AI development now looks like research. Enterprises need infrastructure built for experiments
Classic software had a straight path: write code, test, ship. AI breaks that pattern. Models learn, drift and respond with probabilities, so progress looks less like a sprint and more like a lab - many experiments, few keepers, and constant iteration.
As Ketan Umare, co-founder and CEO of Union.ai, put it: "Research becomes a part of the software development process, and we are not used to doing research in a software community." The goal shifts from "ship every idea" to "lower the cost of trying many ideas."
From prototype to production: treat AI like R&D, not just software
Speaking at AWS re:Invent, Umare outlined how Union.ai helps teams run high-volume experiments across clouds without losing control of data, cost or security. The approach: separate the control plane from the execution plane.
The control plane runs on AWS. Execution happens where your data already lives - in your clouds or on-prem. That lets enterprises keep sensitive data in place while orchestrating training, inference and pipelines across multiple environments as one system.
"We can seamlessly connect these multiple clouds into one common way," Umare said. He added that even AI assistants can generate runnable examples that "YOLO" onto the platform, spin up what they need, and pre-provision resources on the way out.
What the platform handles so teams can move faster
- Python-first workflow: write standard Python locally; with a single action it's containerized and deployed for remote execution.
- Seconds-to-deploy target: aim for a one-second path from local to remote so ideas reach real infrastructure without ceremony.
- Multicloud orchestration: unify training, inference and data movement across clouds and on-prem while respecting data residency.
- Security built-in: manage permissions centrally across projects, teams and clouds.
- GPU-aware scheduling: efficiently pool, queue and allocate scarce accelerators.
- Smart caching: reuse shared results across experiment branches to avoid paying for the same work twice.
- Cost control: queue when capacity is tight and lean on reserved capacity where it makes sense.
Why this matters for IT, dev and research leaders
AI products are non-deterministic. There's no path around experimentation - only better ways to pay for it. Union.ai itself is a 46-person company that says it has quadrupled revenue this year while keeping costs lean, underscoring the efficiency mindset.
- Treat AI like R&D: budget for experiment volume, not just headcount.
- Centralize orchestration: one control plane, many execution targets.
- Keep data in place: bring compute to data and enforce permissions at the platform layer.
- Systematize experiments: version inputs, code, models and metrics so results are repeatable.
- Cache aggressively: deduplicate shared steps across experiment trees.
- Use reservations where stable: mix on-demand and Reserved Instances for predictable workloads.
- Give engineers a zero-friction path to prod: avoid "write Terraform first" for ideas that may not survive the day.
- Instrument cost: make spend visible per project, run and user to steer behavior.
Practical next steps
- Pick one model or agent and wire it into a Python-first, containerized path that deploys in seconds.
- Separate control plane from data planes; enforce data-permissions where the data lives.
- Stand up a GPU queue with caching; track hit rates and wait times to tune instance choices.
- Set an experiment budget and a weekly review: keep, tweak or kill based on metrics, not vibes.
- Upskill teams on AI engineering and MLOps basics to reduce handoffs and toil. See curated options here: AI courses by job.
As Umare said, "there's no way out of doing research, because this is a research-based product." The advantage goes to teams that make experimentation cheap, fast and safe.
Your membership also unlocks:
