The consensus view
Nifty IT down 20%+ this year. TCS and Infosys at 52-week lows. JP Morgan: stuck in the "deflation phase." FY27 guidance: 1.5–3.5% growth, if things go well. Verdict from Bengaluru to Wall Street: the Indian IT services model — selling human effort by the hour — is dying, and AI is the executioner.
Here's what the obituary writers missed: the executioner has a cost problem, and the "dying" companies are the only ones positioned to fix it.
For three decades, Indian IT ran on a simple equation: more work meant more people, more people meant more billable hours, more hours meant more revenue. Headcount was the business model.
AI breaks that equation at the root. When a coding agent produces in an afternoon what a five-person team produced in a sprint, the client's question is no longer "how many engineers do I need?" It's "why am I paying for engineers at all?" Clients now demand productivity-based pricing and insist that AI efficiency gains be passed back to them — what analysts call AI-led deflation. TCS revenue actually declined 0.5% in FY26 in dollar terms.
Anyone selling hours in a token economy is selling ice in winter.
The AI pilots worked. The AI economics didn't. Here's the math nobody puts in the press release:
$30 / M tokens
Most expensive frontier reasoning models (input)
$180 / M tokens
Most expensive frontier models (output)
100s of calls
A single agentic workflow can fire off hundreds of sequential model calls
18% GST surcharge
Indian enterprises pay on top of dollar-denominated API bills
The great enterprise AI paradox of 2026
Everyone has deployed AI, and almost no one is profiting from it. The gains are real — but they're being captured by model providers and cloud vendors, not by the enterprises paying the bills, and not by the integrators implementing the systems. This is the market failure. And market failures are where new business models are born.
$30
Frontier model
per M tokens
→
200× cheaper
$0.14
DeepSeek V4-Flash
per M tokens
For most enterprise workloads — document processing, customer support, code assistance, data extraction — no meaningful quality difference
Open weights mean something even more important than cheap tokens: they can run inside the enterprise's own walls. On the company's own GPUs. Behind the company's own firewall. With zero per-token charges, zero data leaving the building, and zero dependence on a foreign provider's pricing whims.
The catch? Deploying open models is genuinely hard.
You need to choose the right model for the task, size the GPU infrastructure, pick a serving stack (vLLM, SGLang, Kubernetes), quantise, fine-tune on proprietary data, build guardrails, wire it into legacy ERP and CRM systems, and then run it forever.
That is literally the job description of an IT services company.
Indian IT services firms are not knocking on Fortune 500 doors asking for AI business. They are already inside. They run the mainframes, maintain the ERP systems, staff the support desks, hold the security clearances, and know where every skeleton in every legacy codebase is buried.
Early AI revenue numbers (FY26)
A Silicon Valley AI startup selling to a Fortune 500 bank faces an 18-month procurement and security review before writing a line of code. An Indian IT major already has 4,000 people on that bank's floor and knows which COBOL module breaks every quarter-end.
The pieces are already on the table. Here is what the second act looks like:
Every enterprise wants frontier-class AI without frontier-class token bills. The winning offer: 'We'll deploy Qwen, DeepSeek, Mistral, or Sarvam on your infrastructure, fine-tuned on your data, at a fraction of your current API spend — and your data never leaves your building.' This converts a variable, unpredictable token cost into a fixed, ownable asset.
'We cut your claims-processing cost by 40% — we take a percentage of the savings.' AI-led deflation stops being a threat when you're the one delivering the deflation and getting paid for it. The very pricing pressure crushing legacy contracts becomes the product.
The hardest part of enterprise AI isn't the model — it's integration with 30-year-old systems nobody else understands. Indian IT firms hold irreplaceable maps of client legacy estates. Fine-tuning an open model on a client's proprietary codebase creates a bespoke AI asset a generic API can never match.
Banking, healthcare, and defence increasingly cannot send data to US cloud APIs. Every data-residency law, every geopolitical tremor, is a sales pitch for on-premise open-model deployments. Indian firms have credibility in exactly the emerging markets where sovereignty demand is highest.
Models drift. Prompts break. GPUs need optimisation. Managed AI operations is annuity revenue — and running things reliably at scale for decades is the one competency nobody has ever accused Indian IT of lacking.
Speed
The window is open now, but hyperscalers and AI-native startups are racing toward the same enterprise deployment market. Incumbency advantages decay.
Talent transformation
Selling outcomes requires senior AI architects, not benches of junior coders. Retraining a multi-million-person industry is a generational project — even as TCS and Infosys already pivot campus hiring toward AI and cloud skills.
The pricing trap
If firms use open models merely to discount the same old hourly contracts, they'll accelerate their own deflation. The model shift — hours to outcomes — is the hard part, and it demands boardroom courage, not just technical capability.
The bottom line
The market is pricing Indian IT services for the death of the billable hour. Fair enough — that model is dying. But the same market is ignoring a rarer asset: trusted physical and contractual presence inside the world's largest enterprises, at the exact moment those enterprises are desperate for someone to rescue them from runaway AI token costs.
The token economy created the problem. Open source created the solution. The obituary may turn out to be a birth announcement.
The market is pricing the death of the billable-hour model faster than it is pricing the growth of new AI-services revenue. Traditional IT contracts are shrinking due to AI-led productivity gains, and the new outcome-based and open-model-deployment revenue hasn't yet replaced the lost volume. The sentiment is bearish even as the underlying strategic position improves.
AI token deflation refers to the collapse in the cost of AI inference — the price per million tokens has fallen dramatically, making AI tasks cheaper but also making per-hour human labour look expensive by comparison. Indian IT firms built on billing for human effort are directly squeezed: clients demand that AI productivity gains be passed back to them, reducing contract values.
Open models can be deployed inside a client's own infrastructure, eliminating per-token API charges (which can run into millions of dollars annually) and keeping sensitive data behind the client's firewall. Deploying, fine-tuning, and managing these models is exactly the kind of integration and operations work Indian IT firms excel at — turning a cost problem for enterprises into a services opportunity.
Sovereign AI refers to governments and regulated industries (banking, healthcare, defence) running AI models on their own infrastructure rather than sending data to foreign cloud APIs. Data-residency laws and geopolitical risk are accelerating this trend globally. Indian IT firms, with established relationships in Asia, Middle East, and Africa, are well positioned to deliver on-premise open-model deployments to these markets.
The billable-hour model is under severe pressure, but the sector is not dead. TCS, Infosys, HCLTech and peers already have billions in AI revenue, deep enterprise relationships that AI startups can't replicate, and a natural fit for the open-model deployment market. The transition from selling hours to selling outcomes is hard — but the companies everyone wrote off are already inside the buildings where the decision gets made.
Wondering how AI will affect your own career in IT?
Take the free AI Risk Quiz — get a personalised risk score and the skills to build next.
Take the Free AI Risk Quiz →