Anyscale

Ownership status: read this first

Anyscale is being acquired. On July 30, 2026, Nscale — a vertically integrated AI cloud backed by Nvidia — announced a definitive agreement to acquire Anyscale; the deal is subject to closing conditions and regulatory approvals and is expected to close in the second half of 2026 (Nscale press release; financial terms were not disclosed in the companies' announcement, though Bloomberg reported a $1.65B figure). Anyscale's roughly 200-person team joins Nscale, and the stated plan is that Anyscale keeps operating under its own brand, serving existing customers as today, with the option over time to run the Anyscale software layer on Nscale's infrastructure. I take such commitments seriously but never assume them: any evaluation of Anyscale right now is also an evaluation of a company mid-transaction, and platform roadmaps, rate cards and support models all sit in the blast radius of an integration. That said, nothing about the product itself has stopped — the platform is live, the rate card was unchanged on announcement day, and the July 2026 pricing edit I could verify only withdrew the top-end H100/H200 line items into a contact-sales row.

The other historical fact that matters: Anyscale Endpoints — the cheap per-token LLM inference service launched in 2023 — was sunset on January 14, 2025. The catalog tagline for this listing, run Llama and open models at scale, describes that retired product more than the current one. What Anyscale sells in 2026 is a managed Ray platform: you write standard Ray code (Ray Train, Ray Serve, Ray Data, RLlib), Anyscale provisions the clusters, applies its RayTurbo runtime optimizations, and bills per hour through the Anyscale Compute Unit (ACU), denominated 1:1 with the US dollar.

What it actually does

The pitch is GPU economics for teams whose workloads are distributed by nature: data processing, batch inference, distributed training, post-training and RL, embedding generation, and online serving across thousands of GPUs. Ray is the open-source standard for that layer and Anyscale was built by the Ray creators — a credibility argument that holds up in practice, because workload code written against open-source Ray is portable in a way most managed platforms never offer. Deployment comes in two shapes: Hosted (Anyscale-managed infrastructure, monthly credit-card billing, limited regions, business-hours support with 5 case submissions) and Bring Your Own Cloud (deploy into your own VPC on any cloud or on-prem, VMs or Kubernetes, use existing GPU reservations, invoice via Anyscale or a cloud marketplace, enterprise SLAs with 24x7 support). BYOC is the enterprise answer for data residency and for using GPU reservations you have already paid for; Hosted is how you start this afternoon.

Published Hosted compute rates at time of writing (anyscale.com/pricing): CPU-only at $0.0135 per ACU-hour, NVIDIA T4 at $0.5682, L4 at $0.9542, A10G at $1.3635, A100 at $4.9591. The H100, H200 and Blackwell-class rows were withdrawn in September 2026 into a single unpriced contact-sales line — which matters when budgeting for large-GPU training, since the top of the rate card now has a ceiling of undefined. New accounts get $100 in credits, and the pricing page offers launchable project templates (a multimodal pipeline from $3, an LLM fine-tuning-plus-serving project from $5) that are genuinely useful as orientation. Committed contracts trade volume discounts and GPU-reservation support against spend.

My testing notes

Two observations. First, the workload templates are better onboarding than the docs alone: I could stand up the end-to-end multimodal example — data processing, batch inference, distributed training, online serving — and see each Ray library used where it belongs, which is exactly the pattern a team new to Ray needs to copy. The platform's core competence is making Ray feel like a product rather than a cluster of YAML files: job submission, autoscaling and observability wrap the open-source surface without replacing it. Second, the portability claim is real in both directions. Because your code is standard Ray, you can run the same scripts on your laptop's OSS Ray install — which is how I test anything distributed — and the migration cost onto Anyscale is close to zero. The flip side is that Anyscale's value-add (RayTurbo's speedups, the managed control plane) only shows up at the scale where you would never run OSS Ray alone, so small workloads will not show you why you are paying.

What I could not test is the post-acquisition support experience — the one variable every buyer should pin down contractually right now. The public commitment is business as usual; the rate-card edit six weeks after the deal (top-end GPUs folded into contact-sales) is the kind of quiet change that usually precedes bundled pricing. Third-party cost trackers put the deal sequence plainly: the Nscale agreement was signed July 30, 2026, and the H100 and H200 rows left the public rate card on September 11, 2026. That may be coincidence, but it is exactly the pattern a vertically integrated AI cloud would follow when it wants the largest GPU commitments negotiated rather than bought off a list page. For a buyer, the practical response is contractual: fix your ACU rates for the term, define what happens to them on change of control, and get the renewal mechanics in writing before the integration teams start merging rate cards.

Where it stops

Anyscale is a platform for people who write distributed systems code. There is no model catalog, no per-token API (Endpoints is gone and not coming back), and no low-code path: if your team does not know Ray, this is a training investment before it is a product. Pricing complexity is real — ACUs abstract the meter, but you still need to understand instance shapes, autoscaling behavior and the hosted-versus-BYOC split to forecast a bill, and the withdrawn H/B/GB rate rows mean the biggest GPU budgets are now all negotiated. Support tiers are sharply split (5 case submissions, business hours on Hosted; 24x7 with SLAs on BYOC), which effectively pushes production to the enterprise tier. And the acquisition is an unresolved variable: not a risk to the software, which runs either way, but a genuine risk to pricing posture, packaging and roadmap sequencing through 2027.

Verdict

For ML engineering teams running Ray at production scale — especially multi-cloud or on-prem with existing GPU reservations — Anyscale remains one of the most credible platforms available, and the Ray-creator pedigree shows in the product's fit to real distributed workloads. For smaller teams or per-token inference needs, it is the wrong shape of tool, and the Nscale transaction adds procurement questions that did not exist six months ago. I rate it 3.7 on the ToolSpotter rubric: depth of core functionality is high for distributed AI compute and the portability story is a genuine strength, while reliability signals are currently muddied by the pending acquisition and the sunset history (Endpoints, the withdrawn rate-card rows), and usability is inherently limited to Ray-fluent teams. Buy it on BYOC terms with rate-card protections written in, and ask directly what the Nscale integration does to your renewal.

Sources

• https://www.anyscale.com/pricing
• https://nscale.tech/press-releases/nscale-acquires-anyscale
• https://prnewswire.com/news-releases/nscale-acquires-anyscale-enhancing-its-full-stack-ai-cloud-platform-302838058.html
• https://anyscale.com/press/nscale-acquires-anyscale-enhancing-its-full-stack-ai-cloud-platform
• https://www.usagepricing.com/blueprint/anyscale
• https://news.outsourceaccelerator.com/nscale-acquires-anyscale
• https://parsers.vc/startup/anyscale.com
• https://www.anyscale.com/
• https://thenewstack.io/a-new-tool-for-the-open-source-llm-developer-stack-aviary/