Inference Is Rewriting Neocloud Economics After the Training Boom

On May 7, CoreWeave raised the lower end of its annual capital expenditure forecast, citing rising component costs. Shares of the GPU-cloud provider fell more than 9% in extended trading. The sell-off arrived on the same day the company reported quarterly revenue of $2.08 billion, more than doubling year on year, and disclosed that its contracted backlog had swollen to $99.4 billion. The juxtaposition of record demand and nervous markets captured the neocloud sector's central tension in mid-2026: the deals are enormous, but the bill for building the infrastructure to fulfill them keeps growing faster than the revenue those deals promise.

Six weeks before that earnings print, CoreWeave had signed the largest single-customer deal in its history: a $21 billion commitment from Meta Platforms running from 2027 through 2032. Within 48 hours, the company also disclosed a multiyear agreement with Anthropic, the maker of the Claude family of models, Forbes reported. The neocloud now says it serves nine of the ten largest AI labs. CoreWeave followed those wins with a $6 billion commitment from Jane Street, the quantitative trading firm, which also took a $1 billion equity stake. By any measure, the demand signal was unambiguous.

Yet the neocloud proposition is shifting underneath the companies that rode it to prominence. These providers, CoreWeave, Nebius, Lambda, and Crusoe among them, built their businesses during the GPU shortage of 2024 and 2025, when training runs for frontier models consumed every available Nvidia H100 and the hyperscalers could not provision capacity fast enough. The neoclouds stepped into that breach with specialized, GPU-dense infrastructure purpose-built for training workloads. The 2026 question is whether the model that worked for training, large, lumpy, compute-saturated contracts with a handful of well-capitalized labs, translates to a world where inference is becoming the dominant workload.

The F5 2026 State of Application Strategy report, published in early May, found that 77% of organizations now prioritize AI inference over training, RCR Wireless News reported. That proportion has climbed steadily for three consecutive quarters. Inference workloads are different in kind from training: they are latency-sensitive, geographically distributed, often lower-margin, and far more fragmented across customers. A single training cluster for a GPT-class model might consume tens of thousands of GPUs in one location for months. An inference deployment for an enterprise chatbot might need a handful of GPUs in five different metro areas, running continuously for years. The capital intensity and the sales motion are not the same.

This is the third quarter in a row that industry surveys have shown inference pulling ahead of training as the primary AI workload. The shift has implications for every layer of the stack, but it lands hardest on the neoclouds, whose data center footprints were optimized for the training era. Training clusters reward density: pack as many GPUs as possible into a single facility with high-bandwidth interconnect, and sell the whole block to one customer. Inference clusters reward distribution: place smaller GPU pools close to end users, manage multi-tenant utilization efficiently, and compete on per-token pricing. Retrofitting a training-centric fleet for inference economics is a capital-intensive exercise layered on top of an already capital-intensive buildout.

The Deals Keep Getting Bigger

The scale of the commitments flowing to neoclouds is difficult to overstate. CoreWeave's total remaining performance obligations reached $99.4 billion as of its Q1 2026 filing, up from roughly $66.8 billion at the time of its April announcements, according to Morningstar. Nebius, the Amsterdam-based neocloud that emerged from the restructuring of Yandex's international assets, disclosed $46 billion in combined AI cloud deals with Microsoft and Meta, The Motley Fool reported. The company's Q1 revenue surged nearly sevenfold year over year. Its market capitalization, however, still prices in significant execution risk.

Both CoreWeave and Nebius continue to issue debt to fund their data center expansions, a pattern that has drawn scrutiny from analysts who track the sector's leverage ratios. CoreWeave spent roughly $2.60 on capital expenditures for every $1 of revenue it recorded in the first quarter, a ratio that improved slightly from previous quarters but remains far above what any mature infrastructure business would sustain. The neoclouds are effectively borrowing against their backlogs, betting that the revenue will arrive before the debt comes due. The bet is not irrational, Meta and Microsoft are not speculative counterparties, but it leaves little margin for construction delays, chip allocation disappointments, or shifts in customer demand.

CNBC reported in late April that Wall Street analysts have grown increasingly bullish on the neocloud category, but the same report noted that McKinsey had warned the economics of these businesses remain fragile. The core fragility is straightforward: neoclouds are effectively GPU-financing vehicles. They borrow money to buy Nvidia chips, rent those chips to AI labs under multiyear contracts, and use the contract revenue to service the debt. If the residual value of the chips declines faster than the depreciation schedule assumes, or if contract renewals come in below the original pricing, the model compresses quickly.

Component cost inflation is now compounding that fragility. In its May 7 earnings call, CoreWeave specifically cited rising prices for data center components as the driver of its increased capex forecast, Reuters reported. The costs of power infrastructure, cooling equipment, and networking gear have all risen as the global data center buildout strains supply chains. TrendForce estimated that the top nine North American cloud service providers would collectively spend $830 billion on capex in 2026, a figure that captures both the hyperscalers and the neoclouds vying for the same pool of electrical transformers, generators, and fiber.

The TPU Question and the Nvidia Lock-In

One signal of how tightly the neoclouds have bound themselves to Nvidia's ecosystem arrived in early May, when The Information reported that Nebius, Lambda, and CoreWeave had all declined to incorporate Google's Tensor Processing Units into their fleets, despite an active push from Google to broaden TPU adoption beyond its own cloud. The rejection is strategically rational in the short term. Nvidia's CUDA software stack remains the de facto standard for AI workloads, and the neoclouds' entire operational tooling, from cluster orchestration to customer provisioning, is built on Nvidia hardware.

But the TPU question also exposes the neoclouds' limited room for maneuver on cost. Nvidia's pricing power over its cloud partners is formidable. The neoclouds compete for GPU allocations not only with each other but with Microsoft Azure, Amazon Web Services, and Google Cloud, all of whom have deeper balance sheets and can place larger, longer-dated orders. If Nvidia's next-generation Vera Rubin platform commands a significant price premium over the Blackwell generation, the neoclouds will have to pay it or risk losing their performance edge. Google's TPU v6, by contrast, offers competitive inference performance at lower power draw, but adopting it would require the neoclouds to build an entirely separate software and support layer. None have chosen to make that investment.

Lambda and Crusoe, two of the smaller neoclouds, face a variant of the same problem from a different angle. Lambda has built its brand on developer-friendly GPU cloud access, offering on-demand and reserved instances that appeal to startups and mid-sized AI teams. Crusoe has differentiated on energy, colocating its GPU clusters with stranded natural gas and renewable sources to offer a lower-carbon compute product. Both have grown revenue at triple-digit rates, but neither has disclosed a backlog approaching CoreWeave's or Nebius's scale. In a market where the largest customers are writing $21 billion checks, subscale players risk being relegated to spot-market economics: selling whatever capacity the hyperscalers and large neoclouds have not already contracted.

The inference market compounds this scale question. Training contracts tend to be large, concentrated, and multiyear, exactly the structure that supports debt-financed buildout. Inference contracts are smaller, more numerous, and often shorter in duration. An enterprise customer deploying a customer-service chatbot may sign a one-year commitment for a few dozen GPUs, not a seven-year commitment for tens of thousands. Serving that customer profitably requires a different operational model: higher utilization rates across a larger number of smaller deployments, automated provisioning that minimizes idle capacity, and pricing that competes with the hyperscalers' own inference offerings, which benefit from the amortization of enormous general-purpose cloud infrastructure.

Network World reported in early April that neoclouds are beginning to take meaningful market share from traditional data center infrastructure providers, a trend driven partly by the hyperscalers' own decisions to offload certain GPU workloads to specialized providers. This is the paradox at the heart of the neocloud thesis: the same hyperscalers that compete with neoclouds for GPU supply are also their largest customers. Meta's $48 billion in combined commitments to CoreWeave and Nebius makes it both the neocloud sector's most important patron and, through its own Llama model family and open-source AI strategy, a potential competitor for inference workloads down the line.

CoreWeave's adjusted operating income hit a two-year low of $21 million in Q1, Morningstar noted, despite revenue that grew 32% sequentially. The compression came from the front-loaded costs of bringing new data center capacity online, costs that the company expects will moderate as those facilities reach higher utilization rates later in the year. The pattern is common across the sector: revenue scales in step functions as new clusters come online and customers begin paying, but the infrastructure spend precedes the revenue by quarters. The neoclouds are asking investors to underwrite a bridging period whose duration depends on factors largely outside the companies' control: Nvidia's delivery schedules, utility interconnection timelines, and the pace at which AI labs convert their training-scale contracts into steady-state inference consumption.

The inference-cloud market does not yet have a clear margin structure. The hyperscalers are bundling inference with their broader platform services: storage, databases, networking, and the application-layer tools that keep customers within a single cloud ecosystem. The neoclouds, by contrast, largely sell raw GPU compute, which is more substitutable. A customer who trains a model on CoreWeave can run inference on AWS, Google Cloud, or their own on-premises hardware without significant switching friction, provided the model format is portable. That portability is a feature for customers and a structural vulnerability for a GPU-cloud provider whose customer retention depends on performance and price rather than platform integration.

What this means for the neocloud sector over the next two quarters is not a question of demand. Demand exists and is growing. The question is whether the neoclouds can convert their training-era backlog into inference-era operating economics without the kind of margin compression that resets investor expectations. Watch for two indicators in the Q2 2026 earnings cycle: the share of revenue coming from inference workloads versus training, a metric that CoreWeave and Nebius have not yet broken out but that analysts are increasingly requesting, and the trajectory of capex as a percentage of revenue. If that ratio does not begin to decline by the second half of 2026, the market's patience with the bridging argument will be measurable in stock prices.