent construction permit unlocking 1.2GW of self-generating power capacity”.
The court’s 30 October order listed at least 27 turbines on the Southaven site as of 7 April, with a combined capacity of 495 MW, as we calculated in the NAACP suit piece.
The Business Lever: GPUs and Permitted Power
As we read the figures together, the GPU count Musk posted implies a power need larger than the gas capacity xAI has permitted on site, and that permit is shared with a second data center.
The Structural Read
As we read HPE’s QuickSpecs, a GB300 NVL72 rack holds 72 Blackwell Ultra GPUs at 132 kW nominal, so 1.2 million GPUs means about 16,700 racks, by our arithmetic.
As we read DatacenterDynamics’ March report, the 1.2 GW permit covers 41 gas turbines in Southaven that serve Colossus 2 and the upcoming Colossus 3.
As we read the court’s 30 October order, as we calculated earlier, at least 27 turbines with 495 MW of capacity were on the site as of 7 April.
HPE, NVIDIA GB300 NVL72 by HPE QuickSpecs, 8 September 2026
“The rack’s TDP (thermal design power) is 132 kW nominal with an EDPp (electrical design power โ peak) of approximately 155 kW”
Three Implications
FOR XAI’S POWER PLANS By our arithmetic, the rack power for 1.2M GB300-class GPUs is about 2.2 GW against 1.2 GW of permitted on-site gas, a permit that also serves Colossus 3.
FOR DATA CENTER DESIGNERS HPE suggests provisioning bus ways for 192 kW per rack; at that level the same racks would need about 3.2 GW, by our arithmetic.
FOR THE POWER DEBATE Chamath Palihapitiya, quoting Musk’s post, wrote that power is THE binding constraint in AI; the figures above put numbers on that claim for one campus.
The framework says “Companies that solve the power equation will dominate the AI era; those that don’t will watch their chips sit idle.”
As we read it, Musk’s GPU count and xAI’s permitted gas capacity are the two sides of that power equation for one campus. The framework predates Musk’s post and does not discuss Colossus 2.
Musk’s post gives a GPU count, not a power figure, and xAI has not published Colossus 2’s total power draw in what we read. Our 2.2 GW assumes every GPU sits in a GB300 NVL72 rack at HPE’s published figures; Colossus 2 also runs GB200s, as Musk wrote in September, and their racks differ. The figures are rack power only, before cooling and other facility overhead. We did not size xAI’s grid supply, so the 1.8 times comparison covers only the permitted on-site gas capacity.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Musk wrote that Colossus 2 will hold around 1.2 million Blackwells by year-end. At HPE’s published GB300 NVL72 spec, that is about 16,700 racks and about 2.2 GW of rack power by our arithmetic, roughly 1.8 times the 1.2 GW of on-site gas generation Mississippi permitted xAI in March.
A note on sourcing. The GPU count is Elon Musk’s own post; the power figures are our arithmetic on HPE’s published GB300 NVL72 rack specification and assume every GPU sits in such a rack; the 1.2 GW permit is as reported by DatacenterDynamics. Neither Musk nor xAI has published Colossus 2’s total power draw in what we read.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
