July 30, 2026
Kinesis vs Lilac: Compute Platform vs Hosted Inference API
From the outside, Kinesis and Lilac can look similar — both put GPU compute within reach of teams building with AI. Look at where each one sits in the stack and they solve different problems.
Lilac is, first, a hosted inference API. Its live catalog offers a small set of open-weight models — six at last check — behind an OpenAI-compatible endpoint with per-token pricing: no GPU reservation, no infrastructure to manage. Lilac also advertises dedicated GPU infrastructure (H100 through B300) on commitment-based pricing.
Kinesis is a compute and orchestration platform. Rather than consuming AI through someone else's API, you run your own workloads on infrastructure you control — and the platform operates everything underneath them.
The one-line version
On Lilac, you call a hosted model. On Kinesis, you run your own models and applications — and the grid runs everything underneath them.
What running on Kinesis includes
- Any containerized application, using CPU and GPU in any combination. Custom AI training and inference pipelines, regulated workloads such as FinTech and BioTech, and conventional enterprise services running alongside them. As AI becomes part of existing applications rather than a separate stack, that flexibility matters more.
- Hundreds of pre-packaged AI models and applications through the App Gallery. LLM inference works out of the box — but you run the model privately rather than renting a shared, hosted endpoint. For organizations with data residency, IP, or cost-control requirements, that distinction matters: data stays inside your own environment, and inference cost reflects the underlying compute rather than a per-token price.
- Kinesis-managed capacity today, your own hardware tomorrow. Start on vetted datacenter capacity, then attach your own GPUs later — data center, colo, or on-premises. Moving between ours, yours, or a hybrid of both is a topology change, not a re-architecture.
Where the architectural difference shows up
- Hybrid and future-proof. The Kinesis Grid can span AWS, customer-owned hardware, and other providers, so workloads move between environments as requirements evolve — without redesigning the application.
- FinOps. Because the Grid spans multiple providers and regions, workloads can be scheduled on cost as well as availability — reserved capacity for steady-state serving, Spot Instances for burst and batch. Costs stay attributable at the workload level, with no per-token abstraction sitting between you and your actual infrastructure spend. As usage grows, bringing your own hardware into the same control plane gives you direct control over long-term economics.
- DevOps. Standard containerized deployments, declarative topology, and the same operational workflow whether workloads run on Kinesis-managed or customer-managed infrastructure.
- Observability. Utilization, performance, and cost are visible at both the workload and infrastructure levels, so teams optimize performance and cost together. With a managed inference API, much of that operational layer is intentionally abstracted away — convenient until you need it.
Side by side
| Kinesis | Lilac | |
|---|---|---|
| Model | Compute and orchestration platform — run your own workloads | Hosted inference API, plus advertised dedicated GPU infrastructure |
| Model access | Hundreds of models/apps via App Gallery, run privately in your environment | Six open-weight models at last check, shared hosted endpoints |
| Pricing basis | Compute you use — e.g. H100 $2.50/GPU-hr dedicated, no commitment | Per-token for inference; dedicated H100 advertised ~$2.00/hr with monthly commitment |
| Workload scope | Any container: training, inference, and non-AI services side by side | Inference against catalog models; other workloads on dedicated instances |
| Data locality | Your data stays in your environment; placement respects compliance constraints | Prompts and outputs transit Lilac's hosted endpoints |
| Hybrid / BYO hardware | Attach your own GPUs (DC, colo, on-prem) into the same control plane | Not part of the advertised model |
| Getting started | $100 credit, no card required | Pay-as-you-go tokens; subscription credit tiers |
Lilac details from getlilac.com and docs.getlilac.com; Kinesis rates from kinesis.network/pricing. All checked 2026-07-30. Lilac's advertised dedicated H100 commitment rate is lower than Kinesis's on-demand dedicated rate — if committed capacity for raw GPU hours is all you need, compare both.
When Lilac is the right choice
An honest comparison cuts both ways. If what you need is fast API access to a hosted open model — call an endpoint, pay per token, ship — a managed inference API is the shortest path, and Lilac's OpenAI-compatible endpoint and startup-friendly credits make it a reasonable one. The same is true for prototypes where infrastructure control simply doesn't matter yet.
The calculus changes when the model becomes part of your system: when data can't leave your environment, when token bills need to become predictable compute costs, when the same platform needs to run your training pipeline and your app servers — or when you want the option to put your own hardware under the workload without rewriting anything.
Moving from a hosted API to your own inference
The models are the point of overlap: the open-weight models on a hosted catalog can run privately on Kinesis through the App Gallery. The migration is an endpoint change in your application — the model now runs in your environment, at compute cost rather than token price, with the rest of your stack deployable beside it.
FAQ
Is Kinesis cheaper than Lilac? They price different things, so it depends on utilization. Per-token pricing bundles the provider's compute, margin, and idle risk into every token. Running the same open-weight model on Kinesis prices the underlying compute — e.g. an H100 at $2.50/GPU-hr dedicated (checked July 30, 2026) — which typically wins at sustained utilization and loses for occasional calls. For raw dedicated GPU hours, Lilac advertises a lower committed rate (~$2.00/hr H100 with monthly commitment); Kinesis's rate is on-demand with no commitment.
Can I run the same models Lilac hosts? Open-weight models like the ones in Lilac's catalog can be run privately on Kinesis via the App Gallery — same weights, your environment, no shared endpoint.
Do I have to manage infrastructure on Kinesis? No. The platform manages the machine layer — OS, drivers, software stack, failover, scaling. You deploy a container, a Dockerfile, or a GitHub repo. It's more setup than calling a hosted API, and considerably less than operating GPU instances yourself.
What about data residency and IP? Workloads and data stay inside your environment on Kinesis — nothing transits a shared inference endpoint. Placement respects compliance constraints, and you can attach your own hardware for full locality control.
Do I need a credit card to try Kinesis? No. New accounts get $100 in credit with no card required.
Want your models on infrastructure you control? Try Kinesis — $100 credit, no card or see full pricing.