The neocloud sector has enjoyed an extraordinary two years. GPU scarcity created pricing power that no previous cloud infrastructure business had experienced. Hyperscaler capacity constraints drove enterprise customers toward neocloud alternatives at a pace that filled capacity faster than operators could deploy it. Committed offtake agreements with frontier model labs provided revenue visibility that transformed speculative infrastructure investments into fundable infrastructure businesses. The combination of genuine demand, real supply constraints, and available capital produced a market environment that made most neocloud business models look viable regardless of their underlying economics.
That environment is changing. Not abruptly or dramatically, but persistently and in ways that will separate the neoclouds with genuinely durable business models from those whose apparent success reflected favorable market conditions rather than structural competitive advantages. The neocloud sector is about to find out which business models actually work, and the answer will not be flattering to all of the operators who have spent the past two years presenting investor narratives built on demand projections that the current market no longer supports unconditionally.
The Custom Silicon Problem Is Arriving Faster Than Expected
The most significant structural threat to neocloud business models is the accelerating maturity of hyperscaler custom silicon programs in the inference segment. Google’s TPU v8, Amazon’s Trainium 2, and Meta’s MTIA are all advancing faster than neocloud operators anticipated when they made the GPU procurement decisions that define their current capacity profiles. These custom chips are not yet competitive with Nvidia GPUs for all training workloads, but they are increasingly competitive for inference, and inference is the fastest-growing and most commercially valuable segment of AI compute demand.
A neocloud operator whose capacity is entirely deployed around Nvidia GPU infrastructure faces a structural problem as enterprise customers discover that hyperscaler inference costs using custom silicon are declining faster than GPU-based inference costs. The neocloud value proposition that rested on Nvidia GPU performance as a differentiator is being eroded from the bottom up as the workload category growing fastest is also the category where the neocloud’s hardware advantage is shrinking fastest. As explored in our analysis of the rise of inference clouds, the neoclouds best positioned for this transition are those that built differentiated service layers above the hardware. Those that relied on hardware access as their primary competitive argument are now discovering that argument has a shorter shelf life than their financing structures assumed.
The Spot Market Is Telling a Different Story Than 2023
The GPU spot market was the original neocloud business: excess GPU capacity rented at market rates to AI developers who could not get what they needed from hyperscalers. Spot pricing in 2023 and early 2024 reflected genuine scarcity. H100 spot rates reached levels that made GPU infrastructure businesses look like exceptional investments regardless of utilization risk. Those rates have compressed substantially. The scarcity premium that justified spot market pricing has partially dissipated as Nvidia’s production has scaled, as hyperscaler GPU deployments have grown, and as the wave of neocloud GPU procurement over the past two years has added supply to a market where demand growth, while still strong, is no longer outpacing supply at the rate that produced the 2023 scarcity premium.
The compression of spot rates does not make neocloud businesses unviable. It does make the economics of pure spot-market GPU rental significantly less attractive than they appeared during the scarcity period. Operators who built their financial models around 2023 spot rates and signed debt commitments against GPU assets valued at those rates face balance sheet pressure that their committed offtake agreements may or may not fully offset. The neoclouds that are genuinely differentiated from pure spot rental operations have a path through this transition. Those that are essentially GPU rental businesses with more sophisticated financing structures do not.
What the Enterprise Deployment Gap Is Revealing
The most operationally revealing challenge facing neocloud operators is not hardware commoditization or spot rate compression. It is the gap between enterprise customers signing committed capacity contracts and those same customers actually deploying workloads at the utilization levels the contract economics require. Enterprise AI deployment at scale is harder and slower than enterprise AI procurement. The organizational change management, data preparation, compliance review, and model development work that sits between a signed contract and sustained high-utilization deployment takes time that neocloud financial models did not adequately account for.
Operators carrying large committed enterprise contracts that are deploying below target utilization are experiencing a cash flow profile that differs materially from what their financing structures modeled. The GPU assets are deployed. The cooling and power infrastructure is running. The debt is accruing interest. The revenue that the contract entitles the operator to collect depends on the customer deploying workloads, and that deployment is happening more slowly than both parties anticipated. This is not a customer relationship problem that good account management can fix. It is a structural feature of enterprise AI adoption that affects every operator serving enterprise customers at scale, and it will sort the sector into operators whose unit economics can absorb the deployment ramp reality and those whose cannot.
The Differentiation Test Is Coming
The neoclouds that will emerge from this transition with durable competitive positions are already identifiable by their behaviour. They are investing in software capabilities that create genuine switching costs. Instead of leaving deployment challenges to customers, they are building operational services that help enterprise clients move faster. Many are also developing hardware flexibility that allows them to serve inference workloads efficiently as custom silicon changes the performance-per-dollar equation. At the same time, they are diversifying their customer bases away from dangerous hyperscaler revenue concentration and toward enterprise segments where margins are stronger and relationships are stickier.
The companies unlikely to emerge from this transition intact are just as easy to identify. Many still present investor narratives built on 2023 demand assumptions. Others continue to defend pure hardware access as a differentiator in a market where hardware access is becoming commoditised. Some are also carrying leverage ratios that made sense against 2023 spot pricing but become fragile against compressed spot rates and slower enterprise deployment ramps. The neocloud sector is not facing an existential crisis. Instead, it is facing a maturation that will produce a much smaller number of genuinely durable businesses from the much larger number created by 2023โ2024 market conditions. That maturation was always coming. It is arriving now.
