Neoclouds and the Great Unbundling

As of March 31, 2026, CoreWeave reported a contracted revenue backlog of $99.4 billion. Sit with the number for a second. A company that was mining Ethereum a few years ago, that went public in March 2025 at $40 a share to considerable skepticism, carries committed future revenue approaching the GDP of a small country, with quarters like Q2 2025's $1.21 billion, up 207% year over year, and customers like Nvidia placing $6.3 billion orders.

For fifteen years, the cloud industry's foundational assumption was that the bundle always wins: nobody beats AWS because AWS has everything, and everything is what enterprises buy. CoreWeave, Lambda, Nebius, and the rest of the GPU-first neoclouds just falsified that at nine figures a quarter. My claim is that the falsification matters more than the companies. Whatever happens to any particular neocloud, they've taught a generation of buyers to shop outside the bundle, and that habit doesn't reverse.

How the bundle cracked

The opening was supply: AI demand outran hyperscaler GPU capacity, and a buyer who can't get H100 allocations doesn't care how many other services the vendor offers. But the neoclouds kept the customers supply alone can't explain, because specialist economics turn out to be real. A cloud that sells one thing doesn't carry the tax of two hundred others: no army of half-maintained services, no bundle pricing designed for cross-subsidy, datacenters engineered for exactly one workload shape. For training runs, that focus shows up as price, availability, and performance the generalists struggled to match, and sophisticated buyers noticed.

CoreWeave isn't alone in the cohort, which is part of the proof. Nebius arrived by the strangest route available, carved out of Yandex and relisted on Nasdaq as a European-rooted AI infrastructure company. Lambda grew from selling deep-learning workstations into a GPU cloud with its own gravity. Different origins, different balance sheets, one shared bet: that a cloud can be narrow and win. The strongest confirmation came from the incumbents themselves, who quietly became neocloud customers, Microsoft most famously among them. When the bundle buys from the unbundlers, the argument about whether specialists can compete is over; what remains is negotiating the price.

The noticing is the historic part. AI forced the most conservative procurement departments on earth to unbundle one workload, evaluate it on its own merits, and sign with a vendor whose logo their board had never seen. That's a psychological dam breaking. Once a company has split its GPU spend from its general compute, "we buy everything from one cloud" stops being a law of nature and becomes what it always was: one option, with a price.

What unbundles next

Storage went first, quietly, years ago; the Backblazes and Wasabis proved a single-service cloud can undercut the bundle when the service is a commodity. Databases are mid-unbundling now, with specialist Postgres vendors peeling the most valuable managed service out of the bundle one developer at a time. Inference looks next: unlike training, it's latency-sensitive and spiky, which invites both edge specialists and scale-to-zero economics, and my bet is it splits from training procurement within a couple of years.

And then there's the unbundling we have obvious skin in: ordinary small workloads. While hyperscaler attention and capex chase AI factories, the developer buying a container, a database, and a preview environment is nobody's priority, and specialist platforms serve that buyer better than a 240-service console does. The neoclouds proved the top of the market can be unbundled. The bottom is softer.

The tooling is catching up to the habit, too. Multi-vendor procurement only works if comparing and moving stays cheap, which is why every unbundling wave drags a portability wave behind it, and why the lock-in taxonomy matters more in an unbundled world, not less. A buyer juggling four specialist vendors needs the map of what runs where far more than a buyer with one throat to choke ever did.

The steelman: shortage artifact

The case against reading too much into neoclouds is respectable. They may be a GPU shortage wearing a business model: when supply normalizes, hyperscalers reclaim the workloads, and gravity, egress, data locality, enterprise agreements, reasserts the bundle. CoreWeave specifically is a debt-heavy bet with concentrated customers, and a backlog is a promise, with counterparties, not cash. The dot-com era minted specialist infrastructure companies too, and the survivors' list is short. Maybe this cohort is scaffolding: essential during the boom, absorbed or gone after it.

Concede the company-level risk entirely; some of these firms will have ugly years, and consolidation is likely. The ratchet argument survives anyway, because it doesn't depend on who survives. Buyers who unbundled once keep the muscle: the procurement templates exist, the multi-vendor tooling exists, the board slide that says "we evaluate per workload" exists. Habits, unlike companies, don't need to refinance. The bundle can win any given workload back, and it will; what it can't recover is the presumption that it wins by default. Ask the telecom industry how presumptions age: the phone bundle lost its own spell decades ago, the incumbents are all still here, and nobody has bought the bundle out of reflex since.

Which leaves the question where it always lands on this blog: on your side of the table. The neocloud era's real gift to every buyer is permission, permission to ask, workload by workload, "who is actually best at this?", and the only prerequisite for using it is being able to move. When did your team last ask that question about anything other than GPUs?

Related: Three Companies in a Trench Coat, the concentration risk the unbundling chips away at. More about what we're building at light-cloud.com.