Observations on the data centre investment boom

At the heart of the AI investment boom are data centres. This post examines some of the issues associated with them, specifically their financing models and their energy demand. 

The emergence of AI models and applications for inference has resulted in a surge in demand for data centres with high-performance supercomputers. The extent of investments required for these data centres has been described as “one of the biggest capital movements in history.”

It’s estimated that the US has about 20 GW of operational data centre capacity, and another 10 GW is projected to break ground in 2025, of which 7GW will be completed. Globally, too, data centre investments are booming, including in developing countries. The World Investment Report 2025 shows that greenfield investments into data centres in developing countries more than tripled between 2020-24 to $56 bn. 

Technology companies are investing heavily in AI-related hardware as they race to harvest the potential benefits from the promising general-purpose technology and conquer the emerging market. 

The scale of capex by the big US tech companies — Microsoft, Alphabet, Amazon, Apple and Meta — is staggering… Their collective investment splurge amounts to arguably the biggest, and certainly the fastest, infrastructure rollout in history. Arete Research estimates that these companies will spend about $480bn in capex in the next two years, much of it on the 100 data centres they are currently building. Many of those data centres will be powered by Nvidia’s GPUs. 

The company at the centre of this investment boom is OpenAI. With a flurry of deals, the loss-making owner of ChatGPT, with an annual revenue of just $13 billion but over 700 million regular users of its blockbuster chatbot, has signed broad agreements with SoftBank, Oracle, and Nvidia for capital spending of $1 trillion. Even if only a tenth of these commitments materialise, it would be the largest set of corporate investments in a single company in history. 

Nvidia’s CEO Jensen Huang has said that every 1 GW of data centre capacity requires $50 billion of spending on computing hardware, including Nvidia’s GPU processors and networking technology and server racks produced by the likes of Foxconn, HP, Dell, etc. Nvidia’s $100 billion investment in OpenAI alone is estimated to require 10GW of data centre capacity, enough to power 10 million typical US households. Morgan Stanley has estimated that deploying 10GW of AI computing power could cost as much as $600bn, of which $350bn “potentially” goes to Nvidia. Similarly, Stargate’s $400 billion investment is expected to require 7 GW of data centre capacity. In fact, OpenAI’s Sam Altman has writtenthat he wants to add 1 GW of new AI infrastructure every week. 

The vast majority of these investments would go into building data centres to train and run inferences on OpenAI’s ML algorithms. These data centres, in turn, would run the algorithms using Nvidia’s latest generation of GPU chips and Oracle’s cloud infrastructure. 

The deals also highlight companies scrambling to mitigate risks through several innovative emerging models of financing the AI infrastructure. OpenAI itself is buying data centres, cloud infrastructure, and GPU chips through long-term capacity purchase commitments with their respective developers and producers. Under these unique arrangements, the latter would “invest equity” in OpenAI by offering their assets to it to train and run its algorithms. 

The staggering sums involved in these announcements have naturally raised questions.

Who will be on the hook for the data centres built in the hope that AI demand will continue to boom, and where will the cash flow come from to support the mountains of debt that are sure to be involved? How likely is it that OpenAI’s business will support the $300bn in cloud payments it is due to pay Oracle by 2030, or that Nvidia will see the $350bn-$400bn in new chip sales that analysts project from its OpenAI deal?

How would Oracle, already debt-ridden, raise the resources to build the data centres? Would SoftBank find investors willing to assume the massive risks and cough up the amounts it has committed? 

Fundamentally, AI input suppliers (like Nvidia and Oracle) are betting on the expectation of a spectacular boom in the demand for OpenAI’s algorithms. Similarly, financiers like SoftBank see this as the big emerging opportunity. This carries the risk of becoming the mother of all Ponzi schemes if the demand does not materialise. 

Highlighting the risks, a report by Bain has estimated that AI companies need to spend $500 bn annually till 2030 on capital investment to meet anticipated demand. This can be justified only with annual revenues of $2 trillion, which it says the industry will miss by nearly $800 billion.

In recognition of these risks, as investments in data centres grow exponentially, instead of financing from internal reserves, the hyperscaler firms (AWS, Azure, and Google Cloud) are turning towards structured finance mechanisms to expand the envelope of capital and also ensure more efficient sharing of risks. 

Between now and 2029, however, global spending on data centres will hit almost $3tn, according to Morgan Stanley analysts. Of that, just $1.4tn is forecast to come from capital expenditure by Big Tech groups, leaving a mammoth $1.5tn of financing required from investors and developers. The gap will be filled by everything from private equity, venture capital and sovereign wealth to bank loans, publicly listed debt and private credit. But increasingly, the answer is debt… Funding data centres comes not just with the risk that costs overrun, but also that the technology becomes obsolete far quicker than anticipated, requiring new investment that decreases returns for its owner — or forces them to sell at a discount. That means even the deepest-pocketed tech groups may want to share the risk, particularly when debt is cheap and readily available. Deals are being structured in myriad different ways, from structured debt solutions and project finance vehicles to construction loans, asset-backed securitisations and even green bonds to raise money and start building.

Morgan Stanley has estimated global spending on data centres to be $2.9 trillion by 2029, of which only $1.4 billion is expected to come from the Big Tech groups and $1.5 billion is expected from investors and developers. 

The financing sources include structured debt solutions, project finance, construction loans, asset-backed securitisation, green bonds, etc. For financial intermediaries like private equity that are struggling in the face of high interest rates and the absence of exits, data centres are emerging as the new big opportunity. They could also be a major driver of the growth of private credit. 

Essentially, data centre financing models are increasingly shifting from data centre developer firms raising debt against their land and project assets, towards more complex structured financing. They involve primarily long-term capacity commitments and/or land leases by AI algorithm users like the hyperscalers, against which developers raise a combination of equity and debt financing. Input suppliers, such as Nvidia (GPU chips), OpenAI (ML algorithms), and Oracle (cloud infrastructure), agree to share risks by forgoing upfront payments on their products and solutions. This approach diversifies risks among all stakeholders. 

This investment frenzy stands amidst the deep uncertainty about the likely scale of commercial returns of AI-related applications. 

There are wildly different views about how quickly such applications will be adopted and what their economic impact will be. In a much-discussed paper from Goldman Sachs, the MIT economist Daron Acemoglu made a powerful case that the likely benefits of AI would be a lot smaller than investors assumed and would take much longer to realise. In the meantime, there was an asymmetric risk that the technology’s downsides, such as deep fakes, might arrive quicker than the rewards. Acemoglu forecast that AI would only boost US productivity by about 0.5 per cent and GDP by around 1 per cent over the next decade. That is dramatically lower than Goldman’s predictions of 9 per cent and 6.1 per cent, respectively. If Acemoglu’s analysis is correct, the US stock market — including Nvidia — is heading for a messy reckoning.

Notwithstanding the hyperscalers’ generative AI revenues of just $45 billion in 2024 despite the already large investments made on AI models, the prospect of a winner-takes-all market means that the Big Tech firms are unwilling to hold back.

The surge in AI-related investments is now having economy-wide effects in the US. There’s now enough evidence that the AI frenzy is a bubble, and boosting growth in the US economy in an unhealthy manner. In the US, even as private investment has generally tanked due to higher interest rates and rising uncertainty, AI expenditure has propped it up. In fact, but for AI-related spending, the total private fixed investment growth of 3% in the second quarter would have fallen by 1.5%. And even as data centre construction has boomed, residential, manufacturing, and other commercial building work have declined.

Let’s now discuss the energy demands raised by the data centre boom. As discussed earlier, an important aspect of the data centres is that they are energy guzzlers. While the energy demands of AI’s learning phase are spiky, high and volatile, data centres need constant high energy. The patterns of solar and wind, which wax and wane during the day and year, are unsuited to serve this need. In the absence of large storage capacity, given their need to be available at all times, data centres have no choice but to rely on gas turbines. 

Therefore, amidst all the efforts to limit greenhouse gas emissions, the massive energy needs of data centres running AI applications and their training models have resulted in a surge of investments into fossil fuels. In fact, a fifth of all gas power capacity additions in the US is to power data centres. Globally, too, fossil fuels are expected to supply a majority of power for data centres and drive CO2 emissions. 

The Big Tech companies that run these data centres claim they are using clean energy, whereas they are only using green power credits that are backed by actual fossil fuel generation elsewhere. Such credits, which do not have to match the location or time when they’re used, are pure financial engineering and a well-known example of greenwashing. 

As an illustration, the graphic below shows the extent of variations in how Ireland met its energy demand in 2024 (each block represents one day, red to green represents the percentage of green energy from 0% to 100%). 

Even with solar and wind generation, highly polluting sources of energy like oil boilers and open-cycle gas turbines are used to serve peak data centre demand. In countries like Ireland, a major location for data centres, the surge in electricity used by data centres has far outpaced growth in renewable generation since 2019. 

In this context, FT points to a study by researchers at UC Riverside and Caltech in the US of public health costs due to data centres built by Big Tech companies. 

Big Tech’s growing use of data centres has created related public health costs valued at more than $5.4bn over the past five years... Air pollution derived from the huge amounts of energy needed to run data centres has been linked to treating cancers, asthma and other related issues, according to research from UC Riverside and Caltech. The academics estimated that the cost of treating illnesses connected to this pollution was valued at $1.5bn in 2023, up 20 per cent from a year earlier. They found that the overall cost was $5.4bn since 2019.

Closer home, state governments in India are rolling out incentives to attract companies to invest in solar power generation and data centres. Given the country’s macroeconomic stability and the maturity of its regulatory and political environment, foreign investors naturally find these stable and long-term return assets very attractive. But some issues must be kept in mind as we pursue data centre investments. 

Apart from requiring vast amounts of power, data centres are water-intensive too (one data centre in Iowa consumed 1 billion gallons of water in 2024, enough to supply all Iowa residents for five days). See also this. Further, data centres create very few jobs (this WSJ article reports that OpenAI’s 1 million sq ft facility in Abilene, Texas, is projected to employ 100 people full time, one-fifth the number of people who will be working in a nearby cheese packing plant that is a fraction of the size). Another estimate informs that a typical data centre will employ just 40-50 full-time employees.

In the circumstances, any discussion on data centres must be linked to water and energy sources powering them. The hydel energy sources in the Himalayan frontier may not be a good choice given the security concerns of concentrating such critical infrastructure in vulnerable border areas. Wind and solar are, therefore, the obvious choices for a country like India. They require large land extents. 

A recent FT article pointed to a CEEW study that estimates that only 35% of onshore wind and 41% of solar real estate potential in India is located in areas without historical contest over land. 

About 60 per cent of India’s land is farmed, compared with the 37 per cent world average, according to the World Bank, and agriculture is the main means of livelihood for the country’s majority. The Institute for Energy Economics and Financial Analysis estimates India’s net zero goal may require up to 75,000 square kilometres of land for solar energy alone — equivalent to the size of the Republic of Ireland or about 2 per cent of India’s total area. The Ministry of New and Renewable Energy has also calculated that a single megawatt of solar power requires on average four acres of land.

However, this side of the equation must be balanced with the critical importance of data (and therefore data centres) in the emerging digital and AI economy. Besides, there’s the strategically important issue of local storage of data. All this raises several questions and necessitates balancing both sides and moving with caution on data centres. 

Since the data centre industry is in its initial stages in India, it may not be advisable to undertake deep regulation on its location choices for now. But it is important to keep an eye out and guide its emergence in an appropriately geographically diversified and environmentally sustainable manner. The recent experience of excessive concentration of solar renewables in Rajasthan, with its numerous evacuation and other policy problems, is a reminder.

Weekend reading links

1. For all talk of AI focus, it does not appear to be showing up in Infosys's personnel hiring over the last six months. 

Amidst all the investment frenzy in the US and elsewhere over AI, Infosys is spending Rs 18,000 Cr buying back its shares, on top of spending Rs 95,000 Cr on buybacks and dividends over the last five years. 

2. China's dominance of the wind turbines market increased sharply since 2020! (HT: Adam Tooze)

As recently as 2020 the global wind turbine market was still a two-horse race with the US not out of the running. Today, China produces more than double the turbines built by the US and Europe put together.

3. It must remain a matter of big concern that even as the world economy has financialised, the cost of sending hard-earned and pitifully small amount of remittances remains elevated at an astronomically high 7.9% for Sub-Saharan Africa (HT: Adam Tooze). 

Additionally, the cost of sending remittances to Africa remains the highest in the world, which dampens the benefits from migration that accrue to Africa. Remittances are one of the most tangible ways for countries of origin to realize the development benefits of migration. Despite the technological advancements in recent decades, the cost of sending remittances remained at 6.2 percent globally in the second quarter of 2023, more than twice the Sustainable Development Goal target of 3 percent. This is largely due to the fees and foreign exchange margins that migrants and their families must pay in origin and destination countries. SubSaharan Africa was the region with the highest cost of remittances in 2023, at 7.9 percent, whereas South Asia had the lowest cost, at 4.3 percent. Figure 3.3 shows that in 18 of Africa’s 29 core countries and seven of Africa’s nine periphery countries for which data are available, the cost of sending remittances is higher than the global average.

The low rate for South Asia is one of the less discussed successes of India's financial market evolution. 

4. France's public debt has risen alarmingly since the GFC.

5. Adam Tooze points to the scale of Friedrich Merz's fiscal stimulus (via TS Lombard). 

Clearly, Germany is stimulating its economy with vengeance, and it appears to have enough space to do so.

6. Unit economics of AI solutions in India is not very attractive.

Netflix, a video-streaming service, costs as little as $1.69 a month in India, compared with $7.99 in America. For cloud services with a low marginal cost, this is no great sacrifice. But running AI queries is expensive. Processing costs for typical users currently hover at around $0.07 per million “tokens” (the units of data processed by AI models) and the response to a single query can run to hundreds or thousands of tokens. That expense is the same whether the user is in Bangalore or the Bay Area.

7. This sums up the challenge with making money in India.

While India’s large population offers scale, it is a difficult market to monetise. According to digital market researcher Sensor Tower, Indians led the world in 2024, downloading 24.3bn apps and spending 1.13tn hours on them. However, their spending was not even in the top 20, at less than $1bn.

8. Palestine is rapidly disappearing.

9. The Economist has an issue focusing on gig workers, who number 200 million in China (40% of urban workforce) of whom about 84 million rely on platform-based employment (delivering parcels and food, and driving bikes and cars) and another 40 million are freelance factory workers. There are some emerging trends in gig work in China.

Lately gig work in China has spread to its vaunted manufacturing sector. The regimented proletariat is gradually being replaced by millions of casual workers who fill jobs “on-demand”, flitting from one factory floor to another at the direction of giant recruitment platforms. The jobs often require no skills beyond a knowledge of the Roman alphabet. The workers may stick with them for no more than a few weeks or even days. Researchers put their number at perhaps 40m, a third of China’s manufacturing workforce—and more than three times the size of America’s.

One reason for the rise of this gig army is that firms want flexibility. Employers prize the freedom to scale their business up or down, responding to seasonal demand, the vagaries of the market and the shifting winds of geopolitics. Technology has played a role, too. Smartphone apps help match customers’ orders with available delivery drivers; in manufacturing, technology has automated away many tricky tasks that used to need experience. Even as this has created jobs for highly skilled engineers, it has left gaps in assembly, packaging and inspection that any warm body can fill. Flexible employment of all kinds suits many workers. Those who are adept at navigating the platform economy can earn more by job-hopping than they could from a single employer.

This is an important snippet about the gig workers.

The average age of factory gig workers is 26. About 80% are male; 75-80% are single and childless. In manufacturing hubs increasing numbers of young workers sleep in parks and under overpasses.

10. FT reports of failures by subprime auto lender Tricolor Holdings and car parts supplier First Brands Group that raise questions about lending and gatekeeping standards. 

Tricolor had won pristine triple-A ratings as it borrowed in credit markets, while First Brands may have amassed as much as $10bn in debt and off-balance sheet financing and was close to raising even more last month... Both companies made use of asset-backed debt, with Tricolor bundling up subprime car loans into bonds and First Brands tapping specialist funds to provide credit against its invoices. At its core, asset-backed finance is the ability to lend against a specific asset or loan, including consumer credit card balances, leases on railcars and solar panels, aircraft and music royalties...

US investment firms have in recent years pushed deeper into asset-backed debt, often pitching it as a safer product than the loans to junk-rated companies that are their bread and butter. But Tricolor is now being probed over fraud allegations by the US Department of Justice, while some investors have long had questions around First Brands’ financial reporting and use of invoice factoring, with lenders now concerned that they lacked visibility about the scale of off-balance sheet financing... Several large banks have also been caught up in the collapse, including JPMorgan Chase and Fifth Third, which are exposed to losses on hundreds of millions of dollars' worth of auto loans. A second investor who has since sold their position in packaged-up Tricolor loans said they had no idea how potential financial irregularities went unnoticed by JPMorgan Chase, one of the banks that underwrote debt offerings.

These kinds of news are now a recurrent staple of financial markets.

11. Michael Moritz comes out all guns blazing at the decision to levy $100,000 fees for H-1B visas.

Every day the Oval Office seems closer to becoming the equivalent of what the sidewalk outside Satriale’s Pork Store used to be for Tony Soprano: a place where a dubious cast of characters spawns brutish extortion schemes and hit jobs... As usual with the Trump administration, the announcement was chaotic and half-baked... Set aside the drama, the announcement demonstrated yet again the fragile grasp the president and his acolytes have about why the US — especially its technology sector — has worked so well. The large tech companies hire foreign nationals because they possess particular skills. They also retain them to perform tasks in areas where the US has labour shortages.

12. New Zealand appoints Anna Bremen, a Swedish economist who has been the first deputy governor of the Sveriges Riksbank since 2019, to head its central bank, the Reserve Bank of New Zealand. 

13. Akash Prakash has some striking numbers about the AI boom in equity markets in the US.

The Magnificent Seven (Mag-7) holds a 32 per cent weight in the S&P 500. In January 2023, just after ChatGPT was launched, this number was only 18 per cent. Nvidia, with an 8 per cent weight in the S&P 500, now has the largest single-stock weight in the history of the index. Its current market capitalisation is equivalent to 15 per cent of US gross domestic product... If we look at the top 10 companies in the S&P 500 (basically the Mag-7, Broadcom, Berkshire and JPMorgan), they account for a record 40 per cent share of the index and 25 per cent share of corporate earnings. We have never seen such concentration of company size and earnings... Since January 2021, 55 per cent of the entire gain in the S&P 500 was accounted for by the top 10 stocks... In 2023 and 2024, the Mag-7 saw earnings growth of about 35 per cent within the S&P 500, while earnings for the remaining 493 stocks grew only 3 per cent...

The Mag-7 and Oracle account for over 35 per cent of total S&P 500 capex. US hyperscalers (the major tech companies) have doubled their share of private domestic investment since 2023. For these hyperscalers, capex has now crossed 20 per cent of sales, compared with under 10 per cent previously. Even on operating cash flow, they are using over 65 per cent to fund data centre buildouts. To put this in perspective, their capex-to-sales ratio is 20 per cent, and research & development-to-sales is 15 per cent, meaning 35 per cent of sales is being reinvested into growth. Truly unprecedented numbers... At their peak in 2000, telecom companies’ capital expenditure accounted for 0.8 per cent of US gross domestic product. Today, hyperscalers’ capex is already at 1.2 per cent of US gross domestic product (GDP), with the current projection being that this number will cross 1.4 per cent by 2028.

14. Countries that have managed to increase their tax to GDP ratio significantly between 2000 and 2022.

15. Very interesting snippet about the impact of superstitions.

In 1966 — a hinoeuma, or “fire horse”, year under an astrological superstition — the fear of giving birth to a wild, destructive and unmarriageable daughter induced a nationwide collapse in pregnancies... The number of babies born in Japan in 1966 plummeted by 463,000 from the previous year, representing a 25 per cent drop. To reduce opportunity risk, marriages also tumbled by 10 per cent. By the end of 1967, with the threat lifted, births had rebounded by an astounding 42 per cent. On historic charts, the spasmodic V-shape makes 1966 look like a colossal data error... Hinoeuma years, which combine the animals of the Chinese zodiac with 10 celestial signs, come around on a 60-year cycle. The next one is 2026.

16. The Magnificent Seven now make up a third of the US stock market capitalisation. 

Nvidia's $4.3 trillion capitalisation exceeds the $3 trillion value of UK FTSE 100.

17. A China Labour Watch (CLW) report has found that more than half the factory staff assembling iPhones at Foxconn's largest factory at Zhengzhou were seasonal staff known as "dispatch workers", despite Chinese law capping their use at 10% of companys workforce. 

US-based CLW also found that dispatch workers faced staggered payment schedules that withhold part of their wages to deter them from quitting during peak production. These staff were not entitled to the same benefits as full-time employees, such as paid sick leave, paid holiday and social insurance that includes medical coverage and pension contributions. CLW also claimed that there is systematic discrimination in hiring certain ethnic minorities and pregnant women... Foxconn uses the flexibility afforded by temporary contracts to adjust to fluctuating demand cycles and, in recent years, to respond to Apple’s shifting requirements about where iPhones should be made... Dispatch workers get a base salary of Rmb2,100 per month, the minimum wage in Henan, but the bonuses make their salaries competitive in the manufacturing sector. These bonuses are typically paid out after three to four months to ensure retention. Many workers preferred the flexibility of short-term contracts and higher hourly wages. However, many said that they had to work a lot of overtime to bolster their hourly wages, which can be as low as Rmb12 for some workers, but range between Rmb25 and Rmb28 for most, depending on experience levels and hiring cycles. CLW found that many staff work 60 hours per week and others up to 75 hours.

18. Stunning graphic that shows the scale of Nvidia stock's performance.  

Thoughts on Affordable Housing XII

Social rental housing is a significant share of the total housing stock across many European cities. The OECD data puts the share at 8% for the EU as a whole, 34% for the Netherlands, 24% for Austria, 16% for the UK, 14% for France, and 3% for Germany.

Paris, under its Socialist mayor, Anne Hidalgo, is making efforts to buy land and expand its social housing stock to 40% of primary residences by 2035. About 30% will go to people with no or very low incomes, and 10% for those on middle incomes. And it’s a radical plan.

To get there, the city government led by Socialist mayor Anne Hidalgo is expanding its arsenal of interventionist tools that it casts as the only way to correct market failings, especially those wrought by short-term tourist rentals and vacant apartments. Paris is also doing more “pre-emption deals” in which it blocks a planned real estate transaction to instead buy the building itself to convert into social housing. Under its new land use plan, the city has labelled some 800 buildings, one-third offices, as candidates for expropriation. Developers must also include a significantly higher proportion of social housing — half in cities deemed in “hyper-deficit” — in all new-builds. Even office owners are now required to add social housing units when they build or do major renovations…

In Paris, the lack of buildable space, prohibitions on building height, and historical protection rules make it difficult for any developer to build new housing. Private companies and institutional investors have largely been scared off by rent controls, poor yields and rental bans that are slowly being phased in on homes with poor energy efficiency. For these reasons, converting existing structures like offices, car parks and government-owned buildings into new housing has been a welcome solution. Paris is also putting pressure on rich areas, such as the 6th, 7th and 16th arrondissements where the proportion of state-subsided housing ranges from 2 to 7 per cent, compared with 42 per cent in the poorer 19th and 20th arrondissements…Paris has recently set an annual housing budget of around €800mn a year, nearly double that of five years ago, and four times the amount budgeted by the central government.

Sample this illustrative example of how these efforts work.

The car park conversion on rue Nollet in the 17th arrondissement is one of about 40 such projects the city has undertaken, which also dovetails with its efforts to reduce cars in Paris. The city paid for more than a third of the €12.3mn cost with the rest coming from 40- to 50-year-long state-backed loans, without which the conversion would not be economically viable. The agency that built and will operate the building under contract for the state expects a roughly 40-year payback time, far longer than commercial developers would accept.

Paris is part of the French government’s efforts to promote social housing,

In 2000, the government enacted a law, known as the SRU, which required many cities and towns to reach 20 to 25 per cent social housing by 2025. The criteria were set to spare very small towns, but some mayors still think they are too rigid. In addition to billions of euros in government subsidies, the pivotal player has been the Caisse des Depots, a state-owned financial institution, that provides roughly two-thirds of the funding to build. The money for these long-term guaranteed loans comes from the popular regulated savings account called the Livret A, in which most French citizens invest. 

European countries have traditionally pursued policies that favoured the development of social housing units.

Places like the Netherlands and Austria, which have century-old systems that have relied on government-owned developments and non-profit housing associations, have also intervened more in the market than places like the UK or Germany where a bigger role is given to commercial developers… In the UK, after a wave of postwar construction, much of the council housing stock was sold off to the residents in “right to buy” schemes starting in the 1980s. The country now largely relies on a requirement for developers to add affordable units in new-builds, with a typical request in London of a minimum of 35 per cent. But developers can also do less by instead building public facilities like parks or municipal pools…

Vienna and Amsterdam are among the cities that have gone the furthest to reaching over 40 per cent of social housing. From the outset, they avoided a key pitfall: they built homes in city centres instead of on the periphery, thus avoiding the gradual development of poverty traps and high crime areas. European cities that built social housing on the outskirts became plagued with problems. Paris was especially bad. Over time, the banlieues, or suburbs, became home to new migrant communities and gradually lost their original mix of residents from different socio-economic classes. It has left those living there with poor transport links and sub-par schools and public services, leading to the periodic explosions of social unrest. “It was wrong to focus social housing on the edges of cities since it led to segregation and disconnection for the people who live there,” says Tadashi Matsumoto, the head of sustainable urban development at the OECD. He believes housing policies should not just be city-based, but co-ordinated with nearby suburban communities where there is more space. 

But social housing comes with its own set of problems.

But while increasing social housing in cities can help more people on modest incomes find a place to live, there are also downsides when private developers are constrained and the private rental market shrinks, says van Bortel, the Delft academic. In Amsterdam, where three-quarters of rentals are social housing or subject to strict rent controls, the rest of the private market has become very expensive as owners cannot raise rents elsewhere… those who are lucky enough to get subsidised Parisian apartments become loath to leave — only 6 per cent of apartments in the social housing stock change residents each year. The city does not force residents to move even if their incomes increase or if their family size changes.

A few observations:

1. It’s interesting that the cities in rich countries solve their affordable housing problems by constructing social housing, whereas developing countries tend to pursue market-based policies on affordable housing. 

There is no concept of public-supported rental housing in India, and all subsidised housing stock would be a tiny proportion of all urban housing stock. Even in slums, housing supply is market-driven. This is unlikely to change given the fiscal constraints facing governments.

2. Given the scarcity of land in urban areas and the absence of high-rise public housing, the weaker section public housing programs (like the IAY and PMAY) are largely confined to smaller towns and in the far-flung suburbs in the case of the larger cities, where there’s land available to construct single or low-rise units. Even in the largest cities, these low-rise (3-4 floors) colonies are geographically distant (from the city) and socially and economically cut off from it. They have become the Indian version of Parisian banlieues. 

3. In the absence of social housing, there are limits to what can be done to make housing affordable, especially given the already massive and further widening housing affordability gap in the large cities. Other policy concessions like credit subvention schemes (like the CLSS) or planning mandates (like 5-10% affordable housing mandates in large developments), only bridge the gap by a small amount to benefit the Indian middle class. None of these concessions can bridge the gap sufficiently to meet the requirements of even the lower middle-class, much less the poor. This reality must be acknowledged. 

4. In this backdrop, with public housing likely to play a marginal role for the foreseeable future, the only meaningful way to address the problem of affordable housing in Indian cities, especially for the lower-income group of urban immigrants (where household income is less than, say, Rs 20,000), is to bring all policy levers together. 

This would entail making public land available at subsidised rates, increasing the buildable area within the land, and lowering the cost of construction. Given the limited extent of public lands available, the first lever would be limited. However, supply can be augmented by using planning mandates on earmarking land for affordable housing. I’ll blog on this separately. The second lever can significantly offset the land cost by multiplying the extent of buildable area within the available land. 

Reducing the construction cost would require significantly lowering land registration and planning permission fees, GST, subsidising the cost of capital, and expediting all approvals and permissions for all affordable housing construction. This could be supplemented with some property tax concessions for, say, five years. A proposal for a scheme along these lines for state governments is discussed here. The revenue loss from these concessions should be counted against the subsidy that should have been required for social housing. 

5. Finally, like the European cities, there’s a need to focus on rental housing in urban areas. The Government of India came up with an Affordable Rental Housing Policy in 2015 and the Affordable Rental Housing Complexes (ARHC) scheme in 2020. Both need prioritised attention. However, it may not be advisable to leave the catalysis of the market in affordable rental housing to the private sector alone. State governments may have to play a direct role in derisking them by having State Housing Boards and public entities develop, lease, and manage such properties. The Government of India could incentivise with interest subvention and even some direct subsidy for capex. Once derisked, this can become a high-volume real estate segment, one that genuinely serves the Bottom of the Pyramid market.

Weekend reading links

1. Emaar, the Dubai property developer whose major shareholder is Dubai Investment Corporation, is considering expanding outside UAE. And an interesting possibility in India.

Emaar already has a sizeable presence overseas, owning more than 175mn sq ft of land outside the UAE at the end of 2024, excluding a 1.1bn sq ft “economic city” project in Saudi Arabia. Alabbar said Emaar’s current total land bank — an industry term referring to land owned by developers and reserved for future use — was at 1.87bn sq ft, including the UAE. India, where Emaar owned 122mn sq ft of land at the end of 2024, could provide a testing ground for Emaar’s overseas strategy and its subsidiary there was discussing a potential joint venture with local developers, including the Adani Group, Alabbar said. Alabbar dismissed reports that Emaar had been discussing the sale of its business in India to Adani. “We’re not selling,” he said. “We actually were looking for local partners to do a local [joint venture].” Adani did not respond to a request for comment.

2. An illustration of how worsening demographics are going to hit Japan's logistics industry.

3. Private equity funds face declining exits and are resorting to the likes of continuation funds. 

In the first half of this year, private equity firms sold companies they owned back to themselves at a record-setting pace, providing a way out of (or back into, whichever you prefer) some $41bn of investments in the first six months of 2025, according to investment bank Jefferies. That is close to a fifth of all sales in the industry, and is 60 per cent above the level last year, and it comes as private equity groups find themselves sitting on $3tn worth of assets that they are unable to get out of, either by selling to another company or by listing them... Torsten Sløk at private markets group Apollo, noted this week that on top of the steady flow of companies delisting from public stock markets, those companies that do opt for an IPO “are getting older and older”. (Sigh, aren’t we all?) “In 1999, the median age of IPOs was five years,” he wrote. “In 2022, it was eight years, and today, the median age of IPOs has increased to 14 years.”

4. France is facing a pensions crisis. However, the political economy of reform is perilous. 

Not only do French pensioners receive larger cheques from the government than their counterparts anywhere else in the west, they start getting them several years earlier. The result is a situation in which over-65s now have higher average incomes than the working age population — unique both internationally and in France’s own history. Even the rumour of threats to this arrangement is met with mass public outrage and opposition from left and right. Macron’s proposal to nudge the retirement age up towards the lower end of western norms was met with nationwide protests. Barnier’s suggestion of a six-month delay to the latest scheduled increase in pension payments led to the first of two collapsed governments in the past 10 months. Bayrou’s refusal to scrap the same pledges brought about the second. In a particularly stunning statistic highlighted by French political analyst François Valentin, pensions play such an outsized role in the country’s public finances that they accounted for one-sixth of the ministry of defence budget last year, and without them France would not meet Nato’s 2 per cent target for military spending.

5. The latest ASI data show that the share of contract labour in India's organised manufacturing sector workers has reached 42% in 2023-24, up by 8 percentage points in the last ten years. Such contract labour have lower wages and no benefits (paid leave, social security benefits, longer tenure), and are typically hired through third-party agencies. 

Global studies have shown varying levels of contract employment in different countries. A 2023 study pegged it at 10.8 per cent in the US. Latin American countries like Brazil and Argentina have seen numbers ranging from under 10 per cent to 20 per cent at various times, according to a 2016 study, which pegged the number in Europe at 12.3 per cent.

6. Problems with GDP estimations in India. 

7. As Pakistan and Saudi Arabia sign a landmark mutual defence pact, this is an important insight.

Since the 1960s, Pakistan has received more aid from Saudi Arabia than from any nation outside the Arab world, the Brookings Institution estimated. The funding — which was never directly for support of Islamabad’s covert nuclear programme — included direct aid to the government as well as financing for schools, mosques and other Islamist charitable programmes.

8. Trump finally imposes a $100,000 fee for H1B visa applications for skilled foreign workers, the main pathway through which Silicon Valley firms hire engineers and IT professionals. The current visa charge is $215 to register for a H-1B visa lottery and an additional $780 for employers that sponsor visa applicants. Abut 400,000 such visa applications were approved last year, with the majority for those renewing their visas and 75% in 2023 being from India. 

Between October 2022 and September 2023, 72 per cent of the nearly 4 lakh visas issued under the H-1B programme went to Indian nationals. During the same period, top four Indian IT majors with a presence in the US — Infosys, TCS, HCL, and Wipro — obtained approval for around 20,000 employees to work on H-1B visas, as per the latest US Citizenship and Immigration Services (USCIS) data. The share of IT workers in the H-1B program grew from 32% in 2003 to an average of over 65% in the last five fiscal years.  Given the average salary of the H1-B visa holder is $66,000, the fee is certain to almost finish the use of these visas. 

Goes to the point I have been making on Trump's completely transactional policy making. It's important that India keeps this in mind as it navigates a modus vivendi with the US under DJT. Unlike China or Russia, India is not big enough either economically or militarily to deter Trump from acting on his core agenda. Besides, India falls in the firing line on many of those core areas. We should not be surprised if GCCs are the next. 

9. One of the definitive legacies of the Russia-Ukraine war will be the emergence of drones as an important weapon. This is a good summary of how the war evolved in terms of what weapons were being used.

Back in 2022, when Russia started its full-scale invasion, Kyiv had to use its existing Soviet-style kit plus Javelin shoulder-fired anti-tank missiles. Then came western donations of weaponry like Abrams tanks and Himars (high mobility artillery rocket systems). Next, Ukraine’s army of software engineers started using hobby drones, made by Chinese companies such as DJI, first for surveillance, then attacks and defence. Now they are innovating to dramatically extend drone flight range, increase attack capabilities, “swarm” and avoid electronic jamming by using fibre optic cables, balloons and (most crucially) AI. The Russians are doing the same. And that has transformed the nature of war: a world where cheap drones can destroy ultra-expensive ships and planes changes the power dynamics and economics of combat...

Equally startling, while China has been responsible for 80 per cent of global drone production, Ukraine is now racing to become “China free”... Last year it produced more than 2mn drones. It could go above 10mn next year, if it has the funds. That means over half of Ukraine’s drones are now domestically sourced — and China is no longer the only global drone king. This is critical for Ukraine’s defence, and might generate badly needed future export revenues too. Indeed, Ukraine is already considering exporting underwater drones, which it has used to push Russian ships out of the Black Sea so successfully... Nato officials now want to collaborate with Ukraine via partnerships, licensing and private capital investments. They are particularly keen to access the treasure trove of data collected by its drones to train future AI models.

10. A rare example of an infrastructure project completed on time and without much cost overrun is the Thames Tideway Tunnel, a 25 km sewage tunnel passing under Thames. It took nine years of construction and cost £5 billion, and used an innovative financial model, where a surcharge on customer bills will fund the construction cost over 50 years. 

Tideway was project managed by US group Jacobs, while each of the three geographical sections was overseen by a different engineering consultancy. Each employed “hundreds and hundreds” of contractors and subcontractors supplying the labour. All 25,000 people working on the project were fully employed and paid the London living wage as a minimum, in contrast with much of the industry, where staff are employed on zero-hour contracts or by the day. Workers were also kitted out in full Tideway uniform and protective equipment and given training and holidays... Juliano Denicol, director of the Megaproject Delivery Centre at UCL, said Tideway would be studied for decades as an “exemplar case of ‘how’ to deliver multibillion infrastructure programmes”...

Tideway, which as a company was able to lure investors including Allianz, Dalmore Capital and Amber Infrastructure after the government offered guarantees against any cost overruns. The 16mn households served by Thames Water have paid £617mn towards the project since construction began nine years ago, more than the £510mn in equity injected by Tideway shareholders in 2015. The shareholders also loaned the company £764mn at an interest rate of 8 per cent, which has grown to £972.6mn, while the company’s total debts have risen to £4.6bn, with the remaining borrowings held by third-party lenders. Although construction is complete, Thames Water’s households will continue to pay for the project through a surcharge on their bills — now £26 a year — potentially for the 120-year lifespan of the tunnel. The cost — drawn up under a modified version of the Regulated Asset Base model — piles pressure on customers, who are already faced with steep bill increases from the UK’s largest water company, which could yet be renationalised.