Signals Inbox·July 28, 2026·AI Infrastructure

Is Nvidia taking over the AI cloud?

Nvidia is already setting many of the hardware, software and rack-level rules of the AI cloud. But AWS, Azure and Google still own the customers, data and services that turn those systems into cloud businesses.

Q1What would it actually mean for Nvidia to take over the AI cloud?

In this case, Nvidia would be taking over the AI cloud if providers increasingly sold access to an AI platform whose core technology and economics were set by Nvidia.

That definition is more useful than asking whether Nvidia can copy every AWS service. AWS, Azure and Google Cloud cover storage, databases, cybersecurity, identity, analytics, business software and ordinary computing. Nvidia can gain enormous power without rebuilding that whole catalogue.

To judge it properly, we need to check five kinds of control. Hardware control means customers need Nvidia accelerators for their most demanding workloads. Software control means developers stay inside CUDA and Nvidia's libraries. System control means data centers follow Nvidia's designs for racks, networking, cooling and power. Capacity control means Nvidia influences where new clusters are financed and built. Customer control means companies buy their AI computing directly through Nvidia.

Today, Nvidia is strongest in the first three areas, increasingly active in the fourth and still weak in the fifth.

Where Nvidia controls the AI cloud today

Q2Why does Nvidia look like a cloud company now?

Nvidia looks like a cloud company today because it has moved into capacity distribution, partner financing and cloud consumption while still earning most of its money from infrastructure sales.

Its latest reporting change makes the shift unusually clear. Nvidia now divides Data Center into Hyperscale and ACIE, a category covering AI clouds, industrial customers and enterprises. It now treats specialist AI clouds as a distinct source of growth.

The product strategy has moved in the same direction. DGX Cloud gives Nvidia a common environment across partner infrastructure, while DGX Cloud Lepton connects developers with capacity from providers such as CoreWeave, Crusoe, Lambda, Nebius, Nscale, SoftBank and regional operators. Nvidia can help decide how GPU demand is routed without owning every building involved.

Its balance sheet reaches deeper into the market as well. Nvidia's latest filing shows large commitments to cloud services, private investments and partner lease guarantees. The company is helping finance and fill the same clouds that buy its systems.

In practice, Nvidia is building a platform that runs across other companies' clouds.

Q3Is Nvidia already bigger than AWS, Azure and Google Cloud?

No. Nvidia's data-center business is larger than several cloud businesses by quarterly revenue, but those figures measure different stages of the value chain.

Nvidia reported $75.2 billion in Data Center revenue in its latest quarter, including $60.4 billion from compute and $14.8 billion from networking. AWS reported $37.6 billion, Google Cloud reached $20 billion, Microsoft Cloud generated $54.5 billion and Oracle's infrastructure business produced $5.8 billion.

The scale of Nvidia's number is still remarkable. Its Data Center revenue was about twice AWS revenue and nearly four times Google Cloud revenue for the same broad reporting period. It also grew 92% from a year earlier, far faster than the already strong growth reported by the large cloud providers.

Treating that comparison as a cloud-market-share table would be misleading. Nvidia often records revenue when a provider or server maker buys the infrastructure. A cloud operator may recognize the related rental revenue gradually over several years. One large Nvidia system order can appear long before the buyer has earned much from customers using it.

The useful takeaway is blunt: cloud providers currently pay Nvidia first, then spend years earning that money back from customers.

Latest quarterly revenue across the AI infrastructure stack

Q4Is Nvidia really earning cloud revenue?

Only a small part of Nvidia's business currently behaves like recurring cloud revenue.

The latest Data Center quarter came almost entirely from compute and networking. By comparison, Nvidia's filing showed $2.6 billion of remaining performance obligations from contracts longer than one year. That contracted amount equals roughly 3.5% of a single quarter of Data Center revenue.

Deferred revenue reached about $3.1 billion and includes hardware support, software, cloud services, licences and customer advances. Even that broader figure remains small beside Nvidia's quarterly hardware-driven sales.

The margin tells the same story. Nvidia's latest gross margin was about 75%. Cloud operators must pay for buildings, electricity, maintenance, staff, financing and years of depreciation after buying the equipment. Nvidia usually gets paid much earlier in that chain.

So when people call Nvidia the largest AI cloud, they are mixing up the company selling the scarce machinery with the companies renting that machinery by the hour. Nvidia currently captures more value from supplying the cloud than from operating one.

Q5How dependent is Nvidia on the hyperscalers it might disrupt?

Nvidia remains deeply dependent on a handful of large buyers, which limits how directly it can attack the hyperscalers.

Its latest filing says three direct customers generated 21%, 17% and 16% of total revenue. Together, they represented 54% of Nvidia's quarterly sales. Nvidia keeps their identities confidential, and some may be manufacturers or integrators buying for other companies, but the concentration is too large to ignore.

The filing also says one AI research and deployment company contributed a meaningful amount of revenue by buying cloud services through Nvidia's customers. Demand can pass through several layers before it appears in Nvidia's accounts.

AWS, Microsoft, Google and Oracle play two roles at once: they compete with Nvidia in chips and cloud services, while also buying Nvidia systems at a scale few companies can match. A broad Nvidia cloud that consistently undercut them would put some of Nvidia's largest sales at risk.

Nvidia has chosen the less confrontational route. It helps many providers sell more Nvidia-powered computing and lets them compete for the end customer. That way, Nvidia can grow across every major cloud without choosing one winner.

Q6Is DGX Cloud Lepton Nvidia's real cloud strategy?

Lepton is currently Nvidia's clearest cloud strategy because it gives the company reach across providers without the cost of owning every data center.

The marketplace launched with tens of thousands of GPUs from a group of specialist and regional clouds. Developers can stay within Nvidia's software environment while the actual capacity comes from CoreWeave, Crusoe, Lambda, Nebius, Nscale, SoftBank or another partner.

That setup tackles a real weakness in the neocloud market. Most customers know Nvidia, while far fewer know every regional operator selling GPU time. A common Nvidia layer can make the provider underneath feel safer and easier to switch.

If the marketplace becomes widely used, Nvidia can see where demand is building and where capacity is available. It can help developers find systems, give smaller providers access to customers and push common performance standards. The physical operator may then compete mainly on price, location, power and availability.

Nvidia's own spending shows why the partner model is useful. The company has committed tens of billions of dollars to multi-year cloud service agreements, mainly for research and development. It needs huge amounts of external capacity itself, so coordinating partner clouds is more practical than building a global public cloud alone.

Q7Are Nvidia-backed neoclouds becoming a fourth major cloud?

Taken together, Nvidia-backed neoclouds are becoming a major AI infrastructure force, although they still operate as separate companies rather than one unified cloud.

The planned scale has jumped quickly. CoreWeave recently passed one gigawatt of active power, reported more than 3.5 gigawatts under contract and said it is heading toward more than eight gigawatts by 2030. Nvidia and Nebius have announced plans for more than five gigawatts of Nvidia systems by the end of that decade. Nvidia and IREN are working toward as much as five gigawatts.

Those three plans alone exceed 18 gigawatts. The capacity remains planned rather than live, and delivery will depend on power, financing and construction. Still, this is far beyond the experimental clusters that defined the neocloud category a few years ago.

The order books have caught up with the power plans. CoreWeave's latest backlog was $99.4 billion, supported by agreements with companies including Meta and Anthropic. Oracle's infrastructure revenue recently grew 93%, showing that AI demand is also lifting providers outside the traditional top three.

The neoclouds still have separate accounts, software layers, locations, funding structures and customer contracts. Nvidia's marketplace could connect some of that capacity. The ecosystem could become coordinated while ownership stayed fragmented.

Q8Does Nvidia effectively control CoreWeave?

CoreWeave remains independent, yet Nvidia currently has exceptional influence over its technology and expansion.

CoreWeave states in its latest filing that every GPU in its infrastructure is supplied by Nvidia. Its customers have also contractually specified Nvidia GPUs. Changing supplier would mean revisiting customer promises and rebuilding part of the platform, not merely buying different machines.

The financial relationship has deepened. Nvidia invested $2 billion in CoreWeave, and the two companies are working on a buildout of more than five gigawatts of AI factories. CoreWeave is also preparing to deploy Nvidia's Rubin platform as it becomes available.

Yet the risk remains mostly with CoreWeave. It produced $2.08 billion in quarterly revenue but paid $536 million in net interest expense. Interest alone consumed about 26% of revenue. CoreWeave also carried $25.1 billion of principal debt and recorded a $740 million net loss.

Nvidia gets the attractive side of the arrangement. It sells the systems, owns a meaningful investment and benefits from CoreWeave's growth, while CoreWeave handles the debt, leases, power contracts and day-to-day utilization problem.

Q9Is Nvidia financing its own AI demand?

To a degree. Nvidia now supports demand with investments, cloud purchases and guarantees, although the underlying customers are also signing very large independent contracts.

The amounts have become too large to treat as side projects. Nvidia's latest filing lists $27 billion of investment commitments, $30 billion of multi-year cloud service commitments and up to $3.5 billion of partner lease guarantees. Together, those three categories exceed $60 billion before we count $119 billion of manufacturing, supply and capacity commitments.

Some of this money can reinforce Nvidia's own sales. An investment helps a cloud expand, the cloud buys Nvidia systems, and Nvidia may later use or distribute that cloud capacity. A lease guarantee can make a data-center project easier to finance. A long-term cloud purchase gives the provider contracted demand.

Calling all of this circular would go too far. CoreWeave's backlog includes a $21 billion Meta commitment and a multi-year Anthropic agreement. Google Cloud's backlog recently reached $462 billion, while Microsoft reported $627 billion of commercial remaining performance obligations including OpenAI. These customers are committing their own money because they expect heavy AI usage.

The danger is that all these bets can go wrong together. Weaker AI demand could mean slower equipment orders, falling investment values, unused cloud commitments and guarantee losses at the same time.

Nvidia's commitments across the AI infrastructure ecosystem

Q10Can neoclouds survive when GPU scarcity fades?

The best neoclouds can survive lower GPU scarcity, but easy access to Nvidia chips will no longer be enough.

CoreWeave already has enough contracts to keep growing, but its balance sheet leaves little room for a bad year. Its $99.4 billion backlog gives it years of contracted business, and committed contracts generated 98% of its latest quarterly revenue. Demand is real enough to support more than $2 billion in quarterly sales.

The concentration is uncomfortable, though. Its two largest customers produced 65% of revenue. One customer alone represented 45%. CoreWeave must also meet more than $11.7 billion of debt principal payments across the remainder of 2026 and 2027, according to its filing.

As supply improves, customers can compare CoreWeave with AWS, Azure, Google Cloud, Oracle, Crusoe, Nebius and other providers. The winners will need cheaper power, faster deployment, strong software, reliable clusters or better performance per dollar. A provider that only secured GPUs early will lose its edge.

A shared Nvidia marketplace may intensify that pressure. Easier movement across compatible clouds would help the category attract customers, while making individual operators more replaceable. The ecosystem could grow even as weaker members struggle.

Q11Are AWS, Microsoft and Google losing bargaining power?

They have lost bargaining power over the best AI infrastructure for now, but they still control most cloud spending and the routes to enterprise customers.

Synergy Research's latest estimate puts AWS at 28% of global cloud infrastructure spending, Microsoft at 21% and Google at 14%. Together they hold 63% of a market that reached $129 billion in one quarter. Their combined position has barely moved even as Nvidia's revenue has exploded.

Customers are still pouring money into all three. AWS grew 28%, Azure grew 40% and Google Cloud grew 63% in their latest reported quarters. Google said products built on its generative AI models grew nearly eightfold, while Microsoft said its AI business passed a $37 billion annual revenue run rate.

The hyperscalers are spending aggressively to protect that position. Google expects capital expenditure of roughly $180 billion to $190 billion this year. Microsoft added another gigawatt of capacity in one quarter and says its overall footprint should double in two years. Amazon's cash flow has been heavily affected by its AI infrastructure buildout.

Nvidia currently sets many of the technical rules for frontier clusters. The hyperscalers still decide how that computing is packaged with storage, security, data services and enterprise contracts. Their leverage has narrowed around chips, while remaining strong almost everywhere else.

Q12Can custom AI chips break Nvidia's grip?

Custom chips will take meaningful workloads from Nvidia, especially inference, but they are unlikely to break its overall grip soon.

These chips are already running at production scale. AWS says nearly one million Trainium2 chips are training and serving Anthropic's Claude, while Trainium3 is available for newer workloads. Google's seventh-generation Ironwood TPU is generally available and powers both training and inference. Microsoft's Maia 200 is live in its Iowa and Arizona data centers. Meta says MTIA is deployed at scale for ranking and advertising workloads.

A full replacement is unnecessary. A hyperscaler gains bargaining power by moving a large block of predictable work onto its own silicon. Repetitive inference, recommendation systems and internal models are especially attractive because the operator controls the software and can keep the hardware busy.

Nvidia remains harder to displace when customers need portability, frontier performance and a mature development environment across several clouds. A company may accept Trainium inside AWS or TPU inside Google, then still choose Nvidia when it wants the same workload available through multiple providers.

The market is likely to split. Custom silicon will absorb more stable, high-volume computing, while Nvidia stays strongest in demanding general-purpose AI and new workloads where flexibility matters more than the lowest possible chip cost.

Q13Is CUDA a stronger moat than Nvidia's chips?

CUDA is probably Nvidia's strongest long-term moat because it can outlast the advantage of any single chip generation.

Nvidia says more than four million developers use its platform and more than 3,000 applications have been optimized for its technologies. The exact totals come from the company, but the depth of the ecosystem is visible in the libraries, frameworks, deployment tools and engineering knowledge built around it.

A rival chip needs more than impressive benchmark results. It must run common models reliably, connect thousands of accelerators, support popular frameworks, handle debugging and give engineers a practical migration path. Nvidia covers those needs through CUDA, NCCL, TensorRT, NIM and a long list of specialized libraries.

That software reduces the pain of moving from one Nvidia generation to the next. Existing code, tools and staff knowledge usually remain useful. Switching to a different architecture can require weeks or months of engineering even when the alternative hardware is cheaper.

Frameworks such as PyTorch, JAX and vLLM are making hardware less visible to developers, and cloud providers are improving their compilers. Still, large-scale AI performance depends on details below the framework, including kernels, memory, communication and cluster layout. Nvidia's control across those layers keeps CUDA's advantage alive.

Q14Will inference make Nvidia stronger or weaker?

Inference will make Nvidia much larger in dollars while reducing the chance that it keeps training-like market share.

Frontier training favours big, tightly connected clusters where reliability and software maturity can be worth more than a lower chip price. Inference is broader. One system answers complex questions, another ranks adverts, another processes video and another runs a small model close to the user.

That variety gives custom chips room to grow. Meta can tune MTIA for recommendation workloads. Microsoft built Maia 200 around inference economics. Google can run Gemini on TPUs, while AWS can push suitable customers toward Trainium.

Nvidia is fighting back at the system level. Its newer platforms combine accelerators, networking, low-precision computing and software such as TensorRT and Dynamo. The company argues that a more expensive system can still produce cheaper tokens if it serves more requests and wastes less power.

The likely split is straightforward. Nvidia should remain very strong for complex, fast-changing and portable inference. Purpose-built chips will win more of the repetitive jobs where one operator can optimize every layer.

Q15Does Nvidia's networking business change the answer?

Nvidia's networking business materially strengthens its grip on the AI cloud because GPU market share alone understates how much of the system it controls.

Its latest quarterly networking revenue reached $14.8 billion, up 199% from a year earlier. That business now has an annualized run rate near $60 billion, which would already make it a major technology company on its own.

Modern AI clusters need thousands of accelerators to exchange data with very little delay. Slow communication leaves expensive chips waiting. Nvidia addresses that problem through NVLink inside systems and through InfiniBand, Spectrum-X Ethernet, switches and networking software across racks.

Cloud buyers increasingly choose a complete computing system rather than a chip. A rival accelerator can look competitive in isolation and still produce less useful output once networking, memory and software are included.

Nvidia now influences data-center layouts, cooling, power density and deployment schedules because its rack-scale designs shape the building around them. Cloud operators still own the facilities, but more of the technical blueprint comes from Nvidia.

Q16Where does Nvidia still lack control in the AI cloud?

Nvidia still lacks the customer ownership and service breadth that make AWS, Azure and Google Cloud difficult to replace.

Companies choose a cloud for much more than accelerators. They keep data in storage systems, connect databases, manage identities, apply security rules, monitor applications and buy support through one account. Those surrounding services create years of operational dependence.

Microsoft has extra reach through Microsoft 365, GitHub, Dynamics, Windows and its enterprise sales force. Google combines cloud infrastructure with BigQuery, Workspace, cybersecurity and its own models. AWS has the broadest infrastructure catalogue and a huge developer ecosystem.

Recent cloud results show how powerful those relationships remain. Google Cloud said new customer acquisition doubled and the number of deals worth $100 million to $1 billion also doubled. Microsoft reported $627 billion of commercial remaining performance obligations including OpenAI. Nvidia's small recurring revenue base cannot match commitments of that size today.

Governments create another barrier. Sovereign and regulated workloads require local operations, compliance, support and control over sensitive data. Nvidia can supply the systems, while cloud providers and local partners handle the legal and commercial relationship.

For Nvidia, owning less may even be preferable. Its gross margin sits near 75%, while the operators carry the buildings, electricity, debt and idle-capacity risk. The company can dominate the machinery without taking on every burden of running the cloud.

Q17Is Nvidia taking over the AI cloud?

Yes, but only at the infrastructure layer. Nvidia currently controls more of the AI cloud's technical standard and economics than any other supplier, while the cloud market itself remains in the hands of the large providers.

The strongest evidence sits in the technical stack. Nvidia generated $75.2 billion from Data Center in one quarter, including a networking business growing almost threefold. CUDA remains the familiar development environment, and complete Nvidia rack designs increasingly shape how advanced AI facilities are built.

Its influence now extends into capacity. Nvidia invests in neoclouds, guarantees selected facility leases, buys partner capacity and uses its marketplace to connect developers with providers. That gives Nvidia reach across a distributed cloud network without requiring it to own every data center.

AWS, Microsoft and Google still hold 63% of global cloud infrastructure spending. Their cloud businesses are growing quickly, their backlogs are enormous and their custom chips are already handling real production workloads. They control the data, accounts and surrounding software that keep customers attached to a cloud.

Nvidia is becoming the common infrastructure layer beneath the AI cloud, not the next AWS. That position may be more profitable. It can set much of the technical standard, collect money from nearly every provider and leave most construction and utilization risk to its customers.

So the title is partly true today. Nvidia is taking over the machinery and software foundation of the AI cloud, while the hyperscalers continue to own the cloud businesses built on top of it.