Follow ₹100 Through the AI Compute Sector
01 The Product Chosen, and Why
One hour of AI-augmented IT services delivery — a client paying an integrator like Persistent Systems for a software modernization, data engineering, or agentic-workflow engagement — is the right product to trace for this sector, because it's the one where every stage of this series' research actually connects:
Chip fabrication (TSMC) ➔ Chip design and sale (Nvidia) ➔ Hosting/cloud infrastructure ➔ Foundation model API access (OpenAI/Anthropic) ➔ The services integrator (Persistent) ➔ The end client.
02 The Chain, Stage by Stage
TSMC fabricates the actual silicon that becomes an Nvidia GPU, using its advanced process nodes and CoWoS packaging technology. This is the closest thing to a "raw material" stage in this chain — TSMC doesn't design the chip's architecture or own the software ecosystem around it; it manufactures to Nvidia's specification. TSMC's own reported gross margins have historically run in the 50-60% range, healthy but well below what happens at the next stage.
This stage represents the single largest value-capture step in the entire chain. Estimated production cost of a flagship Nvidia data center GPU is roughly $3,320, sold at approximately $28,000 — an 88.1% gross margin at the individual chip level, and Nvidia's company-wide GAAP gross margin runs at 75%.
Nvidia takes TSMC's fabricated silicon and, through chip architecture design and — critically — the CUDA software ecosystem (6 million developers, 25 years of accumulated lock-in) — turns a $3,320 input into a $28,000 product, an 8.4x markup. This happens four layers removed from the end client.
Reliance, Adani, and hyperscalers (AWS, Azure, Google Cloud) buy Nvidia's GPUs (a massive capex outlay — Reliance's $110 billion commitment) and rent out compute capacity per GPU-hour. This is the "distributor" stage. Cloud infrastructure segments report operating margins in the 30-40% range — healthy, but structurally lower than Nvidia's chip-design economics because this stage is capital-intensive rather than pure IP.
Rents compute from Stage 3, trains/serves foundation models, and sells API access per token. This is the most genuinely competitive stage in the whole chain: Anthropic cut flagship prices 3x in 2025, OpenAI weighed defensive cuts in 2026, and open-source models (DeepSeek) undercut both. Margin economics here are the least favorable: frontier labs' compute costs consume a very large share of revenue (in some periods funded by investor capital rather than operating profit).
This is the stage where the client actually transacts, and where audited figures are available, drawn directly from Persistent's FY26 disclosures:
| Line Item | Share of Persistent's Revenue (FY26) | Economic Impact & Flow |
|---|---|---|
| Personnel Expenses | 69.4% | Flows to software engineers, architects, and consultants delivering work |
| Other Expenses (Software licenses, sub-contractors, travel — including AI model API consumption) | 12.2% | Third-party pass-through costs (including AI API tokens in Stage 4) |
| EBIT (Operating Margin) | 15.6% | Operating profit kept by Persistent before tax & interest |
| Below-the-Line Items (Tax, etc.) | PAT Margin: 12.6% | Final net profit margin remaining after all obligations |
This is the honest, audited answer to how ₹100 of client spend splits once it reaches the integrator stage: roughly ₹69 goes to the people delivering the work, roughly ₹12 flows onward to third-party providers (a portion paying for Stage 4 AI model APIs), and the integrator keeps roughly ₹15-16 as operating profit before tax. This is a dramatically thinner margin than Nvidia's 75%, or even cloud hosting's 30-40%.
The enterprise client pays the final ₹100 and receives the finished, delivered outcome — the modernized system, the AI-augmented workflow, or the completed integration.
03 Mapping ₹100 Through the Chain
No single public disclosure gives an exact, audited split of how one client rupee ultimately distributes across all six stages simultaneously — this has to be reasoned through by combining hard figures available at each stage.
| Stage | Approximate Share of ₹100 Reaching Stage | Approximate Margin Captured | Basis |
|---|---|---|---|
| Integrator (Persistent) | ₹100 in ➔ ₹100 recognized as revenue | ~15–16% EBIT (₹15–16 kept) | Disclosed |
| ➔ of which, 3rd Party Software / AI Costs | ~₹12 flows onward | (Passed through, not kept by Persistent) | Disclosed |
| Model Provider (OpenAI / Anthropic) | Of that ~₹12, a portion reaches model layer | Thin-to-negative operating margin | Estimated |
| Hosting / Cloud Infrastructure | A further portion flows on from model provider | ~30–40% operating margin | Estimated |
| Nvidia (Chip Design & IP) | Smaller absolute rupee amount, highest margin | ~75–88% gross margin | Disclosed |
| TSMC (Fabrication Foundry) | Smallest absolute rupee amount in chain | ~50–60% gross margin | Estimated |
04 Which Stage Actually Captures the Most Value — And the Twist
By margin percentage, the answer is unambiguous: Nvidia, the chip-design stage, captures the most value in this entire chain, at a 75-88% margin depending on how it's measured — vastly higher than the integrator stage (15-16%), the hosting stage (30-40%), or the fabrication stage beneath it (50-60%).
This is the finding worth sitting with: the company that never speaks to the end client, never delivers a finished service, and sits four layers removed from the ₹100 transaction actually keeps the largest share of every rupee that ultimately reaches it — because it owns the one asset (chip architecture plus the CUDA ecosystem) that nobody else in the chain can substitute away from in the near term.
1. The AI Model Layer Hype vs Reality: The stage that gets the most public attention and investor excitement — the AI model layer (OpenAI, Anthropic) — is currently one of the worst stages in the chain from a pure margin-capture standpoint, precisely because genuine competition exists there and suppliers are fighting each other on price rather than extracting rent from buyers.
2. The Integrator Squeeze: The stage that actually touches the client and delivers the visible, billable outcome — the IT services integrator — captures one of the thinnest margins of all (15-16% EBIT), not because the work has no value, but because it's the most competitive, most substitutable, most buyer-power-constrained stage in the entire chain.
₹100 spent on an AI-augmented IT services engagement doesn't distribute evenly, and it doesn't distribute the way intuition suggests.
- AI Model Labs: Assuming high value capture, but currently sacrificing margin to compete on price.
- IT Services Integrators: Delivering visible outcomes, but keeping roughly ₹15-16 out of ₹100 as operating profit (thinnest margin in the chain).
- Chip Design & CUDA IP (Nvidia): Sitting four layers back, but converting every dollar of input into ~$8 of output price — a markup no other stage comes close to matching.