The Real Bottleneck Behind the AI Spending Surge

The 725 billion dollar AI capex story is hitting the wrong constraint.

Most Indian enterprises blame stalled AI programmes on GPU supply. The honest test is simple: if your organisation has unused GPU capacity sitting idle, the bottleneck is not GPUs. It is whatever is keeping work from arriving at the GPUs. In our experience, GPU utilisation in private deployments runs below 30 percent. The infrastructure is warm. The work is not arriving.

The Four Real Bottlenecks

1. Data readiness. The PDFs are unstructured, the ERP exports are incomplete, the customer master has six versions of the same customer. Until pipelines, lineage, and quality controls exist, no model is useful. Data engineering is not a precursor to AI work — it is AI work.

2. Talent depth. One head of AI is not capability. Data engineers, MLOps engineers, AI product managers, and embedded domain experts are all required. Most enterprises have one or two of these. Few have all four with meaningful depth. Hiring one senior AI leader and expecting the organisation to follow is how pilots stay pilots.

3. Governance and risk. Legal, compliance, and risk move slower than the technology, and that is appropriate. The bottleneck is not the review — it is the absence of a framework that tells legal what to review. Enterprises with no AI governance framework spend more time in legal hold than those with even a basic one.

4. Change management. The user who has to adopt the tool is rarely in the room when it is built. Pilots succeed on a self-selected sample of enthusiastic early adopters. Rollouts die on contact with the actual workforce — the sceptics, the busy, the unconvinced. The model is fine. The adoption programme was never built.

A Realistic Capex and Opex Split

For an 18-month programme aiming at 3 to 5 use cases in production, the money should go roughly here:

Data plumbing — Recommended: 30–40% | Typical actual: ~10%

Talent — Recommended: 25–30% | Typical actual: ~20%

Platform and tools — Recommended: 15–20% | Typical actual: ~30%

Models and licences — Recommended: 10–15% | Typical actual: ~30%

Governance and risk — Recommended: 5–10% | Typical actual: ~10%

Most enterprises actually run 60 percent on models and platforms and 20 percent on talent. That split predicts the typical outcome: impressive demos, stalled rollouts, and a board asking why the AI investment has not shown up in the numbers.

How to Sequence

Q1: Data plumbing for two priority domains. No model work yet. The deliverable is clean, pipeline-fed data for the domains the first use case will touch.

Q2: Governance framework and a private inference platform. Still no production use case. The deliverable is a paved road — a place models can run, with audit trails and access controls in place.

Q3: First production use case. Narrow, high-volume, low-stake. One team, one workflow, a manual fallback.

Q4: Second and third use cases reusing the infrastructure built in Q1 and Q2. Marginal cost drops significantly because the foundation exists.

Q5+: Scale once utilisation justifies it. Add capacity to a working system, not a speculative one.

The temptation is to skip to Q3. Enterprises that do consistently end up rebuilding Q1 and Q2 inside the use case — slowly, expensively, while the business loses patience and the AI programme loses credibility.

Where to Spend Rs 5 Crore in Year One

If forced to allocate a first-year AI budget of Rs 5 crore, the split that produces the most durable result:

• Rs 1.5 crore — data engineering: two senior engineers, tooling, first two domain pipelines
• Rs 1 crore — core AI team: one MLOps engineer, one AI engineer, one AI product manager
• Rs 75 lakh — private inference platform: own a 30B to 70B class model in-house, on infrastructure you control
• Rs 75 lakh — first production use case end-to-end
• Rs 50 lakh — governance: legal review, audit log infrastructure, framework documentation, team training
• Rs 50 lakh — held for surprises (there will be surprises)

The Rs 5 crore that goes 60 percent into GPU clusters and enterprise AI licences before any of the above exists produces a well-equipped pilot environment. The Rs 5 crore split above produces running systems.

What This Means in Practice

The bottlenecks are addressable. They are also slow. There is no shortcut through data quality, no way to hire a mature MLOps function in a quarter, no governance framework that writes itself overnight.

Enterprises that started the unglamorous work in 2024 — the data pipelines, the internal AI platforms, the governance frameworks — are putting things into production now. Those that are still waiting for AI to mature, for the models to get better, for the use case to become obvious, will still be piloting next year.

The question is not whether to invest in AI. That decision is made. The question is whether the investment goes into the parts of the stack that are actually the constraint, or into the parts that are already warm and waiting.

To discuss private AI infrastructure and the operating model around it, reach out to admin@setidure.com.