Signals Inbox·July 25, 2026·AI Chips

Who can eventually kill Nvidia?

CoreWeave can beat AWS and Azure for some of the world’s largest NVIDIA training and inference jobs, but Amazon and Microsoft still control the broader cloud, the cheaper custom silicon and the balance sheets that decide the long game.

Q1Why is CoreWeave suddenly being compared with Amazon and Microsoft?

CoreWeave is now large enough, fast enough and close enough to the leading AI companies to be treated as a serious competitor in AI infrastructure.

Its latest financial results moved the discussion beyond startup hype. Revenue rose from $1.9 billion in 2024 to $5.1 billion in 2025, then reached $2.1 billion in the latest quarter alone. CoreWeave also passed one gigawatt of active power and reported $99.4 billion in contracted revenue backlog.

The customer list explains the attention. CoreWeave says nine of the ten leading foundation-model providers use its infrastructure. Its recent contracts include Meta, OpenAI, Anthropic, Mistral, Cohere and Jane Street. These are companies buying enormous amounts of computing capacity, not startups renting a few servers for experiments.

CoreWeave grew by focusing on a problem AWS and Azure initially struggled to solve quickly: connecting thousands of scarce NVIDIA GPUs and making them behave like one reliable machine. General-purpose clouds were built to serve everything from corporate databases to streaming applications. CoreWeave designed its facilities, networking and software around large AI workloads from the start.

That narrower focus created the opening. Now Amazon and Microsoft are spending heavily to close it.

Q2What would it actually mean for CoreWeave to beat Amazon and Microsoft?

For this article, CoreWeave beats Amazon or Microsoft when a major customer chooses it for a valuable AI workload despite having access to AWS or Azure.

Winning does not require CoreWeave to become the world’s largest cloud company. AWS and Azure sell hundreds of services across computing, databases, storage, cybersecurity, analytics, business software and developer tools. CoreWeave has neither the reach nor the product range to challenge that entire system.

The useful comparison is narrower. Can CoreWeave become the preferred provider for training very large AI models? Can it win major inference contracts once those models enter production? Can it deploy the newest NVIDIA systems faster and operate them more efficiently? Most importantly, can it keep doing all of that while paying for the infrastructure?

A customer could continue using Azure for identity, data storage and corporate applications while moving a 50,000-GPU training job to CoreWeave. CoreWeave would have won the valuable AI workload even though Microsoft kept the broader cloud relationship.

Q3How far behind are CoreWeave, AWS and Microsoft in scale?

CoreWeave is growing at an extraordinary rate, but AWS currently earns about 18 times more quarterly revenue and has a financial cushion CoreWeave cannot approach.

AWS produced $37.6 billion in its latest quarter, compared with CoreWeave’s $2.1 billion. Microsoft reported $54.5 billion in Microsoft Cloud revenue, although that broader figure includes products beyond Azure and should not be treated as a direct Azure comparison.

The profit gap is more important. AWS generated $14.2 billion in quarterly operating income. Microsoft earned $38.4 billion across the company. CoreWeave recorded a $144 million operating loss and a $740 million net loss.

Amazon and Microsoft can fund new data centers with profits from existing businesses. They can accept lower margins on AI computing, overbuild capacity and wait for demand to arrive. CoreWeave has to raise capital and match most large investments with customer contracts.

A specialist with 5% of an incumbent’s revenue can still win an important product market. But the raw scale gap gives AWS and Microsoft far more chances to recover from delayed facilities, disappointing chips or lost customers.

Latest reported scale and financial capacity

Q4Are CoreWeave’s big customers real wins or borrowed demand?

CoreWeave has now won enough direct customers to prove that it sells more than spare capacity for Microsoft, although outsourced hyperscaler demand played a major role in its rise.

Microsoft generated 67% of CoreWeave’s 2025 revenue. Some of that business reflected Azure needing extra capacity to support demand connected to OpenAI and other customers. In those cases, CoreWeave helped Microsoft deliver its own cloud commitments rather than beating Azure in a clean customer contest.

The newer contracts weaken that explanation. Meta has committed approximately $21 billion for infrastructure extending into the next decade. Jane Street recently agreed to spend about $6 billion on CoreWeave capacity and separately invested $1 billion in the company. Anthropic signed a multiyear agreement, while OpenAI, Mistral and Cohere have also expanded their relationships.

These buyers have different businesses and different reasons for needing computing power. Meta needs large-scale inference. Jane Street trains machine-learning systems for quantitative trading. Anthropic and OpenAI develop frontier models. Their combined demand cannot be explained by one temporary Azure capacity shortage.

The open question is price. Public announcements rarely reveal whether CoreWeave won because its system was technically better, because it could deliver sooner or because it offered generous commercial terms.

Q5Is CoreWeave actually better at training large AI models?

On very large NVIDIA training clusters, CoreWeave can already outperform the general-purpose clouds in ways customers can measure.

The freshest evidence comes from MLPerf Training, an independent benchmarking program with strict accuracy requirements. CoreWeave trained the 671-billion-parameter DeepSeek-V3 benchmark to the target quality in 2.02 minutes using 8,192 NVIDIA GB300 GPUs. When the cluster doubled from 2,048 to 4,096 and then 8,192 GPUs, training time fell from 5.54 minutes to 3.09 minutes and finally 2.02 minutes.

That progression is more useful than one headline speed result. Large GPU clusters often become less efficient as they grow because communication failures, network congestion and synchronization delays consume more time. CoreWeave’s result showed close to predictable scaling across three cluster sizes.

Customer results point in the same direction. Mistral reports 2.5 times faster training on CoreWeave’s GB200 infrastructure. IBM says some enterprise AI workloads trained as much as 80% faster. These are selected customer cases, so the numbers should not be generalized to every workload. They still show that the advantage can reach the model-training team, rather than staying an infrastructure specification that looks good on paper.

CoreWeave concentrates engineering effort on GPU communication, checkpointing, storage throughput, cluster scheduling and failed-job recovery. AWS and Azure handle those problems too, but their engineering organizations also support thousands of unrelated products.

We would not call CoreWeave universally faster. Hardware, model design and software configuration change every result. Still, a company planning one of the world’s largest NVIDIA training runs has a real technical reason to put CoreWeave on the shortlist.

Q6Can CoreWeave stay ahead through each NVIDIA hardware cycle?

CoreWeave can gain valuable early leads with new NVIDIA systems, but it has to recover the cost of each generation before newer and cheaper hardware weakens the economics.

CoreWeave was the first cloud provider to make NVIDIA GB200 NVL72 systems generally available. It is also among the first providers expected to offer the Vera Rubin platform, which NVIDIA says is already in production.

Early access helps. Installation speed matters more. New rack-scale systems require high-density power, liquid cooling, specialized networking and software that can manage failures across thousands of accelerators. A provider may own the hardware for months before customers can run stable production workloads on it.

NVIDIA has strong reasons to support CoreWeave. An independent AI cloud gives customers another way to buy NVIDIA computing without going through Amazon, Microsoft or Google, all of which are developing their own chips. NVIDIA recently invested $2 billion in CoreWeave and is supporting plans to build more than five gigawatts of CoreWeave AI factories.

But NVIDIA cannot afford to hold back AWS or Azure. Microsoft is preparing Rubin-based AI facilities containing hundreds of thousands of chips. AWS remains one of the world’s largest infrastructure buyers. CoreWeave’s early lead will usually be measured in months, not years.

The harder problem comes later. CoreWeave generally depreciates technology equipment over six years, while NVIDIA can introduce several major generations during that period. Hopper was followed by Blackwell, and Rubin systems are already entering the market.

Older GPUs do not become useless. They can continue running inference, fine-tuning, research and smaller models at lower prices. Long customer contracts also help CoreWeave recover costs before market prices fall.

Even so, specialist-cloud pricing for H100s has already fallen from the worst period of scarcity. CoreWeave has to earn its return through high utilization, reliable clusters and software. It cannot assume every GPU will remain scarce and expensive.

Amazon and Microsoft have more places to put aging hardware: internal products, lower-cost services and broad enterprise contracts. CoreWeave needs each generation to remain commercially attractive largely on its own.

Q7Is CoreWeave’s AI focus a real moat?

CoreWeave has built a genuine operational advantage, although Amazon and Microsoft can copy many parts of the system if they decide the market is valuable enough.

Its platform combines bare-metal infrastructure, Kubernetes, high-speed networking, storage, scheduling and direct engineering support. The customer receives a working AI environment instead of assembling separate cloud products and hoping they perform well together.

Experience accumulates quickly. A provider repeatedly operating large training clusters learns where jobs fail, which networking configurations create bottlenecks and how utilization changes under different workloads. CoreWeave can apply that knowledge across customers because those customers often face similar technical problems.

The company can also move faster than a hyperscaler in a few practical ways. It can dedicate a facility to one customer, standardize heavily around NVIDIA and change its roadmap around the needs of a small group of large buyers. AWS and Microsoft have to support older products, internal services, custom chips and millions of general cloud customers at the same time.

The moat is still incomplete. Infrastructure workloads are easier to move than deeply embedded software platforms. CoreWeave now has to turn its operational lead into renewals, pricing power and stronger margins.

Q8Can Amazon and Microsoft crush CoreWeave with their own AI chips?

Amazon’s Trainium chips and Microsoft’s Maia chips could become a bigger threat to CoreWeave than any competing NVIDIA cloud.

CoreWeave buys the core computing technology from NVIDIA. Amazon and Microsoft are trying to design more of that technology themselves, giving them greater control over cost, supply and product integration.

AWS has already moved beyond small experiments. Amazon says Trainium and Graviton together now generate more than $10 billion in annual revenue run rate. It has installed 1.4 million Trainium2 chips, including more than 500,000 in Project Rainier for Anthropic. Trainium3 is running production workloads, and Amazon expects almost all available supply to be committed quickly.

Microsoft’s Maia 200 focuses on inference, the process of running trained models for users. Deployment has begun in selected American regions. Microsoft can place Maia inside Azure and use demand from its own AI products to improve utilization from the first day.

The economic advantage is plain. AWS can design the chip, operate the data center, sell the cloud service and connect it to Bedrock, SageMaker and the rest of its platform. CoreWeave starts by paying NVIDIA’s price for the GPU, then adds facilities, networking, financing and support.

NVIDIA keeps CoreWeave competitive because CUDA remains familiar to developers and supports a broad range of models. Customers may also prefer hardware available from several cloud providers instead of being tied to one company’s chip.

The market will probably split. CoreWeave can remain a strong choice for customers seeking the latest NVIDIA systems and the freedom to move workloads. Amazon and Microsoft should gain share where they can optimize recurring workloads around proprietary silicon. The larger that second category becomes, the harder CoreWeave’s economics get.

Q9Does the shift from AI training to inference make CoreWeave stronger?

Inference expands CoreWeave’s opportunity, but it pushes the competition toward cost, software and global reach, where AWS and Azure have deeper advantages.

Training creates huge but irregular contracts. A model developer may use tens of thousands of GPUs for several months and then release the cluster. Inference continues whenever users talk to a chatbot, generate video, run an AI agent or use an AI feature inside business software.

CoreWeave has shown that it can compete on this workload. Artificial Analysis recently compared 11 providers running Moonshot AI’s Kimi K2.6 model and placed CoreWeave first for the combination of output speed and price-performance. Meta’s large commitment also covers inference capacity, showing that CoreWeave can win production workloads at enormous scale.

CoreWeave has added dedicated inference products and more flexible capacity plans. Customers can reserve predictable infrastructure, use lower-priced spare capacity or choose a specific GPU and software runtime for production.

Inference appears in far more places than frontier training. A consumer application needs low response times across several continents. A bank may need local data storage and specific compliance approvals. A retailer may want its model connected directly to information already held in AWS.

Amazon and Microsoft can also steer predictable inference toward Trainium or Maia, where they control more of the cost. Their own products create steady internal demand, helping keep those chips busy.

CoreWeave looks strongest in large, concentrated inference deployments where performance and dedicated capacity matter. It is much less likely to handle every small AI request made by ordinary enterprises around the world.

Q10Can CoreWeave win normal enterprises, not just AI laboratories?

CoreWeave can become an important second cloud for enterprise AI teams, but AWS and Azure remain far easier choices for running an entire company.

Most companies need more than GPUs. They need databases, access controls, cybersecurity tools, data warehouses, billing systems, disaster recovery and local infrastructure that satisfies regulators.

The geographic difference remains large. A recent expansion took CoreWeave to 49 data centers, including eight in Europe. AWS currently operates 123 availability zones across 39 regions. Microsoft has more than 70 announced regions and over 400 data centers.

CoreWeave has improved its enterprise credentials. Its platform follows SOC 2 and major ISO security standards. It has added identity controls, object storage, inference products and software for operating large clusters.

The realistic path is to leave core applications on AWS or Azure and send the largest AI jobs to CoreWeave. Advanced AI companies already work this way. Conventional enterprises will move more slowly because security, legal and operational teams all get a vote.

Q11Is CoreWeave becoming a software platform or staying a GPU rental company?

CoreWeave is gradually becoming a software platform, although its financial results still look much more like capital-heavy infrastructure than software.

The company now offers tools above the physical GPUs. CoreWeave Kubernetes Service manages clusters. Mission Control helps customers operate AI workloads. CoreWeave Observe tracks system performance. SUNK packages the company’s methods for running large training environments and can extend parts of that operating model across other cloud environments.

Its $1 billion acquisition of Weights & Biases moved CoreWeave closer to AI developers. Weights & Biases is used to track model experiments, evaluate outputs and manage the development process. Those tools reach the people choosing how models are built, not only the infrastructure teams purchasing GPU capacity.

This could improve retention. A company that only rents GPUs can lose a contract when another provider cuts its hourly price. A company that manages experiments, deployments, monitoring and infrastructure becomes harder to remove from the workflow.

The numbers show how early the transition remains. CoreWeave reported a 56% adjusted EBITDA margin in its latest quarter, which looks impressive until the cost of the assets and financing is included. Its adjusted operating margin was only 1%, while the company still reported an operating loss.

Software should eventually produce stronger margins, steadier revenue or greater pricing power. CoreWeave has assembled the pieces. It has not yet shown that they can transform the economics of the business.

Q12How solid is CoreWeave’s $99 billion backlog?

CoreWeave’s $99.4 billion backlog proves that demand is enormous, but a large portion depends on infrastructure that still has to be financed, built and delivered.

The backlog is almost 20 times CoreWeave’s entire 2025 revenue. That gives the company much more visibility than most businesses growing at a comparable speed. It can take a signed customer agreement to lenders and use the expected revenue to support funding for a specific facility.

The stricter accounting measure is remaining performance obligations, which reached $98.8 billion. CoreWeave expects to recognize 36% during the first two years, 39% during the following two years and the final 25% over years five through seven.

About $35.6 billion is expected within 24 months. More than $63 billion sits further into the future, when customers may be using different chips, models and software.

The contracts can also depend on CoreWeave delivering capacity by agreed deadlines. Its filing explicitly excludes some amounts that may be lost through delivery delays, service credits or capacity that can be resold.

CoreWeave backlog by expected recognition period

Q13Is CoreWeave still too dependent on a few customers?

Customer concentration remains dangerous because two customers currently generate almost two-thirds of CoreWeave’s revenue.

In the latest quarter, the largest customer accounted for 45% of revenue and the second-largest for 20%. The concentration has improved from earlier periods, but losing or delaying one major contract could still reshape the entire income statement.

The risk is greater because these customers are sophisticated infrastructure buyers. Microsoft, Meta and leading AI laboratories are building their own systems while renting from CoreWeave. They can move future workloads, negotiate prices or decide that internal capacity has become cheaper.

Microsoft supplied 67% of CoreWeave’s 2025 revenue. Azure is expanding at around 40% a year while Microsoft spends heavily on facilities, NVIDIA systems and Maia chips. It will probably rely less on outside capacity as its own supply grows.

A gradual reduction would be manageable if Meta, OpenAI, Anthropic, Jane Street and other customers continue expanding. The recent wins make that outcome more plausible than it was a year ago.

The ugly scenario combines three events: Microsoft buys less, another major customer delays a facility and CoreWeave has already borrowed against the expected revenue. Then concentration becomes a balance-sheet problem, not a disappointing sales quarter.

Large contracts are unavoidable in this market because only a small number of organizations need billions of dollars of AI computing. CoreWeave does not need thousands of equal-sized customers. It does need enough independent buyers that no single company can decide whether a new data center stays full.

Q14Can CoreWeave finance its growth without breaking the business?

Financing is CoreWeave’s biggest weakness, and one poor infrastructure cycle could hurt the company more than losing a technical benchmark.

CoreWeave spent $7.7 billion on property and equipment in the latest quarter. It ended the period with $25.1 billion in debt principal and another $10.1 billion in operating lease liabilities.

Interest expense reached $536 million for the quarter, consuming roughly one quarter of revenue. The company reported $1.2 billion in adjusted EBITDA but still lost $740 million after depreciation, interest and other costs.

The full-year outlook makes the bet clearer. Management expects $12 billion to $13 billion in revenue while planning $31 billion to $35 billion in capital expenditure. CoreWeave may invest close to three dollars in infrastructure for every dollar of revenue earned during the year.

Much of the spending is linked to contracts, which reduces the risk compared with building data centers without customers. CoreWeave recently secured an $8.5 billion facility backed by computing infrastructure and a customer agreement. Part of the financing received investment-grade ratings, showing that lenders are becoming more comfortable with contracted GPU assets.

The contracts do not remove construction, technology or refinancing risk. CoreWeave still has to complete facilities, maintain high utilization and collect revenue for long enough to service the debt.

AWS and Microsoft can fund similar assets from profitable operations. CoreWeave is funding tomorrow’s revenue with capital raised today. That model creates explosive growth when everything works. It leaves very little room for a delayed campus or a customer that changes its plans.

Q15Can CoreWeave beat Amazon and Microsoft?

Yes, CoreWeave can beat Amazon and Microsoft for some of the world’s largest NVIDIA AI workloads, but it cannot beat either company across cloud computing as a whole.

CoreWeave has already earned that narrower victory. It operates clusters at a scale frontier-model developers care about, has produced leading independent benchmark results and keeps signing contracts with customers that could buy from AWS or Azure.

Its clearest edge is execution. CoreWeave can bring a dense NVIDIA cluster online quickly, tune the whole system around AI and give large customers direct access to engineers who understand the workload. That combination has real value when a delayed training run costs a laboratory weeks of research progress.

Amazon and Microsoft control the stronger long-term positions around that contest. They have vastly more capital, profitable cloud businesses, custom chips, global infrastructure and software already embedded in large enterprises. Those advantages become more important as AI spending shifts toward inference and ordinary corporate applications.

CoreWeave’s financial structure also limits how aggressively it can compete. Heavy debt works while customers keep signing long contracts and every new facility fills quickly. A weaker pricing environment or one delayed customer can affect CoreWeave far more severely than AWS or Azure.

Our judgment is direct. CoreWeave can become the leading independent AI cloud and keep winning selected flagship contracts from Amazon and Microsoft. It can also force both companies to improve their NVIDIA offerings and deploy new systems faster.

Becoming another AWS or Azure is out of reach. CoreWeave’s best future is more focused and still highly valuable: the specialist AI cloud that frontier laboratories, global companies and even hyperscalers use when their own infrastructure is too slow, too rigid or simply unavailable.

Where CoreWeave can and cannot win