West Bengal’s 2026-27 budget, presented on June 22, put artificial intelligence and global capability centers onto the state’s technology agenda. But putting AI in a budget is not the same as building an AI economy. Bengal has heard of technology ambition before. In the 1990s, as India’s IT-services boom was taking shape, the state had talent and Salt Lake Sector V. STPI-Kolkata became operational in 1995-96 with 25 approved units and exports of Rs 7.61 crore.
What it did not build early enough was the compounding ecosystem that Bengaluru and Hyderabad created around anchor firms, export infrastructure, managers, global customers and professional mobility. The sequence was decisive. Bengaluru had already attracted Texas Instruments in 1985; Hyderabad moved through STPI, HITEC City and Microsoft’s development centre in 1998. The gap is visible today. In 2024-25, Kolkata’s STPI software exports were about $1.5 billion, compared with $14.7 billion for Hyderabad and $48.1 billion for Bengaluru. This is not nostalgia. It is the arithmetic of missed compounding. The mechanism is simple. Technology ecosystems do not form merely because a region has educated people. They form when companies, managers, engineers, customers, investors, universities and public policy reinforce one another over time.
Once that flywheel turns, late entrants cannot catch up by announcing parks, missions or incentives. They need a new wedge. AI can be that wedge, but only if Bengal understands what is new about it. AI has not levelled the field; frontier AI is still dominated by capital, compute, research density and US-China scale. But AI has reopened parts of the field where adoption, integration, evaluation, security and domain deployment matter more than training the largest model. Bengal should not try to win yesterday’s IT-services race or today’s frontier-model race. It should compete where AI has to become useful inside real institutions. The cost curve explains why this is a real opening. Stanford’s AI Index found that the cost of querying a GPT-3.5-level model fell from $20 per million tokens in late 2022 to $0.07 by late 2024, a decline of more than 280-fold.
Frontier training remains exp ensive , b ut application-layer experimentation has become dramatically cheaper. New clusters can form around applying models to difficult workflows, messy data and institutional problems. That is why AI is not simply “more outsourcing.” The classic Indian IT model converted labour into export revenue through Y2K remediation, application maintenance, testing , supp or t , BP O, ERP implementation, staff augmentation and cost-efficient coding. It rewarded headcount scale and process reliability. AI attacks the bottom of that pyramid. EY India has estimated that GenAI could raise productivity in India’s technology sector by 43-45 per cent over five years, with software development seeing around 60 per cent improvement and BPO around 52 per cent. For firms that move up the stack, that is opportunity; for firms dependent on volume, it is margin pressure. The new stack is not coding with copilots.
It is data engineering, model integration, agent design, AI governance, cybersecurity, model evaluation, enterprise workflow redesign, domain-specific copilots, production operations and applied product engineering. In the old model, headcount created scale. In the AI model, capability density creates leverage. A smaller team that understands a bank’s risk process, a hospital’s workflow, a logistics network or a government file system may create more value than a large team doing repeatable delivery work. That should define Bengal’s strategy: not building the biggest models, but making AI reliable, safe and deployable for demanding institutions. The next generation of global capability centers should not be old back offices with “AI” pasted on the signboard.
They should host teams that automate finance operations, modernize insurance workflows, monitor cyber threats, evaluate model behavior, build internal agents, clean enterprise data and integrate AI into production platforms. Bengal’s pitch cannot be cheaper seats. It must be exportable applied-AI capability. The most underappreciated wedge is AI assurance. A fluent model is not automatically ready for a bank, hospital, court, factory or government department. Can it be audited? Can it leak private data? Can it be manipulated? Does it hallucinate under stress? Can bias be measured? Can performance be monitored after deployment? Who is accountable when a decision goes wrong? Red-teaming, privacy testing, model-risk management, compliance engineering, security review and incident response are becoming industrial infrastructure.
Bengal does not need to own the largest models to build a global business around making AI trustworthy. Enterprise modernization is the adjacent opp or tunity. Most organizations do not fail at AI because they lack a demo. They fail because data is fragmented, workflows are undocumented, legacy systems are brittle, procurement is slow and users do not trust the tool. Real AI adoption is not a software purchase; it is institutional rewiring. Firms that connect models to actual work will matter more than firms that sell prototypes. If Bengal wants high-value AI work from global firms, it must understand what those firms actually buy when they choose a location. The Bay Area is useful not as glamour, but as evidence of how technology density works. In 2024, it reportedly drew nearly $70 billion of the world’s $134.6 billion in AI funding.
That is not just a capital story; it is a density story. Researchers become founders, engineers leave platform companies to build start-ups, customers become design partners, investors understand technical risk and managers know how to scale. Ideas circulate. Companies such as Microsoft, Google, NVIDIA, OpenAI and Anthropic cluster where this density exists. They do not arrive because a region declares itself futuristic. Google began as a Stanford research project before becoming a company; the lesson is not academic prestige, but the conversion of research into product, talent into teams, and teams into firms. Bengal’s academic base matters only when connected to funded applied labs, industry problem statements, faculty-founder pathways, paid pilots, start-up procurement and commercialization discipline.
There is also a role only government can play: becoming the first demanding customer. The proof of Bengal’s AI ambition should not be a summit photograph. It should be a citizen getting a service faster. Santa Clara County’s AI-assisted work on millions of property deed records shows the point: government AI becomes real when it reduces bottlenecks in messy public records, not when it generates a press release. Land records, mutation workflows, hospital queues, grievance triage, welfare leakage, procurement analytics and municipal routing are precisely the problems through which local AI capability can be built.
The systems must be secure, auditable and human-supervised. A government that buys intelligently can create a market faster than one that merely promotes AI. The talent warning is urgent. Bengal should not prepare young people for the bottom of a shrinking pyramid: commodity BPO, manual QA, low-end coding, data-entry-heavy operations, routine maintenance, generic support and certificate-driven “AI literacy.” The better path is harder but more valuable: data engineering, model evaluation, MLOps, cybersecurity, product architecture, UX, compliance and domain-heavy deployment. The goal is not a workforce that can repeat AI vocabulary.
It is a team that can ship reliable systems. Bengal’s failure in the IT boom was not a failure of intelligence. It was a failure to convert intelligence into density, confidence and compounding. AI offers another opening, but technology openings close quickly. Other regions will specialize, attract capital, build talent pipelines and saturate the markets that now look available. That happened in the IT boom. It can happen again. The next wave will not reward the state that speaks the language of AI. It will reward the state that learns to ship it.
(The writer is a technology leader with experience in the US software industry, including at Microsoft.)