The GPU cluster used to be the dragon hoard: build it, guard it, maybe name a few racks after Greek gods if procurement got sentimental. Now, according to Forbes, Anthropic is looking at a proposed $10 billion compute deal with rival Meta while also carrying a $1.25 billion monthly agreement with xAI. That is not just awkward dinner seating at the frontier AI banquet, it is a signal that capacity has become more important than cloud monogamy. For builders, the lesson is not that everyone should start texting competitors for spare accelerators like college students hunting for a couch. The lesson is that frontier AI infrastructure is splitting into model work on one side and power, chips, permits, leases, and construction risk on the other. The model may be the product, but the outlet in the wall is increasingly the plot. ## Forbes: The Cloud Loyalty Program Has Been Canceled Forbes reports that Anthropic is diversifying beyond traditional cloud providers through several large compute arrangements: a proposed $10 billion deal with Meta, a $1.25 billion monthly agreement with xAI, a 20 year, $19 billion lease with TeraWulf, and a $50 billion commitment for custom facilities with Fluidstack. That list reads less like a vendor strategy and more like a scavenger hunt designed by a CFO who has not slept since ChatGPT happened. Still, the logic is brutally practical: secure capacity wherever capacity can actually be delivered. According to Forbes, the broader pattern is the separation of model development from infrastructure ownership, with construction, financing, and permitting risk spread across partners such as neoclouds and competitors. That is a big deal because the bottleneck is no longer merely who has the prettiest benchmark chart. Forbes also points to Oracle's Project Jupiter for OpenAI as an example of these challenges, noting permitting hurdles, which is the least glamorous possible way for AGI discourse to meet municipal paperwork. ## Nextomoro: Inference Moves Into The Rival’s House Nextomoro adds useful operational texture to the xAI side of the story, reporting that Anthropic announced at its second annual Code w/ Claude developer conference that Claude inference would begin running on Colossus 1, xAI's Memphis data center, "in the next few days." That is the part worth underlining in fluorescent marker: this is inference capacity, not just a trophy cluster for training runs. Inference is where models meet users, latency, uptime, and the depressing reality that every token has a bill attached. Nextomoro also reports that xAI had already moved its own training workloads to Colossus 2, a larger Blackwell based successor facility. Technically, that makes the arrangement less absurd than it sounds at first glance. If one lab has moved training elsewhere and another needs inference throughput, the rival relationship starts to look like airline code sharing, except the planes are GPUs and everyone onboard is asking for chain of thought. ## Tech Jacks Solutions: Do Not Copy The Whale Tech Jacks Solutions frames the Anthropic and Colossus pattern as part of a dual track strategy: hyperscaler distribution plus dedicated training compute. Its warning is blunt and useful, stating that this approach requires frontier lab capital most enterprise AI vendors do not have. In other words, do not look at Anthropic renting vast rival capacity and conclude your company needs a bespoke compute empire before lunch. Tech Jacks Solutions puts the pressure neatly: "Frontier labs don’t control their own destiny when they rent compute by the hour." That sentence is the whole procurement drama wearing sensible shoes. Renting capacity buys speed and flexibility, but it also creates dependency, which is fine until your roadmap, margins, and launch calendar all depend on someone else’s transformers having electricity. ## Epoch AI: The Scarcity Is Systemic Epoch AI's Josh You supplies the zoomed out version, writing that frontier labs do not yet use most AI compute. The same analysis says global AI computing power has grown to the equivalent of around 20 million Nvidia H100s, funded by hundreds of billions of dollars in annual capital expenditures. It also notes Nvidia's AI related sales spiked more than fourfold in 2023, which is a very polite way of saying the shovel seller noticed the gold rush. That context makes the Anthropic strategy feel less like a one off oddity and more like a preview of frontier procurement under scarcity. If Anthropic and OpenAI grow their share of compute, as Epoch AI suggests they may, the pressure on power, chips, facilities, and approvals only gets sharper. For readers building AI products, the practical move is to treat compute like a risk surface: diversify suppliers, understand workload portability, and avoid architectures that assume one vendor will always have infinite capacity at friendly prices. Watch the next frontier lab deals less for brand names and more for delivery mechanics: who has power, who has chips, who can clear approvals, and who can make the economics work without turning every prompt into a tiny mortgage. The model race is increasingly a logistics race wearing a lab coat. Turns out intelligence may scale, but first somebody has to find a socket. ## Sources - Frontier AI Labs Are Renting Compute From Their Competitors - Forbes

Sources