India's AI gamble: training the machines, training itself out of work
On 11 June 2026, two announcements landed within twenty-five minutes of each other and pointed at the same anxious question. At 10:25 UTC, Reuters reported that Tata Consultancy Services — India's largest IT services firm and a bellwether for a sector that employs more than five million people — had signed a partnership with the US artificial-intelligence lab Anthropic to scale enterprise AI deployments. By 10:50 UTC, Al Jazeera's website carried a separate dispatch from Bengaluru: developers were paying workers to record first-person video of themselves doing jobs — packing, sorting, sweeping, lifting — so that AI models could learn to copy the movements and, eventually, replace the people doing them. The two stories are not the same story. Read together, they describe a single trajectory: a country that built its economic reputation on being the world's back office is now being asked to be the world's training data.
The question the rest of the decade will turn on is not whether Indian workers can teach machines to do their jobs. They evidently can, and the rates being offered for footage of routine manual labour suggest a market has already formed. The question is who owns the resulting capability, who captures the productivity gains, and whether the country's political economy is structured to convert this moment into broadly shared prosperity — or into another commodity export, with the value realised on someone else's balance sheet.
A sector on the edge of its own contract
TCS sits at the top of a $280-billion Indian IT industry that grew, for two decades, by selling time. Indian engineers, paid a fraction of their Western counterparts, wrote code, ran back-office processes, and answered help-desk calls for European and American clients. The model was arbitrage: skilled labour, time-zone overlap, English fluency, and a cost differential that made offshoring the default answer to almost any back-office question. The TCS–Anthropic partnership, as reported by Reuters, recasts that contract. Instead of selling hours, TCS is selling the ability to displace hours — at its own clients, in its own delivery centres, and across the Indian IT workforce that built the firm.
This is not yet a layoff story. TCS executives, in the kind of phrasing executives use, told Reuters the partnership would let the company "re-skill" workers for "higher-value" roles. The history of the IT outsourcing sector, however, gives reason to read those phrases with care. The same rhetoric attended the 2003–2008 move into "business transformation services" — a re-labelling exercise that, in practice, meant a thinner pyramid of senior staff supervising a much larger pool of automation, subcontracted labour, and lower-paid domestic contractors. The percentage of the Indian IT workforce on traditional bench projects shrank; the percentage on shorter, project-based contracts grew. Each technological shift was sold as an upgrade. Each, on aggregate, redistributed bargaining power away from labour and toward platform owners.
The Anthropic deal accelerates a transition that the consultancy sector has been negotiating for five years. Indian IT revenue growth, which ran in double digits through the 2010s, has been a story of single-digit annual gains and shrinking operating margins since 2022, as clients routed an increasing share of work to in-house AI tools. The structural bet being placed by firms like TCS is that the only way to defend the revenue line is to be the firm that implements the AI — even if the AI is, by design, a labour-replacing technology. The partners in that bet are, by definition, the AI vendors themselves.
The dataset is the workforce
The Al Jazeera report from Bengaluru describes a more visceral version of the same transaction. Developers are paying Indian workers — the report does not specify rates across the programme, but participants describe small per-clip sums — to record themselves performing tasks that are, in many cases, precisely the jobs those workers currently hold. The footage is fed into specialised AI models with the explicit goal of producing robots that can perform the same tasks without the worker attached. The framing in the piece is that the workers are aware of the dynamic: they know they are training their own replacements. The framing also notes that, for many of them, the per-clip payment is the only way to make rent.
This is the labour question stripped of corporate procurement language. TCS is moving its workers from "human-hour seller" to "AI-implementer" in the optimistic version, and to "redundant" in the pessimistic one. The Bengaluru data-labellers are being moved, by their own labour, from "worker" to "raw input". In neither case is the worker being given a meaningful stake in the system they are enabling. The dataset flows up. The model is trained. The productivity gain accrues to the firm licensing the model. The worker, having delivered the data, is no longer needed for the next unit of production.
This is not a uniquely Indian dynamic. Workers in the United States, the Philippines, and Kenya perform analogous labelling work for large-model training. What is distinctive about the Indian moment in mid-2026 is the convergence of the two flows — corporate IT and physical-task data — into a single national pipeline. TCS signs the partnership that lets Anthropic sell enterprise AI to TCS's own clients. A separate set of developers pays Indian workers to make the embodied-AI counterpart of that enterprise software. The country is producing, simultaneously, the demand-side implementation capability and the supply-side training data.
What the prediction markets are pricing in
The day before the TCS announcement, on 10 June, a Polymarket contract on which US companies the federal government might take equity stakes in showed a 34% implied probability that Washington would take a stake in Anthropic specifically. The contract is a thin read on political possibility — prediction markets aggregate sentiment, not policy — but the figure is useful as a snapshot of how the underlying asset class is being repriced. Frontier AI labs, having burned tens of billions in training costs and compute, are no longer being treated as straightforward venture bets. They are being treated as strategic infrastructure. The US government has, in adjacent sectors, taken direct stakes in semiconductor fabrication, rare-earth processing, and critical-mineral refining. The Polymarket number is, in effect, a market bet that frontier AI is joining that list.
For India, the implication is double-edged. If Washington takes a stake in a frontier lab, the partnership between that lab and TCS becomes a partnership with a US-government-aligned counterparty. The terms on which enterprise AI is licensed to Indian clients — pricing, data-residency, export-control carve-outs — will be shaped in Washington, not in Mumbai or Bengaluru. If the stake does not materialise, the question is whether the private capital structure of frontier AI is sustainable at the scale required to meet demand from a partner like TCS, which has the customer base to deploy the technology at industrial scale from day one.
The prediction-market read is also a read on the broader pattern: AI is moving from a venture-funded software category to a sovereign-asset category. The partners a country chooses, and the terms it negotiates, are no longer purely commercial. They are geopolitical. India, which has historically hedged between the US technology stack and a non-aligned posture in hardware and telecommunications, is being pulled by partners like TCS into deeper integration with a single frontier lab.
The structural frame: arbitrage to dependency
The plain-language version of the pattern is this. India's IT sector grew by selling time cheaper than the West could. The arbitrage closed in two stages. The first stage was wage compression inside the sector, as the cost differential narrowed. The second stage, the one we are now inside, is the displacement of the labour input itself. Indian workers train the models that replace Indian workers. The country captures a share of the value — corporate revenue, tax base, employment at the implementation layer — but a larger share of the value migrates to the firms that own the models and the platforms.
This is not a one-country story. It is the next phase of a global pattern in which low-cost labour is a transitional input to systems that, once trained, no longer need the labour. Bangladesh's garment sector went through a quieter version of this when automated cutting and sewing machinery moved up the value chain. The Philippines' business-process-outsourcing sector is negotiating its own version. India's distinctive exposure is the scale of its IT workforce and the depth of its integration with the US technology stack — both of which make the transition faster and the dependency deeper.
The counter-narrative, the one that comes from the companies themselves, is that every previous technological shift in the IT sector created more jobs than it destroyed, and that the same will be true of generative and embodied AI. There is genuine historical evidence for that claim in the medium-run aggregate — India's IT employment roughly tripled between 2005 and 2022, even as automation ate into specific job categories. The honest reading of the evidence, though, is that the new jobs were not the same jobs as the old ones, the new jobs paid differently, and the new jobs were concentrated in a smaller, more credentialed slice of the workforce. The re-skilling rhetoric is, in this sense, not wrong. It is also not comforting.
What the wire shows, and what it does not
The two source reports and the prediction-market snapshot are the spine of this story. Reuters carried the TCS–Anthropic partnership at 10:25 UTC on 11 June 2026 and separately reported, the same morning, that the final report into the crash of Air India flight 171 was being delayed by unfinished engine analysis — a reminder that India's aviation regulator is also working through the implications of a high-profile accident in the same news cycle. The BBC's coverage, filed late on 10 June, examined the public dispute over the crash's cause while the investigation continued. Al Jazeera's Bengaluru piece, published at 10:50 UTC, supplied the worker-side view of the AI pipeline.
What the four reports together do not show is the policy frame. The Indian government has not, in the visible record of this news cycle, staked out a position on whether Indian workers should be compensated when their data trains commercial models, or on whether the country's AI partnerships should include localisation requirements, IP carve-outs, or domestic compute commitments. The TCS–Anthropic announcement was reported as a corporate deal. The question of whether India intends to negotiate that deal as a sovereign counterparty, the way it has negotiated semiconductor fabs and battery plants, remains open. The fairest read of the evidence is that the country is being treated, by the AI industry and by global capital, as the natural training ground for the next generation of automation — and that the policy response, if there is one, is still being drafted.
Stakes over the next eighteen months
The horizon that matters is short. By the end of 2027, embodied-AI models trained on Indian first-person footage will be commercially deployable in warehouses, retail floors, and light-industrial settings. The TCS–Anthropic partnership will, in the same window, move from announcement to initial enterprise rollouts. The decision India makes in that window — whether to treat its workforce's data as a national resource, whether to require frontier-lab partners to commit to domestic compute, training, and IP, whether to build a public stake in the model layer the way it has built public stakes in banks, ports, and telecoms — will determine whether the AI transition narrows or widens the country's distribution of gains.
The optimistic case is the historical case: India has, in the past, used its labour force as leverage to capture industrial policy outcomes. The pessimistic case is also the historical case: the leverage worked when the labour itself was the scarce input. Once the labour has been converted into a model, the leverage moves with it. The next eighteen months are the window in which that move is still being made. The workers in the Bengaluru footage are training for it. So is everyone else.
This article treats the TCS–Anthropic partnership and the Bengaluru data-labelling programme as separate points on a single labour-displacement curve. The two source wires do not, themselves, make that connection; the synthesis is Monexus's. The Air India 171 reporting from Reuters and the BBC appears in the same news cycle and is included for context, not for thematic fit.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- http://reut.rs/4omqmMx
- http://reut.rs/4ekq1p2