Phala Cloud review
Updated September 28, 2026 · DeployCue Editorial Team · Facts checked September 26, 2026
We earn a commission when you sign up through links on this site. Partner status can affect where a brand appears in our rankings. Partners with a live offer in your country come first, then every other brand by score.
Our verdict
Phala Cloud suits teams that want confidential models and rented NVIDIA machines on a single bill, with the certifications listed in SOC 2 Type I and HIPAA. Its entry GPU hour of $2.38 runs on hourly billing only, which keeps a first short run easy to predict.
Key facts
- Best for
- Confidential model serving and GPU rental on a single bill
- GPU models
- NVIDIA H100, NVIDIA H200, NVIDIA B300↗
- GPU hour from
- $2.38↗
- Billing granularity
- Hourly↗
- Egress fee
- Bandwidth included with each machine, based on its size↗
- Free credits
- USD 20 CVM credits for verified accounts↗
- Locations
- 3↗
- Compliance certifications
- SOC 2 Type I, HIPAA↗
- Legal entity
- Hashforest Technology LLC↗
- Headquarters
- United States↗
- Founded
- 2018↗
- Support channels
- Email, Help-centre, Community↗
- Pricing model
- Hourly billing
- Official website
- phala.com
LLM APIs
- Input price from
- $0.0050 to $3↗
- Output price from
- $0.20 to $15↗
- Models offered
- Phala/qwen3.8-27b-uncensored, Phala/gemma-4-26b-a4b-uncensored, Nvidia/nemotron-3.5-lightning, Z-AI/glm-5.3, Z-AI/glm-5.3-flash, Z-AI/glm-5.2, Z-AI/glm-5.1, Qwen/qwen3.8-27b, Qwen/qwen3.6-27b, Qwen/qwen3-32b, Qwen/qwen3.6-35b-a3b, Qwen/qwen3.5-27b, Qwen/qwen3.5-397b-a17b, Qwen/qwen3-embedding-8b, Qwen/qwen3-vl-30b-a3b-instruct, Qwen/qwen-2.5-7b-instruct, Meta/muse-glimmer-30b, Meta-llama/llama-3.3-70b-instruct, Deepseek/deepseek-v4-flash-0731, Deepseek/deepseek-v3.2, Moonshotai/kimi-k3, Moonshotai/kimi-k2.6, Google/gemma-4-31b-it, Openai/GPT-OSS-120b, Sentence-transformers/all-minilm-l6-v2↗
- Largest context window
- 1,000,000↗
- OpenAI-compatible API
- Yes↗
Facts checked September 26, 2026, 11 sources
Apps and profiles
Pricing
Phala Cloud Qwen3 32B (USD)
input-tokens: $0.12 / million
Phala Cloud Qwen3 32B (USD)
output-tokens: $0.50 / million
Phala Cloud Qwen 2.5 7B (USD)
input-tokens: $0.10 / million
Phala Cloud Qwen 2.5 7B (USD)
output-tokens: $0.20 / million
Phala Cloud Llama 3.3 70B (USD)
input-tokens: $2.00 / million
Phala Cloud Llama 3.3 70B (USD)
output-tokens: $2.00 / million
Phala Cloud gpt-oss-120b (USD)
input-tokens: $0.15 / million
Phala Cloud gpt-oss-120b (USD)
output-tokens: $0.60 / million
Pros and cons
Pros
- GPU hours start at $2.38 on NVIDIA H100, NVIDIA H200 and NVIDIA B300
- Verified accounts receive USD 20 CVM credits for verified accounts, which covers part of a first run
- Hosted models sit behind an OpenAI-compatible API, so existing clients need only a new base URL
- Bandwidth is Bandwidth included with each machine, based on its size, so transfer stays with the machine
Cons
- Support runs on email, help centre and community only, with no live chat or phone line listed
- Machines sit in 3 locations, so a workload is tied to a fixed region
- SOC 2 Type I and HIPAA names a single audit type, so buyers needing a deeper review should confirm scope with the provider
Who it suits
- Teams serving self-hosted open-weight models who need the calls to stay private
- Developers already writing against an OpenAI-compatible API
- Small teams that want a single bill for GPU rental and model endpoints
Who should look elsewhere
- Teams that need a specific older or consumer card, since Salad lists RTX and AMD classes while this range is NVIDIA H100, NVIDIA H200 and NVIDIA B300
- Workloads that must run in a chosen country, because machines sit in 3 locations and SberCloud is worth weighing for a different footprint
Plans, egress and billing
Billing runs hourly, the only model listed, and GPU hours start at $2.38. Bandwidth is Bandwidth included with each machine, based on its size, so there is no separate egress line to reconcile. Verified accounts receive USD 20 CVM credits for verified accounts, which covers part of a first run before the hourly meter takes over.
Core features
Rented machines cover NVIDIA H100, NVIDIA H200 and NVIDIA B300, a tight line at the newest end of the NVIDIA data centre range rather than a long shelf of older parts. Machines run across 3 locations, so a workload is tied to a fixed region instead of chosen per job. The company is US based, and hosted models sit behind an OpenAI-compatible API.
How Phala Cloud compares
Phala Cloud vs RunPod8.6 vs 9.5
Phala Cloud vs ScalaHosting8.6 vs 8.3
Phala Cloud vs Contabo8.6 vs 9.2
Phala Cloud vs Kamatera8.6 vs 9.5
Phala Cloud vs Cloudways8.6 vs 9.2
Phala Cloud vs PlanetHoster8.6 vs 8.7
Phala Cloud vs Acronis8.6 vs 9.9
Phala Cloud vs Aruba8.6 vs 9.3
Phala Cloud vs Decodo8.6 vs 9.1
Phala Cloud vs Hostinger8.6 vs 9.1- Phala Cloud vs MarsProxies8.6 vs 9.0
Phala Cloud vs hide.me8.6 vs 8.7
Phala Cloud vs Proton Drive8.6 vs 8.3
Frequently asked questions
What does a GPU machine cost to start?+
GPU hours start at $2.38 on the NVIDIA H100, NVIDIA H200 and NVIDIA B300 range. Billing is hourly, the only model listed, so a short test run is paid for by the hour. Bandwidth comes with the machine, based on its size, which keeps transfer costs inside the same line item.
Is there anything free to try?+
Verified accounts receive USD 20 CVM credits for verified accounts. That credit sits alongside the hourly meter rather than replacing it, so it can cover part of a first run on NVIDIA H100, NVIDIA H200 and NVIDIA B300 hardware. The credit applies once verification is done, so budget for the hourly rate from the point the account is confirmed.
How are the model APIs priced?+
Model endpoints are billed on their own, with input from $0.005 to $3 and output from $0.20 to $15. The catalogue spans open-weight chat, embedding and vision models, and the longest context listed is 1,000,000. Calls go through an OpenAI-compatible API.
Do I need to rewrite my client code?+
No rewrite is needed for the hosted models. The API is OpenAI-compatible, so an existing client can point at Phala Cloud with a different base URL. The same endpoint fronts the catalogue, while raw NVIDIA H100, NVIDIA H200 and NVIDIA B300 machines are rented separately on the hourly meter.
What compliance does Phala Cloud hold?+
The listed certifications are SOC 2 Type I and HIPAA, which cover security and health data handling. The service runs from a US base, and the provider is registered as Hashforest Technology LLC. Teams with their own audit requirements should check the certificate terms directly rather than assume coverage.
Which regions can I run in?+
Machines run across 3 locations, while the median across GPU listings here is 6. That is a narrow spread, so a workload is effectively tied to a fixed region. The company behind the service is headquartered in the US.
How do I get help if something breaks?+
Support runs on email, help centre and community. A help centre and a community cover common questions, and email reaches the team. There is no live chat or phone line listed, which matters for incidents on a rented machine; the company is registered as Hashforest Technology LLC.
Sources
- cloud.phala.com/about/pricing checked September 26, 2026
Backs: Billing granularity, Egress fee, Pricing model, Support channels
“GPU Rental: whole GPUs set aside for you, rented by the hour with a minimum period.”
- phala.com/about checked September 26, 2026
Backs: Founded
“The founding team starts from one premise: computation should be private without becoming unverifiable.”
- phala.com/confidential-ai-models checked September 26, 2026
Backs: OpenAI-compatible API
“OpenAI-compatible APIs run inside hardware-backed TEEs and return proof of the runtime that handled the request”
- phala.com/gpu-tee checked September 26, 2026
Backs: GPU hour from, GPU models, Locations
“NVIDIA H100 ... On-demand $3.08/GPU/hr ... Slot $2.38/GPU/hr reserved”
- phala.com/models checked September 26, 2026
Backs: Input price from, Models offered
“Sentence Transformers: all-MiniLM-L6-v2 512 context $0.0050/M input ... MoonshotAI: Kimi K3 1.0M context $3.00/M input”
- phala.com/models/moonshotai/kimi-k3 checked September 26, 2026
Backs: Output price from
“output $15.00/M”
- phala.com/models/z-ai/glm-5.3 checked September 26, 2026
Backs: Largest context window
“It supports text input and output with a 1M-token context window”
- phala.com/pricing checked September 26, 2026
Backs: Free credits
“Verified accounts get $20 in CVM credits.”
- phala.com/terms checked September 26, 2026
Backs: Legal entity
“operated by Hashforest Technology LLC”
- phala.com/trust checked September 26, 2026
Backs: Compliance certifications
“SOC 2 Type I: Compliant ... HIPAA: Compliant”
- x.com/PhalaNetwork checked September 26, 2026
Backs: Headquarters
“San Francisco, CA”
Compiled from 11 source pages, checked September 26, 2026. We did not open an account.