Chetan Gupta — Chief AI Officer at Rackspace, previously head of all AI research at Hitachi — argues the enterprise AI conversation has been aimed at the wrong target. Models are a commodity that swaps in and out for commercial and geopolitical reasons; what determines whether AI actually lands is the architecture wrapped around the model: the harness that ties it to an outcome, the orchestration layer that manages many harnesses, the eval layer that lets you swap models without your customers feeling it, and the governance and assurance planes around all of it. “Stop thinking models and start thinking architectures — enterprise architectures for AI” is his one-liner.

From industrial AI to the architecture problem

  • His career tracked the field: PhD in math, HP Labs research, then Hitachi, where he took a bet on industrial AI in 2016-17 (a small North American lab competing with Facebook and Google for talent) that grew into leading all of Hitachi’s AI research globally
  • Industrial AI started as predictive maintenance, repair recommendation, quality and defect detection; deep learning added vision, LLMs added language — and now physical AI: robotics at scale, multiple robots coordinating, virtual worlds for training
  • The geography is telling: the center of gravity for LLMs is North America, but for industrial robotics it is China; Japan is culturally more open to robots (he worked on elderly-support robots where “getting the empathy right was a very hard problem”); his read is commercial robots land first because factories are already task-specific and automated — general robotics for human environments is “anyone’s game”

Why models are not the decision anymore

  • Pre-generative AI, AI reached businesses mediated by a team of PhDs who built and deployed small models into customer workflows; that model of delivery is gone — enterprises can’t staff or fund building frontier LLMs, and everyone now has direct access
  • The problems that replaced it: tokenomics (don’t send “what is the capital of Norway?” to an expensive hosted model when a local one answers), data sovereignty (every query exports your data, context, and processes — your IP, your alpha), jagged capability (great at coding, surprisingly mediocre at writing email — “verbose” and “cliched”), and governance and assurance (deployed models and third-party APIs both need guardrails that match your enterprise, not a model maker’s)
  • On open weights: with Meta stepping back and Chinese open models catching up fast, he expects US open-source models to re-emerge — “the beauty of the United States is that with the right incentive, we really step up”
  • For enterprises already on the journey: “fix a few things” — know which workloads are sensitive (an HR query can leave; your proprietary data should not), keep sensitive work on local or open-weight models in an environment you control, and don’t marry a model family; models will churn for commercial and geopolitical reasons, so marry an architectural way of thinking

Sovereignty, extended beyond nation-states

  • Sovereignty used to mean a nation having its own AI stack; the conversation shifted when people realized every interaction with an external LLM hands over data and context that AI can act on — including building things that compete with you
  • His framing now: sovereignty applies to any entity — enterprise, and eventually individual. “It is not someone else’s constitution that I have to use, but my own constitution from a sovereign standpoint.” People already use these models as therapists and friends; personal data sovereignty is coming down the stack

The architecture, layer by layer (Rackspace’s view)

  • Rackspace goes “from chip to outcome”: AMD partnership plus own data centers, selling private AI environments customers run fully under their control
  • The stack follows familiar design paradigms: compute → data → model layer (LLMs plus your own models in a model library) → inference layer → harness → orchestration → consumption layer, all wrapped by governance and assurance planes
  • The harness is the key construct: “the machinery that specifies the logic, the underlying AI model to use, and the tool sets it could use — through MCP or whatever.” Same model, two different harnesses, two different outcomes. Enterprises run multiple harnesses (coding, HR agents, research) that need an orchestration layer — not so different from how every past stack was abstracted through APIs and specifications

Evals are what let you ride the churn

  • To keep customer experience steady while models turn over, build evals for your own workloads — public benchmarks are only guidelines and don’t represent your workloads, and jagged AI means benchmark winners don’t necessarily win yours
  • An eval layer inside the orchestration framework lets you swap in a cheaper, lighter, safer model when your evals say it holds; it is the modern heir of the industrial “golden dataset” — prove your model works on data relevant to me before you deploy it for my customers

What keeps a Chief AI Officer up at night

  • Architecture design questions (who to partner with per layer, what fits which customer — “no one answer fits all”), building a “mirror org” discipline where Rackspace proves AI on its own workloads before rotating it out to customers, and the underserved layers: governance, assurance, orchestration — especially in sovereign environments, done cost-effectively and reliably

“Stop thinking models and start thinking architectures — enterprise architectures for AI.”

“Don’t marry into any model family… Marry into an architectural way of thinking.”

“You take the same model and you build two different harnesses, you will get two different outcomes.”

“It is not someone else’s constitution that I have to use, but my own constitution from a sovereign standpoint.”

“If you think about large language models, the center of gravity is North America. But for robots, for the industrial world, the center of gravity is China.”