Home/Products

Four ways to buy the same fleet

From one GPU-hour to a managed AI platform.

Every product below runs on the same five HGX B300 nodes in Bangkok. What changes is how much of the operating burden you keep, and how predictable you want the bill to be.

Talk to sales Full price book

Product 01

GPU Cloud

One GPU, one minute, one command.

Billed per second from the moment the instance is reachable. GPUs are passed through, not time-sliced — no MIG partitioning unless you ask for it, no shared NVMe, no hypervisor tax on the memory bandwidth you paid for.

Per-second billing with no minimum and no charge while provisioning
Root on bare metal, or any OCI image from your own registry
Free egress to BKNIX peers and Thai ISPs, 10 TB international included
Persistent volumes that outlive the instance, so checkpoints survive an overnight shutdown

Instance	GPU	GPU memory	vCPU	RAM	NVMe	Network	USD / hour
tac.b300.1x	1 × B300	288 GB	28	256 GB	3.8 TB	100 Gb/s	$8.90
tac.b300.2x	2 × B300	576 GB	56	512 GB	7.7 TB	200 Gb/s	$17.80
tac.b300.4x	4 × B300	1.15 TB	112	1 TB	15.4 TB	400 Gb/s	$35.60
tac.b300.8x	8 × B300	2.1 TB	224	2 TB	30.7 TB	800 Gb/s	$69.90

Product 02

Dedicated Nodes

A whole HGX B300. Nobody else on it.

Eight Blackwell Ultra GPUs, 2.1 TB of HBM3e and 224 CPU cores allocated to one tenant for the length of a term. The NVLink domain stays whole, so the largest open models load without sharding across servers and no inter-node hop lands in the token path.

Sole tenancy on GPU, CPU, memory bus, NVMe and NICs
One predictable line on the invoice, agreed before the term starts
A named machine with a serial number and a rack position, which is what auditors accept
48-hour provisioning from signature, with a spare node held in region

Term	Per node / month	Per GPU-hour
Monthly	$46,900	$8.03
6 months	$40,900	$7.00
12 months	$35,900	$6.15
36 months	$30,900	$5.29

Node specification
GPU	8 × NVIDIA B300 (Blackwell Ultra)
GPU memory	2.1 TB HBM3e aggregate
NVLink	14.4 TB/s all-to-all, 5th generation
Inference	144 PFLOPS FP4
Training	72 PFLOPS FP8
CPU	2 × 112-core x86, 224 cores total

Full specification, fabric and facility detail on theinfrastructure page.

Product 03

Private AI Cloud

Your own AI platform, run by us, inside your perimeter.

For institutions that cannot put their data on someone else's inference API and do not want to build a GPU operations team to avoid it. Single tenant, in Bangkok, with a Thai entity on the contract.

Most “private” AI offerings are a logical partition on shared silicon. This is not that — single tenancy here means the hardware, not the namespace.

Foundation

A first production LLM workload behind your own perimeter.

$34,000 / month

Onboarding from $25,000 · SLA 99.5%

4 dedicated B300, single tenant
Private VPC, no shared data plane
One managed open-weight model, served on vLLM or TensorRT-LLM
S3-compatible object storage, 20 TB
SSO via SAML or OIDC
Business-hours Thai support, 4 h response

Discuss this tier

Enterprise

Most chosen

Company-wide AI with retrieval over internal systems.

$72,000 / month

Onboarding from $45,000 · SLA 99.9%

8+ dedicated B300, single tenant
Multi-model serving with routing and quotas
Managed RAG platform with connectors
Role-based access control, per-department budgets
Full prompt and response audit log, exportable
24×7 Thai support, 1 h response, named engineer

Discuss this tier

Sovereign

Regulated institutions and public sector.

By negotiation

Scoped per institution · SLA 99.95%

Dedicated cage or customer-owned rack
Air-gapped deployment option, no egress path
Hardware and staff vetted to your standard
Evidence pack for BOT, OIC and PDPC examination
Source-available control plane for inspection
Exit plan with full data and weight handover

Discuss this tier

Product 04

Inference API

Change one base URL. Keep your data in the country.

Open-weight models served on the same fleet behind an OpenAI-compatible endpoint that terminates in Bangkok. Existing SDKs, existing libraries, existing observability — the only thing that moves is where the request lands.

Zero retention by default — prompts live in memory for the request and are never written to disk
Never used for training, contractually rather than as a console setting
Thai treated as a first-class case, with published token ratios so you can predict cost in Thai
Free tier of 5M tokens a month, no card required

Model class	Size	Input / 1M	Output / 1M	Typical use
Compact	≤ 9B	$0.09	$0.25	Classification, extraction, routing
Standard	30 – 80B	$0.38	$0.95	General assistants, summarisation
Frontier MoE	200B+ MoE	$0.65	$2.20	Hard reasoning, long context
Reasoning	Long CoT	$0.75	$2.90	Agents, multi-step planning
Vision-language	Multimodal	$0.45	$1.20	Document and image understanding

Cached input −90%

Repeated prefixes billed at one tenth of the input rate.

Batch API −50%

Asynchronous jobs returned within 24 hours.

Provisioned throughput Flat

Reserve fixed tokens per second on dedicated GPUs. From $6,400/month.

Free tier $0

5M tokens per month for evaluation, no card required.

Committed terms

Utilisation is our problem. Discount is how we buy it from you.

A committed term lets us plan power, staffing and the next hardware order. That is worth real money to us, so we hand most of it back.

Term	USD / GPU-hour	Discount	8-GPU node / month	Included
On-demand	$8.90	—	$51,976	No commitment, per second billing
3 months	$7.60	−15%	$44,384	Capacity held, cancel at term
6 months	$6.90	−22%	$40,296	Priority scheduling
12 months	$6.20	−30%	$36,208	Named GPUs, quarterly true-up
36 months	$5.30	−40%	$30,952	Hardware refresh clause included

Talk to us

19 of 40 GPUs are still unallocated.

Tell us the model, the context length and the tokens per second you need. You will get a sizing, a price and an availability date — in writing, within two business days.

Reserve capacity See pricing

Direct line sales@thaiaicloud.co.th · Replies in one business day, Thai or English.