The revenue backlogs underpinning the world’s biggest cloud providers rest on the spending commitments of two pre-profit companies. CIOs should be paying close attention.
A report from The Information, published this week, put a precise number on something that has been quietly building for some time. Together, Anthropic and OpenAI now account for roughly half of the combined $2 trillion revenue backlog held across the four largest US cloud providers – Amazon, Microsoft, Google, and Oracle.
That figure is extraordinary. What it reveals about the structural health of the cloud market deserves scrutiny well beyond the usual AI coverage cycle.
What a revenue backlog actually is
Revenue backlog, in the cloud context, refers to contractually committed future spending that hasn’t yet been recognised as revenue. When Anthropic signs a reported $200 billion, five-year compute agreement with Google Cloud, that commitment sits in Google’s backlog. It’s not earned revenue. It’s a promise – contingent on the AI company continuing to grow, continuing to spend, and continuing to operate at the scale its projections assume.
By themselves, large backlogs are generally a good sign. They indicate forward demand. The problem here is the concentration. When two companies – neither of them currently profitable – represent half of a $2 trillion figure, that’s not a sign of a healthy, distributed market. It’s a sign of a market built on two very large bets.
The dependency runs both ways
On one side, Anthropic and OpenAI need the compute. OpenAI is expected to spend roughly $45 billion on servers in 2026 – up from $17 billion the year before – with the majority of that going to Microsoft. Anthropic projected over $20 billion for the same period, and that figure was set before its annualised revenue surged to $30 billion in April 2026, making a higher spend likely. These are not discretionary purchases. They are existential infrastructure requirements for companies racing to stay at the frontier of model development.
On the other side, the cloud providers have become financially intertwined with these AI companies’ trajectories in ways that are rarely discussed publicly. AWS’s backlog jumped 49 per cent in the first quarter, driven in part by a $100 billion OpenAI commitment. Google Cloud’s backlog is now more than 40 per cent accounted for by Anthropic’s five-year deal alone. The hyperscalers are, in effect, underwriting the AI arms race – and taking on concentrated exposure in return.
By 2029, Anthropic has budgeted a combined total of roughly $200 billion across Microsoft, Google, and Amazon. OpenAI plans approximately $180 billion for that year alone. Both projections assume the companies grow their revenues 20 to 30 times over their 2025 levels within four years.
History suggests that aggressive long-range technology spending projections – particularly in newly formed markets – miss. Sometimes they miss upward. Sometimes they don’t.
What this means for enterprise cloud pricing
Here is where this becomes directly relevant to CIOs and technology leaders who aren’t building frontier AI models.
Cloud pricing has historically followed a predictable trajectory: unit costs fall as infrastructure scales. That dynamic hasn’t disappeared, but it is being distorted. When hyperscalers commit enormous capacity to serve two dominant AI customers, the calculus around resource allocation changes. GPU and accelerator capacity – already constrained globally – gets prioritised toward the highest-revenue commitments. Enterprise workloads that don’t carry the same spending weight get deprioritised in the queue.
This is already happening. Enterprises report that provisioning GPU capacity for AI workloads takes significantly longer than it did 18 months ago, and pricing for on-demand compute has firmed up accordingly. If either Anthropic or OpenAI misses the aggressive growth projections underpinning these contracts, the cloud providers face a revenue shortfall that will need filling. That pressure has historically landed on enterprise pricing.
Resource hoarding and the capacity question
The scale of compute being reserved for AI training and inference has material consequences for everyone else sharing the same infrastructure. Data centre capacity is finite. So is energy availability. The International Energy Agency projects that data centres will consume more than 1,000 terawatt-hours of electricity globally in 2026 – a figure that has roughly doubled in five years.
When two companies plan to spend a combined $380 billion on compute by 2029, the downstream effect on available capacity for other enterprise workloads is not theoretical. It’s a planning constraint. Most enterprise technology teams are not yet modelling it.
Multi-cloud strategies, often pursued for resilience or cost reasons, may prove increasingly valuable for a different reason: they reduce dependency on any single provider’s capacity allocation decisions at a moment when those decisions are being shaped by customers with very different profiles to yours.
The concentration risk most enterprises aren’t stress-testing
There is a scenario that enterprise risk functions are not, in most cases, formally evaluating: what happens to your cloud infrastructure commitments and SLAs if one of the two companies propping up your provider’s backlog materially changes its spending profile?
This isn’t a prediction. Both Anthropic and OpenAI have strong recent momentum – Anthropic’s annualised revenue reached $30 billion in April 2026, while OpenAI runs at approximately $24 billion. The near-term trajectory is positive for both. But the projection-to-backlog ratios are extreme, and the enterprise question isn’t whether you trust these companies. It’s whether your infrastructure strategy accounts for the possibility that the market structure underpinning your cloud provider’s financials could shift meaningfully within your planning horizon.
Vendor risk assessment frameworks – typically applied to SaaS providers and software supply chains – need extending to cloud infrastructure. The scale of dependency this week’s reporting makes explicit is a reasonable prompt to start.
What CIOs can actually do
The answer is not to abandon the cloud. It’s to be more deliberate about how you’re positioned within it.
Genuinely implemented multi-cloud architecture – not just declared on a slide deck – reduces the degree to which a single provider’s capacity decisions affect your workload availability. Open standards, such as Apache Iceberg for data portability, reduce switching friction if conditions change. Negotiating contract terms that include SLA protections and capacity guarantees becomes more important when the provider’s own commitments are concentrated in a small number of customers with vastly different spending profiles to yours.
The cloud remains a powerful model for enterprise infrastructure. The argument here is for going in with eyes open to how the market has changed – not for retreating from it.
This piece pairs with our earlier analysis of cloud-first strategies and their structural limits. The picture that emerges across both is consistent: the economics of enterprise cloud are in a more complicated place than most infrastructure strategies currently reflect.
Q&A: Cloud Concentration Risk – What Technology Leaders Need to Know
Does this mean our cloud provider will raise prices?
Not necessarily, and not immediately. But the conditions for pricing pressure exist. When a provider’s backlog is heavily concentrated in customers with different needs and negotiating leverage to standard enterprise accounts, the pricing dynamics over a three to five year horizon become harder to predict. Contract reviews and multi-year rate locks are worth revisiting with this context in mind.
Should we be reconsidering our vendor commitments?
Reconsidering and renegotiating are different things. The concentration risk this story exposes is an argument for having that conversation with your provider – not for a wholesale exit. The question to ask is whether your current contracts provide adequate protection in a market where your provider’s own commitments are heavily concentrated elsewhere.
What happens if one of these AI companies reduces its cloud spending?
The cloud providers would face a revenue gap that could affect their investment in infrastructure capacity and, over time, their pricing approach with enterprise customers. The more acute near-term risk is operational – capacity that has been reserved for AI training workloads could become more available, but the investment cycle that shaped data centre builds will have already run. Enterprise cloud teams would see pricing and capacity effects with a lag.
Is multi-cloud genuinely viable as a hedge, or does it just add complexity?
It depends on how it’s implemented. Poorly executed multi-cloud – where workloads are duplicated rather than distributed based on capability and cost – adds complexity without proportionate benefit. Well-designed multi-cloud architecture, built around open data formats and workload portability, meaningfully reduces the single-vendor exposure that this situation illustrates. The maturity of open standards has made genuine workload portability more achievable than it was three years ago.
How do we start assessing our own exposure?
Map your critical workloads against your cloud commitments, then layer in your provider’s own customer concentration. What percentage of your strategic cloud provider’s forward revenue depends on Anthropic and OpenAI continuing to grow as projected? That number – now easier to estimate from public reporting – should be sitting in your infrastructure risk register.
Working Through This With Vertex Agility
Cloud concentration risk, vendor dependency, and multi-cloud architecture are live topics in the conversations our teams are having with technology leaders right now. What this week’s reporting has made explicit is something that was always structurally present: enterprise cloud infrastructure is now entangled with the fortunes of a handful of AI companies in ways that haven’t been reflected in most risk frameworks.
Our Cloud Consultancy practice works with organisations to design and implement cloud architectures that hold up when market conditions shift – not just when everything goes to plan. That means honest vendor risk assessment, multi-cloud architecture where it makes sense, and contract structures that provide genuine protection. Our Data Consultancy ensures your data workloads remain portable and your pipelines aren’t locked into a single provider’s capacity decisions. And our AI Consultancy helps organisations move from experimental AI investment toward production-grade adoption that justifies its costs – rather than adding further to the compute backlog with unclear returns.
The cloud remains the right foundation for most enterprise technology strategies. The question is whether yours has been built with enough structural awareness of how the market around it has changed.
Two free resources are worth picking up if this article has raised questions about your own position. The Downtime Defence Audit takes a few minutes to complete and returns a detailed report assessing your disaster recovery readiness, backup strategies, and business continuity planning – precisely the areas that cloud concentration risk puts under pressure. If your teams are actively thinking about workload portability or a move toward multi-cloud, the Cloud Migration Readiness Checklist is a cross-functional resource covering the ten critical areas – from target architecture and observability through to security, data migration, and cutover planning – that engineering, operations, and security teams need to work through together. For a more substantive conversation about your cloud strategy, get in touch with us directly below.