TikTok USDS Wants 4 Gigawatts by 2030 — and the Turbines to Serve It Start Shipping in 2029

CNBC's Oct 23, 2026 report describes a capacity search, not a construction programme. A TikTok US spokesperson said the company \"runs primarily on Oracle Cloud Infrastructure in the U.S. and Oracle will continue to be our primary cloud infrastructure provider.\" The same report points to a TikTok job advert describing a \"flagship B2B AI product at the frontier of large-model inference and creative automation,\" and to the obligation to retrain, test and update the recommendation algorithm on US user data. Leaves the size, the site, the power supplier and the contract value undisclosed

TikTok USDS Joint Venture LLC was established on Jan 22, 2026, with Silver Lake, Oracle and MGX each holding 15%, ByteDance retaining 19.9% and other investors holding about 35%. The board has seven members and is majority American. The recommendation algorithm is licensed from ByteDance, retrained by Oracle, and must be secured in Oracle's US cloud environment, where US user data lives; Oracle is also the venture's trusted security partner and one of the seven directors is an Oracle executive. The structure makes Oracle landlord, security partner and 15% shareholder in the same transaction

Convert the target into electricity and it stops sounding like a software order. 4 GW running flat out draws about 35 TWh a year; at the load factors large AI campuses actually run, roughly 21 to 25 TWh. Berkeley Lab's US Data Center Energy Usage Report puts total US data-center consumption at 176 TWh in 2023, rising to between 325 and 580 TWh by 2028 — 6.7% to 12% of all US electricity. On those assumptions, TikTok's target absorbs 4% to 8% of all US data-center electricity in 2028

Turbines are not scarce in price; delivery slots are scarce in time. GE Vernova's gas power equipment backlog and slot reservation agreements grew from 100 GW to 116 GW in the second quarter of 2026, and management expects at least 125 GW under contract by year-end. In that same quarter the company signed 20 GW of new gas equipment contracts, converted 10 GW of reservations into firm orders, and shipped 3 GW of equipment. That gap between bookings and shipments makes the queue, not the order book, the product being sold

The clearest marker of the lag comes from Doosan Enerbility. Its March 6, 2026 contract with a US client for a data centre covers seven 380MW gas turbines, with deliveries starting in May 2029 and running one turbine and one generator a month. That takes Doosan's US delivery backlog to 12 turbines — about 4.56 GW of nameplate capacity, or roughly 1.1 times TikTok's entire reported target, from one Korean supplier. Mitsubishi Heavy told Nikkei Asia in September that first-half 2026 orders are running at or above last year's pace on US data-center demand. A 4 GW program absorbs about 3.4% of GE Vernova's 116 GW pipeline and roughly 20% of one year's turbine output at the 20 GW run-rate, which means makes TikTok a small customer of a very long queue

Here is where the arithmetic turns uncomfortable. Oracle's first-quarter FY2027 revenue was $19.3 billion, up 30%, with cloud infrastructure revenue up 121% to $7.4 billion and remaining performance obligations up $209 billion year over year to $664 billion. The company delivered 850 MW of additional data-center capacity in the quarter and spent $28.5 billion of capital while running free cash flow at minus $5.4 billion. Those numbers make a four-gigawatt anchor a rounding error next to a $114 billion annualised capital program

ByteDance is not short of money: it secured a $29.6 billion unsecured loan from nearly 30 banks in September, with proceeds going to projects outside China. That money cannot answer this question. ByteDance holds 19.9% of a venture with a majority-American board, whose algorithm must live in Oracle's US cloud, whose compute would be built on US soil, and whose builder must pay for its own generation and grid upgrades under the March 2026 pledge. A Chinese parent's corporate loan funding a majority-American compute platform is exactly the structure the carve-out was built to prevent. Makes project-level US financing the only workable route

The template for that financing is the neocloud model, and it shows its price. CoreWeave spent $20.6 billion of capital over the twelve months to Jun 30, 2026 against $6.9 billion of operating cash flow, ran free cash flow at minus $13.7 billion, paid $1.9 billion of interest on $7.6 billion of revenue, and ended the period with a 0.46 current ratio and net debt at 11.5 times EBITDA. A four-gigawatt program financed that way needs contracted take-or-pay revenue from a creditworthy counterparty, or a balance sheet that can absorb the gap. That comparison puts the counterparty risk on Oracle rather than on the lenders

For a program of this size, power is bought from owners of existing firm generation, because new-build takes years and interconnection queues take longer. Vistra's 20-year agreements with Meta cover 2,609 MW from Perry, Davis-Besse and Beaver Valley, including 433 MW of uprates and more than 15% of contracted capacity as genuinely new supply to PJM, with purchases starting in late 2026 and the full 2,609 MW online by 2034. Constellation's 20-year Illinois agreement with Meta covers 1,121 MW. That structure makes the sellers of firm megawatts the longest-lived winners in the chain

The scale is not an American quirk. Data Center Dynamics reported on Oct 6, 2026 that ByteDance was in talks with local data-centre providers in Ulanqab, Inner Mongolia, to add between five and six gigawatts of compute, and the South China Morning Post reported the same month that ByteDance now takes up roughly one-fifth of China's data-centre capacity. Shows one company running the same power-and-turbine play on two grids with opposite ownership rules