Your 2028 Inference Bill Is Being Financed Right Now
The week's biggest AI number is a letter of intent. The most important one is $12.5 billion of debt in a single data centre — and it should change how confidently you extrapolate falling inference costs.
The week's biggest AI number is a letter of intent. The most important one is $12.5 billion of debt in a single data centre — and it should change how confidently you extrapolate falling inference costs. What happened Two announcements, four days apart, that most coverage got subtly wrong. 24 July — NVIDIA and SK Group announced what the release calls a "$500-billion-plus comprehensive partnership" spanning AI factories and next-generation memory. SK Telecom will build a 2-gigawatt AI factory on NVIDIA's DSX full-stack architecture, deploying Vera Rubin accelerated computing powered by SK hynix HBM4 , with 50,000+ GPUs in the first phase, the first facility online in 2027 , and capacity sold onward partly as GPU-as-a-service. The detail almost everyone dropped: it is formalised through letters of intent . Non-binding. Multi-year. Not committed capex. Several outlets reported $500 billion as though a cheque had been signed. Reporting also placed a $1B NVIDIA investment into Naver (adding roughly 200MW) on the same day, and tallied roughly $950 billion in AI agreements signed across that 24 hours — though that total is a journalists' aggregate across many announcements, not a company disclosure, and we'd treat it as directional colour rather than a fact. 28 July — Meta and BlackRock announced a joint venture to develop and own a 1-gigawatt AI data-centre campus in El Paso, Texas , at roughly $14 billion total development cost. This one has a structure worth reading closely: Element Value --- --- Total development cost $14B BlackRock-managed funds (with GIP, HPS) 80% Meta 20% Meta's contribution Land + in-progress construction, $2.3B BlackRock's cash contribution $4.9B Debt financing $12.5B Distribution back to Meta $1B First capacity online 2028 Tenancy Meta is sole tenant , leasing the entire facility It was announced the day before Meta's Q2 earnings. Why it matters Read the Meta table as a whole and it stops looking like a technology announcement. This is infrastructure finance — the structure used to build toll roads, pipelines and telecom networks — applied to compute. An operator contributes assets and expertise. An infrastructure fund contributes equity. Debt does the heavy lifting. The operator signs a long-dated lease and takes some capital back out. Meta ends up with a gigawatt of capacity while consolidating only 20% of the asset and pulling $1B out the door. Here's the consequence that should reach your planning documents. For roughly two years, "inference gets cheaper on a predictable curve" has been a free assumption. Nobody had to defend it. But the capacity behind that curve is now debt-financed and investor-gated , and it arrives in 2027–2028 . With $12.5B of debt in a single campus, cost of capital becomes a direct input into how fast capacity shows up. That is a materially different world from "a hyperscaler decides to spend its earnings on GPUs," and it means the smooth cost-per-token decline of the last two years is not somethi