Skip to content
Kinesis

Kinesis OneTeam · Share your team’s GPUs without a scheduler project.

More developers than GPUs? Kinesis Grids keep every team moving.

Turn the GPUs you own into one shared Kinesis Grid, with one install command per server: no Kubernetes, no Slurm. People start work without filing a ticket, within the limits you set. Supported data center GPUs split into isolated slices; every other GPU is shared whole, and when GPUs run short, everyone takes turns.

Get started freeSee what your GPUs are really doing

Free to start: no Grid fee on your first 5 servers, and your first 5 users are free.

What keeps teams moving

Less waiting, more shipping.

Everything a shared GPU fleet needs to keep developers moving and every GPU busy, without a scheduler project or a spreadsheet of who gets what.

  • Grids / lab-a100 / Policies
    Fair shareOffFair share

    Allotted time

    4 h

    Apps per person

    2

    Activity

    • Queued for grid capacity (fair share)ana · train-vit
    • Moved off machine — turn ended (fair share)ben · sweep-12
    • Machine reserved for app (fair share)ana · train-vit
    • Admitted from grid queue (fair share)ana · train-vit

    Fair share

    When GPUs run short, apps wait in a queue and take turns. Set the allotted time and apps per person; the Grid moves work on when a turn ends.

  • Grids / lab-a100 / Policies / Quotas
    • Limit GPUs per person4
    • Limit apps per person3
    • Limit instances per personNo limit
    • Grace before stopping15 min

    Exempt apps

    shared-inferencelab-jupyterhub

    Container limits

    CPU 16 coresMemory 64 GB

    Quotas and limits

    Cap GPUs, apps and instances per person on each Grid, set CPU and memory limits per container, and exempt the shared services everyone relies on.

  • Servers / gpu-node-07 / GPU slices
    H100 80GBMIG Slicing On

    Unit

    VRAM

    Assigned app

    3g.40gb40 GBthesis-llm
    2g.20gb20 GBcs231-notebook-04
    1g.10gb10 GBcs231-notebook-11
    1g.10gb10 GBFree

    Fractional GPUs

    Split supported data center GPUs (A100 and H100 class) into isolated slices with NVIDIA MIG, so one card serves a whole class or several jobs. Each slice counts as one GPU against quotas. GPUs that can’t be sliced are shared whole, in fair-share turns.

  • Fleet · All Grids
    • dgx-018× A100 · lab-a100
    • colo-h100-028× H100 · research
    • lab-l40s-034× L40S · teaching
    • ws-rtx-112× RTX 6000 · teaching

    VisionSearchResearchInterns

    Your whole fleet in one view

    Every server across every Grid, with whose work is running on each GPU. Usage by person and cost by project sit one click away.

Illustrative screens · names and numbers are made up

  • Idle GPUs come back

    The Grid spots apps that hold GPUs they aren’t using, warns their owners and frees the GPU for the next team.

  • A Grid per group, if you want

    Give each lab, division or cohort its own Grid with its own admins, or pool everything into one.

Who it’s for

Made for places where many people share a few GPUs.

When many people share a few expensive GPUs, the hard part is keeping everyone moving: a fair turn for every team, room for every job, and knowing who is using what. A Kinesis Grid takes care of all three.

  • Universities and research labs

    One cluster for every lab and course, without a sign-up sheet.

    • Each lab or course gets a quota. Fair share handles the rush before deadlines.
    • Slice large GPUs so a whole class can work at once.
    • Usage by person and cost by project for grant reports and chargeback.
    • Students deploy notebooks and models themselves. No admin tickets.

  • Incubators and accelerators

    GPU access for every startup in the cohort, on the program’s own hardware.

    • Every startup gets its own limits, so one launch week doesn’t starve the rest.
    • Separate projects and roles for each team. Share one Grid or give each its own.
    • Program staff see usage by person and cost by project for the whole cohort.
    • Startups keep the same workflow after they graduate, on their own Grid or on Kinesis capacity.

For developers

Let developers try things without asking first.

From idea to running app without waiting on anyone. Try it, break it, redeploy it, all inside the limits you set.

  • Deploy a container image or a GitHub repo in minutes
  • One-click apps from the gallery, from Jupyter notebooks to open LLMs
  • HTTPS endpoints, DNS and load balancing built in
  • A browser terminal, live logs and metrics for every app
  • Persistent volumes, so data survives a redeploy
  • A REST API and an MCP server, so scripts and agents can deploy too

For managers and finance

Know where every GPU-hour goes.

Spend broken down by project and app, usage by person, and a forecast to month-end. Invoices and hourly exports are finance-ready from day one.

Single sign-on, roles, two-factor authentication and an audit log come standard. Need a dedicated or fully private deployment? See Kinesis Private

Your Grid never places work on community hardware unless you choose Kinesis Batches.

See Grid pricing →
Cost by project
this month
VisionSearchResearchforecast

Illustrative

See what your GPUs are really doing.

Start with a read-only assessment of your servers. Whether or not you go further with us, you keep the utilization numbers.

Request an assessment