Perplexity Killed Its Ad Business in February. Revenue Jumped 50% the Next Month. Here's the Playbook.
On July 30, 2026, London-based Nscale agreed to buy Anyscale — the company built on Ray, the open standard for scaling AI workloads — for approximately $1.65 billion. The deal combines GPU infrastructure with software orchestration in a way no hyperscaler has done from outside the Big Three.
On July 30, 2026, London-based Nscale announced a $1.65 billion agreement to acquire Anyscale, the commercial company built around Ray — the open-source framework that has become the dominant standard for scaling distributed AI workloads. The deal is the most strategically significant AI infrastructure acquisition of 2026, and not primarily because of the price.
What Nscale is buying is the right to call itself the first full-stack AI hyperscaler: a single vendor that owns power generation, data center real estate, GPU clusters, and the software orchestration layer that enterprise ML teams actually use to run their workloads. AWS, Azure, and Google Cloud have never built that integrated stack from outside the Fortune 500. Now a five-year-old London startup with $14.6 billion in valuation and $2 billion in Series C backing is attempting to do it first.
What Nscale Actually Is — and Why It Has the Capital to Try This
Nscale is not a neocloud. It is not renting GPU capacity from hyperscalers and reselling it at a margin. The company builds and operates its own data centers, owns the power infrastructure feeding those facilities, and purchases GPU hardware directly — primarily Nvidia H100 and H200 clusters, with access to next-generation hardware guaranteed through Nvidia's participation as a Series C investor.
Nscale raised $2 billion in its Series C in March 2026 at a $14.6 billion valuation. The round was led by Aker ASA, a Norwegian industrial conglomerate with extensive energy infrastructure assets, with participation from Nvidia, Dell, Nokia, Citadel, Jane Street, Point72, Fidelity, and Linden Advisors. The investor list is notable for what it signals: this is not a venture capital bet on a GPU rental arbitrage play. It is a strategic infrastructure bet backed by an energy infrastructure company, the dominant chip supplier, the dominant enterprise hardware vendor, and a group of financial institutions with long investment horizons.
Before the Series C, Nscale raised $1.1 billion in a Series B in September 2025 — described at the time as the largest Series B in European history. In parallel, the company took on a $1.4 billion GPU-backed loan and a $790 million financing facility in Norway, giving it infrastructure capital beyond what the equity rounds provided. Total funding for the company entering 2026 exceeded $5 billion including debt facilities.
The company's infrastructure model is vertically integrated in a way that traditional cloud providers are not. Nscale controls:
- Power generation: Through partnerships with renewable energy facilities and Aker ASA's energy assets in Norway, Nscale manages power contracts directly rather than paying facility operators for metered power.
- Data center construction and operation: The company builds its own facilities to specifications optimized for high-density GPU compute, rather than leasing space in co-location facilities designed for general enterprise IT.
- GPU procurement: Direct Nvidia relationships at scale give Nscale procurement leverage that co-location operators and second-tier cloud providers cannot achieve.
- Networking: The interconnect fabric inside Nscale facilities is engineered specifically for GPU-to-GPU communication latency requirements, not general enterprise networking.
This vertical integration means Nscale's cost structure at the infrastructure layer is fundamentally different from hyperscalers' cost structure at that same layer — and that structural cost advantage is what makes the Anyscale acquisition strategically coherent rather than financially speculative.
What Anyscale Actually Is — and Why It Was Worth $1.65 Billion
Anyscale is the commercial wrapper around Ray, the distributed computing framework originally developed at UC Berkeley's RISE Lab by Ion Stoica and Robert Nishihara, among others. Ray solves one of the hardest problems in AI infrastructure: how to take a Python workload that runs correctly on one machine and scale it across hundreds or thousands of GPUs without rewriting the underlying code.
The problem Ray solves is not theoretically interesting — it is operationally expensive. Every enterprise ML team attempting to scale training or inference beyond a single node must confront distributed systems complexity: task scheduling, resource management, fault tolerance, and the specific networking requirements of GPU-accelerated compute. Teams that do not use a framework like Ray spend months building and maintaining custom distributed infrastructure. Teams that use Ray spend that time on the actual ML problem instead.
Ray has accumulated adoption across the production ML stacks of companies including OpenAI, Uber, Spotify, and Instacart, making it the default answer to "how do you scale AI workloads in Python" in the same way that Kubernetes became the default answer to "how do you orchestrate containerized services." The framework's governance transferred to the PyTorch Foundation under the Linux Foundation in October 2025, separating the community-owned open-source project from Anyscale's commercial enterprise business.
What Nscale is buying is the enterprise commercial layer above that open-source core:
- Managed Ray clusters: Provisioning, scaling, and managing Ray clusters without requiring ML teams to manage infrastructure directly.
- Workload scheduling and monitoring: The observability and scheduling tools that production ML platform teams need to operate Ray workloads at enterprise scale.
- Enterprise support and SLAs: The contractual guarantees and support response times that enterprise procurement requires.
- The customer base: Coinbase's ML infrastructure, Runway's video AI generation pipeline, and Bedrock Robotics' simulation workloads — with 90%+ enterprise retention.
Anyscale's financial profile at acquisition: $220 million in ARR in 2025, growing revenue 70% in its most recent quarter, approximately 200 employees across the United States, Europe, and India.
At $1.65 billion, the acquisition represents roughly 7.5 times Anyscale's reported ARR — a multiple that reflects the software moat rather than the standalone business. For comparison, Snowflake traded at 20-30x ARR during its hypergrowth phase. Anyscale at 7.5x ARR is not expensive for what it delivers to a full-stack infrastructure play.
The Problem the Deal Is Designed to Solve: The Enterprise AI Cost Stack
To understand why the acquisition is strategically compelling, it is necessary to understand how enterprise AI compute costs currently accumulate.
A typical enterprise running AI workloads in 2026 pays costs across at least three layers:
Layer 1 — Raw compute: GPU instances from AWS, Azure, or Google Cloud. At GPU-intensive scale, this is $3-8 per GPU-hour for H100s, with significant variation based on on-demand versus reserved pricing and regional availability.
Layer 2 — Orchestration software: Tools like Anyscale, Databricks, or custom-built Ray infrastructure to manage how workloads are distributed across those GPU instances. For enterprises buying Anyscale Enterprise, this is a separate contract layer on top of the compute bill.
Layer 3 — Integration and engineering: The cost of the ML platform team maintaining integrations between Layer 1 and Layer 2, managing failure modes, and optimizing resource utilization. At Fortune 500 scale, this team often costs $5-10 million annually in fully loaded compensation.
The Nscale thesis is that full-stack vertical integration collapses all three layers into a single vendor relationship. The company provides the compute (at a structurally lower cost than hyperscalers due to owned infrastructure), the orchestration software (Anyscale), and the management tooling — on a single bill, from a single contract, with a single support relationship.
Anyscale's Azure integration, launched before the acquisition, demonstrated the opportunity at the software layer alone: enterprises reported reducing API costs by up to 90% through intelligent workload routing and resource optimization when using Anyscale on top of Azure. That 90% reduction was achieved purely through software orchestration, without touching the underlying infrastructure costs. On Nscale's owned infrastructure, the combined stack expects to deliver further cost improvements.
| Cost Layer | Hyperscaler Model | Nscale Full-Stack Model |
|---|---|---|
| Compute | Cloud provider markup on GPU hardware | Direct GPU ownership, lower marginal cost |
| Orchestration | Separate Anyscale/Databricks contract | Included in full-stack contract |
| Integration engineering | Enterprise team required | Platform managed by Nscale/Anyscale |
| Data sovereignty | Depends on region/provider | EU-sovereign option available |
| Contract complexity | Multi-vendor | Single vendor |
The Ray Framework as Competitive Moat
The open-source Ray framework is worth analyzing separately from the Anyscale acquisition, because it is the technical basis for the moat that makes this deal defensible.
Ray has approximately 25 million downloads, an active contributor community of over 1,000 developers, and production deployments at some of the most demanding ML infrastructure operations in the world. It is the framework that OpenAI used to scale reinforcement learning from human feedback during the development of GPT-4, Uber uses for its real-time ML prediction infrastructure, and Spotify uses for personalization model training.
That installed base does not move. Companies that have built their ML infrastructure on Ray have invested significant engineering effort in Ray-native code, Ray-native monitoring, and Ray-native deployment patterns. The switching cost from Ray to an alternative distributed computing framework — Apache Spark for ML, Dask, or a custom distributed system — is measured in years of engineering time, not months. This is a classic open-source moat: Nscale has not acquired a proprietary technology that enterprises are forced to use, but rather the commercial service layer above the technology that everyone already uses voluntarily.
The governance separation — Ray under the PyTorch Foundation, Anyscale under Nscale — is strategically important because it means Nscale cannot be accused of "embracing and extending" an open standard in a way that harms the community. The open-source Ray project remains independent. Nscale's commercial advantage is in the enterprise service layer, not in controlling the specification of the framework itself. This is a cleaner structure than what MongoDB or Elastic faced when they modified open-source licenses to prevent cloud providers from reselling their software — Nscale is not modifying the open-source project, it is building a commercial offering alongside it.
Nscale's Full-Stack Vertical Integration Playbook
The Anyscale acquisition is the latest step in a deliberate vertical integration strategy that Nscale has been executing since its founding in 2022. The pattern follows a clear logic:
1. Own the power. Nscale's Norway facilities are powered by renewable hydroelectric energy — giving the company both cost stability (power is the dominant operating cost in AI data centers) and sustainability credentials that European enterprise procurement increasingly requires. Through Aker ASA's energy infrastructure relationships, Nscale has long-term power purchase agreements that insulate it from energy price volatility.
2. Build, don't lease, data centers. Co-location operators optimize facilities for general enterprise IT: moderate power density, standard cooling, mixed workload profiles. GPU-intensive AI workloads require 5-10x the power density of general compute and specialized cooling infrastructure. By building its own facilities, Nscale designs to the workload rather than adapting the workload to existing infrastructure.
3. Secure direct GPU supply. Nvidia's participation as a Series C investor is not just a capital event — it signals hardware access. GPU allocation during the supply constraints of 2024-2025 went to Nvidia's largest and most strategically important customers. Nscale's direct relationship with Nvidia places it in the allocation priority queue alongside hyperscalers rather than behind them.
4. Acquire the software orchestration layer. The Anyscale acquisition completes the stack. ML teams who run Anyscale today are choosing it for the software capabilities; the underlying infrastructure is secondary. By owning Anyscale, Nscale can offer the same software capabilities while routing workloads to infrastructure it controls — capturing margin at every layer.
5. Target EU sovereign AI. The European regulatory environment creates a structural market that hyperscalers struggle to serve: enterprises that need AI workloads processed on EU-domiciled, EU-staffed infrastructure to satisfy GDPR requirements, national security considerations, or EU AI Act compliance obligations. US-headquartered hyperscalers cannot offer the same sovereignty guarantees that a European infrastructure company can, creating a market segment where Nscale has structural advantages that persist regardless of compute price competition.
What AWS, Azure, and Google Cloud Will Do Next
The hyperscalers are not ignoring vertical integration. AWS has Bedrock and SageMaker as its orchestration layer, along with Trainium and Inferentia custom silicon. Google Cloud has TPUs and Vertex AI. Azure has the Nvidia relationship from the OpenAI partnership and Azure Machine Learning. Each has spent billions building managed AI services on top of their infrastructure.
But none of them have Anyscale specifically, and none of them own infrastructure with the European sovereign positioning that Nscale can offer. The competitive response is most likely to take one of three forms:
Acquisition: AWS or Google Cloud buys a comparable ML orchestration platform — Databricks, Weights & Biases, or a smaller Ray-adjacent competitor — to match the software integration story. The challenge is that no other company has Ray's installed base at Anyscale's scale.
Partnership: Hyperscalers deepen existing partnerships with ML platform vendors to offer comparable "one-stop" bundles. AWS has existing Anyscale relationships that will now be complicated by the Nscale ownership.
Organic investment: Building better ML orchestration tools inside existing managed services — SageMaker, Vertex AI — to close the gap with Anyscale's enterprise feature set. The challenge is that the Ray community has already voted with its code contributions on which framework they prefer.
None of these responses is fast. In the meantime, Nscale now has a credible offer for every enterprise ML team that is evaluating whether to remain with a hyperscaler, move to on-premise infrastructure, or try a European full-stack alternative. That is a market conversation that did not exist six months ago.
What Enterprise AI Teams Should Do With This Information
The Nscale-Anyscale combination creates a procurement option that enterprise ML platform teams should evaluate alongside hyperscaler relationships — but the evaluation framework matters.
Full-stack vertical integration has a strong cost story but a real vendor lock-in risk. Enterprises that move ML workloads to Nscale's platform are replacing multi-vendor optionality with single-vendor efficiency. If Nscale's infrastructure availability or support quality falls short, the migration cost is higher than switching between hyperscalers.
The European sovereign AI angle is genuinely differentiated — if an enterprise has EU regulatory requirements that make US-hyperscaler-hosted AI problematic, Nscale is now the most complete alternative to on-premise infrastructure that exists. The combination of owned EU infrastructure and the Anyscale software layer is unique.
For enterprises that are already Anyscale customers: the acquisition changes the commercial relationship but not the technical platform. Anyscale continues to operate under its brand, existing contracts remain in force, and infrastructure optionality (running on any cloud) is preserved. The change to watch is whether Nscale's pricing for the Anyscale-on-Nscale bundle in 2027 creates meaningful incentives to migrate workloads — that is when the strategic implications of the acquisition will be most visible in enterprise procurement conversations.
Takeaway: Nscale's $1.65 billion acquisition of Anyscale is the clearest evidence yet that the AI infrastructure market is consolidating toward vertical integration. The company that owns power, data centers, GPUs, and software orchestration can offer enterprises a cost structure and a contract simplicity that hyperscalers cannot match without fundamentally restructuring their business models. The Ray framework's open-source installed base — 25 million downloads, production deployments at OpenAI, Uber, and Spotify — gives Nscale a software moat that cannot be replicated by acquisition alone. Whether this becomes a defining independent company or a compelling acquisition target for one of the hyperscalers is the question to watch through 2027. Either outcome validates the bet that full-stack vertical integration in AI infrastructure is a fundamentally different competitive position than renting compute by the hour.
Frequently Asked Questions
What is Nscale and how is it different from AWS or Google Cloud?
Nscale is a London-based AI cloud provider founded in 2022 that owns its entire infrastructure stack — from power generation and data center real estate to GPU clusters. Unlike AWS, Google Cloud, or Microsoft Azure, which lease facility space and mix owned and third-party hardware, Nscale builds and operates every layer of the stack itself. The company raised a $2 billion Series C in March 2026 at a $14.6 billion valuation, backed by Nvidia, Dell, Nokia, Aker ASA, Citadel, Jane Street, and Point72. Nscale operates GPU data centers across Europe and North America, with a particular focus on European sovereign AI infrastructure — an increasingly important consideration for EU enterprises that cannot rely on US-domiciled cloud providers for sensitive AI workloads under GDPR and the EU AI Act. The company's vertically integrated model is designed to give it cost and sustainability advantages over hyperscalers that must manage more complex supply chains and multi-vendor data center relationships. Its acquisition of Anyscale on July 30, 2026, extends that vertical integration from hardware into the AI software orchestration layer for the first time.
What is Anyscale and what does the Ray framework do?
Anyscale is the commercial company built around Ray, the open-source distributed computing framework originally developed at UC Berkeley's RISE Lab. Ray is the dominant open-source framework for scaling Python and AI workloads — including data processing, model training, inference, and reinforcement learning — across clusters of GPUs. Major technology companies including OpenAI, Uber, Spotify, and Instacart use Ray in production for machine learning infrastructure. Anyscale packages Ray with enterprise management tools, monitoring, scheduling, and managed cloud services, and sells that bundle to companies that want to run Ray workloads without managing the underlying infrastructure. The company reported $220 million in ARR in 2025 and grew revenue 70% in its most recent quarter before the acquisition. Customers include Coinbase, which uses Anyscale for its ML infrastructure, and Runway, the AI video generation company. In October 2025, the governance of the open-source Ray project was transferred to the PyTorch Foundation under the Linux Foundation, ensuring that the community-owned framework remains separate from Nscale's commercial acquisition — existing Ray users are not required to use Nscale's infrastructure.
Why did Nscale acquire Anyscale instead of building its own software layer?
Building a competitive software orchestration layer from scratch would have taken Nscale three to five years and required recruiting machine learning infrastructure engineers who are extraordinarily difficult to hire in competition with Google, Meta, and Anthropic. The Anyscale acquisition gives Nscale immediate access to 200 experienced ML infrastructure engineers, an existing enterprise customer base with 90%+ retention, an established product with $220M ARR growing 70% quarter-over-quarter, and — critically — the Ray brand that is recognized as the open standard for distributed AI workloads. The acquisition also gives Nscale a defensible position relative to hyperscalers: any competitor that wants to replicate the Nscale model must either build or acquire a comparable software layer, while Nscale has locked in the most widely deployed open-source standard in the category. The $1.65 billion price represents roughly 7.5 times Anyscale's reported ARR, a premium that reflects the strategic value of the software moat rather than the standalone business metrics alone.
How does the Nscale-Anyscale combination reduce costs for enterprise AI?
When enterprises run AI workloads on a hyperscaler like AWS or Azure, they pay multiple layers of margin: the hyperscaler's compute margin, the cost of any third-party orchestration software like Anyscale, and often the cost of integration work to connect those two layers. Nscale's full-stack model collapses these costs into a single vendor contract. Anyscale's integration with Microsoft Azure, launched before the acquisition, demonstrated the opportunity: enterprises using Anyscale on Azure reported reducing their API costs by up to 90% through intelligent workload routing and resource optimization. By running Anyscale on Nscale's owned infrastructure — where Nscale controls power and compute costs directly — the company expects to deliver further cost reductions beyond what was achievable when Anyscale was integrating across third-party infrastructure. The European data center focus adds a sovereignty dimension: EU enterprises that need AI workloads processed within EU borders can use Nscale as a single-vendor alternative to building their own infrastructure or accepting the compliance complexity of multi-vendor hyperscaler setups.
What happens to Anyscale customers and the open-source Ray project after the acquisition?
Anyscale will continue to operate under its brand following the acquisition and existing customers — including Coinbase, Runway, and Bedrock Robotics — will continue receiving service under their current contracts. Nscale has stated that customers will remain free to run Anyscale software on any infrastructure, including AWS, Azure, or Google Cloud, while gaining the additional option to run on Nscale's full-stack platform. The open-source Ray framework is entirely separate from the commercial acquisition: governance transferred to the PyTorch Foundation under the Linux Foundation in October 2025, and Ray remains community-owned and infrastructure-agnostic. Enterprises that run Ray on their own infrastructure or on a different cloud provider are not affected by the Nscale acquisition. The acquisition's impact is limited to the commercial Anyscale enterprise platform — the managed Ray service, the scheduling and monitoring tools, and the enterprise support that sits on top of the open-source core. The deal is expected to close in the second half of 2026, pending regulatory approvals.
Does the Nscale-Anyscale deal threaten Amazon, Google, and Microsoft?
The deal creates the most credible challenge to hyperscaler AI infrastructure dominance that has emerged from outside the Big Three, but it operates on a different competitive axis than a direct pricing war. AWS, Azure, and GCP compete primarily on breadth of services, geographic coverage, enterprise relationships, and ecosystem depth — advantages built over fifteen years that Nscale cannot replicate quickly. Nscale competes on total cost of ownership for AI-specific workloads, European data sovereignty, and the simplicity of a single-vendor full-stack for ML infrastructure teams. The threat to hyperscalers is concentrated in specific use cases: GPU-intensive training and inference for enterprises that are sensitive to cost, enterprises in regulated EU industries that need sovereign compute, and ML platform teams that want to consolidate their orchestration and infrastructure costs. For these segments, Nscale-plus-Anyscale is now a credible alternative where previously none existed outside of on-premise deployments. The longer-term risk for hyperscalers is that vertical integration in AI infrastructure proves as defensible as it has in semiconductors — where TSMC's ownership of fabrication and advanced packaging created a moat that took decades to challenge.