Everyone chose the same move. Now nobody wins.

Every summer for fifteen years, the same four months decided the same trade. May through August is when nearly four in five of the F-1 student visas issued to Indian applicants get stamped, the window that determines who boards a plane toward a US graduate degree, an internship, an H-1B lottery ticket, and eventually, the theory went, a six-figure salary and a green card. In the same four months of 2024, the United States issued 58,694 F-1 visas to Indian nationals. In 2025, it issued 22,149. Two out of every three people who would have made that trip a year earlier did not go.

Nothing about the trade itself had gotten worse in the classroom. The universities were the same. The professors were the same. What changed is that everyone else made the identical bet at the same time, a foreign master's degree became the default move for a generation of Indian STEM graduates, American visa policy tightened underneath all of them at once, and a $100,000 fee landed on the one visa category the whole strategy depended on. A trade that pays well when a few thousand people run it pays nothing when half a million people are running it into a wall built specifically to stop them.

This is not a story about one visa policy. It is a story about a pattern repeating at every altitude of the economy at once: individuals converging on the same degree, startups converging on the same foundation model, corporations converging on the same AI vendor, and nations converging on the same GPUs. Game theory has a precise, unforgiving name for what happens next.

In 1883, the French economist Joseph Bertrand described what happens when two firms sell an identical product and compete only on price: both undercut each other until price equals marginal cost and neither earns a profit. It is the oldest, cleanest result in competitive theory, and 2023 through 2026 has been the fastest live demonstration of it economics has ever produced. The price to buy AI performance equivalent to GPT-3.5 fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a 280-fold collapse in under two years. Depending on the exact task, inference prices at fixed performance have fallen anywhere from 9x to 900x per year. On the hardest benchmark Epoch AI tracks, matching GPT-4's score on PhD-level science questions got 40 times cheaper every twelve months.

The reason is not mysterious. Every major lab converged on the same recipe, transformer architecture, scaled compute, similar training data, and the performance gap between competitors collapsed with it. Benchmark scores that differed by double digits in 2023 sat within fractions of a point of each other by 2024. When your product is indistinguishable from your rival's, Bertrand's century-old result takes over, and it does not care whose logo is on the API.

Follow the compression down the stack and it does not stop at the API. In Indian IT services, the industry that sells implementation of exactly this converged technology to the rest of the world, HCLTech's own chief executive has put a number on what AI does to a renewal contract: a deal that used to be worth $100 million now prices at roughly $80 million, for 25 to 30 percent more delivery effort. On the company's own earnings call he has committed specific productivity numbers to clients at renewal: 35 to 40 percent in software development, 25 to 40 percent in BPO operations, delivered back as lower prices, not higher margins. Brokerage estimates put 12 to 15 percent of the whole sector's revenue at direct risk from the same dynamic.

"We intend to innovate faster than the market to stay ahead of the deflationary curve." C. VIJAYAKUMAR, CEO, HCLTECH · Q1 FY27 EARNINGS CALL, JULY 2026

And the public markets have already priced the lesson in, sharper for whoever grew fastest on the undifferentiated layer. In the six months to February 2026, valuation multiples for horizontal software (the AI-feature-of-the-month category) fell 36 percent, and the decline correlated tightly with how fast a company had been growing. Vertical software, the harder-to-copy, workflow-specific layer, fell only slightly more (39 percent) but with almost no relationship to growth rate at all. Investors, in other words, are no longer paying up for velocity on a commodity. They are paying for whatever cannot be copied.

None of this means the model layer itself is a free-for-all. Economists studying the foundation-model market have found the opposite pressure running underneath the price war: training frontier models requires such extreme economies of scale in compute, data, and talent that the market for building the models themselves is drifting toward an oligopoly of perhaps five labs, even as those five labs cut each other's throats on price. The convergence, in other words, happens on two floors of the same building at once. Whoever owns the floor with the moat keeps the elevator. Everyone building on the floor above it is riding Bertrand's curve straight to the basement.

The US corridor did not close gradually. It fell off a cliff in a single admissions season. Against the 58,694 F-1 visas issued to Indian applicants in the peak May-to-August window of 2024, the same window in 2025 produced 22,149, a 62 percent collapse, and 60 percent below the average of every comparable admissions season back to 2017. The stock of Indian students already enrolled fell too, from 378,787 in February 2025 to 352,644 a year later, the first sustained decline this pipeline has ever recorded. And the policy built to squeeze the exit door shut, a $100,000 fee attached to new H-1B petitions from September 2025, has had a legal history as chaotic as the enrollment numbers: a federal court in Washington upheld it in December 2025, a federal court in Massachusetts vacated the entire proclamation in June 2026, and as of this writing the fee is not being collected while the government appeals. The trade's terminal payoff, the H-1B itself, now carries a coin flip on whether its price just became six figures.

Canada, the second-largest destination and the one marketed hardest as the safer alternative, closed even faster. Study permit approvals for Indian applicants fell from roughly 100,000 in the first half of 2024 to 48,065 in the same period of 2025, a 52 percent drop. The approval rate for Indian applicants collapsed from 69 percent in 2024 to 25 to 27 percent in 2025, and India's share of Canada's entire international-student intake fell from 51.6 percent in 2023 to roughly 8 percent by September 2025. The federal cap behind this is explicit and multi-year: Ottawa's 2026 issuance target of 408,000 permits is 16 percent below the 2024 target, tied to a public goal of pushing temporary residents below 5 percent of Canada's population by the end of 2027. The total stock of study permit holders in Canada fell from over a million in January 2024 to about 725,000 by September 2025, a decline of roughly 28 percent in twenty months.

What did not stop was the money. Indian NBFC education-loan books grew 21 percent in FY26 to roughly ₹78,000 crore, and Crisil projects another 20 percent to ₹94,000 crore in FY27, even as US-linked disbursements crashed 57 percent over the same year. The book is not shrinking; it is reallocating, the US share of outstanding loans fell from 54 percent to 43 percent while the UK's rose from 20 percent to 29 percent, with Germany and Ireland gaining too. That reallocation is presented as resilience. Read differently, it means Indian households are still borrowing at 20 percent annual growth to run the same trade in whichever corridor is still open this year, without much evidence the return on that trade has been recalculated. And the loan book's own numbers hide the real risk: 73 percent of it still sits inside its original repayment moratorium, meaning most of these loans have never yet faced a single EMI. The 0.2 percent delinquency rate everyone quotes is a number about loans nobody has had to repay yet.

A market that has not been repayment-tested is not a safe market. It is an untested one, and the test is coming due on a three- to five-year clock nobody has started counting.

The clearest single number in American labour economics is the college wage premium, and it has stopped moving. It roughly doubled from 39 percent in 1980 to 79 percent in 2000, then plateaued for two full decades; by 2023 it sat slightly below its 2000 level. Over the same period the sticker price of a four-year degree, tuition plus room and board, rose 40 percent in real terms. Researchers at the San Francisco Fed are careful to note this stagnation predates the recent AI wave by twenty years, so it is not an AI story alone. But it is the backdrop AI adoption landed on, and the backdrop is why the newest graduates are absorbing the impact first: recent US college graduates now run an unemployment rate of roughly 5.6 percent against 4.2 percent for the workforce overall, a reversal of the historical pattern in which a degree meant lower unemployment, not higher.

The same compression shows up on freelance platforms, with the same fingerprint every time: the moment a skill becomes commodity-replaceable, both its volume and its price fall together. On Upwork, freelance writing postings fell 33 percent and translation postings fell 19 percent in the fifteen months after ChatGPT's launch; translation hourly rates fell more than 20 percent, the steepest decline of any category on the platform. Peer-reviewed research covering three million postings on a separate global platform found the same pattern at a larger scale: demand for the most easily substitutable skills fell 20 to 50 percent relative to trend. What grew, sharply, were the categories a generic model cannot commoditise: video editing volume rose 39 percent, and demand for chatbot-development work itself rose roughly twentyfold. Even inside the categories that grew, the same study found something the entry-level dossier in this series has now documented twice: the expansion favoured experienced freelancers, and demand for novices fell.

Zoom out to the firm level and the same divide reappears at industrial scale. Revenue per employee at Anthropic sits near $9 million and at OpenAI near $5.6 million, several times higher than Nvidia, Meta, or Google, because both companies own a layer few others can replicate. A typical application built on top of someone else's model reports 45 percent gross margins, well below the 70 to 90 percent a software company used to consider normal, because the inference bill for renting someone else's intelligence is now baked into the cost of goods sold. Ownership of the differentiated layer compounds. Renting it, at any altitude, from a country's education system to a startup's tech stack, gets you the identical payoff curve: falling, and converging on everyone else's falling curve.

The formal result is called Hotelling differentiation, and it says precisely this: two sellers of an identical good compete their profits to zero, but the instant either one makes their product meaningfully different, positive profit reappears, and it grows with the degree of difference. Vertical AI companies that own a proprietary data flywheel or deep workflow integration are living this out in real time, sustaining roughly 60 percent gross margins while wrapper products built on the same rented model watch API costs eat 40 to 70 percent of revenue and churn out 65 percent of their customers within ninety days. Even the bullish investors writing about this concede the moat is not permanent, foundation models keep improving and could still erode any given edge, but for now the split between the two groups is stark and it is widening, not narrowing.

The nation-level version of the same escape has a name too: Taiwan. Facing the same commodity trap in 1970s semiconductors that everyone faces today in AI, the state deliberately declined to compete head-on at the frontier, licensing legacy technology instead and building a pure manufacturing specialty nobody else had bothered to own. Half a century later Taiwan's foundries hold roughly 60 percent of the global market. India's own equivalent white space is sitting in plain sight and mostly unbuilt: an Indian-language AI query costs roughly five times an equivalent English one to run, purely because Hindi and other Indic scripts tokenize inefficiently, which means the market has not yet been commoditised by the same global players competing everyone else's margins to zero. India's conversational AI market, still small at roughly $650 million, is projected toward $5.9 billion by 2034. That gap, not another CS seat, not another master's abroad, not another wrapper on someone else's model, is the one move on the board that a hundred thousand people converging on it does not automatically ruin.

The lesson of every act in this dossier is the same lesson, at every altitude: the crowd is not wrong to want the trade. The crowd is the reason the trade stopped paying.