Ori Industries

What Ori actually is

Ori is a UK-based, AI-native GPU cloud: it rents NVIDIA GPU instances by the minute, sells serverless Kubernetes with GPU autoscaling, hosts inference endpoints with scale-to-zero, and builds private AI clouds for enterprises, sovereigns and telcos. The company — Ori Industries 1 Ltd — was founded in 2018 in London by Mahdi Yahya, raised roughly $178M in disclosed funding including a $176M Series D in February 2025, and currently lists around 54 employees. Its most recent corporate move is significant: Ori has merged with Brookfield-backed Radiant to create an end-to-end AI infrastructure platform spanning land, capital, compute and software, and the group now also sells its platform stack ("AI Fabric") to organisations that want to run their own AI cloud.

The product split matters when you evaluate them. There is the public cloud — GPU instances, serverless Kubernetes, inference endpoints, reserved GPU clusters — and there is the private/sovereign business, where Ori deploys its full stack inside customer or partner facilities. The public cloud's selling points are price clarity and developer ergonomics: published per-GPU rates, billing by the minute against hourly prices, preinstalled ML frameworks and drivers, the ability to suspend and resume workloads, and an explicit "no ingress or egress fees, no hidden fees of any kind" policy. Those are real, checkable commitments, and in a GPU market full of opaque quotes they count for a lot.

The infrastructure story is partner-colocation rather than self-built hyperscale. Ori's first UK cloud region, announced in March 2025, runs at Kao Data's Harlow campus — an NVIDIA DGX-Ready data centre powered by 100% certified renewable energy — and included the first deployment of NVIDIA H200 GPUs in the UK, with GB200-class hardware following. The vendor claims 15+ regions of service worldwide; it does not publish a region list, and the original company description (distributed GPU compute across their own and partner datacenters, with a UK/EU compliance focus) matches what we can verify: UK-anchored, partner-facility expansion, sovereign options.

Pricing: what you really pay

Ori publishes instance pricing outright (ori.co/pricing and /gpu-instances, October 2026), priced per hour and billed per minute:

• NVIDIA H200 (141 GB HBM3e) — $3.50/hour on-demand; 24 vCPU, 240 GB RAM, 3,000 GB storage per single-GPU instance
• NVIDIA H100 SXM (80 GB) — $2.90/hour; 24 vCPU, 240 GB RAM, 3,000 GB storage
• NVIDIA H100 PCIe (80 GB) — $2.90/hour; 30 vCPU, 380 GB RAM, 3,840 GB storage
• NVIDIA A100 (80 GB) — $2.74/hour; 15 vCPU, 180 GB RAM, 300 GB storage
• NVIDIA L40S (48 GB) — $1.55/hour; 15 vCPU, 90 GB RAM, 1,600 GB storage
• NVIDIA V100S (32 GB) — $0.95/hour; L4 (24 GB) — $0.93/hour; V100 (16 GB) — $0.83/hour
• Serverless Kubernetes GPUs — per minute: H200 $0.07875/min ($4.73/h), H100 $0.06525/min ($3.92/h), L40S $0.03488/min ($2.09/h), L4 $0.019044/min ($1.14/h)
• Inference endpoints — dedicated GPUs from $1.27/h (L4) to $4.35/h (H100), with autoscaling and scale-to-zero
• B200 (1,440 GB HBM3e class) — listed as now available, price on application

The structure: 8-GPU and reserved configurations exist in the configurator, reserved instances start at two-week terms, and a Private Cloud option sells 1,000+ GPU clusters reserved for 12 months or longer at pricing Ori claims is up to around 80% below its on-demand rates. How competitive is all this? Very, against hyperscalers — Oracle charges $10/GPU-hour for H100 on-demand, AWS and Azure list higher still, so Ori's $2.90 is roughly a third of hyperscaler list. Against other GPU specialists the gap narrows; Ori sits in the competitive middle of the neocloud market rather than at the absolute bottom. The honest framing of their "81% savings" claim: true versus hyperscaler on-demand list, marketing versus the specialist alternatives you should also compare on /benchmarks methodology.

One structural note: the entry point is genuinely small — a single V100 at $0.83/hour, billed by the minute — which makes the platform usable for experimentation in a way that 8-GPU-node hyperscalers are not. There is no free tier and no documented trial credits; you pay from the first minute.

Performance & infrastructure

The GPU catalogue spans current NVIDIA generations (H200, H100 SXM and PCIe, A100, L40S, L4, V100S) with B200 announced, 40+ advertised GPU configurations, and per-instance specs that are unusually generous — 240 GB of system RAM and 3 TB of storage alongside a single H100, for example, which is more host resources per GPU than most neoclouds attach. Instances ship with OS, drivers and ML frameworks preinstalled, launch from CLI, UI or API, and support suspend/resume for cost control on bursty development work. The serverless Kubernetes layer adds GPU autoscaling (including scale-to-zero inference endpoints), and the reserved-cluster path adds fast networking for multi-node training.

On geography, the verified anchor is the UK: the Kao Data Harlow region (NVIDIA DGX-Ready certified, renewable-powered, close to London and Cambridge) hosting the UK's first H200 deployment from March 2025, positioned explicitly as the start of European expansion. Ori claims 15+ regions globally and has shown US and EU presence in customer deployments, but no public region list exists — we treat the count as a vendor claim. Sovereignty is a stated design goal: the private-cloud offering deploys the whole stack in customer-chosen locations, which is how the UK/EU compliance pitch is actually delivered.

Reliability is where the record is thinnest. Ori runs a public status page and a trust center, but publishes no uptime SLA that we could find, and as a venture-stage company with roughly 54 employees it does not have a decade of operational history. Testimonials (nCompass, Emediately, Framesports, Basecamp Research) praise price, support responsiveness and usability; they are marketing assets, not uptime data. We have not benchmarked Ori, and GPU performance claims belong to NVIDIA's spec sheets rather than the cloud's — what a provider controls is thermals, interconnect, storage throughput and scheduling, none of which Ori publishes independently measured figures for.

Where it falls short

Track record and size, first. This is a young, venture-funded company in the middle of a merger with Radiant — a combination that is strategically sensible (capital, land, compute) but means the product roadmap and pricing you sign up for may change hands. Counterparty risk is real for reserved commitments: a two-week reservation is low-risk, a 12-month private-cloud contract is a different conversation. We would want contractual SLA terms in writing before any production commitment, precisely because none are published.

Second, transparency gaps. Region list: unpublished. SLA: unpublished. Status history: exists but not analysed publicly. The marketing language leans hard ("best-in-industry rates", "81% savings"), and while the price table backs much of it up, claims like those deserve the same scrutiny we apply to any vendor. Third-party ecosystem coverage is thin — little independent documentation, few community reviews, and limited tooling integrations compared with the big clouds; if you use Terraform, expect to check provider support yourself.

Third, scope. Ori is compute and inference, not a cloud platform: no managed database catalogue, no object-storage product line with published SLAs, no marketplace, no global IAM story beyond the platform basics. That is fine for GPU workloads and a limitation for anything needing a full application stack. And unlike hyperscalers, there is no free tier — you cannot evaluate without paying.

Who should pick this

We would pick Ori when GPU price per hour is the dominant cost driver and you want single-GPU granularity with per-minute billing: ML teams iterating on training and fine-tuning, inference products that need scale-to-zero endpoints, UK or European organisations that want H200/H100 capacity on UK soil (the Kao Data region is a genuine differentiator post-2025), and enterprises or sovereigns evaluating a private AI cloud stack from an operator that runs its own. The H100 at $2.90/hour with 240 GB RAM and 3 TB storage attached is our value pick in their lineup.

We would not pick it for a full application platform, for workloads that require a published, credit-backed SLA before production, or for organisations that need the compliance paperwork and operational maturity of a hyperscaler. In those cases, pay the OCI/AWS/Azure premium or pair Ori's compute with infrastructure elsewhere.

Key takeaways

• UK AI-native GPU cloud (founded 2018, London, ~54 staff, ~$178M raised) now merged with Brookfield-backed Radiant into an end-to-end AI infrastructure group.
• Published on-demand pricing: H200 $3.50/h, H100 $2.90/h, A100 $2.74/h, L40S $1.55/h, down to V100 $0.83/h — roughly a third of hyperscaler H100 list, priced hourly and billed by the minute, with no ingress or egress fees.
• Generous host resources per GPU (e.g. 24 vCPU / 240 GB RAM / 3 TB storage with a single H100), preinstalled ML stacks, suspend/resume, CLI/UI/API.
• Serverless Kubernetes with GPU autoscaling and scale-to-zero inference endpoints; reserved clusters from two weeks up to 1,000+ GPU private clouds.
• Verified UK region at Kao Data's Harlow campus (NVIDIA DGX-Ready, renewable-powered, first UK H200 deployment, March 2025); "15+ regions" claimed but not listed publicly.
• No published uptime SLA, no free tier, thin independent coverage: excellent rates for informed buyers, wrong fit where contractual guarantees come first.

Sources

• https://www.ori.co/pricing
• https://www.ori.co/gpu-instances
• https://www.ori.co/get-started
• https://www.ori.co/
• https://finance.yahoo.com/news/ori-selects-kao-data-first-150900890.html
• https://www.prnewswire.co.uk/news-releases/ori-selects-kao-data-for-its-first-uk-ai-cloud-region-302397171.html
• https://platform.tracxn.com/a/d/company/54aaad36e4b0fa0b9faabd50/ori