Infrastructure project delivery and state capability

An important misplaced belief that has become entrenched in infrastructure project delivery is that of substituting public system capabilities with outsourced project management consultants (PMCs) and independent engineers (IEs). 

In India, commentators misleadingly attribute the successes of NHAI or DMRC or the few other successful projects to private participation in general and particularly the adoption of private sector management practices and outsourcing of services to consultants. They overlook the critical importance of the role played by the engineers, planners, and bureaucrats in these organisations, the processes followed, and the autonomy enjoyed by these project entities. 

Echoing these important requirements, Alon Levy at Pedestrian Observations has a very good post on infrastructure project delivery practices, where he makes the distinction between the models of modern management theories-based delivery (followed in US and UK) and the traditional public sector-driven delivery (followed in continental Europe). 

The takeaway is that effective project delivery “requires an active public sector that can supervise consultants and contractors, learn within its own institutions, and assume risk”. Even when project delivery is largely done by private consultants, they are done “under public-sector supervision, with institutional knowledge retained in government agencies even in an environment of privatization”. 

If there’s a common theme to the various elements of Southern European (and largely also French and German) urban rail procurement norms, it’s that they require an expert civil service. Teams of engineers, planners, architects, procurement experts, and public-sector project managers are required to manage such a system, and they need to be empowered to make decisions. This empowerment contrasts with American public-sector norms, in which to a small extent in law and to a very large extent in political culture, civil servants are constantly told that they are dregs and cannot make any decisions. Instead, they are bound by red tape requirements that can only be waived if a political appointee wants to take the risk… 

The idea of listening to engineers and planners is denigrated as siloing, whereas generalist managers with little knowledge are elevated to near-godhood… In contrast, it is less important how many civil servants are hired to supervise contracts than that they have the authority to make judgment calls and that they do not have to answer to an overclass of generalist managers. Italy and France use very large bureaucracies of planners and engineers at Metropolitana Milanese and RATP respectively… Once the civil servants can make decisions and supervise contractors, they can look at bids and score them technically, or delve through itemized lists, or oversee changes and make quick yes-or-no decisions as the builders are forced to vary from the design… making this the most significant single intervention in reducing infrastructure construction costs.

Levy lists out some of the good practices followed by Southern European countries, which have a track record of delivering large projects on time and within budget. While he discusses in the context of metro rail systems, these principles apply to infrastructure projects in general. 

• Technical scoring: infrastructure contracts must be awarded primarily on the technical score of the proposal (50-80% of the weight of the contract) and not on the cost (maximum 50%, ideally about 30%)

• Itemized costs: contracts must have a bill of items, priced based on transparent lists produced by the state, with change orders using the same itemized list to reduce conflict

• Separation of design and construction into two contracts (design-bid-build), rather than bundling into design-build contracts

• Public-sector planning, with the decisions on the type of project and technology made before any designers are contracted

• Flexibility for the builders to vary from the design, so that in practice the design only covers 60-80% of the design, as 100% design is impossible underground until one starts digging

• Moderate-size contracts (tens of millions of dollars or euros to very low hundreds), to allow more contractors to compete

• Limited use of consultants, or, if consultants are used, regular public-sector supervision

The orthodoxy on infrastructure project management today is characterised by certain core principles drawn from management theories - efficiency maximisation (bundling of design and construction), life-cycle cost approach (bundling of construction and O&M), private sector efficiencies (PPPs and long-term concessions), outsourcing services to professional experts (PMCs, IEs), etc. Levy attributes this dominance of modern management theories to the soft power of “English-speaking multinational consultants with extensive experience in megaprojects” and the belief that they knew better than the continental European state. 

As I have blogged on numerous occasions (and written in detail here), contrary to conventional wisdom, it’s technically prudent to separate design and construction in most cases, though with some significant design flexibility during the construction phase; it is financially prudent to separate construction and O&M, so that the O&M contractor does not bear any construction risk; it is cheaper to construct with public finance and then do concessions; and leverage and PPPs must be used only where there are private sector efficiencies to be had and never merely as a means to address fiscal constraintsand avoid large upfront costs. 

A very important, and under-appreciated, aspect is that of effective contract management by the government entity. It’s not uncommon for the consultants employed for the project, including the IE or the third-party quality audit firm, to be captured by the project developer. Given the financial stakes involved, the political economy of infrastructure construction, and the tightly knit ecosystem involving the project developers, consultants, IEs, and other service suppliers, the incentives are strongly aligned to collude at the margins. Strict internal supervision and monitoring become essential to address this problem.

Fortunately, the use of IT applications makes it possible to significantly increase the quality of oversight and remote monitoring of large projects. However, this must go beyond mere videography, biometric attendance, location tracking, GIS mapping, data acquisition systems, and so on, and involve deep data analytics of the data so captured to identify outliers and trends of systematic manipulation. Getting this done and effectively using it to enforce contract and execution discipline is deeply dependent on internal capabilities. 

An important takeaway here is the capabilities of the internal project team and the autonomy given to them. Perhaps the most important requirement for the success of any large project delivery is ensuring that the project team is adequately staffed with engineers, planners, and leaders who are competent and have high integrity. Any relaxation on this, especially in important positions, is a recipe for certain failure. Nothing is more important than this requirement. Unfortunately, this is often the area where governments tend to compromise the most.

Weekend reading links

1. Beneficiaries of George Soros and his Open Foundation.

Among the beneficiaries is Hungary’s Viktor Orbán whose Oxford scholarship was paid by Soros in 1989. Talk about no good deed going unpunished. Another kind of beneficiary is Scott Bessent, the US Treasury secretary, who ran Soros’s hedge fund for many years. Soros was the anchor $2bn investor in Bessent’s own hedge fund, Key Square Group, in 2015.

2. Chinese companies produce many AI tech components.

3. Yogendra Yadav reviews Partha Chatterjee's new book, For a Just Republic: The People of India and the State. 

4. The US tariffs latest update.

5. India's IT industry facts of the day

The top five Indian IT firms had free cash flows of nearly $13bn in the 2023-24 financial year, according to HFS Research. And Infosys said on September 11 it had approved a $2bn share buyback offer — a week before the Trump order. Yet the R&D to sales ratio for India’s IT industry is abysmal: 0.88 per cent on average, according to a 2024 report by India’s Ministry of Corporate Affairs.

6. China moves to restrict Ericsson and Nokia equipment in their telecom networks. 

Chinese state-backed buyers of IT equipment — which include mobile network operators, utilities and other industries — have begun more closely analysing and policing foreign bids. That process has required contracts by Sweden’s Ericsson and Finland’s Nokia to be submitted for “black box” national security reviews by the Cyberspace Administration of China where the companies are not told how their gear is assessed. The reviews by the powerful tech watchdog can stretch three months or longer. Even in cases where the European groups ultimately secure approval, the lengthy and uncertain audits often leave them at a disadvantage to Chinese rivals that face no such scrutiny, the people said... Beijing’s growing sales restrictions have collapsed Ericsson’s and Nokia’s combined market share in China’s mobile telecoms networks to about 4 per cent last year from 12 per cent in 2020.

Amidst these moves, Europeans have been half-hearted in their efforts to restrict Huawei and ZTE. 

Huawei and ZTE have retained 30 to 35 per cent of the European mobile infrastructure market, down only 5 to 10 percentage points from 2020, data from Dell’Oro Group shows. Germany has 59 per cent of installed 5G gear sourced from Chinese groups, according to John Strand of Strand Consult, even though the country plans to phase out high risk Chinese vendors by 2029.

7. FT writes on the wealth of the super-rich

When Forbes magazine released its first global billionaires list in 1987, just 140 names appeared on it. The 2025 version featured more than 3,000 people, worth a collective $16tn. Even allowing for factors such as the rise of China and over three decades of inflation, it is a staggering increase in both numbers and values; the net worth of Elon Musk, judged the world’s richest person in April 2025, was estimated at $342bn — compared with $295bn for the entire class of 1987. Globally, the average wealth of the top 0.0001 per cent of the population grew on average 7.1 per cent a year between 1987 and 2024, compared to 3.2 per cent for the average adult, according to Gabriel Zucman, a professor of economics at the Paris School of Economics and at the University of California, Berkeley... The top 400 wealthiest Americans had a total effective tax rate of 23.8 per cent of income in the years from 2018 to 2020, including individual income taxes, estate and gift taxes, and corporate taxes. In comparison, the rate for the wider US population was 30 per cent, rising to 45 per cent for the highest-earning workers.

Historically, asset-based taxes were the main revenue source for governments. Taxes on income in the UK, for example, were a mid to late-20th century phenomenon closely tied to the emergence of a welfare state. Now, as demographics worsen (with fewer worker and more retired people), the case for wealth taxes is becoming more compelling. 

8. How housing prices in the UK have changed over the last 35 years. 

9. GCCs are cannibalising the business of India's IT services firms.

As GCCs grow, they are eating into the pie of IT services majors, both in terms of business and skilled talent... Out of the 200,000 tech roles in India in FY25, approximately 120,000 were in GCCs, with a 10–15% year-on-year growth, said Vikram Ahuja, co-founder of ANSR, a GCC solutions platform... The real evolution started with traditional companies coming in to set up true capability centres, like [department-store chain] JCPenney, [luxury-superstore chain] Saks Fifth Avenue, [lingerie retailer] Victoria’s Secret…They have no business to be experimenting with this concept. All airlines, hotel chains, car-rental companies are coming. So, it’s become industry agnostic. On the contrary, the more low-tech and the more traditional you are, the more the need [for a GCC]... Lloyds has hired over 2,500 engineers in Hyderabad within 14 months, with 95% focused on tech. At Barclays’ India operations, two–thirds of its tech workforce is now in-house—a stark jump from just about 33% a decade ago. Even Indian lenders are following suit. Just six months ago, RBL Bank achieved a 60:40 split between in-house and outsourced tech talent—a significant leap from the 35:65 ratio of a few years ago.

A major reason for the exit is the low and stagnant wages paid by IT services firms, even as the scope of work expands. 

9. Private equity is taking over India's IB and IGCSE school chains. And results are not good.

Private equity firms are betting big on Indian education, and their playbook mirrors a Western model—optimised for cost control, standardised for scale-up, and centralised for effective management... Both CBSE and state-board campuses face tighter fee caps and myriad state-level approvals. International boards like International Baccalaureate (IB) and Cambridge have a wider fee latitude and can levy “development” charges, creating room for upgrades and margins. The model makes money within India’s nonprofit rulebook. The school usually sits in a Section-8 entity to satisfy K–12 regulations. A for-profit services arm—typically charging 10–15% of school revenue through the likes of management fees, royalties, and infrastructure leases—operates on the side... 

Investors like GSF fall back on the same approach with each school: centralise leadership, trim excess, introduce standardised systems, and make visible infrastructure upgrades. But beneath the surface, the effects of this strategy vary sharply... While fee hikes have remained within the standard 5–10% range at premium schools like Sancta Maria (Rs 7–8 lakh in annual fees)—already operating near permissible ceilings—some lower-fee campuses could see steeper increases... Infrastructure investment varies significantly by operator and campus... The international school Manthan in Hyderabad saw 60–70% staff attrition after a 100% ISP acquisition... At TIPS Coimbatore, acquired by Globeducate, 20% of the staff left after the founder exited... Student numbers, too, fell by nearly a fifth at Glendale and Oakridge in the years after their acquisition... Pre-acquisition pay rises of 8–15% have been slashed to 2–7% under new management—standard practice in the West, but a sharp adjustment in Indian schools... The result: well-trained, experienced teachers leave, and classroom quality drops... The same Western-school playbook that made these operators successful abroad doesn’t map cleanly onto India’s hyper-competitive, founder-led education landscape.

How JVs with US MNCs strengthened Chinese manufacturing (and weakened the US)

Amidst all the concerns about China’s weaponisation of its manufacturing dominance, we tend to overlook the critical role played by Western multinational corporations in helping China develop those capabilities. 

I had blogged earlier about how Apple and the iPhone made China. As this Bertelsmann Institute report shows, China has been the standout beneficiary of the era of globalisation spanning the last three decades. 

In this backdrop and as US companies shift their operations away from China, a new paper by Jaedo Choi, George Cui, Younghun Shim, and Yongseok Shin reveals how joint ventures by US companies in China have strengthened Chinese firms, reduced US welfare, and benefited large US firms at the expense of small firms and the real wages of workers. 

They show that, notwithstanding the risks of technology leakages, the US companies, by and large, voluntarily complied with the Chinese policy that explicitly or implicitly mandated technology transfer by MNCs through JVs. This ultimately led to over-investment in the JVs and excessive technology transfer to China, resulting in profit losses for other US firms due to intensified competition from China. 

One novel finding is that US firms experienced more negative outcomes in industries with more joint ventures in China… Our quantitative analysis shows that there are indeed too many joint ventures in equilibrium relative to the social optimum for the US… we find direct positive spillovers from MNEs to Chinese parent firms (or partners) of joint ventures… Following the formation of joint ventures, Chinese parent firms experienced significant growth in sales, capital, and exports. Furthermore, their patenting activities became more similar to those of their MNE partners, indicating a direct diffusion of technology between partners through joint ventures… we find evidence of indirect spillovers to other Chinese firms. In industries with more FDIs (joint ventures and wholly foreign-owned enterprises), even the Chinese firms that were not a party to a joint venture grew faster and more technologically advanced… we find that in industries with more FDIs into China, US firms experienced more negative outcomes in terms of sales, employment, investment and innovation.

They use a two-country growth model to estimate the impact of such joint ventures. 

Once a joint venture is established in China, the probability of technology diffusion from the US leader firm to the Chinese leader firm increases, consistent with our empirical finding of direct spillovers. As a result, the surplus from a joint venture includes not only the flow profit of the joint venture firm but also the value of the higher probability of technology diffusion to the Chinese leader firm. Additionally, Chinese fringe firms, which do not participate in any joint venture by construction, benefit indirectly. This is because there is an additional source of technology diffusion—the joint venture firm itself—within the industry, and the Chinese leader firm is likely to have higher productivity after forming the joint venture… 

The entry of a new joint venture firm immediately intensifies competition in the industry. The stochastic technology diffusion to the Chinese leader and the fringe firm further intensifies competition over time. The US leader takes all these competition effects into account when making the joint venture decision. It also partially captures the profit flow of the joint venture and the spillover benefits to the Chinese leader through bargaining. However, it ignores the negative effects of heightened competition on the profits of its domestic competitor, the US fringe firm…

US leaders benefit from joint ventures in the short run through lower trade costs for serving the Chinese market and lower wages in China. They also partly capture the value of technology transfer to Chinese leader firms through bargaining. Over time, however, Chinese firms catch up faster due to the technology diffusion facilitated by these joint ventures, and the heightened competition negatively affects US leaders. Nevertheless, the present discounted value of US leaders’ profits is higher with joint ventures—otherwise, they would not invest in them. For US fringe firms, leader firms’ joint ventures have only a negative effect on their profits, through intensified competition from China. Because US leader firms ignore this negative effect on US fringe firms, there may be too many joint ventures relative to the US social optimum.

They point to important impacts of JVs on innovation and comparative advantage.

On the one hand, the increased probability of technology diffusion to China means that profits from successful innovations are smaller and shorter-lived, which may reduce innovation efforts. On the other hand, the option to form a joint venture makes US leaders innovate more, because their innovation increases profits from the joint ventures and the fees they receive from Chinese leaders through bargaining. In our quantitative analysis, the former dominates in the medium to long run, so US leaders innovate less with joint ventures. For Chinese leaders, technology diffusion serves as a substitute for their own innovation efforts, and they innovate less with joint ventures.

Furthermore, in the model, the value and hence the likelihood of forming joint ventures for US leaders are higher when the US-China technology gap is larger, which we confirm in the data. Since joint ventures reduce the technology gap between the US and Chinese firms through technology diffusion, they have the effect of eroding the US comparative advantage and terms of trade, reducing the gains from trade for the US.

The authors use their model to calculate the short and medium-term impacts of a ban on JVs from 1999.

We find that prohibiting joint ventures increases US welfare by 1.2 percent in units of permanent consumption. For the US, leaders’ profits fall by 22 percent in present value terms, while fringe firms’ profits increase by 4.9 percent. The total profit of the corporate sector declines. Yet, the real wage increases by 2.9 percent due to higher labor demand in the US, leading to the overall welfare gain. The ban has a transitory negative effect, because US firms cannot immediately benefit from lower wages in China and reduced trade costs via joint ventures. However, this effect is outweighed by medium-run benefits, as the US maintains its technological advantage over China for longer, driven by higher innovation efforts and less technology leakage to China.

As for China, when the US bans joint ventures, Chinese leader firms compensate for reduced technology diffusion by increasing their own innovation efforts. However, China’s productivity growth is substantially delayed, and the absence of joint ventures reduces China’s welfare by 10.3 percent in units of permanent consumption. In China, the profits of both leaders and fringe firms, as well as the real wage, are lower without joint ventures from the US.

The role of the JV model of technology transfer is an under-appreciated aspect of China’s economic growth. More the technology gap, the greater the economic benefits to the host country for the investment through the JV strategy. 

A JV mode of technology transfer, especially in the scale and manner that China has achieved (the iPhone is the best illustration), is not replicable. No other country has the political system and economic advantages that enable its firms to undertake the kind of hard bargaining required to extract JVs that involve technology transfer. There are at least two aspects.

Only an authoritarian one-party system like China can summon the level of coordination and discipline required to force such JVs on foreign multinational corporations. Complementing this, the Chinese firms, too, (with the guidance of the Government) have shown an unmatchable level of commitment and enterprise to use the JV as a springboard to vault them into global leadership in their industry. It’s this combination of the two, government and private sector, working together, that made these JVs succeed in technology transfers and quickly move Chinese firms up the value chain. This is Public Private Partnership (PPP) with Chinese characteristics. 

On the face of such evidence, it is not tenable for the US (and European) firms to continue with such JVs even as China tightens its dominance in manufacturing and the Cold War intensifies. The restrictions imposed by the US on multinational corporations’ activities in China in strategically important sectors must be viewed in this context. This should also be another reminder to all remaining supporters of continuing normal trade and investment activity with China (academic scholars who promote free trade, and corporate interests intent on maximising their private gains from such trade). 

The normal rules of the game on the economy and trade do not apply when faced with a political and economic system like that in China. Even those arguing that China will gradually vacate the lower-value-added manufacturing and move up the value chain are most likely wrong, since, given the continental size and diversity (in terms of stages of development) of the economy and its unparalleled competitive advantages, China is unlikely to vacate them in any meaningful manner for a long time. 

The other takeaway from this is that while it may not be possible for other countries to emulate China’s approach to JVs with foreign companies in full, it’s useful to make them an important instrument of their industrial policies. No country is closer to China in this respect than India. It has many of the economic advantages that China offers to foreign MNCs, which make it acceptable enough to agree to such JVs. 

In areas like defence equipment, aeroplanes, renewable energy generation, metro railways, smart meters, telecom equipment, surveillance cameras, and so on, the Indian market is among the largest for foreign firms. It provides the government sufficient bargaining power to single-mindedly pursue the objective of maximising technology transfers through JVs. It should show the coordination, discipline, and persistent pursuit required to achieve the objectives. 

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.