The most efficient AI compute on earth.
Available anywhere.
think Grid delivers the Think AI Fabric as capacity. Dedicated bare-metal nodes with ILM already running, hosted in Riyadh, in one all-inclusive monthly rate. No capital outlay, no hypervisor, no egress charges.
Same silicon. A different machine entirely.
A hyperscaler sells you a slice of a server: virtualised, shared with strangers, and metered on the way out. think Grid gives you the whole node, every gigabyte pooled as one, with ILM already running on it. The chips are the same. Nothing else is.
Hyperscaler instance
A slice of a shared server. Your throughput moves when theirs does.
think Grid SuperNode
The whole node, on dedicated bare metal. Nobody else is on the machine.
Four walled accelerators
Will not fit
384 GB on the invoice, 96 GB per wall. Anything larger has to be split by hand.
One pooled memory space
Fits, 144 GB spare
Every gigabyte in one space. Large models load whole, and the rest stays usable.
One model per accelerator
Typical
You rent the whole accelerator and use a quarter of it. The rest is billed and idle.
ILM, running on every node
92.3%
Models share silicon, large ones shard across it, and training runs beside serving. Included, not extra.
Hyperscaler instance
1,597 GB/s
Memory bandwidth per accelerator on a listed instance.
think Grid SuperNode
1,792 GB/s
Twelve per cent more memory bandwidth per accelerator, on the same silicon generation. Models are fed faster, so the compute waits less.
AWS and Azure
2,770 km
Their nearest 384 GB regions are Mumbai or Europe. Every request and every byte of training data makes the trip.
think Grid, Tier III Riyadh
In Kingdom
The data stays where it started, under the jurisdiction it started in. Where a hyperscaler does sell in-Kingdom capacity, it is still virtualised and shared.
Every way to buy the same AI compute.
Monthly cost for 384 GB of PRO 6000 Blackwell, at on-demand list price, 730 hours. Priced from each vendor's own published rates on 13 August 2026. Ours is all inclusive; theirs is compute only.
| Provider | Configuration | Region | Machines | Per month |
|---|---|---|---|---|
| think Grid | think SuperNode, bare metal, ILM included | Riyadh | 1 | $15,000 |
| Google Cloud | g4-standard-192 | Dammam | 1 | $15,765 |
| Azure | 2 x NC288ds_xl_RTXPRO6000BSE_v6 | East US | 2 | $16,060 |
| Google Cloud | g4-standard-192 | Mumbai | 1 | $17,081 |
| Azure | 4 x NC144ds_xl_RTXPRO6000BSE_v6 | East US | 4 | $18,630 |
| AWS | g7e.24xlarge | Mumbai | 1 | $19,759 |
| AWS | g7e.24xlarge | Frankfurt | 1 | $20,573 |
Compute only for all three hyperscalers: storage, egress and support plans are billed separately and are excluded. Spot, committed-use, reserved and Dev/Test rates are excluded. Azure NCv6 is offered in no Middle East or India region, so its lowest-priced region is shown. Sourced from the AWS EC2 price list, the Azure Retail Prices API and the Google Cloud SKU list, retrieved 13 August 2026, and cross-checked against each vendor's public calculator. The think Grid rate is our own all-inclusive published rate. Ask us for the workings behind any figure.
Our think SuperNode, as a service.
Four Blackwell accelerators that typically require multiple air-cooled servers, condensed into a single liquid-cooled deployment. ILM serves foundation models alongside real-time inference from the 384 GB pool. Two SuperNodes give full N+1 across every model you are running.
| Software | ILM orchestration platform, licensed for the term. Included. |
|---|---|
| CPU | AMD Threadripper PRO 9985WX, 64 cores / 128 threads. Included. |
| RAM | 512 GB (4 x 128 GB) R-DIMM ECC DDR5. Included. |
| GPU | 4 x NVIDIA PRO 6000 Blackwell, 96 GB GDDR7 ECC, 1,792 GB/s each. Included. |
| Storage | 4 x 8 TB PCIe Gen5 NVMe, up to 14.8 GB/s read, 13.4 GB/s write. Included. |
| Networking | Think AI Constellation single-port OSFP NDR 400Gb/s PCIe Gen5 NIC. Included. |
| Cooling | Think AI SuperNode custom liquid cooling system: sealed, dry heat rejection. Included. |
| Hosting | Tier III Riyadh facility, redundant power, diverse connectivity. Included. |
| SLA | Premium: 24/7, 99.9% uptime, one hour Severity 1 response. Included. |
| Maintenance | Preventative service, thermal validation, firmware and ILM updates. Included. |
| Onboarding | Delivery, installation, commissioning and integration support. Included. |
| SLA upgrade | Mission-Critical: 99.95% uptime, 30 minute Severity 1 response. USD 1,750 per month. |
Production inference
Multi-model concurrency on one node.
Voice and agents
Low-latency pipelines, close to your users.
Fine-tuning
Train while serving, on the same node.
High availability
Two nodes, full N+1 across every model.
All prices in USD per node per month. Saudi VAT additional where applicable. Hardware remains the property of Think AI throughout and is returned on termination. Any official quote is valid for one week.
One monthly rate. Nothing bolted on afterwards.
No capital outlay, no facility obligation, no egress charges. The hardware remains ours to carry; the capacity is yours to use.
$39per GB of VRAM, per month
The lowest cost per gigabyte of any hyperscaler, on the same silicon, in every region they offer it. Everything in the rate: the node, the facility, ILM, storage, egress and the SLA.
| Provider | Region | Per GB | vs think | |
|---|---|---|---|---|
| think Grid | Riyadh | $39.06 | — | |
| Google Cloud | Dammam | $41.05 | +5% | |
| Azure | East US | $41.82 | +7% | |
| Google Cloud | Mumbai | $44.48 | +14% | |
| Azure | East US, 4 x NC144 | $48.52 | +24% | |
| AWS | Mumbai | $51.46 | +32% | |
| AWS | Frankfurt | $53.58 | +37% |
Monthly cost for 384 GB of NVIDIA PRO 6000 Blackwell divided by the memory it buys, at on-demand list price, 730 hours. Competitor rates normalised for block storage, a support plan at 8% and 5 TB of egress, because our rate includes all three. Published on-demand rates, August 2026, for every region these providers offer this silicon in. See the rate card.
One node, carrying the work of three.
Most systems in production run several models at once: a reasoning model, task models, an embedding model, a reranker. The usual default is one model per accelerator, and much of the VRAM and compute goes unused but is still paid for. think Fabric solves this, placing them together, three to four per accelerator.
| Twelve models in production | Machines | Monthly equivalent | Per model |
|---|---|---|---|
| One per accelerator, AWS Frankfurt | 3 instances | $61,718 | $5,143 |
| One per accelerator, AWS Mumbai | 3 instances | $59,278 | $4,940 |
| One per accelerator, Google Cloud Dammam | 3 instances | $47,296 | $3,941 |
| think Grid SuperNode with ILM | 1 node | $15,000 | $1,250 |
Model mix as measured on a think AI node: a 32B reasoning model, a 7B, a 4B and a 1.1B, co-located at three per accelerator, the conservative end of the measured three to four. Hyperscaler figures are compute only, at the rates above. Term rates reduce the think Grid figure further.
Why no one else can do this
Renting silicon by the hour makes the accelerator the product, so stranded capacity is yours to pay for and there is no commercial reason to reclaim it. We sell the workload, not the card, which is why we built the layer that makes one machine carry the work of three.
think Grid, promised
The lowest monthly cost for 384 GB of this silicon we can find anywhere, including inside the Kingdom. 43.6 W/L of true SuperNode cooling density against an industry ~7. Every figure drawn from published, dated documentation, ours included, and open to scrutiny.
Capacity this month, not next year.
Tell us your models and throughput. We will size the node, quote the term, and have you serving from Riyadh.
Contact Think AI