Picture the next Meta hyperscale campus under construction somewhere in the American Midwest. Floor plans stacked with Nvidia GPU racks. Now add standalone CPUs to the blueprint — and you have the chip order CNBC reported this week.

Meta has agreed to buy millions of Nvidia AI chips for a data center build-out that, for the first time at this scale, includes standalone CPUs, CNBC reported. The expanded agreement deepens a supply line running between two of the AI infrastructure boom's biggest buyers and sellers.

The headline sounds routine. The detail is not. Per CNBC, Meta is buying discrete Nvidia central processors — not just accelerators — a procurement pattern that pulls Nvidia past its GPU franchise and into the general-purpose compute stack Intel and AMD have owned for decades.

What's actually in the deal?

The CNBC report crystallizes three terms: millions of AI chips, a data center build-out, and standalone CPUs. The network's headline summary did not publish per-unit pricing, wafer counts, or delivery cadence. What it commits Meta to is volume at a scale that, even at conservative average selling prices for current-generation AI silicon, sweeps the contract into the multi-billion-dollar tier alongside other leading Nvidia buyers.

For Nvidia, "millions" places Meta into the same customer stratum as Microsoft, Google, and Amazon. Adding standalone CPUs broadens the deal from accelerator racks into full server nodes — where Nvidia will now face off directly against the x86 incumbents.

Why does the CPU line matter?

Nvidia's central processor work — anchored by its Grace line and any successors — has historically stood in the GPU business's shadow. Pulling those parts into a Meta rack shifts the strategic geometry. Sockets that, until recently, would have gone to AMD EPYC or Intel Xeon now flow to Nvidia. CNBC's "including standalone CPUs" is the single line that will draw the sharpest attention from AMD and Intel.

Hyperscalers already running custom Arm silicon in production now have a second commercial-grade Arm alternative on the table from Nvidia. The competition has been building for years. Volume is what changes.

What about Meta's in-house silicon?

Meta's custom AI accelerator program — the Meta Training and Inference Accelerator, or MTIA — has been generating chips now shipping in production. The Nvidia expansion does not displace that roadmap. It supplements it, giving Meta a faster path to scaled training and inference capacity while MTIA yields ramp on internal workloads.

Hyperscalers rarely run a single-vendor compute strategy. They blend merchant silicon from Nvidia, AMD, and Intel with homegrown parts. CNBC's report hints Meta is leaning harder into the merchant side of that mix for the current cycle.

What the supply chain reads from this

A multi-million-unit Meta order lands hard on the upstream wafer chain. Most of Nvidia's accelerator and CPU silicon comes off TSMC's advanced nodes, with advanced packaging — CoWoS variants from Amkor, ASE, and SPIL — already running at full utilization through 2025. A Meta order of this reported magnitude extends that bottleneck rather than relieving it.

Networking follows compute. Nvidia's Spectrum-X Ethernet and Quantum InfiniBand switches typically ship alongside its compute platforms. If Meta's fabric orders scale proportionally, longtime default partners Arista and Broadcom should expect Nvidia to take share on these specific racks.

What to watch next

Meta's next earnings call will put hard dollars on the announcement. Investors will press management on capex guidance, the split between Nvidia, AMD, and MTIA in the build pipeline, and the timing of new campus deliveries. Analysts will also ask whether Meta plans to commercialize any of this capacity — a question Mark Zuckerberg has so far deflected in detail.

For now, the headline is the story: millions of chips, an expanding Nvidia partnership, and a CPU line that just got considerably more interesting for everyone competing with it.