A speculative note for institutional fixed income portfolio managers in India
19 August 2026
Fixed income managers are professionally allergic to speculation, and rightly so. Fixed income, as an asset class, rewards those who harvest small and predictable edges such as carry, roll-down, auction tactics, and some limited spread trades and relative value. It punishes those who bet the book on grand macro narratives. That instinct is correct ninety-five percent of the time.
This note is about the other five percent.
TL;DR: Nominal rates will rise, credit spreads will widen, and the spread between rating categories will widen even more. Data center debt will be worth owning – but be selective about which data centers.
AI is causing a lot of changes in the physical world – and those changes are happening fast enough to matter within the maturity of bonds already sitting in your HTM book. Multiple highly capable companies are racing to build artificial general intelligence (AGI) – systems that match or exceed human performance across most economically valuable cognitive work. On a growing set of benchmark tasks, frontier AI models already exceed the performance of expert human knowledge workers. Whether AGI arrives in 2027 or on any particular date, the capital expenditure is not hypothetical. Hyperscalers are committing hundreds of billions of dollars annually to compute infrastructure. Data center campuses of 2 gigawatts and above, with announced designs reaching 5 gigawatts – each consuming as much electricity as a mid-sized Indian state’s residential load – are under construction. And critically for us, India is now fully onboard this change. Reliance has announced a roughly ₹10 lakh crore AI infrastructure program anchored on multi-gigawatt data centers in Jamnagar. AdaniConneX already has 2GW capacity, and plans to expand to 5GW encompassing Adani Green Energy’s 30 GW Khavda project in Gujarat. Google is building a gigawatt-scale AI hub in Visakhapatnam with a $15 billion commitment; Microsoft and Amazon have announced investments of similar order. The Government of India (GOI) expects over $200 billion of data center investment in the coming years. GOI has already offered a tax holiday till 2047 for foreign cloud providers using India based data centers. The current 10 year GSec (6.94% 2036) will live its entire life inside this transition. We posit that this transition will change equilibrium interest rate, GOI’s fiscal position, the distribution of corporate default risk and the duration of insurance liabilities. And we suddenly realize that it is no longer a speculation, but the duty of care to understand and adapt to this change. In this note, we anticipate the effects of the “boring scenario” where the AI technology works and then diffuses into the economy, and then the economy reorganizes around AI, the way the economy reorganized around the electricity, the internet and the smartphone.
Part 1: AI’s diffusion into the economy
A – AI raises productivity of invested capital
While the investments into AI technology itself might have more misallocation, the general entrepreneur will use AI to make more productive capital investments and time allocations. And it is important to distinguish between the risk-reward profile of the limited partner at a VC firm, and the general AI-assisted entrepreneur. Surely a few data center projects will stall, and perhaps a few gigawatt scale data centers will not utilize their full capacity, and a few AI startups will fail entirely. But, that capital is Risk Capital of limited partners and private equity, who stand ready and able to absorb those losses. Those losses might cause increased risk aversion in the US, but we think that the losses will not spill over any pain into the Indian economy. What we are truly optimistic about is the general utilization and allocation of capital by the micro and small enterprises. The excess returns of these AI-guided investments will likely get competed away and gains will accrue to consumers via lower prices. The disinflationary effect of these AI-precipitated cost reductions will be muted by the demand pressure on energy. So, total factor productivity will rise because there will be fewer bad investments, and less capital wasted on doomed ventures, mispriced projects and vanity capex, thereby raising the returns on invested capital. Some of the AI guidance might end up wasting money, but the modern entrepreneur will be all in. Moreover, structurally, India has a massive infrastructure deficit, which also will compete with the AI-led investment for capital.
Overall, we think the investment-demand shock is large enough and frontloaded enough that the real rates rise by 100-200 basis points on the net over the next decade.
B – Depreciation of human capital and appreciation of physical and financial capital
When a substitute for human cognition becomes available at a marginal cost approaching the price of electricity, the scarcity value of human cognition falls, and the scarce complements – compute, power, land, grid connections, water, and the financial capital that mobilizes them – appreciate. This is the deep logic behind why data centers, energy assets, and “dry powder” all gain relative value.
We cannot resist highlighting Moravec’s paradox: AI finds the hard human things easy (reasoning, coding, analysis, writing, etc.) and the easy human things hard (general labor, plumbing, nursing, driving on Indian roads, laying transmission cable, etc.). Cognitive white-collar labor depreciates early and physical labor depreciates much later. Wages for electricians and welders may rise during the buildout decade even as wages for junior analysts and programmers fall.
India’s roughly $250+ billion a year of exports encompassing software services, BPO, and GCCs, plus a large share of remittances earned by NRIs abroad sits exactly in the depreciating segment. US Immigration visa situation will worsen, and hope from India-UK’s CETA is also dim. This is not a marginal exposure; IT services are one of the largest sources of India’s current account resilience and the anchor employer of its urban middle class. We are carefully watching headcount and revenue-per-employee at the large IT firms.
The macro-financial cascade is: services export growth stalls, then current account widens, then rupee depreciates structurally, leading to imported inflation (oil, gold, GPUs), leading to RBI defending INR with higher rates, leading to higher GSec yields, especially at the long end where currency risk premia are highest.
C. Fiscal stress for the government
AI stresses GOI budgets from both sides.
On the revenue side, human cognitive labor depreciates and so, the personal income tax base erodes. India’s direct tax collections lean heavily on the salaried, urban, services-sector professionals (with salaries of INR 1 lakh per month and higher) most exposed to AI-led depreciation in cognitive labor. Corporate tax buoyancy will disappoint too: the largest AI-rent earners are foreign, and GOI has already exempted foreign cloud providers’ India-routed global income from tax until 2047.
On the expenditure side, more jobless youth demand more government dole.
So what can GOI do? It has SLR and the IRDAI investment regulations in its pocket. And that is precisely what GOI has done by allowing full FDI in the insurance sector. GOI is creating a structural demand for its paper. So, the sovereign yields will be artificially suppressed, increasing the credit spread materially. And CIO’s will realize eventually that the government debt offers the poorest real value and will have no choice but to extract every basis point from the discretionary portfolio that they control.
D. Weaker Capital Controls
India’s capital controls are among the most institutionally entrenched in the world: FEMA, the LRS cap, a 30% tax plus 1% TDS regime on crypto that has pushed volumes offshore but kept domestic rails legible, and a central bank actively building the digital rupee precisely to preserve monetary sovereignty. Capital controls don’t need to fail – they only need to become leakier. If global real rates rise 100-150 bps and if the Indian savers gain even a partial, gray-channel access to those returns, Indian real rates also will rise.
E. High Real Rates with inflation under control
AI is disinflationary via lower marginal cost of services, but inflationary due to its insatiable appetite for energy. Food inflation will remain hostage to monsoons for years to come – the rain-fed belts of the Vindhyas, Bundelkhand, Chota Nagpur, and the Deccan plateau will see to that.
Our view is that the policy rate ends the decade moderately higher than today’s levels (100 -150 basis points higher) in nominal terms, but substantially higher (250+ basis points) in real terms.
F. A boost for India’s bond market
India is targeting 500 GW of non-fossil capacity by 2030, has launched a Nuclear Energy Mission with small modular reactors on the roadmap and amendments underway to open the sector to private capital, and is watching hyperscalers sign power agreements that would have seemed fantastical five years ago. Power generation, transmission, and water infrastructure feeding them, all need massive capex, which are excellent debt investments. Data centers will also raise high quality debt. We think that these debt fundraises are great for the Indian debt market and will expand the private placement ecosystem of India.
G. AI Datacenter Debt – the good and the bad
Wait a minute.
First, the good news. In line with the Jevons Paradox, as the unit cost of AI inference plummets, total consumption of AI will surge rather than shrink. Driven by cheaper energy, algorithmic advances in AI, and accelerating enterprise adoption, overall compute demand will increase further. Why? Look at the success of the Chinese Kimi K3 AI model, a 2.8 trillion parameter open weights and open source model, for which there is already a waitlist to purchase the paid subscription. We surmise more such heavyweight models are on the way, and US and European enterprises will use cloud compute to adopt them and customize the weights. One enterprise success story will get emulated by others and the race (to consume cloud compute and use AI) will be on. India’s gigawatt AI data centers will come online after 2030 and take the combined benefit of customized ASIC silicon, algorithmic optimizations in AI and customized parameter weights.
Now, the bad news. India’s gigawatt AI data centers will mostly not be used for training, and will be used for inference. That is a risk for AI related data center debt investments in India. OBBBA Act’s 100% Bonus Depreciation pulls US labs’ training capex homeward, and US export controls on frontier chips discourage US AI labs (OpenAI, Anthropic, Meta, etc.) from setting up training clusters on Indian soil. Inference does not need a big data center and can be done reliably on small, regional data centers.
Our view? The risk exists specifically for the ultra-large gigawatt scale data center debt. The debt of small, edge scale data center of 10-100 MW capacity will be fine.
H. Banks, insurers and financial services become more efficient
Advances in AI and better adoption can collapse the cost of underwriting, documentation, reconciliation, claims processing, fraud detection, and compliance. The credit losses will also fall as underwriting improves. Banking and insurance are highly competitive industries, and gains of efficiency will get competed away via lower costs for the consumers. A few early adopters will have better margins temporarily and might even gain market share. However, as homogeneous models trained on similar data get utilized to underwrite risk, errors stop being idiosyncratic. When every lender’s model shares the same blind spot, the entire system piles into the same mispriced risk and discovers it together. There is a risk here, but of a different kind: fewer small losses, and rarer but larger, synchronized ones.
I. Annuity liabilities lengthen with increasing longevity
In India, currently a 65 year old can expect to live to be nearly 78. If AI-accelerated medical research adds even three to five years to this, every annuity book and every pension obligation lengthens in duration. Life insurers, which are the dominant buyers of 30+ year GSecs, will need more ultra-long duration. This creates a peculiar and important dynamic at the long end of the curve: fiscal supply pressure pushing long yields up, colliding with an intensifying structural insurance bid pulling them down. The result, we will argue, is a flat – possibly even inverted – curve beyond the 20 year maturity. We draw the reader’s attention to the US treasury curve, which is essentially flat from 20 year to 30 year maturity.
J. Deployment rents, not model rents
India does not have any companies that develop frontier AI models, or any companies that own leading-edge semiconductor fabrication plants. Although India’s solar power is now at par with the cheapest in the world (partly thanks to China) but that is not a very defensible position. Still, with investments in the physical infrastructure, India can position itself to earn deployment rents.
Part 2: Implications for the sovereign yield curve
With assumptions in mind, we build a base case scenario and two alternative scenarios.
2.A: The base case of “limited steepening” (probability ~ 60%)
Short End – Overnight to 2 years:
The RBI has to defend the rupee, raise rates in tandem with the US Federal Reserve, keep inflation checked and prevent excessive steepening. All four objectives can be achieved with one answer – raise short term rates. AI-driven services disinflation will partially offset the increase in energy costs, while weather disruptions such as El Nino will stoke food inflation via poor monsoons and floods. We think RBI policy rate ends the decade at 6.5% with real short rates being 50-75 bps higher.
Belly : 5 years to 20 years:
This is where the repricing concentrates. Three forces stack:
i) rising global neutral interest rate (r*) transmitted through the now-open FAR channel;
ii) rising domestic investment demand as the data center/energy capex wave competes with AI-powered MSME loan demand for domestic savings;
iii) more borrowing from the Center and the States as fiscal uncertainty grows.
We would pencil in the 10-year GSec yield at 8.50% by the early 2030s and at least 100 basis points higher in real terms than today. The 2s10s and 5s15s segments steepen meaningfully.
The ultra-long end (30–40 years):
Here the fiscal supply pressure collides with the structural insurance bid. Life insurers with lengthening annuity books will keep this part structurally bid.
We expect the 15s30s segment to flatten even as 5s15s steepens leading to a kink forming around the 15-year point. In plain terms: the curve gets steeper where free-market forces dominate and stays suppressed where regulated demand dominates.
2.B: The bear case: “The Twin Shock” (probability ~30%)
This is the scenario where human capital depreciation hits hard and fast. Using AI, US based corporations insource cognitive work, via offshore GCCs and onshore staffing. In this scenario, the Indian IT services revenue declines fast. The current account deteriorates in tandem. We are assuming a status quo on the export tariffs, but there might be a more hostile tariff regime in the future, further deteriorating India’s current account. If the rupee comes under sustained pressure, RBI will be forced to defend with materially higher rates (forex swap facility won’t be enough). In this case, the 10 year nominal yield rises above 9.25% and the 30 yields hit 10.5%+. We are tracking IT-sector headcount and revenue-per-employee, and we encourage you to track it as well. For us, it is a meaningful rates indicator now.
2.C: The bull case: “Deployment Dividend” (probability ~10%)
India’s deployment bet pays off asymmetrically: competitive electricity rates (largely from cheap solar) and 5+ GW of data center capacity (and growing) make India the inference hub for the Global South. This is a contest India can win on trust, data sovereignty, and access to US GPUs that Chinese clouds are increasingly denied.
AI diffusion through Indian MSMEs lifts trend GDP growth toward 8%+ and GCCs keep growing, while the debt/GDP ratio falls through denominator growth as it did for the US after WWII. Real rates rise (this is a high-investment scenario too), the rupee is stable, and the GSec curve shifts up modestly while flattening, as one would expect from a benign, growth-driven rate rise.
The scenario’s Achilles heel: India’s cheap electricity rests on a solar and storage supply chain whose upstream technology China controls. China has already restricted exports of wafer and battery manufacturing technology once. A supply chain that a geo-political and strategic rival can throttle is not a foundation; it is a hostage.
Real yields rise in each scenario.
Part 3: Corporate credit: Only one scenario
3.A: Credit Spreads Widen
Let’s revisit Robert Merton’s structural model, which describes the corporate bond as a zero coupon risk-free bond minus a put on the firm’s (borrower’s) assets, while the equity is a call on the firm’s assets. As per Merton’s model, the lender has written the put, and is therefore short the volatility on the borrower’s assets.
We are quite sure that AI will dramatically increase the volatility of businesses’ outcomes. How? AI compresses time to market and erodes moats built on proprietary knowledge. Due to AI, a few new entrants will rapidly gain market share, while a few incumbents will perish. More firms will grow to 10x, and more firms will quietly become obsolete, faster than in any prior decade. Equity holders benefit from the variance increase because they’re long the call on the firm’s assets. Bondholders are short the volatility on firm’s assets.
So, credit spreads widen on average because the tail of distribution of firm outcomes becomes fatter. Statistically speaking, firm outcomes (both good and bad) fall further away from the mean, and kurtosis increases.
3.B: Spreads between ratings also widen
Rating dispersion also widens for the same reason. A credit rating is, crudely, a measure of distance-to-default under an assumed volatility of firm value. When the volatility of firm outcomes rises economy-wide, the same distance-to-default translates into a higher default probability, and the effect is convex. Default probability increases a little for a AA rated firm, which has higher distance to default, but increases a lot for the BBB rated firm which has a shorter distance to default. The BBB borrowers might get pushed out of the private placement / EBP market entirely, and back into the banking system (more on that some other day).
We also expect downward rating transition rates (probability of migration from a higher rating to a lower rating per unit time) to reverse course: after a decade of decline in India, they will start rising, and then rise at an accelerating pace. Ratings are calibrated to historical default studies from a lower-variance world; they encode incumbency as safety. In an AI transition, incumbency is the risk in many sectors, so expect the market to reprice credits 12-24 months before the agencies act.
Moreover, corporate credit (with its limited upside and full downside) will have to compete with both sovereign debt (which is risk free) and equity (which has full upside). Corporate credit will appear structurally less attractive relative to GSecs and equity.
Insurers and banks must open up to the debt boom of AI Capex. Our recommendation is to build private-credit and project-bond origination capability for the AI-infrastructure complex now, while the competition is still thin. Data centers are among the most bond-friendly assets ever invented: enormous upfront capex, ten-year contracted cash flows with investment-grade counterparties, tangible collateral, and inflation-linked escalators. We recommend funding small, regional data centers in the 10-100 MW range.
In closing
We hold these views with conviction about direction and humility about magnitude and timing. Real yields rise in every scenario without invoking catastrophe; credit spreads widen; and the dispersion between rating categories widens most of all. What a CIO or CRO should actually do about this – duration, curve placement, and relative value – deserves its own note, and that is the subject of our next essay.
We are always open to a conversation.
Gaurav Singhal, CFA, FRM
https://www.linkedin.com/in/gsinghal01/
info@qari.in