#345 Anecdata*
Observations from Silicon Valley and Sri Lanka
India Policy Watch #1: The Coming AI Tsunami
Insights on current policy issues in India
—RSJ
Over the past couple of weeks, I spent some time in Silicon Valley trying to get some sense of where exactly the AI economy is headed first hand. Jensen Huang’s “five-layer AI cake” has become the dominant framework for thinking about the industry. Energy, Chips, Infrastructure and Models stacked one upon the other. Then, the tooling and applications at the top. My own interest was primarily in the model and application layers because that is where India’s opportunity appears most plausible. We are unlikely to dominate frontier compute. We do not have the capital base, hyperscale infrastructure or industrial policy architecture required for that race. But applications and workflow systems are where most traditional Indian IT services players are pinning their hopes on, and I wanted to understand if there was real play there.
Fairly quickly in my visit and in some of the early meetings with venture funds, PEs and investors, I could see a mismatch between technological momentum and financial reality. Nobody I met doubts that AI capabilities are advancing at a remarkable speed and are going to be transformative. That debate is effectively over. But the amount of capital that’s already been pumped in across the 5 layers, and that’s still waiting on the sidelines, reminded me of the dotcom boom era. One leaves the Valley with the distinct sense that the entire ecosystem is pricing in civilizational transformation while still struggling to explain current profit-pool opportunities.
This matters because technology gains show up in different forms over time that can’t be predicted, while capital gets deployed on bets that are visible now. The core technology could change society with massive capital being deployed on it, but the returns on capital deployed could be diffused across sectors. Examples from the early 20th century echo this. Railroads transformed America and destroyed enormous amounts of investor capital. Airlines change mobility radically while remaining structurally weak businesses for decades (possibly the worst sector in terms of shareholder returns in the past century). Telecommunications reshaped the world but commoditised much of the underlying infrastructure. Everyone else who built on telecom rails made money. AI may well become foundational infrastructure for the next era of capitalism while simultaneously proving to be one of the great overinvestment cycles in technology history.
Part of the reason for this is the usual problem of extrapolating the current trend into the future instead of planning for scenarios. The industry still behaves as though compute will remain scarce. Current valuations require sustained pricing power at the model and infrastructure layer. Yet most technology markets eventually move toward abundance and price compression. Open-source models will continue to improve, and over time, inference costs will get optimised. The better the models become, the harder it may become to preserve economic rents around foundational intelligence itself.
The AI infrastructure companies I saw already understand this. The focus on optimising storage and compute, and the kind of funding and research talent going there, suggests this will happen sooner. But at the overall ecosystem level, what I understood was that companies were trying to industrialise intelligence before fully understanding how it would ultimately be monetised. The metrics right now seem to be tokens used per engineer rather than anything that creates tangible economic value.
At the application layer, something more significant is happening. The best way to understand most AI application companies is not to think of them as software firms in the conventional sense. They were attempting to solve problems that I refer to as the coordination problem inside enterprises.
Every large organisation consists of three broad layers. There is the customer-facing interface on one side with people, stores and internet and mobile apps funnelling customers into the enterprise. There are systems of record on the other side that include core databases, ledgers, ERPs and transactional systems, which build trust and fidelity within the enterprise for the customers.
Between these two layers sits the actual enterprise. That’s the series of processes and workflows with checks and balances that ensure the customers coming in turn into economic value for the firm. That middle layer is essentially a coordination machine. Operations, procurement, compliance, finance, audit, legal, and HR are all coordination systems. Documents flow from one team to another with incremental value addition. This is a division of labour with specialised teams having experts and their own network of systems that only they can access. This is how organisations run and manage to control their execution. Exceptions are escalated, and information is reconciled across systems and people. Large enterprises are sprawling networks of coordinated actions.
The SaaS era tried to solve this problem through standardised workflows. Software companies encoded organisational processes into applications and asked enterprises to adapt themselves to those systems. This worked well enough for two decades. But much of enterprise coordination still escaped formal systems and remained in meetings, emails, spreadsheets, and institutional memory.
The AI application layer is now attempting something more ambitious. Instead of merely digitising workflows, these systems are trying to orchestrate workflows dynamically. Agents retrieve information, reason across systems, route tasks, generate outputs and coordinate actions across organisational boundaries. The industry uses the word “agentic” almost obsessively today because the enterprise agent is becoming the central abstraction of this cycle.
I saw platforms describing themselves as AI operating systems where one could define chains of agents and sub-agents in natural language and construct autonomous workflows in weeks. Sometimes days. The abstraction layer itself is changing, and soon the workflow becomes adaptive and self-improving. An enterprise agent is essentially a software construct attempting to replicate the operational behaviour of middle layers inside organisations. It will learn the workflow from its code, gather context with practice and then replicate the workflow.
But the AI agent need not be split into departments. It can work across systems, resolve ambiguity and most importantly, act autonomously over time with perhaps a human in the loop. The application layer of AI is gradually becoming the coordination layer of the enterprise itself.
This has deep implications for India’s IT services industry and the broader Indian economy.
For three decades, India benefited from a global labour arbitrage built around large-scale white-collar process execution. Much of the services industry grew around managing coordination complexity for global enterprises. Teams maintained workflows, reconciled systems, processed documents, managed tickets, supported applications and handled the operational exhaust generated by large organisations.
Agentic systems are now coming directly for that layer.
The real disruption is not coding copilots or chatbots. It is the possibility that large parts of middle and back-office coordination itself become autonomously executable. What will a bank look like if the coordination layer is running with autonomous agents? There will be a significant deflation of mid and back office across the services industry.
The phrase “AI will augment humans” remains broadly true in the long run. New professions will emerge, of course. Human demand has historically expanded alongside technological progress. But the transition will take time, and it will be painful. The political economy will struggle with this. We may be entering a period of significant deflation of work before new forms of work emerge at scale.
India has a limited presence in the energy, chip and infrastructure layers of the AI stack. The model layer itself is evolving at such speed that even many people inside the Valley struggle to keep pace. Frontier labs are now pursuing increasingly autonomous systems where code writes code, workflows recursively improve themselves, and software development itself becomes partially automated. Whether such ambitions fully materialise is secondary. The direction of travel is unmistakable.
India can certainly attempt sovereign frontier models and should probably do so for strategic reasons. But it is difficult to see how such efforts can compete directly with the concentration of capital, talent, and compute now embedded in the Valley ecosystem.
That leaves the application layer as India’s most plausible opening. There is a genuine opportunity there. India has deep enterprise process knowledge, large software talent pools and long experience integrating technology with operational workflows. One can imagine entirely new categories of AI-enabled service providers emerging from India. Firms combining domain expertise, workflow redesign and agentic orchestration into globally competitive operating models. But the great deflation may overwhelm the opportunity creation cycle in the near term. That’s a real possibility.
Large Indian enterprises themselves may require dramatically fewer white-collar workers over the next three to five years if autonomous coordination systems mature at the pace currently visible in the Valley. Entire layers of repetitive managerial and process-heavy work may shrink simultaneously across sectors. Banking, insurance, IT services, operations, support functions, audit, compliance and internal coordination roles could all face pressure.
India already faces significant underemployment disguised beneath services growth and platform work. Quick-commerce delivery networks and gig labour have partially masked deeper weaknesses in formal employment generation. AI now threatens to arrive precisely at a moment when the economy has still not fully absorbed its existing white-collar workforce.
The next disruption may therefore not resemble earlier automation cycles that primarily affected manufacturing labour. This one may hollow out sections of the educated middle-class services economy itself.
I left Silicon Valley with the sense that the direction of history is now largely set. The technology is real, and the disruption is coming. The only uncertainty is the speed at which institutions, labour markets and political systems are able to adapt.
That uncertainty may become the defining economic and social question for India in the next decade.
Related: Check out this Puliyabaazi with Sidu Ponnappa on this theme.
India Policy Watch #2: An Evergreen Puzzle
Insights on current policy issues in India
—Pranay Kotasthane
While RSJ was in Silicon Valley, I spent the last week as a tourist in Sri Lanka. For an Indian public policy analyst, travelling to other countries always throws up comparisons, puzzles, and questions. We notice differences in governance outcomes and can’t help but wonder why we are the way we are.
In fact, this instinct is how the famous ‘M document’ came into being. VP Singh happened to travel to Malaysia twice in a matter of months and was confounded by the pace of change happening there. He then asked Montek Singh Ahluwalia what India could do to transform itself in this way. Ahluwalia prepared the memo that would go on to become the basis of the economic reforms in 1991-92.
Sri Lanka in 2026 provokes a similar reaction, though in a different dimension. This is a country that has faced a devastating civil war, a catastrophic economic collapse, and a destructive environmental disaster in the last 40 years. It is heavily dependent on aid, and one can see the markers of international aid everywhere—a water filtering machine at the main airport carries a ‘donated by Japan’ sticker, China’s infrastructural footprint is unmissable wherever you go, and India’s first response in environmental crises is well-appreciated.
Despite these challenges, the Sri Lankan State doesn’t seem to be as omniabsent as its Indian counterpart. The roads, not just in the capital, but even in rural areas, are uniformly marked and well-designed. There are no signs of the white-topping boondoggle, and the roads are largely pothole-free. Roads are peppered with well-marked zebra crossings even in remotely populated areas. Vehicles stop graciously and let pedestrians use these crossings without honking endlessly. The footpaths in Colombo are stroller- and wheelchair-friendly, yet it is hard to spot a two-wheeler misusing them to beat traffic snarls. Not once over the week did I see people driving on the wrong side of the road. It’s one thing to see these things happen in the richer countries, and completely another to see some of it happen in a locale that’s superficially not different from our own.
Even in formerly war-torn areas like Trincomalee, roads are well-marked, lanes are painted, and people largely follow traffic rules. Public libraries—even if battered and forlorn—dot small towns. The civic infrastructure of a country facing multiple crises looks better maintained than that of comparable Indian towns. What explains this?
An easy explanation is endogeneity. Sri Lanka is roughly 1.5 times as rich as India on a per capita basis. More money implies better public services. Broadly true, but the comparison breaks down when we decompose India into smaller fragments. States like Goa and Karnataka are far richer on a per-capita purchasing power parity basis than Sri Lanka is, and yet you don’t see the kind of public service provision in our richer states.
Another explanation could be generational literacy. Sri Lanka achieved near-universal literacy by the 1970s. India's literacy gains are more recent. But even here, India’s most literate states are no match for Sri Lanka in rule-following and public service provision.
Another instinctive reaction is that Sri Lanka simply doesn’t have megacities. Colombo’s city proper has roughly 700,000 people; Kandy has 125,000. Managing a 125,000-person town is obviously easier than managing a 13-million-person agglomeration like Bengaluru. But this is a weak objection because there is no reason Bengaluru must be governed as a single, monolithic unit. It could be decentralised into borough-level governments with genuine authority over roads, waste, and traffic, as in London, Tokyo, or even New York. Scale is not a natural constraint on Indian urban governance. It is a political constraint.
Beyond grand explanations, there might just be some micro-institutional work that is doing the heavy-lifting. For example, Sri Lanka’s Road Development Authority (RDA) is a single national agency responsible for the design, construction, and maintenance of the national road network. It sets and enforces uniform standards across the country—lane widths, marking patterns, signage placement, speed indicators, etc. When a road is built or resurfaced anywhere in Sri Lanka, it follows one set of specifications.
Road governance in India is fragmented across at least four tiers: NHAI handles national highways; state PWDs handle state highways and major district roads; municipal corporations or councils handle urban roads; and panchayats nominally handle rural roads. Each tier has its own lack of standards. A beautifully asphalted NHAI stretch gives way to a cratered state highway, which feeds into an urban road where the white-topping contractor has left things half-finished for months. No single authority owns the standard across the full network.
The fragmentation can also explain why Indian civic norms around road use are weaker. When road design is inconsistent, drivers rationally stop paying attention to markings altogether. Why would you stop at a zebra crossing when the last three you saw were painted over, half-erased, or placed in locations where no pedestrian could possibly cross? Sri Lanka’s consistency in road design has created a feedback loop. Uniform markings lead drivers to learn to respect them, which in turn leads pedestrians to trust them. India’s inconsistency produces the opposite loop: unreliable markings lead to a rational disregard and norm erosion.
At this point, it’s tempting to reach for cultural explanations. Perhaps Sri Lankans are simply more civic-minded, more rule-following, more invested in public order. Perhaps there’s something about the society—Buddhist values, island-nation cohesion, a smaller and more homogeneous population—that makes people respect zebra crossings and public libraries in ways that Indians don’t.
I’m sceptical. Sri Lanka is no stranger to the pathologies we usually attribute to Indian culture. Corruption is endemic. Communal tensions between Sinhalese and Tamils produced a full-blown civil war. Political dynasties have dominated the country. Clientelism, patronage networks, ethno-religious mobilisation—Sri Lanka is no stranger to these. If ‘culture’ were the explanation for poor public services, Sri Lanka should look no different from India. And yet, the roads are better. The zebra crossings work. The public libraries are visibly present. This suggests that the difference isn’t in the culture of the people but in the structure of the State.
The Indian State optimises for avoiding the worst-case outcomes, such as runaway inflation, civil wars, large-scale violence, and recurring macroeconomic crises. Performance and delivery are merely good-to-have features. Other countries like Sri Lanka optimise for uniformity. A small, relatively homogeneous state can afford to centralise authority, impose uniform standards, and focus institutional energy on delivery.
Whatever the precise reason, why we, as citizens, demand and expect so little from the Indian State while venerating those who hold its reins remains an evergreen puzzle.
HomeWork
Reading and listening recommendations on public policy matters
[Book] Samanth Subramanian’s This Divided Island is a terrific book on Sri Lankan politics and society.
[Puliyabaazi] The next episode is on the invisible infrastructure created by Engineering Talent, Non-Profit Trade Bodies, and Public Libraries.
*Anecdata = Anecdote + Data
Love the honest assessment of India's AI position. The coordination layer insight is excellent — agentic systems aren't replacing individual workers, they're replacing the middle layer that orchestrates across functions. That's where India's services industry lives, and the vulnerability is real.
Two threads I'd add. First, topology matters. AI bites more easily where work is legible — codifiable, decomposable into tasks. Much of India's formal services economy fits that description. But India also runs on enormous illegible infrastructure — informal supply chains, relationship-driven transactions, trust-based coordination — where AI doesn't land as easily. So that's an interesting defence in some ways, and an opportunity.
Second, I know you based your notes on your visit so perhaps China was out of sight, out of mind. India's DPI infrastructure — Aadhaar, UPI, ONDC — could lend itself to China's deployment-and-integration model than to Silicon Valley's engagement-and-spectacle model. The real question for India may be whether its institutional scaffolding can absorb AI into existing infrastructure before the deflation you describe arrives.
Your closing line — "the only uncertainty is the speed at which institutions adapt" — is the whole game. I explored this from a different angle a few days ago, asking what determines whether AI's disruption leads to a three-day work week or 40% unemployment. Same technology, two futures. The difference is institutional: https://rajeshachanta.substack.com/p/the-doorman-and-deepseek
I was also in Sri Lanka in March, and had very similar observations! SL is very similar to Karnataka in population density, diversity, terrain, per-capita GDP (nominal and PPP), etc. Economically there are regional imbalances, similar to Bengaluru vs other districts, but nowhere did we see the kind of squalor that one sees everywhere in India. Even in the tiniest places (we were in 12 different places) toilets were clean, public places were being continually swept, there were no drunks (in spite of the many liquor shops), no loud horns, etc.
Culturally, I feel that one reason for the uniformly better experience is the large number of international tourists that visit Sri Lanka (2.5m vs 0.5m in Karnataka). Apart from the money, they bring high expectations. Once the citizen get used to the higher standards, they will not accept anything less. We in India have grown up without any standards, and we seem to accept it as inevitable.
Lack of pride in our surroundings seems to be a common problem here. Our leaders tweet about how nice the streets are, and cycle when they go abroad, and don't do anything about our own streets. We saw one family at a Buddhist temple in Kataragama, from grandparents down to kids, sweeping the ground and picking up all the leaves. I was comparing it to one of our temples which we are trying to make zero-waste, where we see everything from diapers and shoes, to packets of oil thrown in the trash.
I am sure there are many other reasons. There is the elephant in the room: caste, for which there may be no equivalent in Sri Lanka.