Meta's $27 billion Nebius deal is a ceiling, not a commitment

Only $12 billion of the $27 billion is committed

The five-year agreement, announced 16 March 2026 and recirculated by the wires this week↗, dedicates $12 billion of NVIDIA Vera Rubin capacity to Meta, with deliveries from early 2027 across Finland, the UK and the US ¹². The other $15 billion is an option spread over future clusters: Nebius offers that capacity to third-party customers first, and Meta commits to buy whatever remains ¹².

The headline figure is a maximum, not a purchase guarantee ¹. It sits on top of an earlier Meta arrangement worth about $3 billion over five years, announced 11 November 2025 ¹.

The converts fund the gap

Nebius closed $5.75 billion of convertible senior notes on 24 August 2026↗, upsized from a $5.0 billion pricing, after raising $775 million of secured debt in July ³. Balance-sheet debt already stood at $10.1 billion on 30 June ⁴. In the June quarter Nebius spent $5.65 billion of capex against $582 million of revenue, and spending has outrun sales in every quarter since late 2024 ⁴. Microsoft's September 2025 contract, $17.4 billion rising to $19.4 billion, is the anchor Nebius says let it borrow more cheaply ⁵.

Meta buys last, by construction

The fine print costs Meta nothing until capacity goes unsold. The $15 billion covers clusters Nebius had originally planned to sell to other AI customers, which makes Meta the residual bidder on whatever others decline ¹⁶.

Meta is renting everywhere while building its own data centers: $14.2 billion with CoreWeave in September 2025, expanded to $21 billion in April, on top of the Nebius agreements ⁶. A July Bloomberg report that Meta might rent out its own excess compute↗ knocked Nebius down 17 percent in a day ⁶. Management's answer on its May call was that several customers compete for every GPU it brings online ⁶.

A $100 million patch for idle GPUs

The week's genuinely new event is smaller. On 1 October Nebius bought Inferize, a 17-person Tel Aviv team founded in January 2026, on undisclosed terms↗; CTech estimated the price at $100 to $150 million ⁷⁸.

Inferize snapshots a warmed GPU serving engine and restores it in seconds, claiming 10 to 30 percent GPU cost cuts ⁸. Its own benchmark shows a 20-second start against a roughly 23-minute cold start, numbers no independent test has checked ⁸.

"Keeping spare GPUs running is the price of being ready for demand," chief executive Guy Bortnikov said ⁷.

Underwrite $12 billion, not $27

$12 billion is contracted. The $15 billion becomes revenue only where better buyers fail to appear, which is the opposite of an anchor tenant. The Vera Rubin racks arrive from early 2027 on a convert-funded balance sheet, and NBIS equity plus the August convert holders carry the utilization and residual-value risk on that fleet. The idle-GPU tax Inferize was bought to cut is the carrying cost of the gap in between.