AMD and Meta each signed a $14 billion data-center deal this week. The $14 billion means something different each time.

The important change this week is not that two more AI companies signed big data-center deals -- that happens most weeks now. It's that AMD and Meta separately arrived at the same headline number, $14 billion, through completely different financial mechanics, on completely different timelines, for completely different reasons. Read past the headline and the two deals barely resemble each other.

The AMD deal: a chipmaker becomes a tenant, and a shareholder

AMD signed 15-year leases (with three five-year renewal options) for 530 MW of data-center capacity across five Core Scientific sites: 377 MW directly to AMD in Pecos and Hunt County, Texas, and Muskogee, Oklahoma; 152 MW to an AMD-affiliated neocloud at Auburn, Alabama, and Dalton, Georgia. Delivery is staggered -- Pecos and Auburn come online in the first half of 2027, Muskogee and Dalton in the second half, Hunt County by the first half of 2028, with the full 530 MW live by the end of 2028. AMD also holds a reservation right on up to an additional 1,925 MW at the same sites through December 28, 2028, which is where reporting's "up to 2.5 GW" framing comes from -- that capacity is optioned, not contracted.

As consideration, AMD received a warrant for up to 30 million Core Scientific shares at a $23.47 strike price, with roughly 6.5 million shares vesting immediately at signing. The $14 billion figure attached to this deal, per Core Scientific's own management presentation, is potential base contracted lease revenue over the full 15-year term -- the landlord's expected topline, not money AMD is spending upfront.

The Meta deal: an 80/20 joint venture, not a lease

Meta and BlackRock's $14 billion is a different kind of number: the total projected development cost of a 1-gigawatt AI-optimized campus in El Paso, structured as a joint venture in which BlackRock-managed funds own 80% and Meta retains 20%. Meta contributes roughly $2.3 billion in land and construction-in-progress assets; BlackRock puts in about $4.9 billion in cash, partly backed by $12.5 billion in debt financing raised for the project. Meta then leases capacity back from the venture under an initial four-year term with four additional four-year extension options -- a structure that can run 20 years. The campus is expected to support more than 4,000 construction jobs at peak and roughly 300 permanent staff once operating, with capacity coming online in 2028.

Same $14 billion, different denominator

Divide each headline number by what it actually buys and the two deals stop looking like variations on a theme. Meta's $14 billion is a construction budget for 1,000 MW: about $14 million per megawatt of build cost, financed mostly by BlackRock's balance sheet rather than Meta's. AMD's $14 billion is Core Scientific's own estimate of what it can collect in rent across 530 MW over 15 years -- worked out, at roughly $1.76 million per megawatt per year, to a landlord's revenue forecast, not a construction line item. One number describes what it costs to build the building. The other describes what the building might earn once it exists. Neither is a number AMD or Meta is writing a check for today.

What the two deals do share is a financing pattern that keeps recurring across this build-out: neither AMD nor Meta is fully funding or owning the facility it needs. AMD is a tenant with an equity kicker; Meta is a 20% owner leasing back its own campus from the majority holder that financed it. It's the same asset-light logic behind the [up to $250 billion guarantee Nvidia is reportedly weighing for OpenAI's Ohio data center](#/article/nvidia-openai-250-billion-ohio-data-center-financing) -- compute demand has outrun any single company's appetite to carry gigawatt-scale construction on its own balance sheet, so the infrastructure gets financed by whoever specializes in financing infrastructure, and the AI company signs a long lease instead.

The actual constraint hasn't moved

Neither deal changes what's actually scarce. Full delivery on both lands in 2028 -- these are power and construction timelines, not chip-availability ones. AMD is reserving capacity years before it needs to fill racks with Instinct GPUs; Meta is locking in a site and a grid interconnection years before the campus can run a single training job. The bottleneck this quarter is still power and buildable land, not silicon -- which is exactly why the deals are structured as long-dated leases and joint ventures rather than purchase orders.