The number that moved me this past week was not one million. It was eighty to a hundred and twenty weeks โ the current US lead time for a large power transformer.
While the timeline shouted about a million GPUs rising in Memphis, I was watching the substation. That habit was mined in Lagos, not learned in a terminal: in 2020 I spent three months in a rented apartment tracking fifteen thousand Uniswap V2 pool transactions by hand while the crowd priced the headline. The loud metric is almost never the binding one. The quiet one โ the queue, the permit, the delivery window โ is where the timeline is actually decided. While the crowd shouted, I watched the exit.
So when a crypto outlet ran "xAI details expansion plans" above a body that contained no details, I did not read it as reporting. I read it as a transmission. A number had been released into the narrative economy with no enclosure around it โ no date, no denominator, no citation, and a summary that repeats the article almost verbatim, which is the signature of a relay with near-zero editorial cost.
That is the thing worth analyzing. Not the article. The number.
Start with the baseline, because the baseline is the only part anyone can verify. Colossus 1 came online in Memphis in July 2024 with roughly 100,000 H100s, stood up in 122 days, and was later expanded toward 200,000 mixed H100/H200 accelerators. Musk called it the most powerful AI training cluster in the world, and for a window of months that claim was defensible. Colossus 2 is the next phase: GB200 Superchips, more land, more power, and the figure that escaped containment โ over one million GPUs.
For readers of this column, the stakes are not abstract. The compute narrative has been financialized. DePIN compute networks, GPU tokenization, decentralized inference markets โ all of it is priced, in significant part, off the perceived physical ceiling of centralized AI. Every headline about a gigawatt campus is functionally a headline about the scarcity that the decentralized version is supposedly arbitraging. If the underlying number is soft, the trade built on top of it is soft in exactly the same place.
And the number is soft. What travels is not a fact but a migration.
Restore the caliber first. In early 2025, on X, Musk described Colossus 2 at roughly 555,000 GB200 Superchips and noted explicitly that each GB200 carries two NVIDIA GPUs. Multiply: 555,000 ร 2 โ 1.11 million. That is the origin of "over one million." Not fabrication, precisely. A unit conversion performed in public and then dropped on the floor.
The headline is a Superchip count doubled into a GPU count, and the doubling is where all the ambiguity lives.
Now run the physics, because the physics does not care how the number was told.
Power. A GB200 NVL72 rack draws roughly 120 kW. Per-superchip TDP sits in the 1,200โ1,400 W band; an H100 SXM is about 700 W. A million accelerators implies an IT load of 0.66โ0.9 GW, and at a PUE of 1.2โ1.3, a facility load of 0.8โ1.2 GW. MLGW's system peak is on the order of 2.5โ3 GW serving roughly 650,000 customers. One gigawatt is not an increment to that system. It is a third of it. In utility planning language this is not demand growth; it is restructuring, and it gets scheduled against everyone else's queue position.
Capital. Rack-scale GB200 systems run $30,000โ$40,000 per GPU all-in once you count the Grace CPU, the NVLink switch fabric, liquid cooling, optics and integration. A million units lands at $30โ40 billion of hardware before a single kilowatt-hour is purchased. Blended self-build assumptions put the range at $25โ40 billion.
Revenue base. Third-party reporting places xAI's 2024 revenue in the low hundreds of millions at best, largely from X integration and API surface. That is not a rounding gap against the capital plan. It is two orders of magnitude.
Which gives the honest ordering of constraints: power and firm capacity first; environmental permitting second, since Shelby County sits in an ozone marginal nonattainment area and xAI ran mobile gas turbines during the transition โ Solar Turbines class, roughly 5โ16 MW each, ten-plus units, 100โ250 MW of nameplate โ with the nitrogen oxide implications and the litigation from NAACP and the Southern Environmental Law Center that follow from that; capital third; chip supply fourth, and there the bottleneck is CoWoS-L advanced packaging capacity at TSMC rather than wafers, which means xAI is competing for the same allocation as five larger buyers; cooling fifth, because GB200 mandates liquid, and Memphis summer humidity is not a design assumption anyone chose for elegance.
GPUs rank fourth or fifth on that list. The headline leads with the item that is least scarce.
There is a second layer the transmission never approaches: effective compute decay. Node count scales linearly and MFU does not. Stragglers, checkpoint rollback, failure recovery โ at a million accelerators, a 20โ30% model FLOP utilization rate is not a rounding error, it is a multiple. The million is a gross weight. The number that actually trains a model sits downstream of interconnect efficiency, and xAI has chosen Spectrum-X Ethernet over InfiniBand, which buys an open ecosystem at the cost of harder collective-communication engineering at extreme scale. That tradeoff does not survive a relay piece, because a relay piece has no mechanism for reporting a tradeoff.
The same flattening applies to cumulative versus instantaneous. "One million GPUs" is far more plausibly a procurement arc across generations โ H100, then H200, then GB200, then GB300 โ than a simultaneously powered fleet. Media relaying erases that distinction, and anyone pricing an AI compute glut, or its absence, off this number is pricing a category error rather than a capacity figure.
Then the regulatory shadow, which is silent in every version I have read: a frontier training run at 10ยฒโทโ10ยฒโธ FLOPs crosses the 10ยฒโถ reporting threshold under US executive order 14110 and triggers dual-use foundation model obligations. The number that gets repeated as a boast is also a disclosure trigger. Nobody frames it that way, which tells you the framing is doing work.
Here is my own anchor for why this pattern is familiar. In 2024 I spent two months modeling BlackRock's spot-Bitcoin entry against long-term holder behavior for a piece called "From Speculation to Settlement." The transferable lesson was not about flows. It was this: institutional capital does not rewrite a narrative, it changes what the narrative is permitted to ignore. The ETF absorbed the volatility question and moved on. This buildout absorbs the physics question and calls it settled. Same cognitive move, different balance sheet.
Institutional empathy is not the same as institutional candor, and the gap between them is where retail narrative risk accumulates.
Which brings me to the contrarian read, because the consensus crypto interpretation of Memphis is that it is fundamentally bullish for decentralized compute. Centralized AI hits the wall, the spillover finds DePIN. I would push back on that, gently but firmly.
The crypto market prices GPU count. The scarce asset is firm capacity and interconnection. A million accelerators is a press release, and press releases are reproducible by any competitor with a banker and a board. A gigawatt of firm, permitted, non-interruptible load with an approved air quality permit is a moat โ three to seven years of queue, docket and hearing, and no amount of capital compresses it.
Here is the inversion worth sitting with. The reason containerized, off-grid, load-following compute is genuinely interesting is not ideological decentralization. It is that it skips the interconnection queue. Strip the whitepaper language and behind most decentralized compute pitches there is a backlog someone is routing around. Not sovereignty โ permitting arbitrage. The thesis is not trustless. The business case is "we don't need a transformer that arrives in 2027."
One more asymmetry, and it connects to something I have watched on-chain for years. Governance turnout across this sector sits below five percent on most proposals. The physical analogue is a Shelby County air quality hearing. The mechanism is identical: the quorum is smaller than the externality, and that is how a gigawatt gets approved by default. The whales who show up are not voting on your behalf; they are voting on their own queue position.
So watch the megawatt number, not the GPU number. When TVA and MLGW publish the firm capacity actually attached to Colossus 2, you will have the one figure that cannot be re-based, doubled, or retold โ the one that settles what the last twelve months of headlines were actually about. Until that number lands, treat "one million" as a transmission rather than a fact, and treat the speed of its migration as the real data point: how quickly an unverified figure becomes an anchor that portfolios, policy and pitch decks are quietly built around.
I do not trade tokens; I trade timelines. The ledger is cold, but the pattern is warm โ and the pattern here is that we are pricing the noise and ignoring the queue.