CoreWeave's $99 Billion Backlog Bets on Inference, Not Training

On May 7, CoreWeave reported first-quarter revenue of $2.08 billion, more than double the same period a year earlier, and disclosed a contract backlog that had swelled to $99.4 billion. The number was enormous by any standard, roughly equivalent to the annual GDP of Slovakia, and it arrived barely a month after the company signed $21 billion in new commitments from Meta and a separate multiyear deal with Anthropic inside a single 48-hour window. Shares fell 11 percent the next day. The sell-off was not a rejection of the backlog. It was a reaction to second-quarter guidance that came in below Wall Street estimates and to a capex trajectory that chief financial officer Nitin Agrawal described, on the company's earnings call, as 'higher than previously communicated' thanks to rising component costs on next-generation Nvidia systems.

The divergence between the backlog and the stock chart captures the central tension now running through the neocloud sector. These companies, CoreWeave, Nebius, Lambda, Crusoe, and a growing roster of smaller entrants, built their businesses as a fast-moving solution to the GPU supply shock that followed the launch of ChatGPT in late 2022. They raised debt against Nvidia hardware, built purpose-built data centers, and rented compute to AI labs that could not get enough capacity from Amazon Web Services, Microsoft Azure, or Google Cloud. Three years later, that stopgap has become a permanent layer of the AI infrastructure stack. CNBC reported in late April that Wall Street analysts have turned broadly bullish on the category, even as McKinsey warned that the underlying economics remain fragile.

What has changed is the workload mix. The GPU shortage of 2023 and 2024 was driven primarily by training, the massively compute-intensive process of building large language models from scratch. Training runs are spikey, multi-month affairs that consume tens of thousands of GPUs and then stop. Inference, by contrast, is the steady-state work of running those trained models to answer user queries. It is lower-margin per GPU-hour but far more predictable, and as AI labs shift from building models to deploying them at scale, inference is becoming the dominant compute workload. Every major neocloud is now racing to position itself as the inference layer of choice, and the financial engineering required to get there is becoming as sophisticated as the data-center engineering that launched the category.

The numbers tell the story of that pivot. In early May, The Next Web reported that Nebius had paid $643 million to acquire Eigen AI, a 20-person spinout from MIT that specializes in maximizing the number of tokens a single GPU can process during inference. The price, more than $32 million per employee, was not for the team alone. It was for software that Nebius believes can differentiate its inference offering from both hyperscaler clouds and rival neoclouds at a moment when raw GPU availability is no longer a sufficient competitive moat. 'Inference economics will determine who survives the next phase,' Nebius CEO Arkady Volozh said in a statement accompanying the deal.

Inference economics will determine who survives the next phase., Arkady Volozh, CEO of Nebius Group, on the acquisition of Eigen AI, May 2026

CoreWeave's own numbers underscore why the inference market matters. The company disclosed in its Q1 filing that committed contracts now stretch as far as 2032, with Meta alone accounting for roughly $35 billion in total obligations across two agreements. That kind of duration is unusual in cloud computing, where one-year and three-year commitments are the norm. It reflects the fact that large AI labs are locking in inference capacity years ahead of demand, a signal that they expect inference workloads to be both enormous and durable. For CoreWeave, the commitments provide revenue visibility that traditional cloud providers rarely enjoy. The tradeoff is that the company must build the physical infrastructure to serve those contracts, and it must do so at a moment when the cost of Nvidia's next-generation Vera Rubin systems is pushing capex higher than analysts had modeled.

That capex pressure is not unique to CoreWeave. The Motley Fool noted in late April that both CoreWeave and Nebius continue to issue debt to fund data-center construction, a financing model that worked smoothly when GPU-backed loans were novel and interest rates were lower. In the current environment, with component costs rising and investors scrutinizing margins more closely, the debt-fueled expansion strategy is beginning to show strain. CoreWeave's Q1 adjusted loss of $1.11 per share, reported alongside the revenue beat, was wider than analysts had forecast, driven in part by depreciation on its rapidly expanding GPU fleet.

The competitive landscape is also shifting in ways that could compress neocloud margins. Google has been aggressively promoting its custom Tensor Processing Units, or TPUs, as an alternative to Nvidia GPUs for inference workloads, offering them through Google Cloud at prices designed to undercut GPU-based competitors. In a report published by The Information on May 6, reporter Anissa Gardizy detailed how Nebius, Lambda, and CoreWeave have each resisted adopting TPUs in their fleets, betting instead that the Nvidia software ecosystem, particularly the CUDA programming framework and the inference libraries built on top of it, represents a switching cost that customers will continue to pay a premium to avoid. 'Neoclouds are Nvidia shops by conviction, not convenience,' one industry analyst told The Information.

That conviction carries risk. If Google's TPU price-performance advantage widens, or if Amazon's custom Trainium and Inferentia chips gain traction for inference, neoclouds tied exclusively to Nvidia could find themselves on the wrong side of a cost curve. For now, the market is validating the GPU bet: nine of the ten largest AI labs are CoreWeave customers, according to the company's disclosures, and none have shown meaningful signs of shifting inference workloads to custom silicon. But the history of cloud infrastructure suggests that cost advantages, once they reach a certain threshold, overwhelm ecosystem loyalty. Watch for any AI lab that signs a large-scale TPU inference commitment in the second half of 2026. It would be the cheapest signal that the Nvidia lock is weakening.

Crusoe, the smallest of the three named neoclouds, is pursuing a differentiation strategy that sidesteps the GPU-versus-TPU question entirely. The company, which began as a bitcoin mining operation that captured flared natural gas to power its rigs, has repositioned itself as a 'clean compute' cloud for AI, emphasizing its access to low-cost, stranded energy sources. Crusoe has not disclosed a backlog on the scale of CoreWeave or Nebius, but its pitch is resonating with enterprise customers that face internal carbon-reduction mandates and cannot easily square multi-gigawatt AI deployments with their sustainability targets. Whether 'green GPU compute' commands a price premium in a commodity market remains an open question, and one that will not be answered until Crusoe reports its own public numbers.

Lambda, for its part, has positioned itself as the developer-friendly alternative, the neocloud for AI engineers who want a cloud experience that feels more like using a Linux workstation and less like navigating an AWS billing console. At Nvidia's GTC conference in March, Lambda announced it would be a launch partner for Nvidia's Vera CPU platform and the STX system architecture, placing it among the first neoclouds to offer the next-generation silicon. Being early to new Nvidia hardware matters in a market where AI labs compete on model performance and will pay a premium for faster iteration cycles. Lambda's strategy is to win on speed to silicon rather than on scale, a bet that the inference market will fragment across multiple providers rather than consolidating around the largest player.

The Hyperscaler Counterpunch

The neocloud thesis rests on a premise that is not yet fully tested: that AI labs will prefer to rent compute from a specialist rather than from the same hyperscalers that are themselves building competing AI models. Amazon, Microsoft, and Google have each invested billions in their own AI efforts and have every incentive to steer compute demand toward their own platforms. For now, the capacity shortage has papered over that conflict. When AWS cannot provision 50,000 GPUs fast enough, a check to CoreWeave is the pragmatic solution. But as hyperscaler capacity catches up, and all three have massive data-center buildouts underway, the question of 'who can the neocloud not afford to lose as a customer?' becomes urgent. CoreWeave's reliance on Meta for roughly a third of its backlog is both an asset and a concentration risk that would alarm any credit analyst in a more mature industry.

The hyperscalers are also moving to close the inference-optimization gap that companies like Eigen AI exploited. AWS has tuned its SageMaker inference endpoints for large language models, Azure has deeply integrated with OpenAI's inference stack through what amounts to a captive customer relationship, and Google Cloud has its own inference-optimized serving infrastructure for Gemini and third-party models. If the hyperscalers can match neocloud inference economics while offering the broader platform services, databases, identity management, compliance tooling, that enterprises want, the neocloud value proposition narrows to speed and simplicity, which are easier for large competitors to replicate than cost advantages rooted in specialized hardware or software.

The enterprise market is the next battlefield, and it is where the neocloud model faces its hardest test. AI labs are comfortable renting raw GPU capacity and managing their own infrastructure tooling. Fortune 500 companies, by and large, are not. CIO Dive reported in January that neoclouds are beginning to size up enterprise opportunities, building managed service layers on top of their bare-metal GPU offerings to appeal to customers that want AI capabilities without the operational burden of managing clusters. But building an enterprise sales and support organization is expensive, and the margins in managed services are different from the margins in raw capacity leasing. This is the third quarter in a row that neocloud executives have name-checked the enterprise on earnings calls without reporting a material revenue contribution from that segment.

Network World observed in April that neoclouds are beginning to take market share from traditional data-center infrastructure providers, not just from hyperscalers, as enterprises that once leased cage space in Equinix facilities now direct that spending toward GPU-equipped neocloud capacity instead. That shift implies that the addressable market may be larger than the hyperscaler overflow that launched the category, but it also invites competition from a broader set of players, including colocation giants that can add GPU hosting to their existing real-estate portfolios.

The financial engineering behind neocloud expansion deserves scrutiny as the sector matures. CoreWeave's $8.5 billion debt facility, arranged in early 2026 and detailed by CoinTelegraph, represents a new asset class that Wall Street calls 'ComputeFi': loans collateralized against Nvidia GPUs and the future cash flows they are expected to generate. The model migrated from crypto mining, where ASIC-backed loans were common during the bitcoin boom, and it works as long as GPU resale values hold and AI compute demand continues to grow. A sustained downturn in AI infrastructure spending, or a generational leap in chip efficiency that devalues older GPUs, would test the collateral assumptions that underwrite these loans. No such downturn is on the visible horizon, but the memory of the crypto-mining credit cycle, which ended with lenders seizing depreciated hardware, is less than three years old.

Neoclouds are Nvidia shops by conviction, not convenience., Industry analyst quoted by The Information, May 6, 2026

What distinguishes the inference-cloud market from the training-cloud market that preceded it is the shape of demand. Training is project-based: a lab raises capital, runs a training job for three to six months, and the GPUs go quiet until the next model cycle. Inference is product-based: every user query that hits ChatGPT, Claude, or Gemini generates a small but continuous stream of GPU work that scales with end-user adoption. For a cloud provider, inference revenue is stickier and more predictable, which is why CoreWeave's 2032 contract duration is not a fluke but a structural feature of where the market is heading. The providers that can lock in inference contracts now are securing revenue streams that will outlast the training buildout cycle. The ones that remain dependent on training jobs will see their revenue oscillate with model-release calendars.

The TPU question is a microcosm of a larger strategic choice facing every neocloud: whether to optimize for cost or for ecosystem compatibility. Google's pitch is straightforward, TPUs are cheaper per inference token, and the gap widens at scale. The neocloud counterargument, as The Information's reporting made clear, is that moving an inference stack from CUDA to TPU software is not a trivial engineering exercise; it requires rewriting model-serving code, retesting for correctness, and accepting a different set of performance characteristics. For AI labs that measure latency in milliseconds and serve billions of queries per day, the switching cost is real. But it is not infinite, and if the cost delta grows wide enough, some lab will take the leap. That lab's experience will be the template for the rest of the market.

CoreWeave's May 7 earnings report, for all the stock-market drama it generated, was not a referendum on the neocloud thesis. Revenue doubled. The backlog hit a record. Nine of ten top AI labs are paying customers. The guidance miss was narrow, a few hundred million dollars on a multi-billion-dollar base, and largely attributable to the timing of data-center completions rather than any softening in demand. What the market was pricing in, however, was the recognition that neoclouds are no longer valued on backlog alone. They are being valued on margins, on the sustainability of their debt structures, and on their ability to manage the transition from a supply-constrained training market to a more competitive inference market where hyperscalers and custom silicon are also competing for the same workloads.

What to Watch for in the Second Half

Three indicators will determine whether the neocloud sector holds its current trajectory or begins to face the fragility that McKinsey warned about. The first is component costs on Nvidia Vera Rubin systems: if those costs continue to rise faster than neoclouds can pass them through to customers, margins will compress and debt-service coverage ratios will weaken. The second is enterprise adoption: if a Fortune 500 company signs a nine-figure inference deal with a neocloud before the end of 2026, it validates the enterprise thesis. If none do, the neoclouds remain dependent on a small number of AI-lab customers that have outsized negotiating power. The third is the TPU beachhead: if any major AI lab announces a production inference deployment on Google TPUs at meaningful scale, it becomes a reference architecture for cost-conscious customers and a direct challenge to the Nvidia-only strategy that has defined the category.

For now, the neocloud sector is in a position that few infrastructure businesses ever reach: enormous demand visibility, a backlog measured in decades, and a customer base composed of the best-capitalized companies in the technology industry. The question is whether that position is a durable competitive advantage or a temporary artifact of a GPU shortage that is slowly being solved. The inference market will provide the answer, one token at a time.