Millennial Portent

8 July 2026 machine-learning, fiction

Charles nailed it

The shopping center at Noosa civic has a small table that sits just outside the supermarket. On this table, people orphan their old books and adopt others. No money changes hands, there is no registration, it’s the kind of thing that you see more often in rural communities. It’s a welcome distraction from the weekly grocery run and I’ve formed a habit of parking the trolley to browse.

This table is a gold mine. You’ll find old books lacking barcodes hand-bound in 400 gsm. You’ll find weird and wonderful titles like “Jesus lived in India”, filled with scribbled notes to friends documenting the shelves travelled by the book and the countries travelled by its owners.

You’ll also find the odd “Best Science Fiction of” for a random year in a previous decade. These are the kinds of books I borrow on occasion, and when the mood strikes, chance a read through.

One such find had me rifling through the best-of wallet belonging to the year 2000, where I found a short story called “Antibodies” by Charles Stross.1 This was a spooky read, one of the early clarion calls about birthing AI gods, but most of that happens in the background, and it takes a while for the setting to click.

It’s a good read with a good through-line, but there was one line that survived and surprised from 26 years ago, when this was written, that had me incredulous. Every now and then, I suspect a sci-fi author of being a time traveller.

The narrator is newspaper skimming, which is a tidy bit of world-building, and they gloss across “a piece about Indian sweatshop software development being faced by competition from code generators, written to make western programmers more productive”.

Cue raised eyebrows, guffaws, monocles plopping into drinks et cetera.

Yes, this was a good one. Was it so predictable? In 2000 software off-shoring was becoming a thing, it was a boom industry, and I suppose a story about AI would paint a world in which AI could compete with that sort of thing.2 So I’ll just go ahead and say Mr. Stross called it back then.

It made me think. This is a tension I’m seeing in my work life, and it hasn’t fully played out yet. Now, I’d like to dodge the hand-wringing and to avoid those who like to clutch at pearls, so let’s ignore that there are countries involved with different interests and consider the economics.

The headline article

Filling in as best we can what might’ve been the content of the headlined article, I’ll paint the current picture in 2026. If you’re familiar with the off-shoring industry dynamics you can save yourself 2 minutes and skip to the next part.

When “Antibodies” went to print, Indian software exports were worth about $4 billion a year; while the paperback yellowed, the trade grew to something closer to $600 billion.3 The line the narrator skims past now has six million careers riding on it.

The shape of the industry matters more than the size though; junior work is the lowest-hanging fruit for the AI labs, and the off-shore trade has spent decades growing mostly that. The trade runs on a pyramid: graduates hired by the stadiumful, billed out cheaply for the routine work, and pointed at a distant onshore counterpart who owns the relationship.4 The promise is that the survivors climb; the revenue is the bottom of the pyramid, sold by the hour.

Those bottom layers are the ones “code generators” are good at, and you can watch it happening from either end.5 One of the majors cut its graduate intake by three quarters in a single year. The five largest firms spent nine months of fiscal 2026 adding a net 17 employees between them; from stadium to minibus.6 Twelve thousand layoffs landed with the explanation that “this is not because of AI but to address skills for the future”.7 ¿Por qué no los dos?

The money, meanwhile, is going the other way. The same firms that shut the fresher gate are stocking up on compute, and delivering more with the people they kept; the industry’s own review has revenue growing nearly three times faster than headcount, and credits, in as many words, “AI-driven efficiency gains”.8 Field experiments on the client side find the same lopsidedness: hand out code assistants and the juniors speed up by a third while the seniors barely notice.9 The machine is good at the bottom of the hierarchy, which is the bit that’s on sale.

Around 100 years ago something similar happened with the telephone network. Between 1920 and 1940 the switchboards were automated out from under the operators, who were young, cheap, and usually on their first job; in the converted cities, between half and four fifths of those positions evaporated.10 The engineers upstairs kept their careers, and the rung below them stopped existing. This is similar but different, and I don’t think it’s confined to the first floor.

The part that hasn’t played out yet is the second-order effect: the pyramid doubled as the industry’s training pipeline, and the freshers grinding through the boilerplate were the next generation of seniors. The models say you take the output bump now and pay later, when nobody is left who learned the job the hard way,11 and the early trials back them up: juniors who delegate to the model can barely explain the code they just shipped.12 Both ends of the offshoring arrangement are dismantling the ladder they climbed. That’s going to make things interesting in a few years, but we never needed another switchboard operator so perhaps that’s okay.

Nobody in this story is waiting politely to be phased out, so consider the rational moves. The outsourcers’ opening gambit is selling the disruption itself, and the biggest of them now report AI revenue in the billions, though none will say what counts.13 The snag is that the clients got the same memo, so they arrive at renewal expecting the productivity as a discount, and a $100 million deal comes back as an $80 million deal.14 The analysts have already drawn the survivor: a diamond rather than a pyramid, thin on juniors, thick with machine-wranglers through the middle, margins to be announced.15

So what does an offshore centre sell when the code itself has been commodified? Hazarding a guess: the clock. Somebody has to babysit the AI through the midnight of whoever owns it, keeping the agents fed overnight or manning the security desk, and a workday that runs while yours sleeps is the one input that never got cheaper.16 No analyst tracks this as a category yet, so it’s conjecture, but the early sightings fit: a very large deal in which the vendor’s humans mind the client’s machines, and a security-operations trade that has always sold the night shift.17

The old edge, “we have great software engineers”, stops differentiating the moment the marginal engineer is a model, and the repricing is visiting every trade that ships work down a wire: freelance writing postings fell by a fifth within months of the chatbots arriving, the crowd end of data-labelling collapsed once its workers were caught delegating to the machine, and translators now spend their days marking the machine’s homework.18 The analysts call it the move from labour arbitrage to compute arbitrage, which is a polite way of saying the work is being off-shored again, to the cloud this time.19

Context is king

Which brings the catch-22: inside the model’s capability frontier everyone performs like a senior, while outside it the model will walk you off a cliff and compliment your hiking form on the way down.20 Directing the machine is therefore a judgment problem, and the judgment is not a coding skill. Engineers better than Claude still exist in niche pockets, but peel back the layers and the differentiator is mostly context, and even context is a wasting asset, because the tools bottle the seniors’ know-how which gets copied like the company owns it (which it does).21 The unbottled remainder, what the business needs this quarter and why, tends to live onshore, in organisations that haven’t spent decades moving it.

Tends to, because in some organisations offshore is entrenched: the mature captives own products and compliance end to end after twenty years, and their context genuinely lives abroad.22 So the question in the boardroom shifts from “should we off-shore” to “which shore”; the cloud is on the shortlist now, and where your context lives does much of the deciding. My suspicion is that a large part of the industry is about to be off-shored to the cloud, and that software is only the first cargo.

Now, where does it stop? In Charles’s future, all this was happening at the same time as the deus ex machina23 arrives, so we’re given a “what next” that amounts to civilisational sunset with a shadowy happy-ever-after. Industry never got to play it out.

Assuming for a minute that we’re not on a similar timeline to the story, cloud-shoring doesn’t stop with re-shoring the off-shore. Context first gets bottled, then manufactured at a rate no human can keep up with. Each model release reaches fruit that hung a little higher; the ranks thin from the bottom up. Executives are left with the rational economic decision of either preserving human accountability for legal reasons or just paying the risk-adjusted insurance premium and trusting Babbage’s vision of a perfect, machine-run organisation.24

We land in the factory of the future being the machines, the man and the dog.25 At least the dog has a job, we can smile at that.

Another 26 years?

I’m thinking then, that my response to seeing this millennial portent take form, is probably to not rely so solidly on a job in the future. Speaking of good fiction, Tchaikovsky’s “Service Model”26 is a good read. It spells out a future where jobs cease to exist, and those that own things become the overclass, served by murderous robot butlers. It’ll be interesting to see how the picture it paints holds up in another 26 years if we’re still around. If I land on the right side of history then I’ll be able to ask my murderous robot butler to pop down and fetch me a copy.

Footnotes

  1. C. Stross, “Antibodies,” Interzone, no. 157, July 2000. Reprinted in G. Dozois, Ed., The Year’s Best Science Fiction: Eighteenth Annual Collection. New York: St. Martin’s Griffin, 2001. The story was shortlisted for the 2001 Theodore Sturgeon Memorial Award. ↩

  2. I’ll concede there’s a bit of a bait-and-switch here, “code generators” aren’t called out as AI specifically; it’s more a silhouette of the present day than a portrait of Claude, but the shadow cast is recognisable in function and form. ↩

  3. The fiscal 2000 figure is NASSCOM’s, as reported in C. P. Chandrasekhar, “ICT in a Developing Country Context: An Indian Case Study,” UNDP Human Development Report Office background paper, 2001. Current figures are from NASSCOM, Technology Sector in India: Strategic Review 2026, Feb. 2026: $315 billion total revenue, $246 billion of it exported, 5.95 million direct employees. The global market size is the softest number here; Statista puts 2025 IT outsourcing near $590 billion and Precedence Research near $660 billion, and the spread is definitional rather than a disagreement about facts. ↩

  4. The freshers come from engineering colleges in annual campus rounds and are billed out at marketplace rates of roughly $18 to $40 an hour for testing, boilerplate, maintenance, and support, against $40 to $50 for seniors and $50 to $70 for architects; something like three quarters of revenue is attributed to these junior-heavy layers. The bench, salaried engineers between assignments, has shrunk from a customary 80 or 90 days to 30 or 40, which is an indicator in its own right; see “TCS layoffs signal collapse of IT sector’s traditional talent pyramid,” Outlook Business, 2025. ↩

  5. On the western side, payroll microdata from ADP puts employment of developers aged 22 to 25 down nearly 20% from its late-2022 peak while the over-30 cohorts hold flat; E. Brynjolfsson, B. Chandar, and R. Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, Nov. 2025. The authors are explicit that they do not claim the whole effect for AI. The U.S. Bureau of Labor Statistics reads the same split into its 2024-34 projections, with software developers (SOC 15-1252, a heading that also covers QA analysts and testers) growing 15% while computer programmers (SOC 15-1251), the occupation code for coding to someone else’s spec, shrink 6%. The caveat: the junior collapse dates from late 2022, which is also when the cheap money ended, the pandemic over-hiring corrected, and a US tax change (Section 174, amended from 2022, requiring development salaries to be amortised over five years) made developers dearer to book; SignalFire’s talent reports name the end of cheap money as a co-cause and decline to apportion blame. Brynjolfsson et al. offer three replies: AI-exposed occupations are on average less sensitive to interest rates, the declines survive excluding the tech sector, and with firm-time fixed effects the significant declines date from 2024, after the correction. ↩

  6. Infosys hired more than 50,000 graduates in fiscal 2023 and 11,900 in fiscal 2024, per company filings. The net-17 figure is from “AI May Reset Indian IT’s Pyramid Model, But Not Fresher Hiring,” Analytics India Magazine, 2026, which tallies the five largest firms over the first three quarters of fiscal 2026 against 17,764 for the same period the year before. Fresher targets ticked back up after the fiscal 2024 trough, and interest rates do reach Mysore through every American procurement budget, so the level of any one year’s intake is a noisy, partly macro signal. The steadier signal is composition: senior hiring continues while the fresher gate stays shut. ↩

  7. K. Krithivasan, chief executive of TCS, quoted in “TCS announces 12,000 layoffs by FY26 in strategic workforce realignment,” People Matters, Jul. 2025. The cuts amount to about 2% of a workforce of roughly 600,000 and fell on the mid and senior levels. ↩

  8. NASSCOM, Strategic Review 2026: revenue up 6.1% in fiscal 2026 against headcount up 2.3%, with AI services counted at $10 to $12 billion. The industry still added about 135,000 people in the year, but the growth increasingly sits in captive Global Capability Centres, where multinationals internalise the offshore function and cut out the vendor, rather than in the outsourcers themselves. Of the AI revenue, TCS discloses an annualised run rate of $1.8 billion, Infosys $275 million, and HCLTech $146 million. ↩

  9. Z. Cui, M. Demirer, S. Jaffe, L. Musolff, S. Peng, and T. Salz, “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers,” SSRN working paper 4945566, 2025. 4,867 developers across Microsoft, Accenture, and a Fortune 100 firm, output measured as completed pull requests, a 26% average lift split as 27% to 39% for juniors against 8% to 13% for seniors; the authors had no access to the code itself, so quality goes unmeasured. The counterweight is J. Becker, N. Rush, E. Barnes, and D. Rein, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” arXiv:2507.09089, 2025, in which sixteen experienced developers working in large, mature codebases were 19% slower with the tools while believing themselves 20% faster; the help runs out about where the seniority begins. ↩

  10. AT&T mechanised more than half the American network over the period, and operator employment for the incoming cohorts fell 50 to 80% in the converted cities. J. Feigenbaum and D. P. Gross, “Answering the Call of Automation: How the Labor Market Adjusted to Mechanizing Telephone Operation,” The Quarterly Journal of Economics, vol. 139, no. 3, pp. 1879-1939, 2024. The incumbents bore the damage: a decade later they were likelier to be in lower-paid work or out of the workforce, while the cohorts who would have taken the job found other first jobs at somewhat lower wages. The operators were domestic rather than offshore, so the geography does not transfer; the mechanism, automation arriving at the cheap entry tier and removing it without much disturbing the tiers above, is the part that does. ↩

  11. E. Ide, “Automation, AI, and the Intergenerational Transmission of Knowledge,” CEPR Discussion Paper DP20940, 2025. An overlapping-generations model in which novices acquire tacit knowledge by working beside the most productive experts. Automating entry-level work raises output immediately and can lower long-run growth and welfare even when novice employment holds up, because it changes who the novices learn beside. ↩

  12. “How AI Assistance Impacts the Formation of Coding Skills,” Anthropic, Jan. 2026. Fifty-two junior engineers, an unfamiliar asynchronous Python library, a comprehension quiz afterwards scoring 50% for the AI-assisted group against 67% without, nearly two letter grades; no significant difference in speed. Interaction style carried the effect: engineers who used the model to ask questions retained the material, while those who delegated wholesale did not. ↩

  13. The largest of them discloses an AI revenue run rate that grew from $1.5 billion to $2.3 billion annualised across fiscal 2026, about 6% of its total, with peers reporting $620 million and $275 million; none of the three publishes classification criteria. TCS Q3 FY26 results, Jan. 2026; “TCS reports $2.3 billion AI revenue as growth stabilises in FY26,” Forbes India, Apr. 2026. In December 2025 Microsoft signed Copilot deployments of 50,000+ licences each with TCS, Infosys, Wipro, and Cognizant, which is the pivot in a single image: the majors as deployment channels rather than labour suppliers. ↩

  14. The $100 million line is HCLTech chief executive C. Vijayakumar’s, who flags a deflationary impact of 3 to 5% on fiscal 2027 guidance; P. Ganguly, “Indian IT braces for AI-deflation as pricing pressure reshapes growth,” Forbes India, May 11, 2026. ISG measures contracted price-performance improvements at renewal running at double the historical average, while hedging on how much is AI; S. Jones and M. Rose, “AI Isn’t the Only Reason Providers Are Feeling Pricing Pressure,” ISG Index Insider, Oct. 31, 2025. Wipro booked more deals in fiscal 2026 for flat revenue, and Motilal Oswal models 12 to 15% of sector revenue at direct displacement risk, against roughly $4 billion of disclosed AI revenue across the top five. ↩

  15. HFS Research, Horizons: Agentic Services, Apr. 2026, and A. Venkatesan, S. Gupta, and N. Jhunjhunwala, “Services are starting to scale beyond headcount, signaling an early pivot to SaS,” HFS Research, Jan. 14, 2026. Revenue per employee has begun decoupling from headcount, with roughly one orchestration role created for every five automated and a margin trough in between; outcome-based pricing remains unsolved, with practitioners calling 2026 a pilot year. ↩

  16. The nearest established trade is managed security, where 24/7 coverage has always been the pitch; India’s managed security services market sits near $1.75 billion and growing. HFS calls the emerging arrangement “human-on-the-loop”, agents executing while humans handle the judgment calls, in the same Agentic Services report. ↩

  17. HCLTech’s $1.14 billion, five-and-a-half-year managed infrastructure deal with Mercedes-Benz, signed 2026, has AI handling routine tickets, anomaly detection, and first-line support, with the vendor’s people managing the AI layer. ↩

  18. O. Demirci, J. Hannane, and X. Zhu, “Who Is AI Replacing? The Impact of Generative AI on Online Freelancing Platforms,” Management Science, Jan. 2025: automation-prone writing and coding posts down 21% within eight months of ChatGPT, image work down 17% after the image models. X. Hui, O. Reshef, and L. Zhou, “The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market,” Organization Science, vol. 35, no. 6, pp. 1977-1989, 2024, adds the sharp detail that the more experienced freelancers were hit harder. Amazon closed Mechanical Turk to new customers in July 2026, after studies found a third to a half of its workers already delegating tasks to LLMs; the annotation work that survives is expert-tier. Slator’s 2025 industry data has LLMs among 89% of top benchmark performers in translation, with working linguists reporting most of their time now goes to review. ↩

  19. P. Bendor-Samuel of Everest Group, “From Labor Arbitrage to AI: The Next Big Disruption in Tech Services,” Forbes, Aug. 13, 2025, puts AI efficiency gains at 30 to 40% against offshore’s customary 20 to 25%. HFS Research declared the labour-arbitrage model past its shelf life in April 2026. ↩

  20. F. Dell’Acqua et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality,” Organization Science, Mar. 2026. 758 consultants randomised on GPT-4: inside the frontier the lowest-skilled gained 43% on quality against 17% for the strongest, while outside it AI users were 19 percentage points more likely to land on the wrong recommendation. Neither juniors nor seniors reliably located the frontier, and a follow-up found juniors fresh off the tool giving systematically wrong advice about what it could do. ↩

  21. E. Brynjolfsson, D. Li, and L. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, vol. 140, no. 2, pp. 889-942, 2025. 5,172 support agents at one firm: novices gained 34% while experts gained roughly nothing, because the assistant had encoded the experts’ tacit knowledge; agents with two months’ experience began performing like agents with six. The authors flag the crowding-out risk, that experts who defer to the tool stop generating the know-how future versions would bottle. ↩

  22. NASSCOM and Zinnov count about 1,760 Global Capability Centres in India at $64.6 billion revenue and 1.9 million employees, with 89% of the largest operating at what Zinnov calls transformation maturity, owning global product and compliance mandates end to end, and global leadership roles based in India compounding at roughly 40% a year. ↩

  23. In both senses of the word. ↩

  24. C. Babbage, On the Economy of Machinery and Manufactures. London: Charles Knight, 1832. The book treats the factory as a machine to be analysed and perfected, extends the division of labour to mental work, and pairs with the calculating engines Babbage built to squeeze human error out of computation. In fairness to Babbage, his factory kept its humans; they were components to be costed rather than removed, which is close enough for the executives in question. ↩

  25. “The factory of the future will have only two employees, a man and a dog. The man will be there to feed the dog. The dog will be there to keep the man from touching the equipment.” Commonly attributed to Warren Bennis, who used it in On Becoming a Leader (1989) prefaced with “someone said”; Quote Investigator traces the earliest print appearance to Datamation, Nov. 1978, as a joke circulating in the British Post Office Engineering Union, and leaves the author anonymous. ↩

  26. A. Tchaikovsky, Service Model. New York: Tor Books, 2024. ↩

@misc{hollows2026millenni,
  author = {Hollows, Peter},
  title  = {{Millennial Portent}},
  year   = {2026},
  month  = jul,
  url    = {https://dojo7.com/2026/07/08/millennial-portent/}
}