Nov 17, live webinar: ECS to EKS without freezing the roadmap for a quarter

Who Sits Between You and the GPU? Microsoft's Neocloud Delays and the Case for Your Own Cloud Account

Bloomberg reported on November 5, 2026 that Microsoft is tightening oversight of neocloud partners Nebius, IREN and Nscale after delivery delays. Here is what the public contracts say, how many companies sit between you and a rented GPU, and when to deploy GPU workloads in your own cloud account instead.

Romaric Philogene
CEO & Co-founder
NOV 17, 2026 · 18 MIN
Who Sits Between You and the GPU? Microsoft's Neocloud Delays and the Case for Your Own Cloud Account

Key points:

  • Bloomberg reported it, and Microsoft has not confirmed the details. On November 5, 2026, Bloomberg (Brody Ford) reported that Microsoft is tightening oversight of data center projects run by Nebius, IREN and Nscale after construction and delivery delays, and is reconsidering further third-party deals. Microsoft did not confirm specifics; a spokesperson said it manages partners "against our requirements and customer demand."
  • The public contracts are large and still mostly in delivery. Nebius signed about $17.4 billion (up to $19.4 billion) through 2031, IREN about $9.7 billion over five years, Nscale about 200,000 NVIDIA GB300 GPUs across four sites, and Lambda an undisclosed "multibillion-dollar" deal. Nebius now says only about half of its Microsoft capacity in New Jersey ships by year-end, the rest in Q2 2027.
  • A resale chain stacks counterparties. Between your workload and a GPU there can be a chip vendor, a server OEM, a site operator, a power utility, a neocloud, a hyperscaler and a reseller. Every link can slip, and when capacity is short, the party closest to the hardware decides who gets served first.
  • Renting through a chain still makes sense for short, flexible work. Experiments, burst inference and one-off training runs fit. Multi-year, take-or-pay capacity for production does not, unless the contract is in your name and you have a second provider ready.
  • Your own cloud account does not create GPUs, but it removes layers. You hold the contract, the IAM, the logs and the exit. Pair it with standard Kubernetes on two providers and a delayed data center becomes a redeploy instead of an outage.

Qovery · Infrastructure control plane
Build with Claude Code, Deploy with Qovery
Learn more

Microsoft has more GPU purchasing power than almost any company on earth, and according to Bloomberg it still could not get some of its rented capacity on time. If that happens to Microsoft, it can happen to a 40-person AI startup renting the same GPUs three links further down the chain.

I wrote about where to find GPU capacity in 2026 a few weeks ago. This post takes the next question: once you find capacity, who sits between you and the GPU, and how much of that chain should you accept?

What did Microsoft reportedly do with Nebius, IREN and Nscale?

Bloomberg reported on November 5, 2026 that Microsoft is tightening oversight of data center projects involving IREN, Nebius and Nscale, and is reconsidering whether to sign more agreements with third-party infrastructure providers (Bloomberg, "Microsoft Sours on Neoclouds Amid Data Center Delays, Costs", via Yahoo Finance). The story is attributed to people familiar with the matter. Treat the details below as reported, not confirmed.

Here is what the report says, provider by provider:

  • IREN. Its Texas campus ran into cabling issues, a generator fire in April and hardware delivery delays disclosed in August. Microsoft and IREN reportedly negotiated a revised timeline that could push completion back by weeks or months. IREN declined to comment.
  • Nebius. Microsoft reportedly sent its own workers to Nebius's New Jersey site, which was fined for running generators without permits. TechCrunch puts that fine at $1.1 million, levied on DataOne, the third-party site operator (TechCrunch, November 12, 2026).
  • Nscale. Microsoft reportedly walked away from several Nscale projects, including a roughly $3 billion deal tied to a UK data center that was held up by grid power. Nscale told Bloomberg the UK withdrawal did not affect its other deals with Microsoft.

Microsoft's side of the story is in its own numbers. Bloomberg reports that Microsoft added more than $200 billion of new data center lease commitments over the last year, and that CFO Amy Hood told investors on the October 28 earnings call that delays in some third-party capacity were disrupting the company's spending plans. A Microsoft spokesperson said the company manages relationships with infrastructure partners "against our requirements and customer demand."

The market read it fast. IREN shares fell 18% on the day of the report (Investing.com, via Yahoo Finance), and Bloomberg put the drop for both IREN and Nebius at about 20%.

I want to be fair to these teams. Building gigawatt-scale data centers in two years, against grid queues, permit offices and an NVIDIA allocation schedule, is brutally hard. Nebius CEO Arkady Volozh said it plainly on his Q3 call: "Building at this scale is hard, and delays are part of our industry right now" (TechCrunch). I have shipped infrastructure late too. Nobody in this story is careless. The trouble comes from how many companies a single rented GPU now depends on.

Which Microsoft neocloud contracts are public?

Four Microsoft neocloud contracts have public terms: Nebius, IREN, Nscale and Lambda. CoreWeave, the largest, predates them and is covered further down. Here is what each company disclosed, plus what has been reported since.

ProviderAnnouncedDisclosed valueDisclosed GPUs and sitesDelivery schedule (as disclosed)Status reported in Nov 2026
NebiusSept 8, 2025About $17.4B through 2031, up to about $19.4B with options (The Register)New data center in Vineland, New JerseyTranches during 2025 and 2026About half of the Microsoft capacity by end of 2026, the rest in Q2 2027 (TechCrunch)
IRENNov 3, 2025About $9.7B over five years, 20% prepaid (TechInformed)NVIDIA GB300 at the 750MW Childress, Texas campusPhased through 2026Revised timeline that could add weeks or months (Bloomberg)
NscaleOct 15, 2025Not disclosed (Nscale)About 200,000 GB300: ~104,000 in Texas, ~12,600 in Sines (Portugal), ~23,000 in Loughton (UK), ~52,000 in Narvik (Norway)Sines from Q1 2026, Texas from Q3 2026, UK from Q1 2027Roughly $3B UK deal dropped; other deals unaffected per Nscale (Bloomberg)
LambdaNov 3, 2025Not disclosed, "multibillion-dollar" (TechCrunch)Tens of thousands of NVIDIA GPUs, including GB300 NVL72Not disclosedNot named in the Bloomberg oversight report

Two details in that table matter for anyone renting GPUs.

The first is the supply chain inside each deal. IREN disclosed a separate agreement to buy the GPUs and related infrastructure from Dell for about $5.8 billion (TechInformed). Nebius's New Jersey site is run by a third-party operator. So even the "direct" contract between Microsoft and a neocloud already depends on an OEM, a site operator and a utility.

The second is the financing. Nebius announced $40 billion of planned 2027 capital spending and told investors that customer prepayments and asset-backed financing could cover up to 85% of it (TechCrunch). That model works when projects ship on time. It gets tight when a large customer's capacity slips by two quarters, because the cash for the next site depends on the last one going live.

Lambda is a different case. It has worked with Microsoft for more than eight years, and nothing I found reports problems with its deal. I include it because it is public and because it shows the same pattern: a hyperscaler renting GPU capacity from a specialist instead of building all of it.

Who sits between you and the GPU when you rent AI compute?

In a resale chain, the GPU you rent can pass through five to seven companies before it reaches your workload, and each one is a counterparty that can be late, change terms or reprioritize. Here is the longest version of the chain as it exists in late 2026:

  1. NVIDIA allocates the chips.
  2. A server OEM (Dell, Supermicro, HPE) builds and ships the racks. IREN's $5.8 billion Dell order is this link.
  3. A site operator or landlord owns the building and the permits. DataOne in New Jersey is this link for Nebius.
  4. A utility delivers the power. Grid access is what Bloomberg says sank the UK Nscale project.
  5. A neocloud (Nebius, IREN, Nscale, Lambda, CoreWeave) runs the GPUs as a service.
  6. A hyperscaler (Microsoft here) buys that capacity and resells it inside its own cloud or to its own large customers.
  7. A reseller or marketplace (a GPU broker, a model API, an aggregator) slices it again for smaller buyers.

Your team sits at position eight.

When the chain works, nobody notices it. When capacity is short, the party closest to the hardware decides who gets served first, and position eight is rarely first. Microsoft has enough weight to send its own engineers to a neocloud's site. A startup buying GPU hours through a broker does not even know which site its GPUs sit in.

Microsoft executive Mat McBride gave the cleanest description of this I have seen. Bloomberg quotes him saying, "We've had to pinch our nose and borrow GPUs from some of our competitors." If the second-largest cloud on earth borrows GPUs from rivals, a capacity plan with a single supplier is optimistic for everybody else.

What counterparty risks does a GPU resale chain add?

A resale chain adds five risks that a direct contract in your own cloud account mostly removes: delivery risk, priority risk, concentration risk, financial risk and exit risk.

Delivery risk. Every physical link can slip. The IREN timeline moved because of cabling, a fire and late hardware. The Nebius timeline moved partly because of a permit fine and neighbor lawsuits over noise, per TechCrunch. None of those are software problems your team can fix, and none of them show up in a GPU price list.

Priority risk. When a delayed site finally comes online, the capacity goes to whoever the contract says first. Microsoft's contracts with Nebius and IREN are for Microsoft. If you rent the same GPUs indirectly, you are downstream of Microsoft's own needs, and Microsoft said a year ago that it expected its cloud capacity to stay constrained through at least June 2026 (Directions on Microsoft).

Concentration risk. Neoclouds often depend on one huge customer. CoreWeave's IPO filing showed Microsoft accounted for 62% of its 2024 revenue (The Register). That is normal for a company scaling this fast, and CoreWeave has been diversifying since. But it means the provider's roadmap, site choices and on-demand inventory follow its anchor customer. A CoreWeave spokesperson told Bloomberg the company continues to have "a great relationship" with Microsoft, and I have no reason to doubt it. It still means your capacity depends on someone else's relationship.

Financial risk. Gartner predicted in October 2026 that cost pressure and a lack of a differentiating business model will force over 80% of neocloud providers to cease independent operations by 2030 (Gartner). "Cease independent operations" includes acquisitions, which are often fine for customers. It also includes shutdowns, and those are not fine if your model weights, your images and your network setup live only on that provider.

Exit risk. The longer the chain, the less you control when you need to leave. If your GPUs come through a broker, you may have no direct contract with the operator, no access to the underlying cluster and no notice period you can enforce.

None of this is a comment on any one company's quality. Nebius reported Q3 revenue of $981 million, up 571% year over year, and $7.4 billion in cash (TechCrunch). These are serious businesses. The risk comes from the structure. More links mean more places for a delay to start, and less say for the buyer at the end.

Does renting AI compute through a resale chain still make sense?

Yes, for short, flexible and interruptible work, and no for multi-year production capacity that you cannot move. The deciding factors are how long you commit, how much of your revenue depends on the capacity, and whether you can redeploy elsewhere within a day.

Here is how the main ways of getting GPUs compare, from the most layers to the fewest:

OptionWho you contract withCompanies between you and the GPUTypical commitmentWhat happens if a site slipsBest for
GPU marketplace or broker (RunPod, Vast.ai, NVIDIA DGX Cloud Lepton)The marketplaceMost (marketplace, then a neocloud or data center, then the physical chain)Hourly, no termYou get different GPUs, or none, with little noticeExperiments, batch jobs, burst inference
Hyperscaler capacity backed by a neocloudThe hyperscalerHyperscaler, neocloud, site operator, OEM, utilityPay as you go, or enterprise commitmentThe hyperscaler absorbs it; you see quota or region limitsTeams already standardized on that hyperscaler
Direct neocloud contract (managed Kubernetes, reserved nodes)The neocloudNeocloud, site operator, OEM, utilityMonths to years, often prepaidCredits or a revised date per contract; you plan around itLarge training runs, price-sensitive inference at scale
Hyperscaler reservation in your own accountThe hyperscaler, in your nameHyperscaler and its supply chainDays to months (e.g. EC2 Capacity Blocks up to 6 months)You hold a dated reservation and can see it in your accountPlanned training windows, launch capacity
On-demand or flex capacity in your own accountThe hyperscaler, in your nameHyperscaler and its supply chainNoneNo reservation, so you queue or move regionsElastic inference, development
Multi-provider setup, your accounts, one platformEach provider, in your nameSame as above, per providerBaseline on one, burst on anotherShift the workload to the other providerProduction inference that cannot stop

A few concrete numbers for the own-account options. AWS EC2 Capacity Blocks for ML reserve 1 to 64 GPU instances (up to 512 GPUs) for up to 6 months, booked up to 8 weeks ahead, on P6, P5 and P4d instances. On Google Cloud, GKE flex-start with Dynamic Workload Scheduler runs GPU nodes for up to seven days with discounts of up to 53%, and a single request can ask for up to 1,000 VMs. Azure offers on-demand capacity reservations that hold VM capacity in a region without a term commitment, billed whether or not you use it.

Your own Azure account does not escape the Microsoft story: some Azure GPU capacity is exactly the neocloud capacity Bloomberg wrote about. Your own account changes who you contract with and what you control, and it adds no GPUs to the region.

Here is how I would decide:

  • Renting through a chain is fine when the job is under a few weeks, can restart from a checkpoint, and you have no data residency constraint.
  • Sign directly with a neocloud when the price gap is large and real, you get managed Kubernetes, and the contract has delivery dates with credits or an exit if they slip.
  • Keep production in your own account when the workload earns revenue every hour, touches customer data, or must sit next to your databases.
  • Never sign a multi-year prepaid commitment for capacity you cannot move, with a provider you cannot leave, for a workload that has no second home.
Keep the GPU contract in your name.
Qovery runs Kubernetes and GPU node pools inside your own AWS, GCP, Azure or Scaleway account, or on a cluster you already rent from a GPU cloud. One control plane, several providers, no extra platform team.

What does deploying GPU workloads in your own cloud account protect you from?

Your own cloud account removes the contractual and operational layers of the chain, while the physical layers stay. It cuts out the reseller and gives you a direct line to whoever runs the hardware, and that is where most of the downstream risk sits.

What you gain:

  • A contract in your name. Reservations, credits and committed-spend discounts count toward your own agreement with AWS, Google Cloud, Azure or Scaleway. Our AWS EDP checklist covers how to sign one and keep an exit.
  • Visibility. You see your quota, your reservations and your capacity errors directly, instead of reading a broker's status page.
  • Control of the data path. Weights, prompts and logs stay in your VPC, under your IAM and your audit logs.
  • A clean exit. If your images, Helm charts and pipeline target standard Kubernetes, moving to another provider is a redeploy.

What you do not gain:

  • More GPUs. Hyperscalers are short too. Google said in April 2026 it was "compute constrained in the near-term", and Microsoft has said the same about Azure. An own-account setup with one provider and no reservation still waits in the same queue.
  • Immunity from the physical chain. Your hyperscaler buys from the same NVIDIA, OEMs and utilities.

So the useful posture is your own accounts, plural. One provider where you hold baseline capacity, at least one more where you can burst or fail over, both running the same containers on Kubernetes. That design treats any single supplier, neocloud or hyperscaler, as replaceable. Our GCP capacity article walks through the same idea when the constraint is your primary cloud.

How should you size capacity commitments when supply is this uncertain?

Commit to the capacity you are sure to use, and buy flexibility for everything above it. Long commitments are where a slipped data center hurts most, because you pay for capacity that arrives late, and the work you planned for it has to wait.

A sizing approach that has held up with the teams I talk to:

  1. Measure your floor. Look at the last 90 days of GPU hours for production inference. The lowest steady level is your baseline. Commit to that, not to your forecast.
  2. Match commitment length to certainty. A 6-month Capacity Block for a planned training run is easy to justify. A 5-year take-or-pay contract for a product that is 8 months old is mostly a bet on your roadmap.
  3. Read the delivery clauses before the price. Ask what happens if capacity arrives a quarter late. Is there a credit? Can you terminate the undelivered tranche? Is the prepayment refundable? Microsoft could renegotiate IREN's timeline; your bargaining power is usually smaller, so get it in writing.
  4. Ask where the GPUs physically are. Which site, which operator, which power source? If a provider cannot or will not tell you, assume you are deep in a chain.
  5. Price the second provider too. Keep a smaller footprint warm on another cloud. The cost is a cluster, a pipeline target and a few idle nodes, and it buys you the ability to move.

How do you set up multi-provider GPU fallback without doubling your platform work?

Run the same container on standard Kubernetes with GPU nodes on two providers, keep model weights where both can reach them, and deploy to both from one pipeline. For a small team, the hard part is keeping two clusters operated and up to date.

The checklist:

  • One image, one serving stack. Pin the CUDA base image and the inference server (vLLM, SGLang, TGI or Triton) so the same digest runs on both providers. Our Lambda and CoreWeave migration guide has a GPU SKU mapping table if the two providers offer different cards.
  • Weights in portable storage. Store them in object storage that both clusters can read without large egress bills, or mirror them into each provider and version them.
  • Two clusters, both warm. A cluster you have never deployed to is not a fallback. Run a small replica set on the second provider all the time, even at one node.
  • Traffic you can shift. Put the model behind an endpoint that can route to either cluster, with health checks, so failover is a weight change.
  • Private networking where data crosses. VPN, a cloud interconnect, Tailscale or Cilium Cluster Mesh, depending on volume. The GPU capacity guide compares them.
  • One pipeline, one set of permissions, one audit log. Two providers should not mean two deployment systems.

For scheduling batch and training jobs across clouds (SkyPilot, Ray, Kueue and others), see our multi-cloud GPU orchestration comparison.

Where does Qovery fit if you want GPUs in your own account?

Qovery is an Agentic DevOps platform. It runs Kubernetes clusters and GPU node pools in your own cloud accounts and deploys your services to them from one control plane. Qovery never hosts your applications or resells GPUs; the contract with each provider stays in your name, which removes the reseller link from the chain above.

What that looks like in practice:

  • Your accounts, any of them. Qovery creates and manages clusters on AWS, Google Cloud, Azure and Scaleway, and attaches Kubernetes clusters you already run, including a managed Kubernetes cluster rented from a GPU cloud.
  • GPUs per service. You request GPUs on the service that needs them, and Qovery provisions GPU node pools on the cluster where the service runs. On AWS, node provisioning uses Karpenter.
  • Clone to where capacity is. Clone an environment to a cluster on another provider and deploy it with the same variables, pipeline and settings. That is the multi-provider fallback without a second platform team.
  • Guardrails stay on. RBAC, audit logs and OPA-based policy tokens apply on every cluster, so adding a provider does not open a hole in your permissions.
  • Bring your own network. Qovery has no built-in cross-cloud networking. It is configurable with a VPN, a cloud interconnect, Tailscale or Cilium Cluster Mesh, and Qovery deploys on top.

The Free plan includes one cluster of your own, which is enough to test a GPU service in your account before you decide anything.

How do I set this up with Qovery from Claude Code or Cursor?

Install the Qovery Agent Skills once, and your coding agent can create the cluster, deploy your inference service with GPUs, and clone it to a second provider, asking you to approve each step.

Install the skills globally for Claude Code, Cursor, Codex, OpenCode and Gemini CLI:

BASH
curl -fsSL https://skill.qovery.com/install.sh | bash

Or scope them to one project:

BASH
curl -fsSL https://skill.qovery.com/install.sh | bash -s -- --project

Add the Qovery MCP Server to Claude Code. It uses OAuth and is read-only by default, so it can inspect clusters and deployments without changing anything:

BASH
claude mcp add --transport http qovery https://mcp.qovery.com/mcp --callback-port 4242

Only when you want the agent to make changes, add write access. This also needs write access enabled in the Qovery Console:

BASH
claude mcp add --transport http qovery "https://mcp.qovery.com/mcp?read_write=true" --callback-port 4242

The skills (qovery-onboard, qovery-deploy, qovery-troubleshoot, qovery-optimize, qovery-speedup, qovery-preview, qovery-terraform) do the deploying, and the MCP Server operates what already exists. Then ask in plain language, for example: "deploy this vLLM service with one GPU to my AWS account with Qovery, then clone the environment to my GCP cluster." Swap the providers for Azure, Scaleway or a Kubernetes cluster you already rent.

Start with the AI agent quickstart or the Claude Code integration.

Frequently asked questions
Is Microsoft really tightening oversight of Nebius, IREN and Nscale?

Bloomberg reported it on November 5, 2026, citing people familiar with the matter, and Microsoft has not confirmed the specifics. Microsoft's CFO did say on the October 28, 2026 earnings call that some third-party capacity deliveries were delayed, and Nebius has since said only about half of its Microsoft capacity in New Jersey will ship by the end of 2026, with the rest in Q2 2027.

Does renting AI compute through resale chains still make sense after the Microsoft neocloud delays?

It makes sense for short, restartable and flexible work such as experiments, batch jobs and burst inference, where losing a provider costs you hours. It makes much less sense for multi-year production capacity, because every company in the chain is a counterparty that can be late, and the buyer at the end of the chain is served last when capacity is short.

Should teams deploy GPU workloads directly in their own cloud account instead?

For production, yes, ideally in two accounts on two providers running the same containers on Kubernetes. Your own account puts the contract, quota, IAM and logs in your name and makes leaving a redeploy. It does not create GPUs, so pair it with reservations for baseline capacity and a second provider for fallback.

What are Microsoft's public neocloud contracts worth?

Nebius disclosed about $17.4 billion through 2031, up to about $19.4 billion with options. IREN disclosed about $9.7 billion over five years with 20% prepaid. Nscale disclosed about 200,000 NVIDIA GB300 GPUs across Texas, Portugal, the UK and Norway without a dollar value, and Lambda disclosed a "multibillion-dollar" deal without a figure.

What should I ask a GPU provider before signing a capacity commitment?

Ask which site and operator will host your GPUs, what happens if delivery slips by a quarter (credits, termination of undelivered tranches, refundable prepayments), and whether you get managed Kubernetes so your workload stays portable. Then make sure a second provider can run the same image before you sign anything longer than a year.

Romaric Philogene
About the author
Romaric Philogene

Romaric founded Qovery to make Kubernetes accessible to every engineering team. He writes about platform strategy, developer experience, and the future of cloud infrastructure.

Next step

Keep the GPU contract in your name.

Qovery runs Kubernetes and GPU node pools inside your own AWS, GCP, Azure or Scaleway account, or on a cluster you already rent from a GPU cloud. One control plane, several providers, no extra platform team.