Google deal at $920 million a month highlights how AI compute pricing is hard to compare

Alphabet’s Google has reportedly signed a contract for AI compute worth $920 million per month, with delivery running through the middle of 2029, according to a market report from 24/7 Wall St. The post says the agreement involves access to around 110,000 Nvidia GPUs, positioning the cost as a straightforward announcement of demand pressure in the AI infrastructure buildout.

The same report compares that figure with an earlier deal attributed to Anthropic, which it says pays $1.25 billion per month for access to 220,000 GPUs. On those two reported monthly totals and GPU counts, the market post calculates that Google’s effective price per GPU is a little over two times higher than the Anthropic-per-GPU figure.

The arithmetic has intuitive appeal, but it also raises a key question: if the GPUs are Nvidia hardware in both cases, why would the unit economics diverge so sharply? The 24/7 Wall St analysis argues that one reason for the gap could be differences in the “mix” of GPU types, noting that the contract specifics are not publicly available. Another possibility raised is that Anthropic could be benefiting from a bulk-buy or volume discount effect.

The post also points to timing and market conditions as a driver, describing the AI data center buildout as moving toward a more extreme phase in which supply and demand constraints intensify. In that framing, even as large hyperscalers use custom silicon, Nvidia GPUs remain a central option for scaling frontier-model training and deployment, which can keep pricing power with GPU and systems providers.

Notably, the report does not claim that Nvidia negotiated a particular pricing structure tied to one customer versus another, and it does not provide details on performance characteristics, network bandwidth, power delivery, or software stack terms. Instead, it uses the limited public numbers it cites, and then focuses on what might explain the premium, including potential contract terms that do not show up in a simple GPU count.

For Nvidia, the relevance of such deals is straightforward: Nvidia’s data center GPUs and the broader AI computing ecosystem underpin much of today’s training and inference infrastructure. If the market is effectively pricing GPU access at increasingly premium rates, that generally supports the idea that supply-constrained compute remains valuable even for very large buyers.

Still, there is substantial uncertainty around the exact comparability of these contracts. The report itself acknowledges that the share of different GPU models is not disclosed, which matters because “GPU count” can mask variation in capabilities and configurations. It also does not establish whether the deals include equivalent quantities of supporting infrastructure such as storage, interconnect, hosting terms, or capacity guarantees.

What to watch next is whether the companies involved provide more detail on capacity, hardware configurations, and service terms in future announcements or filings. More transparency could clarify whether Google’s quoted compute comes at a genuine premium for equivalent hardware and conditions, or whether the apparent difference is mostly explained by how the workloads and GPU varieties are structured.