The AI Boom Has a Gigawatt Accounting Problem

Today opens on EE Times’s “The AI Boom Has a Gigawatt Accounting Problem”, which states $100bn. $100bn is stated. The copy does not say whether it is a queue request, a signed commitment, or a plan, so the figure is not yet an instrument a committee can use. A desk can underwrite a stated figure. It cannot underwrite a mood, and it cannot treat a request as if it were a signature.

Set the lead against the other filings that actually state a number. TheElec’s “LG Electronics Signs 5GW Cooling Deal for North American AI Data Centers”, which states 5 GW 5 GW is written as committed. That is the figure that can move a book, and it still has to be read against whether the copy also stated the site, the date, or the megawatts. TechCrunch’s “AI computing startup Lambda to raise $4B ahead of planned IPO”, which states $4bn and $14.5bn That figure is stated, and the copy does not say whether it is a request, a signature, or a plan. Of the 114 stories filed today, 25 state a price or a megawatt. The rest can be accurate and still leave supply where it was. A committee that reads the unnumbered file is reading noise.

20 filings from earlier in the week stay on the book, and only because each is still large: a blockbuster, at least $250 million, or at least 300 MW. SiliconANGLE’s “Reflection AI debuts open-source Beam model with 501B parameters”, which states $25bn. Financial Times’s “Wall Street banks launch record $60bn chip deal for Broadcom and Anthropic”, which states $60bn. Converge Digest’s “Lambda Closes $1 Billion Loan for Contracted GPU Deployments”, which states $1bn and $1.01bn. They do not lead. Set that against this morning. Capital already spoken for and a figure that just filed are different risks, and the essay leads on the increment.

Attention and risk are not the same map. Today put 114 United States stories on the file, and 20 from earlier in the week remain because they are still large. The thickest desk is Models & Labs, with 56 stories. That is where the file spent its attention. Of the 134 stories on the file, 44 state a price or a megawatt. The chart counts stories, and a story on two desks is counted twice. It is not a measure of megawatts. A busy desk with nothing a committee can take is still a quiet desk for anyone who has to sign.

On Power & Sites, TheElec’s “LG Electronics Signs 5GW Cooling Deal for North American AI Data Centers”, which states 5 GW 5 GW is written as committed. That is the figure that can move a book, and it still has to be read against whether the copy also stated the site, the date, or the megawatts. On Chips & Supply, TheElec’s “​NEXTIN Wins 25.44 Billion Won Wafer Inspection Equipment Order From SK hynix” The copy states neither a price nor a megawatt. Without one of those, the filing cannot change what a committee is able to approve. On Technology, TheElec’s “​Microchip, Navitas Unveil 800V Power Reference Design for AI Servers” The copy states neither a price nor a megawatt. Without one of those, the filing cannot change what a committee is able to approve. On Capital, TechCrunch’s “AI computing startup Lambda to raise $4B ahead of planned IPO”, which states $4bn and $14.5bn $14.5bn is stated. The copy does not say whether it is a queue request, a signed commitment, or a plan, so the figure is not yet an instrument a committee can use. On Policy, Utility Dive’s “States, renewable developers back MISO’s accelerated interconnection queue plan” The copy states neither a price nor a megawatt. Without one of those, the filing cannot change what a committee is able to approve.

A forecast stays for the day it filed and then goes, however round the number. A projected market is not a constraint. The constraint is a figure someone has requested, signed, or left unnamed.

A megawatt filing still has to land in a hall that can take the rack. The kilowatt per rack is how many of those megawatts fit in the row, and the square feet of the building are a separate reading from the same facilities guide.

The rack a hall is being built for is not the chip on the earnings call. NVIDIA shipped GB200 NVL72 in 2024 and GB300 NVL72 in 2025, both 72 GPUs on the same MGX footprint. Vera Rubin NVL72 is in production and ships in the second half of 2026, with 72 Rubin GPUs and 36 Vera CPUs. It is built to run 100 percent liquid-cooled at a 45°C water inlet, on a liquid-cooled busbar rated to 5,000 A. A hall already on liquid cooling does not redesign the CDU plant to take it. That is NVIDIA’s rack note of 16 March 2026.

The electrical job is the one that changes the drawing. An MGX 800 VDC power rack, also the second half of 2026, takes the building’s existing AC and hands 800 VDC to the compute racks in the row. The building electrical system stays. The row power center, expected in 2027, is an overhead 800 VDC busway up to 2 MW per row. Kyber, which NVIDIA’s 13 October 2025 OCP note dated to 2027, puts 144 GPUs in one rack on up to 18 vertical blades and moves the in-rack parts from 54 VDC to 800 VDC. The same copper carries over 150 percent more power, which is how a 200 kg busbar comes out. NVL576 is eight racks of 72, not one chassis. Feynman on Kyber is a scale-up to 1,152 GPUs, and NVIDIA has not dated that rack. A hall still on the drawing board is the one that takes medium voltage straight to 800 VDC. The 2 MW figure is the row bus.

The kilowatt is what one of those racks asks of the hall. NVIDIA’s DGX GB rack guide states about 120 kW of consumption. The GB300 NVL72 reference states up to 142 kW for the full rack. The DSX facilities guide states a 330 kW cabinet TDP for Vera Rubin NVL72, up from an MGX Gen 1.1 cabinet at 198 kW. A megawatt of that stated rack power is 8.3 GB200 racks, 7.0 GB300 racks, or 3.0 Vera Rubin racks. Kyber and Feynman have no stated kilowatt. The same DSX guide sizes an example site at 972,500 square feet of building for a 250 MW-class IT load, which is 3,890 square feet per megawatt and 2.57 megawatts per 10,000 square feet of that building.

Henry Hub is $3.194 a MMBtu. Run through a 7.2 MMBtu/MWh combined-cycle heat rate and a 1.3 PUE, the fuel in a powered kilowatt is $21.82 a month. The capacity-weighted lease ask, from the leasing book and not from this morning's wire, is $160.62. The lead GPU converts a MMBtu into 792 dollars of rental revenue.

The price that can stop a project is not the cheap one. Fuel at $21.82 a kilowatt-month is a small number beside a lease ask of $160.62. That gap is the trap. The electron looks cheap until the project cannot get one, on a date, at a site. An interconnection filing outranks a note about cheap gas, because the filing is about the right to draw and the note is about the commodity.

Rackline's scarcity index is 78 out of 100, which it calls tight, and development pressure is 79. Read the score as a compression of those prints, next to today's filing. It is not a forecast of rent, and it is not a price. A tight score with a queue full of requests is a market that has announced more power than it can deliver.

Amazon is 33 percent of the latest quarter's combined capex. That split is the risk, not the size of the total. A pause by the largest buyer would move the queue behind it, and the queue would not see the pause until the next cash-flow statement. The total is the context. Q2 2026 capex for Microsoft, Alphabet, Amazon and Meta combined was $165.0 billion, up 27.1 percent on the quarter before it and up 87.1 percent on the year. Four times that quarter is a $660 billion run rate, against guidance of $735–750bn. This is corporate capex, cash paid for property and equipment. None of the four discloses a data-center share, and Rackline does not estimate one.

Inside the latest quarter, Amazon is 33 percent, Alphabet is 27 percent, Microsoft is 22 percent, Meta is 18 percent. The chart is the last four quarters, company by company, so a reader can see who moved and who did not. The guidance trail is the other half of the same concentration. The four-company total, on bases the companies do not state identically, was re-marked as follows: January 2026, $640bn and up; April 2026, $695–725bn; July 2026, $735–750bn.

Trailing four quarters of the same cash capex come to $510.8 billion. The peak quarter on file is Q2 2026 at $165.0 billion, which is the quarter just printed. A separate table, trailing four quarters through Q1 2026, sets that spending against operating cash flow. Amazon's capex in that window was $151.0bn, 102 percent of operating cash flow. The four together were 67 percent, on $433.9bn. A share above 100 means the buildout spent more cash than the business produced. That table stops before the latest quarter, and Rackline is not going to extend it.

Today's wire can announce. The 45-day book is the control on that announcement. It keeps what was signed, or reported as signed, and it drops a target, a tender, and a story that only says a company is in talks. It is led by $4bn — TechCrunch, “AI computing startup Lambda to raise $4B ahead of planned IPO”; $8.2bn — The Next Platform, “AMD Buys World Labs For $8.2 Billion To Get AI Model Cred”; $3.36bn — Data Centre Review, “Nscale raises another $3.36 billion ahead of its IPO”; $37.5bn — Converge Digest, “Flex Secures $2B Investment for Axiom at $37.5B Enterprise Value”. A figure from this morning that cannot get onto this list did not change who owns or controls capacity.

The bars are the money the filings actually state, in the currency they state it, ranked by millions only so the lengths can be compared. Rackline does not convert pounds into dollars to make the chart neater. A deal that states megawatts and no price stays in the sentences and off the bars, because a megawatt and a dollar are not the same commitment.