Workers and startups are helping train AI to replace them

Data annotation work is increasingly moving up the value chain, from tagging and labelling data to replicating the work of semi-skilled (on the factory floor) and skilled (consultants, analysts, lawyers, engineers, and doctors) workers. 

Startups sell data to AI labs, which use it to train and refine their AI algorithms and develop software products/solutions that replicate the work of these workers. In other words, the semi-skilled and skilled workers, or at least some among them, are feeding their time and skills into the AI algorithms that seek to replace them and their kind. Both the training startups and those workers offering their services to them are basically helping make themselves redundant. 

On this, the FT has a very good film about how Indian startups are paying factory floor workers and gig workers (and even people in their homes doing regular household chores) to use cameras and record their work. Data annotation is becoming the new BPO for India’s IT industry. 

The Ken has an article that raises the possibility that for all the attention and hype around robotics, Indian startups might remain stuck at the lowest end of the robotics value chain - data collection. 

India was the back office for the IT boom. It became the annotation and reinforcement-learning labour pool for the generative AI boom. It is now emerging as the behavioural data factory for the physical AI boom... Building the robot is only half the problem. Building the intelligence behind it is much harder. That requires data. Vast amounts of it. Unlike large language models, which were trained on the equivalent of hundreds of years of human reading scraped from the internet, robotics companies are working with barely a fraction of that in video... What they need is meticulous, first-person recordings of humans interacting with the physical world, carefully collected, annotated, and painstakingly structured. 

So the industry turned to India. Across the country, workers are recording themselves doing everyday chores for data-collection firms, which then sell that footage to companies such as Tesla, Figure AI, and Agility Robotics to train their humanoids. Indian startups see this as a moment to claim a seat in the global AI value chain. The country has over 260 robotics startups, and investors are beginning to pay attention... The footage being recorded by Indian workers becomes proprietary once it leaves the country. The datasets assembled from it are accumulating on foreign servers. The foundation models trained on them are owned by foreign companies.

The NYT has an article about how startups like Handshake, Mercor, and Surge in the US are paying skilled workers to collect data on their work. 

Mercor and a handful of similar start-ups are the primary middlemen in a supply chain of “human data” that may power the next generation of A.I. As OpenAI, Anthropic and other major ventures compete to become the industry’s dominant platform, the market for premium data that has been vetted by experts is exploding…They need mathematicians to annotate proofs, lawyers to mark up briefs and professors to grade essays… To use the parlance of the industry, data labeling has moved up the “value chain,” and the start-ups that offer this service have become some of the fastest growing in Silicon Valley… The data-training start-ups see a lucrative opportunity in recreating workplaces in miniature: controlled environments in which their gig workers can evaluate and reproduce emails, memos and slide presentations in context. The information emerging from such a setup, the companies boast, will help shrink the gap between what A.I. models can accomplish and what office workers actually do from one minute to the next, as ideas and instructions flow between meetings, documents and applications…

To keep improving their models — to make them more useful, more sophisticated, less prone to hallucination and mistakes — A.I. companies heavily refine what goes into them. That’s post-training, and it includes buying data from vendors like Handshake and its competitors… Deeptune, a start-up that makes “training environments” with simulations of the software programs, like Slack and Salesforce, that many workers toggle between all day long to get their work done. The idea is to painstakingly create a mirror image of, say, an investment bank so that A.I. can observe every interaction…

It may turn out that once OpenAI, Anthropic and others have taught their models to perform a certain job, their need for more training data in that area could sharply decline. In this way, Mercor, Scale, Handshake and their peers are much like the elite freelancers they employ: making money today, but in danger of being dropped tomorrow… People sign up for data-training gigs for a variety of reasons. The main one is, of course, money… Though the labor is unpredictable and rates vary… the workers who cobble together enough shifts can generate meaningful income. People might sign up because they have been laid off, or because they can’t find enough work in their field. They might do it because they’re eager to get “A.I.” on their résumé, or because they need extra cash in retirement… Many people who contract for these companies understand that this is a short-term opportunity, a brief chance to train the models to automate jobs before they themselves are automated out of the job of training models.

I asked Claude to generate a visualisation of this market landscape, including an assessment of the Indian landscape. The numbers are clearly estimates and must be validated (though at a ballpark they appear alright). 

The unit economics of the data chain shown below for a garment worker in India is instructive. She gets roughly ₹400 a day to wear the camera (or $0.60 per hour); the startup pays the factory ₹450–500 per hour; US-based startups like Human Archive price data at $1–10 per hour; and once annotated and packaged, it sells to global robotics labs at $15–50 per hour. That is a 25–85 times markup, and every rung above the worker is owned outside India.

The graphic also shows that the vast majority of AI workers are doing the BPO equivalent, whereas the vast majority of funding is going to those building the data centres. India has 170-odd AI startups that have raised $2.6 billion in total and over 260 robotics startups, but the genuine model/product builders are a tiny set, and the majority have rebranded annotation as an AI line of business (iMerit, Objectways, Awign, Karya, Deccan AI, Human Archive, Egolab, Neo Cambrian, Humyn Labs, RoBoEra, etc.). 

It must also be highlighted that in the majority of cases in India, the data goes from the garment worker to an Indian data aggregator to a robot-brain lab in San Francisco, and comes back as a robot/humanoid. The frontier LLM labs are not in that loop. This also means that none of the emerging governance conversations about frontier models - safety frameworks, export controls, model-access negotiations - touches the mainstream data collection work being done in India. India is negotiating hard for access to frontier language models while simultaneously handing over, for ₹400 a day, the training substrate for the physical models that will actually displace its manufacturing workforce. Those are two different conversations, and only one of them is being had.

Further, as the Times article highlights, while these annotation startups are flourishing now, they may not be sustainable ventures. Once experts teach the models to do something, their services are no longer needed in the same way, and the vendors themselves need the models to keep improving to show they add value, while needing them to remain imperfect so clients keep coming back. 

I asked Claude for historical precedents and got this:

Frederick Winslow Taylor’s explicit programme, from the 1890s, was for management to “gather in all of the great mass of traditional knowledge which in the past has been in the heads of the workmen.” Skilled machinists were stopwatched; the Gilbreths filmed them with chronocyclegraphs — a literal 1910s head-camera. Workers cooperated because they were paid piece-rate bonuses to do so. The tacit craft was decomposed into instruction cards and handed to cheaper, unskilled labour. Outcome: enormous productivity gains, the collapse of the craft wage premium, a machinists’ revolt, congressional hearings in 1911–12, and Taylorism banned in US government arsenals by 1915. It took roughly fifty years and the postwar labour accord before the gains were broadly shared… 

In the 1990s American hospitals routed physician dictations to transcriptionists in Bengaluru and Chennai; it was unglamorous work, but India was good at it. That corpus is precisely what trained speech recognition. The industry peaked and then largely evaporated. Compensation to the transcriptionists: zero… most startups in this space risk meeting the same fate as the transcription companies of the 1990s.

In this context, I am reminded of the claim made by Daron Acemoglu and Simon Johnson in their book Power and Progress that the trajectory of technological progress is a political choice made by society and should not be left to corporations and technocrats. Their central claim is that the direction of technology is a social choice, not a technical destiny, and that redirecting it requires countervailing power rather than better-intentioned technocrats. 

The problem, though, is that globally, and especially due to the Trump 2.0 regime, the rule makers have surrendered agenda-setting to Big Tech and AI Labs. Closer home, India has almost no leverage over the direction of frontier AI. Instead, its leverage is confined to the terms on which its labour and data enter the supply chain, and not to bending the technology’s arc. 

In the circumstances, what can a country like India do?

Here are some thoughts for consideration. One, a statutory floor rate for training-data contribution and an industry-led collective licensing body for data work are both administratively feasible and could increase value capture (from the worker’s current share of 1-2% of the value created) without banning anything. A comparator is the model of SoundExchange (US) or PRS (UK) in the music industry, which acts as a government-designated clearinghouse that collectively licenses music, collects usage fees, and distributes royalties to creators, effectively removing the burden of individual licensing. This model would also subtly frame the market in terms of treating data as labour, and not as mere raw material. 

Second, on the regulatory side, it may be useful to revisit the DPDP Act provision that permits employers to process worker data without explicit consent under “employment purposes”. Instead, there should be purpose limitation, or restrictions on repurposing training data for other activities, and consent requirements of all involved. 

Third, public spending on AI innovation and procurement preference could be made conditional on the recipient retaining licensing rights to datasets collected from Indian workers rather than doing work-for-hire. This would frame the collection of data as an input and not a product, and industrial policy could price it appropriately. Fourth, there is the argument about extending statutory instruments like the gig worker welfare boards or the Code on Social Security present in some states (Rajasthan, Karnataka, etc.) to cover data work. It could help build countervailing power. 

But pursuing these agendas can be costly. This being a global market, prohibiting or putting too onerous terms on value capture and the entry of data into the supply chain will backfire by moving the work to Vietnam, Ethiopia, or the Philippines. Besides, for the Indian workers, already facing an acute scarcity of jobs, the choice isn’t really on offer, and ₹400 a day is ₹400 a day. There is a collective action problem here which calls for multilateral engagement through a forum like the ILO. 

But this should not mean that we sit back helplessly and allow the market dynamics to play out. Instead, before enacting any of them, there must be a public debate on the merits or otherwise of these proposed measures. What are their respective costs, and what can be done to mitigate them? What versions, if any, of these measures should be enacted? Such debates are essential to make informed and collective social and political choices.

The public debate is important since the agenda-setting process here, like with any technology change, pushes certain considerations to the forefront while also marginalising certain others. Almost always, the former represents the interests of the corporations and elite beneficiaries of the change, and the latter represents those of the vulnerable and voiceless. Therefore, such agenda-setting debates are a purely political activity, with profound social implications. 

It is also important since there is the distinct likelihood that India could spend the next five years as the world’s back office for the third time, and when the juice has been sucked out and value captured, there could be nothing left standing that India owns. 

PS: In this context of collective action problems, it is worth taking inspiration from one very impressive and encouraging breakout (which has not received the level of attention it deserves) from South Korea. It is a tribute to the maturity and wisdom of the country’s corporate and political system and the robustness of its democracy that Samsung and SK Hynix agreed to share 10% of their windfall profits from memory chip sales, with no ceiling on payouts, with their employees for the next ten years. Sample this.

Samsung Electronics... agreed last month for employees to share the chipmaker’s blockbuster profits from an AI-led boom... SK Hynix... handed employees a similar profit-sharing deal last year... Samsung is also going to give Won500mn loans at low rates to employees... Samsung and SK Hynix together control much of the market for the advanced memory chips used in AI servers. Employees at both companies are in line for average annual bonus payouts of Won600mn, which compares with a national average salary of about Won50mn... district of Hwaseong... expected to gain corporate income tax receipts of Won1tn to Won1.3tn from Samsung alone this year, an extraordinary sum for a city authority whose annual budget is about Won3.5tn.

Weekend reading links

1. FT long read on how senator Deborah O'Neill, as chair of the Parliamentary joint committee on corporations and financial services, has single-mindedly exposed and brought the knees the Big Four auditing and consulting firms in Australia. 

Deborah O’Neill has led the charge against KPMG in Australia over a client confidentiality scandal that prompted the departure of the firm’s chair, chief executive, chief operating officer, audit leader and a senior partner over the past month... comes on the heels of a similar implosion at PwC. The rival Big Four firm came unstuck when a data leak led to the exit of senior management... An EY employee was charged with accessing the bank details of Prime Minister Anthony Albanese while working on contract at Australia’s biggest bank. Meanwhile, Deloitte partially refunded the Australian government after admitting that it used AI to compile a report...

In 2023, the Labor senator forced the publication of emails that implicated PwC partners in the tax leaks scandal. PwC partners were caught sharing secret government tax plans that one of them had obtained from his work on an advisory board in Canberra, in the hopes of winning business in the US. This year O’Neill used parliamentary privilege to air allegations made by a KPMG whistleblower that had been inadequately investigated by the firm. KPMG has now been exposed as having used confidential information from existing audit clients to try to win new business from rivals — some from PwC as its audit customers looked to switch in the wake of that firm’s woes...

She entered parliament in 2010 having spent her career in education. O’Neill soon discovered that some Big Four consultants acted like some of her former pupils — copying the answers from the back of the book and then marking their own work, as she puts it — and used her role to put the leaders of the firms under pressure.

2. Soumaya Keynes points to a fascinating study by Rebecca Diamond of Harvard University of the use of GLP drugs that appears to show increased confidence and employment rates among women in the US. The study finds, using data gathered between 2021 and 2023, that the poorest third of women in the US suffered an obesity rate 14 percentage points higher than the richest third, whereas the gap was negligible for men. 

The study's main findings:

After 18 months, women using GLP-1 drugs who start off without a job enjoy employment rates 27 percentage points higher than otherwise similar non-users. Women who start off with a job see their employment rate fall slightly, and although the data is too noisy to pick out effects on their earnings, it looks like their household income rises by 10 per cent. That second effect is a bit surprising, and possibly explained by parallel developments in these women’s love lives. Diamond estimates that GLP-1 drugs give single women a dramatic 29 percentage point increase in their chances of coupling up. On average, their new partners are richer than them, giving their household income a bump. Which could explain why a few of the women losing weight then feel able to drop out of work.

The study also points to a more disturbing consequence.

So far at least, GLP-1 drugs are disproportionately used by the rich. In Diamond’s study two-fifths of the women paid for the drugs out of pocket, at a median cost of $275 a month. Research based on Voy prescriptions shows how, adjusting for relative obesity rates, uptake is skewed towards more affluent areas. If obesity becomes an even stronger signal of economic disadvantage, the stigma attached could grow.

3. The rise of London's King's Cross area as perhaps Europe's AI capital.

Two decades ago, King’s Cross was central London’s most neglected district. Today, it is home to the main foreign outposts for several of the world’s wealthiest companies, from Big Tech giants Google and Meta to their richly funded AI challengers including OpenAI, Anthropic and Jeff Bezos’ Prometheus. AI researchers and entrepreneurs are packing out the area’s canal-side cafés so densely that venture capitalists prowling for their next deal are struggling to prevent their coffee meetings from being overheard by rivals.

For many, this resurgence can be traced back to one individual: Sir Demis Hassabis, the DeepMind co-founder and Nobel laureate who stayed in London to build his AI lab following its sale to Google in 2014 for £400mn. “Demis keeping DeepMind in London and resisting the gravitational pull of [America’s] West Coast is the most important thing that has ever happened to the London tech ecosystem,” says Tom Hulme, a tech investor at Alphabet’s GV venture capital unit... Just as PayPal helped launch the careers of a generation of Silicon Valley founders and investors including Elon Musk and Peter Thiel, a “DeepMind mafia” in London is pulling in billions of dollars to AI start-ups founded by Hassabis’s former lieutenants, including David Silver’s Ineffable Intelligence and Tim Rocktäschel at Recursive Superintelligence...

It was Hassabis’s pursuit of an “artificial general intelligence” capable of scientific research and a wide range of human tasks that in many ways kick-started the current AI boom. DeepMind’s sale to Google prompted Elon Musk to set up a research lab to counterbalance the internet group’s dominance of AI; that lab was OpenAI. DeepMind went on to make a series of AI breakthroughs including AlphaGo, which in 2016 beat the board game Go’s world champion Lee Sedol, and AlphaFold, which used deep learning to predict protein structures with superhuman speed.

King's Cross has emerged as the Canary Wharf of tech in UK, and this description is apt and underlines the continuing importance of personal interactions and connections. 

The density of AI talent in King’s Cross was why the government’s scientific research agency Aria took a “very conscious decision” to base itself there rather than Whitehall, says Pippy James, its deputy chief executive. “We were definitely inspired by Kendall Square in Boston,” she says, referring to the area surrounding MIT where Google, Microsoft and Amazon, as well as biotech companies Moderna and Novartis, have offices. “Value creation comes from those serendipitous collisions.” In recent weeks there has been a steady stream of American AI companies announcing moves into the area. OpenAI and Anthropic have signed leases for tens of thousands of square feet in King’s Cross. Others moving in include Bezos’s “physical AI” company Prometheus, Cursor, the AI coding company that recently agreed a $60bn sale to SpaceX, AI agent group Perplexity and open model developer Reflection. Google, which already has many researchers and engineers in the area, plans to start moving staff into its vast new “Platform 37” office this summer after almost a decade in development.

4. This is a striking factoid about the importance of chips now. 

With SpaceX going public, the list of the 10 most valuable US public companies is entirely made up of tech companies for the first time. Of those, three are semiconductor specialists. But with chips a key ingredient in AI, the other seven are also now all designing their own chips.

Some stats about the global chip squeeze.

... Elon Musk’s xAI to rent out spare capacity in its data centres. In recent weeks, Anthropic, Google and start-up Reflection AI have agreed to pay a total of around $2.3bn a month — or $28bn a year. This looks like a big return on Musk’s data centre investments. As of March this year, xAI’s total capital spending over its lifetime totalled $26.5bn... this surge in demand has already prompted a huge increase in supply, both of planned chipmaking capacity and newly minted chip stocks. One sign is the $600bn that memory chipmakers Samsung and SK Hynix said this week they plan to invest in Korea. Another is the $29bn that SK Hynix hopes to raise when its American depositary receipts begin trading in the US next week... TSMC has said it will boost its capital spending by as much as 37 per cent this year, as it did in 2025. But those increases follow two years of retrenchment and would leave 2026 capex only around 50 per cent higher than 2022. Contrast that with the biggest buyers of AI chips. Seven of the largest data centre operators are planning to spend an astounding $848bn this year, at least five times what they spent in 2022, according to a calculation by the newsletter Exponential View.

5. On the new bonds issued by SpaceX.

The bonds enjoyed very robust demand at the point of issuance, but some see that as a problem in itself. Allianz’s chief investment officer has described the market’s willingness to hand money over to Musk as a clear sign that we have moved from “a healthy boom, a stretched boom . . . into bubble territory”. Ominously, the bonds have weakened since they launched.

6. The AI LLMs scorecard. 

Where will Sarvam stand?

Also national scorecard.

7. The low-margin business of mobile phone assembly. Amber Industries which makes air conditioners for eight of the top 10 brands and more than a quarter of all ACs made in the country, now proposes to assemble smartphones for Oppo. 

Now, the company plans to sub-lease a part of Oppo India’s factory in Noida, set up SMT lines, and start assembling phones there... But smartphone assembly is one of the toughest businesses in electronics manufacturing... The target is to eventually assemble about a fifth of Oppo India’s volumes, scaling from roughly 8 million handsets in the first year to nearly 15 million in the second... Amber expects Ebitda margins of 1.5–2%, excluding benefits from the government’s production-linked incentive (PLI) scheme. That’s well below the 8.8% Ebitda margin generated by its broader electronics business in FY26, and the 7.1% operating margin recorded by its AC-heavy consumer-durables segment... Take Dixon, for instance. The company already accounts for nearly one-fifth of India’s smartphone output. Even at that scale, smartphone assembly, aided by PLI incentives, generates Ebitda margins of only about 3%.

8. AI boom compared with historical episodes.

9. On the rise of non-compete clauses in OECD countries and their adverse impact on productivity.

About 30 per cent of employers surveyed by the OECD said they had increased their use of the clauses in the past five years... It estimates that a 10 percentage-point increase in the prevalence of non-compete clauses in an industry was associated with a 1.9 per cent decline in the level of labour productivity, with workers stuck in sub-optimal jobs and firms less able to gain new skills. In many countries, non-competes have spread into parts of the labour market where the original justification of protecting sensitive information and firm-specific information is “weak or absent”, the research found, pointing to their use among entry-level fast-food staff in the US, manual workers in Italy and childcare workers or yoga instructors in Australia.

10. The market concentration in DRAM chips.

A market in which a monopolist owns 100 per cent gets an HHI of 10,000, a duopoly scores above 5,000, and a perfectly competitive market approaches zero... The US DoJ considers anything between 1,000 and 1,800 points to be “moderately concentrated”. Per Counterpoint Research, as of the first quarter of 2026 the memory market is 38 per cent Samsung (South Korea), 29 per cent SK Hynix (South Korea), 22 per cent Micron (US), 8 per cent CXMT (China), 2 per cent Nanya (Taiwan), and 1 per cent everyone else. This gives us an HHI of 2,838... Both Samsung and SK are two of Korea’s largest conglomerates (the so-called chaebols) which benefit from cosy relations with the state, so viewing them as fierce competitors in the same memory market might be wrong-minded in this case... And if the two companies function as a single economic entity in the global DRAM market, we should probably count them together for HHI purposes. And doing this we get a much higher HHI reading of 5,042. That’s above 5,000 —the HHI of a perfect duopoly.

11. China’s excess capacity in manufacturing requires something similar to what was done by the former Prime Minister Zhu Rongji. 

In the late 1990s and early 2000s… under Zhu’s slogan of “zhua da, fang xiao” or “grasp the large, let go of the small”, Beijing retained its grip on key strategic industries while relinquishing control of a vast sea of smaller companies and factories…Thousands of mines, steel mills and other industrial sites were shut for good. An estimated 30mn to 40mn workers lost their jobs. The process was deemed painful but necessary: not only in setting up China’s accession to the World Trade Organization in 2001, but in freeing Beijing from supporting uneconomic industries… 

Over the past 15 years, as China’s share of global manufacturing surged to around one-third, the share of lossmaking industrial businesses jumped from about 10 per cent in 2010 to nearly 25 per cent last year, according to the MERICS China Overcapacities Monitor. This dynamic exists across everything from steel and cement to cars, computer chips and robots. Take the automotive sector for example. Domestic car sales last year totalled 23.9mn against estimated production capacity of 45mn to 50mn. Sales are highly concentrated among a clutch of leading companies. According to HSBC, more than 70 per cent of EV sales — including plug-in hybrids — are being soaked up by 10 brands, leaving 47 others jostling for the remainder. In the shrinking market for petrol and diesel cars, 10 brands have about 70 per cent of sales and 73 others compete for the rest.

12. South Korean capitalism and windfall profits sharing - SK Hynix and Samsung edition.

Soaring global demand for high-bandwidth memory chips used in AI systems has propelled SK Hynix and Samsung Electronics to record earnings. This week Samsung announced quarterly operating profit of Won89.4tn ($59.7bn). The windfall is being shared with employees. Last September, SK Hynix agreed to pay workers 10 per cent of annual operating profits for a period of 10 years. Samsung followed with a similar arrangement in May after its union threatened strike action. With both firms expected to earn hundreds of billions of dollars this year, average bonus payouts per memory chip worker could reach about Won600mn ($400,000) at Samsung and even more at SK Hynix. Such amounts are staggering in a country where the average worker earns Won50.6mn per year, according to Korea Enterprises Federation data. 

The Bank of Korea has warned of potential inflationary pressure as a result, and towns where many semiconductor workers live are undergoing property price jumps. Competition is intensifying for places at universities offering semiconductor “contract” programmes that guarantee jobs at Samsung Electronics or SK Hynix upon graduation. Admission scores required for some such courses now exceed the average for natural sciences at Seoul National University, the country’s top-ranked university, and are just below those needed for medicine... The boom is also reshaping Korea’s marriage market. Matchmaking agencies, which are known for using meticulously harsh metrics to rank clients, are now giving higher points to chip workers.

13. Data centres are consuming massive amounts of power and water.

Data hubs already devour more electricity globally than all but 10 countries. About 448 terawatt hours last year if you’re interested. The AI boom means that amount is on track to roughly double within four years... By 2030, they could be using enough water to meet the basic needs of all 1.3bn sub-Saharan Africans for a year, UN researchers estimate.

And it is provoking backlashes. Sample this from the US.

An unprecedented 75 US data centre projects worth around $130bn were blocked or delayed in the first three months of this year, nearly as many as in the whole of 2025, says the Data Center Watch research group. It reckons active opposition group numbers have grown from 396 at the end of 2025 to 833 by the end of March.

14. The Strait of Hormuz squeeze was not as bad as earlier episodes. 

15. Transformers are at the heart of power transmission, distribution, and use. Thanks to the AI and data centre boom, transformer prices have gone over the roof.

Specialist electrical steel — essential for transformer cores — is produced by only a handful of global suppliers, many of whom are struggling to keep up with the surge in demand. Market growth, price volatility and limited mining capacity have also strained supplies of copper, crucial to the conductivity of windingsaround a device’s core. But one of the most acute bottlenecks is the shortage of skilled workers needed to carry out complex, labour-intensive manufacturing tasks...

With up to 80,000 different designs, most transformers still have to be built largely to order, taking three to six months to make. The most demanding stage is the windings, when copper wire is applied around a transformer’s core — a “beautiful” process, according to Bruno Melles, an engineer now leading Hitachi’s global transformer business. Each winding is unique and is “still a human manual activity that we’re very proud of”, he says. The number and pattern of these windings dictate voltage — fewer turns lead to lower voltage while higher-voltage devices can have multiple windings stretching hundreds of kilometres. Once complete, the assembly is placed in a protective outer metal shell, where oil acts to insulate and cool the device...

The world’s biggest transformer manufacturers have reported tens of billions of dollars in backlogs in the first quarter of 2026... US developers are turning to imports. The EU, Mexico, South Korea and Brazil are the biggest suppliers of power transformers to the US, together accounting for more than three-quarters of imports by value last year.

Into this mix come innovations in the form of modular solid-state transformers that use modern power electronics and respond dynamically to changing power needs, enabling real-time monitoring and control (which legacy transformers with their steel and copper cannot do). 

The loss of financial market discipline

This post will provide a framework for thinking about the erosion of financial market discipline. 

The FT has an article on an inflating earnings bubble in the financial markets. 

Analysts are now forecasting a 25 per cent increase in S&P 500 company earnings for the coming year, according to Bloomberg data… Ben Inker, co-head of asset allocation at GMO, said forecasts for the next two years were “rising at an exceedingly high rate, nothing we have seen outside of a crisis recovery”. Consensus estimates for coming-year profits have risen by almost 20 per cent in six months, the biggest such jump since 2021…

Capital Economics analysts warned this week that “AI-related equity markets may be approaching a point where earnings expectations and capital expenditure assumptions become difficult to sustain” and a correction in these could “trigger a broad equity market pullback”. Michel Lerner, head of UBS’s investment analytics platform HOLT, said “shares in the AI food chain are priced to maintain supernormal profits” and warned of an “earnings bubble” forming in the market. While exceptional profits appear likely to keep being delivered in the immediate future, he said, “the likelihood of sustaining these levels of profitability and growth is incredibly low”.

This earnings bubble supplements an already inflated valuations bubble. 

The BIS Annual Report has an excellent graphic that shows that stocks are pricing an earnings bubble, underpricing risk, and household exposure has increased sharply.

The implied long-term earnings growth for the largest corporations sits well above recent historical benchmarks, with US stocks often trading at large premia to peers in other major markets. These implied rates often exceed even the elevated growth that some of the technology firms have delivered in their relatively short lifetimes. As these firms mature and command a larger share of the market, sustaining such high growth could become increasingly challenging… Risk premia on the largest US stocks have compressed markedly since the COVID-19 pandemic, with the distribution shifting clearly to the left. This points to growing investor complacency and reduced compensation for risk-bearing… A major equity market correction could have larger macroeconomic consequences today than in the past. Household equity exposures have grown over the past few decades, both relative to total wealth and income. A large correction in valuations could have more pronounced wealth effects and sharper consumption pullback than in the past. And with US stocks accounting for an outsized share of global equity markets – about 64% of the MSCI Global index – the wealth impact from a US-led repricing could propagate globally.

And, the report says, all this is fuelled by a complex web of private circular transactions involving hyperscalers, chip makers and AI labs. 

Chip makers and hyperscalers take equity stakes in AI labs or neocloud providers, who in turn commit to multi-year purchases of chips or computing power. Data centre construction is increasingly outsourced to third parties that lease facilities back to hyperscalers on long-dated contracts with embedded exit clauses. The terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times. Together, such arrangements account for a sizeable share of sector-wide financing and forward revenue. A sharp repricing of equity risk could prompt a reassessment of corporate credit risk and lead to tighter credit conditions more broadly… 

Any tightening in credit conditions could expose existing vulnerabilities in the less transparent private credit space, whose reach has expanded among middle market and small firms… A larger shock, whether from a renewed inflation surge or a sharp AI-led repricing, could trigger a more widespread credit crunch… The growing role of private credit also raises concentration risks. Direct lending funds, dominant players in the private credit ecosystem, have quadrupled their lending to the AI and information technology (IT) sectors in the past five years, to about 15% of their portfolios. These loans tend to be larger than those in other sectors, while their terms such as tenor and pricing remain broadly similar, raising questions about lending standards and risk pricing. 

In this backdrop, it is useful to ask the question whether financial markets have lost their disciplining powers. What is it about modern financial markets that makes them underprice risk? 

For sure, greater access to information has bridged information asymmetries, increased transparency, and reduced uncertainties and risks. And this has, in turn, lowered price discovery frictions and increased market efficiency. However, this has gone alongside trends in the opposite direction. Already, the complexity of financial instruments obscures risks and distorts price discovery. But other market practices have emerged in the last two decades that distort incentives and erode the market discipline. Some of them are consequences of the aggressive expansion of central bank monetary policy toolkits in the aftermath of the Global Financial Crisis. 

There is a moral hazard created by the market participants’ internalising the belief that too-big-to-fail institutions or industries will always be bailed out. The institutions themselves come to view this as an insurance against their risk-taking. Well-intentioned actions of central bankers like forward guidance and commitments to backstop against shocks (Greenspan Put and Draghi’s “whatever it takes”) create similar moral hazard across markets. The gradual erosion of gatekeeping standards (ratings inflation, audit failings, index inclusion - e.g., for SpaceX) distorts incentives and leads to underpricing of risks. Excessively bullish market guidance, like with AI stocks now, fuels a cycle of irrational exuberance that amplifies bubbles. Finally, the inherent dynamics of financial markets bake in the fear of missing out among the institutional investors. 

The fundamental basis for any efficient market is that information flows facilitate price discovery, and the costs of the actions of market participants are internalised. Based on this, an analytical framework on financial market discipline can be built on three pillars - information quality (participants must get accurate signals); price discovery (prices reflecting the underlying value); and skin in the game (losses must fall on risk takers). All three must hold simultaneously for market discipline. 

The three erosions are not independent. Mechanical demand distorts prices; distorted prices reward gatekeepers who validate them; validated prices attract more backstops when they wobble. Each pillar’s erosion masks the erosion of the others. By the time discipline is missed, all three have been hollowed out simultaneously. This is why the current market environment can display record-high valuations (Price Discovery pillar broken), narrow risk premia (Skin in the Game pillar broken), and record-optimistic analyst forecasts (Information Quality pillar broken) - all at the same time, without triggering any traditional warning system. The system that would normally catch such a configuration has been progressively disassembled.

Interestingly, each erosion channel began as a reasonable response to a specific problem. Deposit insurance solved bank runs. Forward guidance solved communication ambiguity. Passive investing solved active-fund underperformance. Rating agencies solved information asymmetry. None of them is wrong in isolation. The problem is what happens when all the individual protections are stacked simultaneously, and all of them are pursued to their extremities. The system loses its balance.

The market’s self-correcting machinery is gradually replaced by a suite of external supports, and the participants who set prices become smaller and smaller relative to the participants who follow rules. At some point, arguably reached, the market is no longer a disciplinary mechanism at all. It is a policy instrument with an ambiguous owner. That is why the current earnings-bubble concern is qualitatively different from previous bubbles in that there is no obvious mechanism left through which the market disciplines its own excess.

So what can be done to restore the disciplining powers of the market?

Quite simply, it should be about restoring accurate price signals, consequences and costs of risk taking, and the trustworthiness of information, and all by going back to first principles. 

The restoration of price discovery requires both thickening the market (a high concentration of buyers and sellers) and increasing liquidity (allowing assets to be transacted without significantly impacting the price). The internalisation of the costs of risk-taking demands rolling back moral hazard-inducing backstops and bailouts, or at the least ensuring these interventions are priced appropriately. Finally, restoring the quality and credibility of information requires separating information provision from the transactions side of financial intermediaries, and letting the market guide with information. An agenda is outlined below. 

This must be complemented with cross-cutting reforms such as restrictions and greater oversight on revolving-door personnel movements, enhancing the white-collar prosecutorial capabilities of regulators, enhancing the supervision of non-bank financial institutions (NBFI), macroprudential capital surcharges for risk concentration, etc. 

The problem, though, is that these are all very difficult reforms even at the best of times, but particularly difficult now. The silver lining is that a big crash can create the conditions for such reforms.

AI for organisational and bureaucratic reforms

The debate rages about the extent of AI’s likely impact on the economy and human lives. So far, there has been an apparent lack of commercial value creation to justify the gigantic and exponentially increasing volume of AI investments. 

I blogged here on the distinction between horizontal and vertical use cases of AI, with success on the latter being limited, here cautioning about the likely impact of AI on development and in developing countries, and here on some possible high-impact use cases for AI in lower-income countries. 

The most common horizontal use of AI is in personal productivity improvements. Claude, ChatGPT, etc., are already having large effects on personal productivity. But its translation into vertical use products is muted. 

On this, John Burn-Murdoch points to the work of Mert Demirer, Leon Musolff and Liyuan Yang to make two important points. The first point is the “disconnect between reported increases in coders’ output and the apparent lack of a corresponding boom in product or value creation,” which creates a very steep funnel between inputs and outputs. 

The study by MIT’s Mert Demirer and co-authors tracked software developers’ work before and after they adopted AI tools. Importantly, they measured this at several different levels, from the amount of code written, to the number of discrete files edited, to the number of projects or features worked on, to actual releases of new software. They found an explosive impact at the top of this funnel — coders created or edited almost 300 per cent more files — but that boost was halved to 150 per cent by the time they got to the number of discrete pieces of work submitted for review, and that in turn shrunk fivefold to a roughly 30 per cent uplift in the number of full software releases.

The authors also found little evidence of AI-assisted increases in software development, leading to increased consumption of Apps. 

This brings us to the second point made by the authors about AI’s impact - it is likely to be fully realised only when new organisational structures, markets, and business models emerge.

But Demirer and his co-authors feel the more likely explanation is that current organisational structures and marketplaces are not set up to take advantage of real underlying gains. That view is supported by the evidence from past technological revolutions, where the real jumps in productivity and job displacement came from new companies and processes rather than incumbents grafting new technology on to existing workflows. In the case of electricity in the late 19th and early 20th century, productivity gains were modest where factories simply replaced giant steam engines with giant electric motors but left the rest of the machinery and layout unchanged. The boom arrived decades later when engineers fitted individual workstations with their own small motors.

In this backdrop, Jack Dorsey and Roelof Botha have an insightful article on organisational impact. Specifically, they claim that AI’s productivity-enhancing value can address the fundamental coordination problem in large organisations that manifests in the form of a trade-off between span-of-control limitations (which add organisational layers) and speed of information flows. They argue that AI sharply increase people’s span of control, thereby reducing organisational layers and hastening decision-making. 

The first organisational models emerged in the military to organise large numbers of soldiers into a coherent and effective fighting unit. It involves a hierarchical chain of command that allows for a span of control and a seamless flow of information and instructions. The model then entered the corporate world through the US railroads in the 1840s and 1850s, which borrowed West Point-trained engineers from the US Army. They trace the evolution of the modern organisational form,

In the mid-1850s, Daniel McCallum of the New York and Erie Railroad created the world’s first organizational chart to manage a system stretching over 500 miles with thousands of workers… McCallum’s chart formalized the same hierarchical logic the Romans had used: layers of authority, defined reporting lines, structured information flow. It became the blueprint for the modern corporation… Frederick Taylor (1856-1915), often called the “Father of Scientific Management,” optimized what happened within that hierarchy. Taylor broke work into specialized tasks, assigned them to trained experts, and managed through measurement rather than intuition. This produced the functional pyramid organization - a structure optimized for efficiency within the information routing system that the military had pioneered and the railroads had commercialized… 

In 1959, McKinsey’s Gilbert Clee and Alfred di Scipio published “Creating a World Enterprise” in the Harvard Business Review, providing an intellectual framework for a matrix organization that combined functional specialties with divisional units. Under the leadership of Marvin Bower, McKinsey helped companies like Shell and GE implement these principles, balancing central standards with local agility. This became the “professional” or “modern” corporation that propelled the postwar global economy… The McKinsey 7-S framework, developed in the late 1970s by Tom Peters and Robert Waterman, distinguished the “hard Ss” (Strategy, Structure, Systems) from the “soft Ss” (Shared Values, Skills, Staff, Style). The core idea was that structural elements alone were insufficient. Organizational effectiveness required alignment across cultural traits and the human factors that determine whether a strategy actually succeeds.

They suggest that AI makes it possible to solve the fundamental coordination problem within large organisations that necessitate hierarchical formations. 

For the first time, a system can maintain a continuously updated model of an entire business and use it to coordinate work in ways that previously required humans relaying information through layers of management… In a traditional company, a manager’s job is to know what’s happening across their team and relay that context up and down the chain. In a remote-first company where work is already machine-readable, AI can build and maintain that picture continuously. What’s being built, what’s blocked, where resources are allocated, what’s working and what isn’t. That’s the information the hierarchy used to carry. The company world model carries it instead… 

In a conventional company, the intelligence is spread throughout the people and the hierarchy routes it. In this model, the intelligence lives in the system. The people are on the edge… The edge is where the intelligence makes contact with reality… the edge doesn’t need layers of management to coordinate it. The world model gives every person at the edge the context they need to act without waiting for information to travel up and down a chain of command… Everything else the old hierarchy did, the system coordinates, and everyone is empowered, with a role that’s much closer to the work and the customer.

They identify three roles - Individual contributors (ICs) who are deep specialists and experts who build and operate system capabilities; Directly Responsible Individuals (DRI) who own specific cross-cutting problems or opportunities and customer outcomes; and player-coaches who replace the traditional manager whose primary job was information routing, who do both building and handling people. 

All this makes great sense and points to how corporate organisational models are likely to emerge as the application of AI progresses. There will be frontier firms in a few sectors that will lead the way for others to follow. 

AI applications are a promising opportunity to address inefficiencies and coordination failures in public bureaucracies, too, and improve the quality of public administration. 

For a start, it has the potential to restore internal capabilities, which have eroded steeply. Over the years, thanks to practices like outsourcing all analytical and documentation work to consulting firms and the hiring of individual consultants (most notably now, the system of Young Professionals, YPs, in governments), there has been a complementary erosion of in-house expertise. The capabilities to articulate proposals for internal deliberations and file circulation have atrophied. Given that bureaucracies run on deliberations and files, this trend is an underappreciated aspect of state capability weakness. 

AI provides an opportunity to reverse these trends and develop internal capabilities. The primary reason for the reliance on external expertise is the extent of analytical work and documentation required during the deliberative process (everything from a concept note on the proposal to reports for appraisals, and Cabinet Notes). The bureaucratic leaders who are burdened with a multiplicity of tasks, work under tight timelines, face increased fetters from oversight agencies and courts, and must rely on an increasingly enfeebled internal bureaucracy. In the circumstances, they prefer to outsource the thinking and documentation to outsiders. I have blogged earlier on the perils of this approach. 

AI tools like Claude are excellent at analytical work and the generation of these documents in response to clearly articulated prompts. It becomes a simpler proposition if bureaucrats can quickly and easily obtain a draft concept and supporting documents, and then scrutinise, validate, and refine it before circulation for approval. AI tools can then become a force multiplier for bureaucratic leaders, who are now constrained by their limited bandwidth and acute dependencies. 

This would also empower bureaucratic leaders, or at least some among them, and could enhance the quality of their engagement with the decision-making process. Besides, by minimising the drudgery of the bureaucratic process, it would also allow bureaucratic leaders to apply their minds and exercise judgment more effectively, thereby improving the quality of decisions and policy design and implementation. It would also lower decision-making delays.

It should therefore become a priority of the National Informatics Centre (NIC) (or an AI division within it) to develop or license AI application that is embedded in the e-office software and enables officials to sift through large documents and generate proposals/presentations, circulation notes and reports using prompts. This has transformative potential for productivity improvements, not only stopping the erosion of internal capabilities but also helping rebuild them. 

If this can be done, it opens up opportunities for far-reaching administrative reforms. The current bottom-heavy pyramid can be rationalised to make it fit-for-purpose.

A major inefficiency is the presence of multiple layers within the administrative system. It is a widespread practice across governments to have YPs, and those recruited as data entry operators originate the note file (a task earlier performed by the clerical staff). The note then gets circulated across several layers, often seven or eight till the approver. This can be radically pruned down to no more than three or four, including the approving authority. 

Such de-layering is especially relevant for technical ministries and departments whose activities are more amenable to AI-based support. Such ministries should have a separate administrative staffing plan, one that takes into account the role that AI can play in generating documentation and considerably reducing any drudgery associated with analytical work. 

As a general illustration, there are perhaps three kinds of activities in any department - shared services (HR, procurement, establishment issues, statutory matters, etc.), administration of departmental programs, and analytical and technical work. There are significant low-hanging likely process-efficiency improvements in all three, and substantive value-addition potential in the third activity. 

This would also necessitate a reassessment of public recruitments. The advent of AI applications means that, unlike in earlier times, apart from merely documenting the issues in a note file, the case worker (the ASO or SO) can now be expected to do some analysis and provide comments. This also means that a smaller base can serve the clerical roles (the entire paraphernalia of clerical cadres can be collapsed into just two functional levels - maker and checker), and their educational qualifications and skills must reflect the requirements for the revised scope of work. I’ll blog separately on this. 

The increased use of AI applications to analyse and document, and a compact and delayered deliberative process captured in the file circulation can also increase the quality of collective engagement and ownership of the bureaucracy in decision-making. It lets (and nudges or forces) everyone contribute meaningfully to the process instead of being passive pass-throughs of instructions and note files. It presents the opportunity to shed reliance on outsourced expertise and build back state capabilities.

This is deep work and, even in the best case, is likely to be adopted only by a few units in the first phase. The objective should be to create the conditions that encourage the emergence of these lighthouses and channel them to diffuse change more widely.