For Microsoft customers, that distinction has become practical rather than theoretical. Microsoft is cited as a major customer or partner of Nebius, Lambda, and Crusoe, while CoreWeave’s scale and NVIDIA-heavy fleet make it a frequent alternative for organizations that need large dedicated clusters. The buyer’s question is therefore less “who has the cheapest GPU?” than “which capacity is actually available, under what interruption and commitment terms, with what deployment risk?”
MarkTechPost’s August 21 survey collects public rate cards, financial disclosures, company announcements, and SemiAnalysis ClusterMAX ratings. The comparison has real value as a starting ledger, especially because several providers do publish prices that can be checked. But its headline rankings blur a critical divide: a low on-demand price can be available only for preemptible capacity, while a multi-gigawatt “contracted” figure can include infrastructure that has not yet been energized, populated with GPUs, or made available to ordinary cloud customers.
Rate cards show a real price spread — with important qualifiers
The clearest public pricing result is that Nebius, Lambda, and Crusoe undercut CoreWeave on several currently shipping NVIDIA GPU generations. MarkTechPost lists Nebius at $3.85 per H100 GPU-hour on demand, Lambda at $3.99, and Crusoe at $3.90, compared with CoreWeave’s $6.16. For Blackwell B200 hardware, Lambda’s listed $6.69 per GPU-hour is below Nebius at $7.15 and CoreWeave at $8.60.
Those figures matter for teams running flexible development, fine-tuning, batch inference, or short-lived training workloads. They do not settle the cost of a production deployment. Instance topology, storage, network performance, scheduler behavior, reserved-capacity terms, support, regional availability, and the ability to keep an 8,000-GPU job running all affect the bill far more than a headline H100 rate.
The most striking discount figures in the comparison are for interruptible capacity. Nebius’s listed preemptible H100 price of $2.15 per GPU-hour and CoreWeave’s North American spot price of $2.46 look compelling for experiments that can checkpoint often and restart elsewhere. They are poor substitutes for a contract-bound production training run unless the workload architecture is designed around interruption.
Lambda is a useful counterexample. Its B200 rate is the lowest published on-demand figure in the group, according to MarkTechPost, but the company does not publish a spot tier. That can make Lambda simpler to evaluate for teams that want predictable self-service pricing, while limiting the discounted capacity options that can drive down costs for fault-tolerant jobs.
Crusoe’s rate card deserves separate treatment because it is the only one in this group that publicly lists AMD Instinct MI300X pricing, at $3.45 per GPU-hour, and offers MI355X through sales. That is meaningful for organizations standardizing on AMD-compatible software stacks or wanting negotiating leverage against NVIDIA-centric supply. It does not mean AMD is a transparent drop-in replacement for every CUDA workload. Migration effort, model-library maturity, container compatibility, distributed-training behavior, and operational familiarity should be priced into the decision before a lower GPU-hour figure becomes a savings claim.
CoreWeave’s premium is tied to service depth, not merely silicon
CoreWeave is the most expensive listed option for several mainstream NVIDIA configurations in the MarkTechPost table. It is also the provider with the strongest third-party quality signal cited in the article: SemiAnalysis’s ClusterMAX 2.0 awarded CoreWeave Platinum status, while the other providers sit below it or were not rated.
That distinction is more valuable to a large enterprise than it may look on a comparison chart. A sophisticated managed cluster is not a basket of interchangeable GPUs. The operational difference lies in whether the provider can deliver a large fabric, run it reliably, support Slurm or Kubernetes at scale, replace failed components quickly, and sustain training throughput across a multi-week job.
CoreWeave’s second-quarter results, reported by the company and covered by outlets including The Motley Fool and Tom’s Hardware, show the scale behind that premium. The company reported $2.575 billion in quarterly revenue, 112% year-over-year growth, 1.5 GW of active power capacity, and roughly $104 billion in revenue backlog at June 30. It also said it had added more than $25 billion in net new commitments early in the third quarter.
The caution is financial and physical rather than demand-related. CoreWeave’s reported quarterly net loss was $626 million, including $640 million in net interest expense, and it raised 2026 capital-expenditure guidance to $35 billion to $39 billion. The company’s backlog and growth indicate buyer demand, but they also require an enormous build-out of data centers, electrical systems, networking, and GPU supply. For procurement teams, a provider’s capacity roadmap should be treated as a delivery schedule to audit, not as capacity already in hand.
Contracted gigawatts are not deployed cloud capacity
Crusoe’s June 9 announcement is the most obvious reason to separate contracted power from live service. The company said it had 4.9 GW of contracted AI infrastructure across data-center projects and Crusoe Cloud, with a development pipeline above 40 GW. That is a substantial commercial milestone, and it reinforces Crusoe’s role as both cloud provider and AI data-center developer.
But Crusoe’s own definition matters: the 40 GW figure includes contracted projects, sites under active tenant negotiation, and sites in advanced development. The 4.9 GW figure is also broader than a count of immediately usable, customer-accessible GPU capacity. Those are not defects in the announcement; they are different stages in the infrastructure pipeline.
A procurement team that treats a 4.9 GW contract figure as proof of near-term B200, H200, or MI355X availability risks confusing real-estate and power commitments with provisionable compute. The correct follow-up questions are more concrete: Which campus? Which building? Is the power energized? How much is allocated to the customer? Which GPU generation is installed? What networking fabric is complete? What date is backed by contractual remedies if delivery slips?
The same discipline applies to every provider in the comparison. CoreWeave’s active 1.5 GW is more immediately meaningful for current service capacity than its larger contracted total. Nebius’s stated plans to reach 5 GW of contracted power and deploy more than 1 GW annually beginning in 2027 indicate aggressive expansion, but they are future-facing infrastructure targets. Groq’s planned move from 54 MW to more than 200 MW by 2027 similarly describes a build-out trajectory rather than an instantly available GPU cloud.
Nebius and Lambda are the clearest alternatives for accessible NVIDIA capacity
Nebius emerges from MarkTechPost’s ledger as the most aggressive public-price competitor to CoreWeave. It offers lower listed H100, H200, and B200 prices, publishes preemptible rates, and is the only provider in the comparison with a public on-demand B300 price. For teams needing newer NVIDIA hardware without committing to an opaque sales negotiation, that transparency is an advantage.
According to MarkTechPost’s reading of Nebius’s quarterly materials, its AI-cloud annualized run-rate revenue reached $3 billion, while the company has been selling a mix of short-term premium capacity and longer-term contracts. The publication also notes that Nebius disclosed unusually high pricing for scarce short-term capacity. That is the important contradiction behind a cheap rate card: the public price may be available for self-service supply, while urgent, dedicated capacity can command a very different price through a negotiated contract.
Lambda occupies a different position. Bloomberg and Lambda confirmed in May that Michel Combes became chief executive while co-founder Stephen Balaban moved into the CTO role. Lambda has also disclosed a $1 billion senior secured credit facility and a multibillion-dollar Microsoft agreement involving tens of thousands of NVIDIA GPUs, including GB300 NVL72 systems.
For an enterprise buyer, Lambda’s appeal is its blend of self-service GPUs and cluster offerings designed to bridge the gap between an individual instance and a giant hyperscale reservation. Its limitation is disclosure: as a private company, Lambda does not publish SEC filings or quarterly financial results comparable with CoreWeave and Nebius. Buyers should therefore seek operational evidence directly in a contract review, including current regional capacity, hardware delivery schedules, support commitments, security controls, and exit terms.
Groq belongs in the comparison, but not in the GPU-price ranking
Groq is the outlier. It runs an inference cloud around its LPU technology and charges by tokens rather than by GPU-hour, so placing it beside GPU clouds in a pricing table is inherently approximate. A token rate can be the right economic metric for a low-latency application; it cannot be directly converted into the cost of training a foundation model on an eight-GPU B200 node.
Groq confirmed in June that it had raised $650 million to expand its inference cloud, operated 13 data centers, served more than five million developers, and expected to scale toward 200 MW by the end of 2027. The company said its current strategy followed a December 2025 non-exclusive technology licensing agreement with NVIDIA, and that NVIDIA’s LPX platform incorporates Groq inference technology.
MarkTechPost reports that Groq intends to add NVIDIA GPU capacity through the NVIDIA Cloud Partner program. Until that service has a published rate card, regional availability, and defined instance specifications, it should be evaluated as planned capacity—not as a substitute for CoreWeave, Nebius, Lambda, or Crusoe GPU clusters today.
For Windows and enterprise IT teams, the practical shortlist begins with workload type. Use spot or preemptible rates only for restartable jobs; use published on-demand rates for exploratory budgeting, not final total-cost comparisons; and require a provider to distinguish active capacity from contracted, under-construction, and merely planned megawatts. In the neocloud market, the most expensive mistake is not paying too much per GPU-hour. It is signing for capacity that cannot arrive when the project needs it.