2 million GPUs and the next phase of enterprise AI

AI’s next phase isn’t models. It’s infrastructure.

If you’re planning AI beyond pilots, this is the shift that matters this quarter: the bottleneck is moving from “which model?” to “what workflow, on what infrastructure, at what cost per outcome?”

Recent announcements, including AWS and NVIDIA’s plan to deliver 2 million additional GPUs across AWS’s global infrastructure through 2027–2028, show how seriously the market is betting on agentic and physical AI. The headline GPU count is easy to report; the more useful question is what this signals for your own AI roadmap.

Why infrastructure now matters more than the model

A single AI request can now trigger multiple model calls, database queries, tool executions, retrieval steps, and verification loops. As systems become more autonomous, compute consumption becomes less predictable. The price of an individual model call tells only part of the story. The real cost sits across the entire workflow: retrieving data, processing context, running models, executing tools, verifying outputs, and sometimes repeating the process. Initiatives like the expanded AWS–NVIDIA collaboration reflect this broader infrastructure requirement, extending beyond GPUs into CPUs, networking, data processing, vector indexing, and robotics. The model may be the visible part of an AI application; the infrastructure around it increasingly determines whether the application is economically viable.

What this means for your AI plans

For most businesses, the answer is not to start buying GPUs. Cloud infrastructure at this scale will remain largely invisible to the end user. The more important question is whether AI workflows are being designed with their infrastructure requirements in mind.

As you review your roadmap, ask:

• What does the workflow actually require?
  Not every task needs the largest model or continuous autonomous execution.
• What does one completed outcome cost?
  Measure the economics of the entire workflow rather than individual model calls.
• Is the data infrastructure ready?
  Retrieval, indexing, storage, and context management can become bottlenecks long before the model does.
• What happens at production scale?
  A workflow that works for 100 test cases may behave very differently at 100,000 executions.
• Who owns the outcome?
  Infrastructure can scale automatically; accountability cannot.

These are the kinds of questions Saguna works through with clients when moving from pilots to enterprise‑scale AI, ensuring that infrastructure investments map directly to measurable business value.

How we approach AI, cloud, and product delivery

We treat AI as an infrastructure and workflow challenge, not just a model‑selection exercise. Our focus is on designing systems that are economically viable at scale, aligned with business outcomes, and built to evolve as usage grows. That means starting with the workflow and the data, then choosing models and cloud patterns that support reliable, cost‑efficient execution.

In practice, this means working with clients to clarify the end‑to‑end workflow, measure cost per completed business outcome, and embed AI into existing products and platforms in a way that amplifies human judgment rather than replacing it.

Key elements of this approach include:

• Workflow‑first design: Define the business process and success criteria before selecting models or infrastructure.
• Outcome‑based economics: Measure cost per completed outcome instead of focusing only on token usage or model calls.
• Data and context as core: Treat retrieval, indexing, storage, and context management as integral parts of the AI system.
• Production‑minded scaling: Start with a small, real use case in production and scale only what demonstrates measurable value.
• Managed where commoditized, owned where it matters: Use managed cloud and AI infrastructure for standard layers, while retaining ownership of the data, workflows, and logic that create differentiation.

Where this is headed

The AWS–NVIDIA announcement is ultimately less about GPUs than about the direction of AI itself. As AI moves from experimentation into production, infrastructure becomes part of the strategy. Agentic systems will require more compute and more complex execution environments, while physical AI will extend those requirements into the real world.

The companies that benefit most will not necessarily be those spending the most on AI. They will be the ones that understand what infrastructure their workflows actually require, what those workflows cost, and whether the resulting output creates measurable value.

Saguna Consulting’s work across artificial intelligence, cloud enablement, technology transformation, and product development is built around this implementation challenge: helping organizations move from AI experiments to production systems that are reliable, cost‑efficient, and tied to clear outcomes.

AI is no longer just a software capability being added to the business. It is becoming infrastructure and companies need to start designing for it accordingly

‍