U.S. National Science Foundation

NSF Awards: 1925001 & 2505560

Helping Researchers Onboard to the Commercial Cloud

A CloudBank workshop for Rice University research IT

Jose Cols, Rob Fatland, Anthony Arendt

UW eScience Institute · CloudBank

February 2027

  • San Diego Supercomputer Center
  • UC San Diego
  • University of Washington
  • UC Berkeley
  • Strategic Blue

Agenda

  1. 1. Cloud and CloudBank14 min
  2. 2. Flowchart18 min
  3. 3. Live demo: AI-assisted deployment14 min
  4. 4. Thinking like a cloud engineer18 min
  5. 5. Common pitfalls13 min
  6. 6. What comes next10 min
  7. Wrap-up3 min

Part 1

Why cloud, and why CloudBank

What the commercial cloud adds to research and teaching

Why cloud?

Alongside your HPC cluster, the cloud lets researchers:

Bring compute to the data

Analyze large datasets where they live, without downloading them

Scale up on demand

Rent GPUs or large-memory machines, and pay only while they run

Run in parallel

Many simulations, parameter sweeps or fine‑tuning runs at once

Use and deploy LLMs

Call hosted models through an API, or serve your own on GPUs

Share reproducible environments

Collaborators and students work in the same setup

Host research services

Web apps, APIs and live model endpoints that others can reach

Your cluster and the cloud work together

HPC cluster

  • Shared queue, fixed hardware
  • Free at point of use
  • Tuned for tightly coupled MPI
  • Admins manage the software

Commercial cloud

  • On demand, any hardware
  • Pay per hour used
  • Managed services and APIs
  • Researcher controls the stack

The friction

What we've heard from researchers

  • I couldn't get a cloud account on my grant without a university credit card.

  • I just don't have enough time to learn all this. Can you give me a jump start?

  • Setting up my first experiment in the cloud took forever, and it was different for every platform I tried.

  • I got a $12K bill last month. It's opaque, and I don't have time to wade through it.

Paraphrased from conversations with researchers

CloudBank: A Managed Service to Simplify Cloud Access

Researcher

with an ACCESS or NAIRR Pilot award

CloudBank

  • Opens your accounts
  • Pays from your award
  • Tracks your spending

Commercial clouds

  • AWS
  • Google Cloud
  • Microsoft Azure
  • IBM Cloud

where your work runs

Run by UC San Diego (SDSC), UC Berkeley and UW eScience · funded by NSF since 2019

Sources: SDSC news, April 2025 · NSF news · ACCESS: CloudBank Research

What you get with a CloudBank account

Billing, handled for you

  • No credit card

    NSF pays through your ACCESS or NAIRR Pilot award; no purchase order

  • One bill, several clouds

    One account across all the clouds CloudBank offers

  • A spending cap

    Alerts as you spend, and tips on idle resources

  • Dollars, not credits

    Also pays for marketplace software and AI model access

Accounts, handled for you

  • Set up for you

    Your cloud accounts are created after onboarding

  • One login

    Your ACCESS account signs you in to each cloud

  • Collaborators

    Add or remove people in one place

  • Room to grow

    Add another cloud to your project later

Two ways to use CloudBank

CloudBank Research CloudBank Classroom
For Research projects of any size Courses taught with Jupyter notebooks
You get Your own accounts on AWS, Google Cloud, Azure and IBM Cloud A ready-made JupyterHub with grading tools; students just log in from a browser
To start Apply for an ACCESS or NAIRR Pilot allocation Apply for an Explore ACCESS allocation; CloudBank sets up your hub

Sources: ACCESS: CloudBank Research · ACCESS: CloudBank Classroom

A shared hub replaces setup day

Laptop setup

  • First hour lost to installs
  • Different versions on every machine
  • Needs a capable personal laptop
  • Every student downloads the data

Shared hub

  • Teaching starts in minute one
  • One identical environment for all
  • Any browser, any device
  • Data waits on the server

Part 2

Guiding a researcher through the flowchart

The conversation we have with every first-time cloud user

Role-play: a researcher asks for the cloud

  • Environmental engineer studying hurricane damage
  • Wants to fine-tune a model on storm imagery
  • Wants a public dashboard of the results
  • Says: “I think I need the cloud.”

Start with the research problem, not the platform

  1. Is the cloud the right tool?

    Nocampus HPC, or ACCESS and NAIRR on-premises resources

  2. Is CloudBank the right route?

    Noa departmental or institutional cloud account

  3. Allocation

    ACCESS (Explore, Discover, Accelerate) or the NAIRR Pilot

  4. Onboard

    CloudBank opens accounts on AWS, Google Cloud, Azure or IBM Cloud

  5. Work

    build, run and watch spending

  6. Exit

    archive, publish and close accounts

Eligibility and recommended uses

Who
U.S.-based researchers and educators, graduate student or above
What
Projects with a scientific focus, or that support a scientific study

Good fit

  • Modest-scale AI workloads
  • Testing software scalability
  • Unique hardware or cloud services
  • Distributed-cloud and network studies
  • Hosting tools, datasets and gateways

Better elsewhere

  • Large-scale compute
  • GPU-intensive workloads

Use on-premises ACCESS or NAIRR resources instead

Sources: ACCESS allocations policy · CloudBank recommended uses

Where LLM work runs best depends on the task

LLM taskClusterCloudWatch out for
Call a hosted modelNot neededAPI, billed per useData leaves campus
Run an open modelGood for batch jobsGood for always-on servicesGPU queues and quotas
Fine-tune or trainBest for large runsGood for small fine-tunesGPU hours add up fast

Grant award tiers

TierRangeAverageTypical use
Exploreup to $10,000$5,000Grad students, first-time users; 1-page proposal
Discover$10,001–$37,500$18,750Multi-user app or scaling up; 3-page proposal
Accelerate$37,501–$75,000$56,250Established gateway, many users, multiple years
  • Hundreds of Explore awards a year
  • Discover and Accelerate are competitive
  • Two installments; renewable after 1 year

CloudBank through the NAIRR Pilot

A second way to a CloudBank account, for AI research

How it differs from ACCESS

Focus
AI research and education
Who
Everyone on the project must be U.S.‑based, not just the lead
Awards
Now cash awards; early awards were in credits
Review
Longer: monthly cycle, decision by the end of the next month

Requesting CloudBank

  1. Prepare Cloud cost estimate, grant number, CVs for PI and co-PIs, an ACCESS account
  2. Request Three-page proposal at nairrpilot.org; reviewers match it to CloudBank
  3. Onboard Within 1–2 weeks: orientation and a monthly cost estimate, then accounts open

Teaching? Request a NAIRR Classroom project for a CloudBank JupyterHub.

Sources: NAIRR Pilot resource requests · CloudBank: NAIRR Pilot Research · NAIRR: CloudBank onboarding · CloudBank: NAIRR Pilot Classroom

From application to account in under a week

  1. Estimate costs

    Use the cloud cost calculators to estimate expenses.

    Cloud provider
  2. Request allocation

    Create an ACCESS account and request credits.

    ACCESS
  3. Exchange credits

    Turn ACCESS credits into CloudBank dollars.

    ACCESS
  4. Onboard

    Follow the instructions in your onboarding email.

    CloudBank
  5. Start working in the cloud

    Cloud provider

Tags show where each step happens. Cloud providers: AWS, Google, Azure, IBM.

Plan the exit before the first instance

  1. TagLabel every resource from day one
  2. MonitorReview spending alerts every week
  3. ArchiveMove keepers to cold storage
  4. ClosePublish outputs, then shut down

Awards end on a date; nothing shuts itself down when they do.

Part 3

From zero to a working cloud deployment

A live walkthrough: account, use case and AI-assisted setup

What you will see in the demo

  1. Log inACCESS identity opens the console
  2. Use caseA workflow a Rice lab runs
  3. Ask AIDraft the poorly documented steps
  4. VerifyRun, tag, then shut down

Use case: hurricane data where it lives

  • NOAA radar and satellite data in public cloud buckets
  • Analyze in the same region; download nothing
  • Serve results as a small web API

Generative AI fills the gaps provider docs leave

Treat its answer as a draft: run it, read the errors, ask again.

Ask for the guardrails up front

I have an AWS account through CloudBank.
Give me AWS CLI steps to run a Python API
on a small spot instance in us-east-1.
Tag everything project=rice-demo and stop
the instance automatically after 4 hours.

Assistants often skip tags and auto-stop unless you ask for them.

Activity

6 min

Hands-on: deploy it with an AI assistant

  1. Open the demo notebook on the shared hub
  2. Ask an assistant for the deployment steps
  3. Run them and fix one error together
  4. Tag, then stop everything you launched

Deliverable

A working endpoint, then a stopped one

Part 4

Thinking like a cloud engineer

Design around storage, compute and network, then choose the abstraction

Every cloud design starts with three resources

Answer these before choosing a service; the cost follows from the answers.

Choose how much of the stack you manage

Managed infrastructure

vLLM on a GPU node

  1. You manage: Your requests
  2. You manage: Code and model choice
  3. You manage: Scaling and patching
  4. You manage: OS and software
  5. Provider manages: Hardware and network

Managed platform

A managed model endpoint

  1. You manage: Your requests
  2. You manage: Code and model choice
  3. Provider manages: Scaling and patching
  4. Provider manages: OS and software
  5. Provider manages: Hardware and network

Managed service

A chat API

  1. You manage: Your requests
  2. Provider manages: Code and model choice
  3. Provider manages: Scaling and patching
  4. Provider manages: OS and software
  5. Provider manages: Hardware and network
  • You manage
  • Provider manages

One AI task at three levels

  1. ServiceCall a hosted chat API
  2. PlatformDeploy to a managed endpoint
  3. InfrastructurevLLM on your GPU node

Same task each time: more control and more responsibility at every step.

Switching levels can be a one-line change

from openai import OpenAI
client = OpenAI(base_url=URL, api_key=KEY)
reply = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": q}])

Point URL at a hosted API or at your own vLLM server; the rest stays.

Source: vLLM documentation, docs.vllm.ai

SkyPilot: describe the job, not the servers

resources:
  accelerators: A10G:1
  ports: 8000
setup: pip install vllm
run: vllm serve $MODEL --port 8000

sky launch finds an A10G, installs vLLM and serves the model on port 8000.

Source: SkyPilot documentation, docs.skypilot.co

When one node is not enough

  1. LaptopPrototype on a data sample
  2. One nodeThe levels you just saw
  3. RaySpread tasks across many nodes
  4. Multi-nodeDistributed PyTorch training

Go deeper with the take-home tutorials and the Cloud Clinic recordings.

Build, learn, then advise

  1. BuildRun a tutorial from GitHub
  2. LearnGo deeper with Cloud Clinic
  3. AdviseHelp a researcher pick a level

CloudBank credits let you build on a real account, not a sandbox.

Activity

2 weeks

Try one example at each level

  1. Call a hosted chat API from a notebook
  2. Deploy an open model to a managed endpoint
  3. Serve it with vLLM on a GPU node
  4. Compare cost, effort and control

Deliverable

A one-page note for your researchers

Part 5

Pitfalls we see again and again

And how research IT can head them off

Six pitfalls and how to head them off

PitfallWhat it looks likeWhat research IT can do
Credits never usedAward granted, nothing launchedBook an onboarding call in week one
Learning curve underestimatedMonths before a first resultOffer a starter template and office hours
Zombie instancesBills grow after the work endsRequire tags, alerts and autostop
Leaked access keysUnknown instances appear overnightUse SSO and short-lived credentials
Wrong serviceA big VM doing a managed service's jobReview the design before launch
Wrong storageHot storage for archives; heavy egressSet lifecycle rules; compute in-region

Rule-of-thumb cloud costs

What common resources cost, to the nearest round number

Compute

10¢/hour
Laptop-class virtual machine 2 CPUs, 8 GB
$1/hour
Powerful workstation 16 CPUs, 64 GB
$1/hour
NVIDIA A10G GPU machine Spot: about half. Ask for quota first.

Storage and data

2.3¢/GB/month
Object storage you use often 1 TB ≈ $23 a month
0.1¢/GB/month
Cold archive storage 1 TB ≈ $1 a month
9¢/GB
Downloading data out of the cloud 1 TB ≈ $90

Left on for a month, a $1/hour machine costs about $730. Turn it off when you are done.

Prices: AWS on-demand, US East (N. Virginia), October 2026. Amazon S3 pricing · Amazon EC2 on-demand pricing

Keep access keys out of code

  • Automated scanners search public repositories for keys
  • Prefer SSO and short-lived credentials
  • Require MFA on every human login
  • Delete keys nobody is using

Bring cloud habits, not cluster habits

Cluster habit

  • One big machine for everything
  • Shared filesystem holds all data
  • Copy results back home
  • Queue jobs and wait

Cloud habit

  • Right-size, or use managed services
  • Object storage with lifecycle rules
  • Analyze in the same region
  • Launch, finish, shut down

Activity

3 min

Spot the pitfall

  1. Listen to three short scenarios
  2. Name the pitfall in each one
  3. Suggest one guardrail that prevents it

Deliverable

One guardrail per scenario, shared aloud

Part 6

What comes next: open, shared, reusable

From single projects to a marketplace of research resources

CloudBank by the numbers

Awards from October 2025 to October 2026

  • 292 research projects awarded
  • $2.24M allocated to research
  • 78 CloudBank Classroom courses

Research awards by tier

TierProjectsAverage
Explore234$4,423
Discover56$19,689
Accelerate2$52,094

Top research fields

  1. AI and intelligent systems
  2. Computer science
  3. Materials engineering

plus 40 more fields

Source: CloudBank allocation records, October 2025 to October 2026

From one lab's project to everyone's resource

  1. BuildWork in a cloud project
  2. PublishGitHub code; Hugging Face models, data
  3. DiscoverA shared catalog across institutions
  4. ReuseAnother lab runs it unchanged

Open science is the exit plan that keeps paying off.

Shared LLM tools for research computing

  • Prompt libraries for common cloud tasks
  • Assistants that know CloudBank's processes
  • Agents that tag, budget and shut down
  • Reviewed, versioned and shared like code

Activity

3 min

What would Rice share first?

  1. Name one Rice dataset, model or tool
  2. Pick where it would live and who maintains it
  3. Name what stops it being shared today

Deliverable

A shortlist of Rice resources worth sharing

What to take back to your researchers

  • Ask the fit question before the platform question
  • Route through CloudBank: no card, one bill
  • Let AI draft setup; you verify it
  • Choose the level, then scale
  • Set guardrails early; share what you build

Learning & support

You don't have to figure out the cloud alone

Learn on your own

Ask a person

  • Office hours

    Monthly on Zoom, 10–11 am Pacific. Bring a draft proposal or a billing question.

  • Community forum

    Ask a question or search earlier answers

    community.cloudbank.org

Write to us help@cloudbank.org

Share Your Feedback

Help us improve this workshop — your response takes about 2 minutes.

Thank You

Apply for credits and try one example at each level.