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.

Pooled VRAM per node 384 GB GDDR7 ECC, one memory space
RiyadhLive
London, UKPlanned
Helsinki, FinlandPlanned
Mumbai, IndiaPlanned
Portland, USPlanned
Johannesburg, South AfricaPlanned

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

Tenant
You
Tenant
Tenant

A slice of a shared server. Your throughput moves when theirs does.

think Grid SuperNode

You

The whole node, on dedicated bare metal. Nobody else is on the machine.

Four walled accelerators

240 GB model
96
96
96
96

Will not fit

384 GB on the invoice, 96 GB per wall. Anything larger has to be split by hand.

One pooled memory space

240 GB model
384 GB pooled

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

Riyadh Mumbai

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

Riyadh 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.

1,792 GB/s Memory bandwidth per accelerator Against 1,597 GB/s on a listed hyperscaler instance of the same generation.
Included Storage, support and egress Billed separately by AWS, Azure and Google Cloud.
1 node To reach 384 GB Azure needs two machines or four, and the pool never sits in one.

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 Gridthink SuperNode, bare metal, ILM includedRiyadh1$15,000
Google Cloudg4-standard-192Dammam1$15,765
Azure2 x NC288ds_xl_RTXPRO6000BSE_v6East US2$16,060
Google Cloudg4-standard-192Mumbai1$17,081
Azure4 x NC144ds_xl_RTXPRO6000BSE_v6East US4$18,630
AWSg7e.24xlargeMumbai1$19,759
AWSg7e.24xlargeFrankfurt1$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.

$15,000 Per node, per month Monthly rolling, 30 days written notice. SKU: GRID-SUPER-S-A.
4 PFLOPS FP8 dense 4x NVIDIA PRO 6000 Blackwell, 96 GB each.
384 GB Total VRAM GDDR7 ECC, pooled as one memory space by ILM.
400G Constellation NIC Additional nodes join the same fabric at the same rate.
SoftwareILM orchestration platform, licensed for the term. Included.
CPUAMD Threadripper PRO 9985WX, 64 cores / 128 threads. Included.
RAM512 GB (4 x 128 GB) R-DIMM ECC DDR5. Included.
GPU4 x NVIDIA PRO 6000 Blackwell, 96 GB GDDR7 ECC, 1,792 GB/s each. Included.
Storage4 x 8 TB PCIe Gen5 NVMe, up to 14.8 GB/s read, 13.4 GB/s write. Included.
NetworkingThink AI Constellation single-port OSFP NDR 400Gb/s PCIe Gen5 NIC. Included.
CoolingThink AI SuperNode custom liquid cooling system: sealed, dry heat rejection. Included.
HostingTier III Riyadh facility, redundant power, diverse connectivity. Included.
SLAPremium: 24/7, 99.9% uptime, one hour Severity 1 response. Included.
MaintenancePreventative service, thermal validation, firmware and ILM updates. Included.
OnboardingDelivery, installation, commissioning and integration support. Included.
SLA upgradeMission-Critical: 99.95% uptime, 30 minute Severity 1 response. USD 1,750 per month.

Monthly rolling

USD 15,000

Per node, per month. 30 days written notice.

Talk to us

1 year, billed monthly

On application

12 month term. Discount applies.

Talk to our team

1 year, paid upfront

On application

Prepaid term. Our best rate.

Talk to our team

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.

ProviderRegionPer 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 Frankfurt3 instances$61,718$5,143
One per accelerator, AWS Mumbai3 instances$59,278$4,940
One per accelerator, Google Cloud Dammam3 instances$47,296$3,941
think Grid SuperNode with ILM1 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