It's Not the Job That's Dying. It's the Discount.

For two years now we have been arguing about whether AI kills jobs, and the data keeps refusing to cooperate. Prime-age employment in America sits near all-time highs. Noah Smith's Labor Day post laid out the case with characteristic thoroughness: the job-killer everyone in San Francisco believes in "keeps stubbornly refusing to kill jobs," software developers are near their highest share of the US labor force ever, translators are still employed, and firms that adopt AI tend to hire rather than replace.

I believe that argument as far as it goes. But this autumn I keep noticing the thing actually being repriced, and it is not a job title. It is an address.

But this autumn I keep noticing the thing actually being repriced, and it is not a job title. It is an address.

Start with what is already on paper. In August, Business Standard reported from Bengaluru, citing people familiar with the matter, that Visa had laid off about 1,400 employees at its technology centre there, nearly 40 per cent of its India workforce, in its first major retrenchment since the centre opened in 2015. About 500 engineers and 900 non-engineers. The roles were mostly in product and technology teams, and the cuts reached mid-level staff, including product managers. The paper's framing: the company "looks to embrace artificial intelligence (AI) to improve productivity and efficiency."

The detail I keep turning over is the distribution. Visa plans to cut 2,600 jobs worldwide. India holds roughly a tenth of the company's people, about 3,500 of some 34,000, and the paper reports that the losses there "account for more than half of the planned global layoffs." Restructuring 101 says cut where the payroll is fattest. These cuts landed disproportionately where the discount was. And the neighbours are doing it too: Mastercard cut 1,400 jobs, PayPal 4,800, Block 4,000, Intuit 3,000, in what the paper describes as financial technology firms globally "reducing their workforces rapidly as AI-driven efficiency gathers pace in 2026."

Now look at a single day: October 8. The Trump administration suspended Microsoft, Adobe and six other companies, including several of India's biggest IT services firms, from the program that lets H-1B workers apply for green cards, over alleged visa abuse. Vice President Vance singled out Microsoft as the biggest abuser of the program among American firms and explained his language carefully. H-1B holders are like "indentured servants," he said, because "when somebody comes in on an H-1B, they are completely subordinate to that company. If they lose that job, they have to leave the country."

Hours after that speech, President Trump gave Satya Nadella the National Medal of Technology and Innovation. Decorate and squeeze, in one afternoon. I don't read that as incoherence. I read it as a bargain being struck in public: the American state is all-in on the AI buildout, and is charging the industry for the privilege in labor pipelines. Vance said as much in his own way. The message to Microsoft was not to harm the company, he said, but to see it thrive "by employing and empowering your fellow Americans."

And on that same October 8, the world's largest IT services employer reported its quarter.

The coincidence of these headlines is not my point. My point is that they are one story, and that the machine deciding it is not the one the argument is watching.

So here is my claim, with its test up front, because a busy week is a coincidence, not a proof. The arbiter of who does tech work is no longer wages, and no longer visas. It is the cost of supervision. The offshoring bargain ran on one calculation: specifying work, checking work and catching mistakes was expensive, so it paid to move the doing to where the doers were cheap, and to pay managers to supervise across the gap. When AI makes specifying and checking cheap, the cheap-doer advantage shrinks and the address gets repriced.

Watch what happens next, because the signature is specific. Work changing employers is ordinary outsourcing and happens every quarter; that alone proves nothing, and I won't pretend otherwise. What would be specific is contracts repriced around what the machines output, with supervision itself sold as a product. And what would sink me: the displaced work showing up in American headcounts, or the captive centers simply rehiring the same roles at the same prices.

Now listen to that employer talk for an hour.

TCS, the Indian services giant that employs nearly 600,000 people, held its earnings call on October 8. In an interview with CNBC-TV18, chief executive K. Krithivasan was asked whether roughly half of long-running IT contracts had already been through what the industry calls AI-led deflation. That is "more or less the right way to look at it," he said, and "the estimate of about 50% is approximately right." The average price cut per project, he said, is "probably in the range of 10% to 20%." When does the new AI work outweigh the shrinking of the old? He would not draw a line. "The traditional business will see some shrinkage, while the new technology or service line will grow much faster," he said. "That is a natural process."

That is the first half of the signature: contracts repricing themselves around what the machines can do. Here is the second half. Among the launches TCS announced on the call: "domain-specific agentic control planes." Its chief operating officer, Aarthi Subramanian, described what enterprises need as agents scale: "orchestration, access controls, security, observability, cost controls, and policy enforcement," plus management of AI token costs. Her promise to customers: "to deliver the right level of intelligence at the right cost."

Read that back as a job description. Orchestrating work, controlling access, watching what subordinates are doing, enforcing policy, and keeping the cost of intelligence at the right level is a fair description of what managers do. TCS has taken supervision, which used to be headcount inside the client, and turned it into a product line. That is my mechanism, wearing a name on a slide.

TCS has taken supervision, which used to be headcount inside the client, and turned it into a product line. That is my mechanism, wearing a name on a slide.

So where does the displaced work go? Not into American hiring, so far. But do not count that as my proof; I already said why. What is worth watching is what happens to the desks themselves. Best Buy's captive center in India is being handed to TCS and transformed, in the companies' own words, into an "AI Capability Center." Krithivasan calls deals like that "a new category of transformation partnerships," and says TCS and its clients are "building repeatable value platforms that will industrialize AI at scale." Meanwhile TCS hired about 4,500 people in the quarter and its workforce passed 598,000. The desks in Bengaluru are not emptying. The name on the badge is changing, and so is the price. An address does not have to die to be repriced.

My favorite detail of the month: on that earnings call, an analyst asked the CFO whether rising subcontracting costs were "a proxy for visa costs," noting the company had filed fewer visas since "the sharp jump in visa fee last year." Think about where the two great stories of the month just met. On the call where an industry is repricing its contracts around machine output, an analyst is asking whether the delivery model's costs are rising because of immigration policy. The CFO's answer was a careful piece of fence-sitting: "Some part of it could be linked to that also, but not completely."

Why do the aggregates miss all of this? Because the aggregate is the wrong instrument. American employment statistics count American jobs. The Census surveys showing firms hiring count American firms. Even the strongest version of the optimistic case carries its own caveat: it measures numbers of companies, not numbers of jobs. The thing I am describing lives exactly where those instruments cannot see: headcount at an Indian subsidiary, contract structure, the difference between a captive center and a vendor's. The chart everyone cites, software developers as a share of the US labor force, is near an all-time high and steeply rising. It is true, and it is looking at the wrong continent. "AI refuses to kill jobs" and "the offshore staffing model is visibly shrinking" are not contradictions. They are two needles on two different gauges.

Now the honest part, because the boring explanations deserve better than a wave of the hand. The visa crackdown is protectionism with a Microsoft edge; a plain political story needs no AI at all, and the same afternoon's medal ceremony shows how political it is. TCS's own demand language is "selective and uneven," and management blamed a drop in India revenue on a single project deferral, so some of this is an ordinary slowdown in a flattering costume. Krithivasan himself noted that when customers hunt for savings, "it can lead to vendor consolidation." Sometimes the story is just firms merging suppliers and cutting heads, because firms always do. And the Visa numbers come from unnamed sources; the AI gloss is the paper's, while the company memo speaks only of "efficiency" and reinvesting in "highest-potential opportunities." All of that is live.

The way to tell these apart is the one test I named: follow the work. Cost-cutting and protectionism shuffle work between spreadsheets. The collapse of a supervision arbitrage moves it between business models, and what moves in the TCS numbers is pricing and product, not just headcount. One caution flag worth noting: TCS shares are down more than 29% over the past year even as its AI revenue crosses a tenth of its business. The market has not priced this with any confidence either.

The part I keep turning over is who bears the transition. Vance is not wrong that the H-1B holder is structurally subordinate; the tie to one employer is real. But India's Foreign Ministry called his comments "unwarranted" and "deeply offensive," and given the colonial history of indenture, you can see why. Roughly 70 percent of the H-1B visas approved in 2025 went to people from India. The exit route to America is also what made the Bengaluru bargain bearable, and it is closing at the exact moment the work itself is being repriced by software. Neither move was decided by the engineer at the center of it. She appears in neither country's headlines.

If you want one filter for the next year of AI-and-jobs coverage, take this one: stop watching occupations and start watching arbitrage. Watch employers and contract structures instead of layoff counts. Watch who owns the supervision. When checking is expensive, doers are cheap and managers cost a fortune. AI is inverting that ratio, and that is what is happening to the discount that built an industry. Not the job. The discount.