At 3 a.m. Manila time the number landed in my feed like a bad tick: "Musk projects at least 13 gigawatts of AI compute by end of next year." I read it three times. No publication date. No direct quote. No split between contracted capacity, capacity under construction, and energized IT load. Two paragraphs, one of which was a copyright notice. The substantive payload was three sentences, all sourced to a single voice, all unverifiable against anything I can pull from a node.
This is a headline echo โ a second-hand reheat of an X post, dressed as reporting. It contributes exactly one thing to the record: confirmation that the string "13 GW" has entered the propagation chain. The article is not the object of analysis. The number is. And the number is interesting for a reason that has nothing to do with artificial intelligence.
It is a unit swap. Chip counts are auditable; gigawatts are not. That single substitution is the most important forensic detail in the entire episode, and almost nobody covering it noticed.
Context first, because you cannot audit what you have not defined. A compute claim lives or dies on its unit of measure. When a builder says "I have two hundred thousand GPUs," I can, in principle, cross-check that against purchase orders, data-center floor plans, satellite thermal signatures, and the customs manifests of the accelerator supply chain. When the same builder says "I have thirteen gigawatts," I can cross-check almost nothing. Gigawatts describe a state of the physical world โ energized, cooled, connected load โ and the word covers four completely different stages of a project that differ by a factor of five to ten.
That ambiguity is the whole game. Contracted capacity is a signature on a power-purchase agreement. Capacity under construction is steel and concrete. Planned capacity is a slide in a deck. Energized IT load is the only number that produces tokens. A claim that does not specify which of the four it means is not a forecast. It is an option on a narrative, priced at zero and sold as vision.
I have spent eighteen years watching people confuse the map for the territory, and I have learned that the tell is always the unit. In 2017, as a junior quant in Manila, I manually audited the Zilliqa genesis-block contracts and found an integer-overflow bug in the sharding transaction-batching logic. The fix delayed mainnet by two weeks. That episode taught me a habit I have never dropped: cite the contract address and the version number, or the claim does not exist. A number without a provenance is a rumor with a decimal point.
The provenance problem here is severe. The phrase "by end of next year" resolves to 2026 or 2027 purely on the basis of a publication date the source material does not give me. That is not a minor omission. It is the difference between a stretch and a fantasy. I flag it now so that every conclusion below carries its correct weight: the arithmetic is confidence-A, the physical constraints are A/B, and the verdict on intent is inference, graded B/C.
Now the arithmetic, which is where the argument actually ends. Thirteen gigawatts is an electrical dimension, not a silicon dimension. Run it through the standard conversions and the scale becomes absurd on its face. At full load, 13 GW draws about 114 terawatt-hours a year โ roughly 2.6 percent of total US generation in 2024, which was on the order of 4,300 TWh. At a realistic 60 percent utilization, it is 68 TWh a year, which is about 39 percent of all US data-center consumption in 2024 and roughly 15 percent of the entire global data-center load. In generation terms it equals thirteen one-gigawatt nuclear units, or about twenty-six 500-megawatt combined-cycle gas turbines. In human terms it is the residential demand of roughly ten million homes โ Georgia and South Carolina combined. Against a US peak load near 740 GW, it is a 1.8 percent instantaneous addition: a medium-sized country stapled onto the grid.
Translate that into hardware and the absurdity compounds. An H100 SXM at 700 watts thermal, loaded with CPU, networking, and cooling overhead, runs about 1.2 to 1.5 kilowatts per GPU at the facility level. Thirteen gigawatts is therefore somewhere between 8.7 and 10.8 million H100-equivalents. If you prefer the Blackwell generation, a GB200 NVL72 rack draws roughly 120 kilowatts for 72 GPUs โ about 1.67 kilowatts per GPU โ which lands near 7.8 million accelerators. At roughly $40,000 per GPU for GB200-class parts, the accelerator bill alone clears $310 billion. Civil works, substations, and cooling at $10 to $15 million per megawatt add $130 to $195 billion. The all-in capital requirement is $440 to $510 billion โ roughly the combined 2025 capital expenditure of Microsoft, Google, Amazon, and Meta, executed by a single company in a single year.
Set that against known progress and the gap stops being a rounding error. Public reporting puts xAI's Colossus 1 in Memphis at the 150 to 250 megawatt class, scaling from 100,000 toward 200,000 H100/H200 units. Colossus 2 is targeted at roughly the one-gigawatt class. Thirteen gigawatts is a thirteen-fold jump โ a full order of magnitude. Even at the most aggressive construction pace on record โ Colossus 1 reportedly stood up 100,000 accelerators in 122 days โ a build of this size is a three-to-five-year continuous program, not a next-year event.
The physical chokepoint is where the claim actually breaks, and it is the part the market refuses to price. Large power transformers carry lead times of two to three years. Switchgear and HVDC equipment are just as tight. Interconnection queues at PJM, ERCOT, and MISO run two to four years for large loads as a matter of routine. The only path that can land inside twelve to eighteen months is behind-the-meter mobile gas turbines, 30 to 50 megawatts each โ which means 260 to 430 units, plus fuel supply and emissions permits for every one of them. Gigawatt-scale campuses take twenty-four to thirty-six months from site selection to energization. If 13 GW of energized capacity were to exist by the end of next year, more than 10 GW would have to be under active construction today. Nothing in the public record supports that. That single sentence is the decisive evidence, and it is why I am confident the number is not energized load.
Here is where my own history colors the read. During the 2020 DeFi summer I built a Python script to track Uniswap V2 liquidity pools across more than 500 tokens, and found that 60 percent of new pairs showed wash-trading patterns before their public listing. The lesson was not that the volume was fake. The lesson was that volume could be manufactured faster than it could be verified, and that the manufacturing was itself the product. I see the same structure in power claims. A megawatt can be counted at the PPA stage, again at the construction stage, and again at the energized stage, and each count is technically true and collectively a lie. Following the exit liquidity to its cold storage is easy when the ledger is public. Following a gigawatt to its substation is not, because the substation does not publish a block.
Which brings me to the thing hashrate taught me. Bitcoin's network is the one large-scale compute claim in this industry that is continuously, publicly, and adversarially auditable. You do not take a miner's word for its capacity. You read the difficulty adjustment. The chain confirms all. When miners began converting their power contracts into AI and HPC hosting deals โ a real and accelerating trend โ they inherited the same disclosure problem that now sits at the center of this story. A signed megawatt is not a lit megawatt, and the market keeps paying for the signature. In 2026 I trained an anomaly-detection model on five years of on-chain data to flag wash-trading across new Layer 2 networks, and it surfaced a $50 million synthetic-volume scheme at a major venue. The method that caught it was not cleverness. It was triangulation โ cross-referencing one signal against an independent one until the manufactured part stood out. Applied here, triangulation means checking power claims against satellite thermal imaging, cooling-water discharge volumes, substation load telemetry, and the accelerator supply chain. Against those instruments, a bare gigawatt figure is a rumor in a nice font.
The metadata holds the provenance the price ignored. That is true of NFT collections with broken IPFS links, and it is true of a compute forecast with no publication date. When the article cannot tell me whether "next year" is 2026 or 2027, it has forfeited the ability to be wrong โ and a claim that cannot be wrong is not a claim. It is marketing.
Now the contrarian turn, because the lazy read is to call this a lie and stop. That would be a mistake, and it would miss where the actual money is. The engineering truth of the 13 GW figure is almost certainly "no." The demand signal it carries is almost certainly "yes." These are not contradictory. The declaration's function is not to describe a build schedule; it is to move capital and to trigger a race. A number this large, dropped by the most aggressive player in the field, does not need to be true to be effective. It needs only to be believed by competitors, who then accelerate their own power and equipment procurement to avoid being left behind. That is a self-fulfilling power arms race, and it is already visible in sold-out turbine order books running to 2028 and 2030.
If the demand signal is real, then the beneficiaries are not where the headline points. The downstream AI companies are competing on a resource whose upstream suppliers are the actual bottleneck. Heavy-duty gas turbines, large transformers, high-voltage switchgear, liquid cooling and CDUs, and behind-the-meter generation are the scarce inputs, and their order books are the hard evidence of the race. Upstream certainty is far higher than downstream certainty. The AI capability side faces a different problem entirely: the marginal return on raw pre-training compute has been falling for two generations. Going from GPT-4 to its successors, a tenfold increase in training compute produced benchmark gains smaller than the previous step. The binding constraint on capability has migrated from pre-training FLOPs to post-training method, data quality, and productization. Thirteen gigawatts buys more inference capacity and cheaper tokens per unit. It does not buy a capability leap, and the article's implication that it does is the second unit error in the piece.
Correlation is not causation, and the most dangerous correlation in this narrative is the one between headline capacity and equity value. A gigawatt announcement correlates with a funding round. It does not cause one to be well spent. I watched this movie in 2022. When Luna collapsed, I executed our fund's emergency protocol and liquidated 40 percent of our high-risk DeFi positions within hours, then built a correlation matrix that exposed the hidden leverage links between Celsius and Three Arrows Capital. The matrix did not predict the collapse. It revealed that the entities everyone treated as independent were the same trade wearing different names. The AI power complex has the same topology now: the same handful of turbine makers, the same transformer shortage, the same interconnection queue, the same political permission, all financing each other's narratives. When one node stumbles, the correlation is not zero.
The soft constraint nobody models is political. Data-center load has already pushed capacity prices up in PJM and lifted residential bills in Georgia and Virginia. A single 13 GW load, wherever it lands, will intensify that conflict and invite backlash โ the least-priced risk in the entire expansion. The geography gives it away. Thirteen gigawatts cannot all sit in Memphis. It implies cross-state and possibly cross-border power procurement, Gulf partnerships, ERCOT behind-the-meter generation, and nuclear PPAs that do not energize until the 2030s. Every one of those paths runs through a permitting authority, which means the real dependency is political convenience, not engineering.
What the article never asked is the only question that matters: how many megawatts of this are energized today? The absence of that question is the absence of journalism. And the choice of "at least" is not a floor. It is an unfalsifiable ceiling โ a phrase engineered so the number can never be caught being too small, while quietly implying it might be larger, which is precisely the shape of language that precedes a raise.
So here is what I am watching next, and it is not the next tweet. I am watching for the first disclosure that separates contracted from energized capacity in megawatts, because that is the number that produces tokens and the number a serious operator will eventually have to publish. I am watching transformer and turbine order books as the honest leading indicator, since equipment does not lie about its lead time. I am watching the next PJM capacity auction for the price of the political constraint. And I am watching the miner-to-AI conversions, because those operators already hold the one asset this entire narrative is really about: an energized, permitted, grid-connected load. The code doesn't care how loud the announcement is. The grid doesn't either.
The question is not whether Musk can build thirteen gigawatts. The question is whether the market can still tell the difference between a signature and a substation โ and how much it is willing to pay before it learns.