#287 Inflection Points

Understanding DeepSeek Breakthrough, Making Sense of Technological Sovereignty, and Parsing the Budget Announcements

Global Policy Watch: It’s Getting Real

Insights on global issues relevant to India

— RSJ

As I write this, Nirmala Sitaraman is starting with her Union Budget address on the floor of Lok Sabha. I will return to it at the end of this section when I read the transcript later this afternoon. Before that, there are a couple of topics to discuss.

I had marked 2025 as the year where AI use cases and AI competitive intensity ramp up significantly as part of my list of predictions. I hadn’t bargained for how dramatically this would come true in the first month of the year itself. What were the odds that an unknown startup (to the outside world) from China with a hedge fund quant billionaire as its founder would upend the global AI race? Zero, I would say. Yet, DeepSeek, a research-focused lab, released an open-source reasoning model last week whose performance on quantitative, coding and reasoning tasks was on par with OpenAI-o1 while costing about 6-7 per cent of o1’s price. This is a 2-year-old Chinese research lab, a side project of a hedge fund team run by a billionaire founder, that worked on older generation Nvidia H800 chips because of US restrictions on the export of advanced chips to China to develop its R1 model by spending less than $6 million. And it performed on par. Till a couple of weeks back, all such estimates of costs to train models by the Big Tech firms in the US were in billions. China just knocked off a few zeros on AI costs while doing the same to the market cap of Silicon Valley AI firms during the week.

There’s the usual scepticism that showed up soon on the real costs of the DeepSeek platform, the way they have used the output of other proprietary AI models to train R1, and the censorship of China content. But no one could question the outcome, which was for all to see. DeepSeek published all its research and released its models in open-weights forms last week. This allows developers to build on top of it and use tokens at significantly lower costs. There are two significant cost elements for an AI platform—the cost of training the model and the cost of inference, which is that of using the model. Through a mixture-of-experts approach (ensemble learning), they have optimised for the number of parameters that are called into use whenever there is a query because only the ‘relevant’ expert who can help with a part of the query is called in to help. This isn’t a new technology, but they have used it in a manner that hasn’t deteriorated the output quality. The other thing where the quant fund background of the team has helped is they have been able to use quantisation of computing very effectively (and aggressively). The effective use of these two means they could use the old-gen chips and significantly smaller computing power to run the platform. On the inference side, they have used reinforcement learning (RL) to significantly reduce the time and cost of a token. In the earliest version of ChatGPT, the algorithm would read the sentence completely like humans and then infer the context and meaning and generate an answer. The newer versions of OpenAI did something called inference time learning. That is instead of answering once at the end of the sentence, the algorithm would answer multiple times during the time the question is put to it. This required significantly more training and compute effort but with much better outcomes. What DeepSeek has been able to do is to use reinforcement learning innovatively to blow apart the time and cost of training. Now this was always possible to do in a limited test environment. DeepSeek has scaled it, opened it up for others to use, and democratised AI in a single fell swoop. The commercial uptake of AI will be an order of magnitude different in 2025 because of this.

What does this mean for the new ‘arms race’ around the world? Broadly, there are four things to track as things evolve fairly rapidly here:

First, this isn’t a Sputnik moment, as some have called it. DeepSeek’s breakthrough doesn’t mean China has caught up or beaten the U.S. on AI. There was an expectation that the cost of compute would fall for AI platforms in some Moore’s law equivalent pattern. My guess was a halving of the token costs every six months. DeepSeek has got us to a price point we would have reached maybe four years later. From a technology standpoint, this will be replicated in less than six months by every existing AI platform. There was a self-serving aspect to the U.S. Big Tech talking up the billions it would spend on its AI platforms. It was to make it prohibitive for any upstart to think about challenging them (who would sink such large capital?) and convenient to therefore price it steeply because customers would believe that AI costs will be high because of such large initial investments. DeepSeek has blown this quasi-cartel to smithereens with R1. It is a significant tech accelerator. It has shown China can replicate its genius of reverse engineering a product or a platform developed by the West at a fraction of the cost, even in AI. This is a remarkable leap in its software capabilities. But it isn’t a new frontier of technological advancement. There is a difference.

Second, this accelerated decay of the cost of computing triggered by DeepSeek will mean we will see a rapid commoditisation of the core GenAI capabilities this year. The US AI companies can no longer justify the need for vast amounts of capital, data centre infra and advanced chips to do what DeepSeek is already doing at a tenth of the cost. So, we will see an unprecedented uptick in the commercialisation of AI across the board. All the benefits and fears of AI that were held back because the cost of experimentation was high have been brought forward by four years.

Third, the real impact is beyond the market of AI use cases and the correction in the valuation of AI-driven tech stocks. The geopolitical repercussions of this will be more long-term and real. It was already clear that the AI race wasn’t any more about technology companies competing among themselves. The Trump administration’s investment in Stargate and the export ban on high-end Nvidia chips were already clear signs that this had become a geopolitical battle. DeepSeek R1 has shown that unless immediately challenged, it can become the standard for open AI models worldwide. As I have written before, in any new technology, the country that sets the global protocol or standard builds a lasting moat. Till last week, when the Biden administration came up with its AI governance guidelines, it was the US that was surging ahead with defining the protocol of AI usage for the world. With DeepSeek, China has shown the world that there is an alternative and they needn’t fear being categorised by the US into Category 2 or 3 nations for access to AI platforms. That apart, the ability of China or any other country to use these open-source models to direct capital and resources for surveillance, cyber warfare and conventional military applications will no longer be constrained by how the US views them. The American exceptionalism on current AI capability is quite likely over. They will have to innovate fast on the frontier of technology and hope China isn’t able to reverse engineer it once again in order to gain back that sense of being exceptional. Easier said than done.

Lastly, for India, the DeepSeek breakthrough should be seen as for what it is. The democratisation will help us if we can ramp up our ability to build use cases for ourselves and the world using the current AI platforms. We should be asking what our large IT services industry can do to maximise this opportunity. I feel the industry will figure this out for itself because it has been good at riding every large wave of technology over the last 30 years. The other question to mull over is something that India has been traditionally good at - how to use constraints of lack of capital and access to cutting-edge technology to come out with a creative solution that works as well? This was once feted as the Indian jugaad, as much as I hate that term. But that’s what drove our space or nuclear programs, to quote just two examples. We are usually good at squeezing the maximum performance out of things or rigging a new way of using an outdated technology to get an outcome close to that of a cutting-edge technology. Can we do this in AI like DeepSeek? This is a policy question, really. Token budget announcement of setting aside a few hundred crores won’t solve this.

Of course, the real game changer here will be if China develops its chips for model training. How far is Huawei from that? We don’t know yet. It has built AI chips for inference that are edging close to Nvidia. But model training is a different ball game altogether. I won’t bet against China announcing something on this in the next 12 months.

DeepSeek is the first battle of this arms race. The real war is about to start.

---

Global Policy Watch: What Makes a Country Technologically Sovereign?

Insights on global issues relevant to India

— Pranay Kotasthane

DeepSeek’s breakthrough has sparked an interesting conversation on technological sovereignty. By imposing escalating export controls and compute rationing rules, the US seeks to convert its dominance in a key enabling technology into an unassailable lead over other nation-states. The underlying logic is that AI sovereignty is a zero-sum contest like the control over land, moon, and natural resources. And like any other zero-sum contest, it necessitates denial strategies—I can have more only if you have less. Eventually, the US began thinking of AI sovereignty as an end in itself, a goal worthy enough to divide the world into three tiers of GPU access.

Despite these restrictions, technology diffusion did happen even though government intervention blocked chip diffusion. In a world where the research, hardware, and software ecosystems are global, the logic of technology sovereignty was called into question. If a dominant player like the US couldn’t stamp its authority over AI—because it does not control all the factors that go into making it—then who can?

In this context, I started looking for alternate conceptions of technological sovereignty. For instance, a paper titled Technology sovereignty as an emerging frame for innovation policy, posits technological sovereignty as merely a means to enhance a State’s agency in the international system and never an end in itself. The authors argue:

technology sovereignty should be conceived as state-level agency within the international system, i.e. as sovereignty of governmental action, rather than (territorial) sovereignty over something..

the pursuit of technology sovereignty should be understood as the attempt to safeguard public agency in the domain of technology and innovation, i.e. the ability to act independently in the face of institutional and economic boundary conditions and, in some cases, third parties' adversarial actions. Importantly, this ability to act independently does not result from the nations' domestic capacity alone, but also from their embeddedness into a robust and reliable network of international relations and partnerships (Binz and Truffer, 2017) to access necessary inputs. In fact, the quality and reliability of this external network may in most cases be at least equally constitutive for the degree of global technological agency that a nation can muster than its own, domestic capacities. [Edler Et. al.]

Seeing technology in these instrumental terms means that technological sovereignty is about gathering the necessary capabilities to preserve a state’s survival and prosperity. As authors of another important paper, Technological Sovereignty as Ability, not Autarky, explain:

a polity is technologically sovereign, if it possesses the technological abilities necessary to maintain political and economic sovereignty.

This view of technological sovereignty offers hope to many countries playing the catch-up game. The fact that a breakthrough like DeepSeek came out of China and not other places shows it had the base technological capabilities that allowed it to act independently in the face of chip restrictions. It could reasonably protect its economic and political sovereignty (for now) despite not having the kind of dominance that the US enjoys.

What these base technological capabilities are is essential to analyse. China’s embeddedness in the Western scientific and innovation ecosystem allowed it to access ideas that could circumvent hardware constraints. Even though it remains far behind the US regarding the technology itself, it has demonstrated enough to preserve its political and economic sovereignty.

DeepSeek should also motivate other developing countries to focus on asymmetric capabilities. A goal of amassing all components of a technological supply chain at home is expensive and futile because interdependencies—whether on materials, machines or external collaborators—don’t go away. Instead, the line of attack for preserving one’s sovereignty is to dominate in just one or two segments of the technological supply chain, deterring other players from acting against your interests. This one leverage point is more impactful than hoping to become good across the entire value chain.

Seeing technological sovereignty through this lens of ability allows us to reflect on India’s position. Would buying 18000 GPUs (as the government plans to do) increase India’s economic and political sovereignty in any way? If not, should the government focus on strengthening a segment where India has a comparative advantage? If technological autarky is impossible, how can India increase its technological ability? These are some deep-seeking questions for us to consider before we launch ourselves into rhetoric over AI sovereignty.

---

India Policy Watch: Economy? What Economy?

Insights on current policy issues in India

—RSJ

Quick Note on the budget.

If you do some basic back-of-the-envelope stuff, you would have known this budget would be a non-event. I mean, if you take the 6.5 per cent revenue growth for FY26, then estimate the standard sources of revenues and the costs, the non-negotiable spends on defence, subsidies and the vertical allocation to the states and the desire to bring the fiscal deficit below 4.5 per cent as per the fiscal consolidation roadmap, what are you left with? Maybe around ₹2 trillion, which is about 0.6 per cent of the GDP, to allocate some big spending in new areas.

So, all you should have expected is a few headline-grabbing moves like raising the income tax threshold and some signature new schemes. That’s precisely what has been announced. And that would keep the partisans in thrall for the next month. But nothing beyond that. The income tax breaks will boost consumption, but my old question remains. More money in consumer’s pockets doesn’t necessarily mean more consumption. If that were to be true, China’s consumption would have been strong for many years now. The decision of what to do with the extra money in pocket—spend or save—depends on how consumers view their future. So, we will see how this ₹1 trillion (I’m sceptical of this calculation, too) tax break really translates to a consumption boost during the next 12 months.

That apart, I have a sneaky feeling for the past six months that, the PM has lost interest in the economy, specifically in economic performance being a variable for winning votes. We will grow at 6.5 per cent one way or the other seems to be the prevailing wisdom. The new Hindu rate of growth, I guess. Middle-class voters, for whom economic performance matters, are mostly beholden to the PM and the BJP, in that order. And what the Lok Sabha elections 2024 and the subsequent Maharashtra elections have shown is that a focus on timely welfare schemes plus the usual set of emotive issues can work in favour of the party to get the poor and marginalised votes.

So, why queer the pitch with painful reforms that are difficult to sell to people (like the experience of farm reforms showed). So, in possibly the most benign 12-18 month period (Bihar and Bengal elections will anyway be identity-driven) where there was a possibility of real reforms, we have no signs of any such push forthcoming. Instead, I expect more legacy-cementing efforts from the PM during this time and whatever is left of the agenda to replace the Nehru-Gandhi veneration with the idols from the parivaar pantheon. I hope I’m proven wrong.

---

India Policy Watch: Budget Takeaways

Insights on current policy issues in India

— Pranay Kotasthane

Here are four disjoint thoughts from the budget season, in no particular order.

One, the Economic Survey highlighting deregulation as a driver of India’s economic growth is a welcome development. Indian public policy has disproportionately favoured programmes over policies. Even though there weren’t any great attempts at deregulation in the budget itself, the Chief Economic Advisor’s advice will hopefully help prepare the ground for another set of reforms. These lines from the Preface are worth reflecting on:

Above all, underpinning specific policy efforts will have to be the philosophical approach to governance. “Getting out of the way” and allowing businesses to focus on their core mission is a significant contribution that governments around the country can make to foster innovation and enhance competitiveness. The most effective policies governments - Union and States - in the country can embrace is to give entrepreneurs and households back their time and mental bandwidth. That means rolling back regulation significantly. That means vowing and acting to stop micromanaging economic activity and embracing risk-based regulations. That means changing the operating principle of regulations from ‘guilty until proven innocent’ to ‘innocent until proven guilty’. Adding layers of operational conditions to policies to prevent abuse makes them incomprehensible and regulations needlessly complicated, taking them further from their original purposes and intents.

“Getting out of the way” is not easy for societies that are still structured around communities, groups, and kinship. These societies are largely hierarchical in nature. Various institutional forces propelled people in Western societies to go out and build relations with strangers. This process started as early as the first millennium. As a result, dealing with strangers and building networks and communities with them became necessary. Scale became inevitable and easier. One has to trust to deal with and form relationships with strangers. Written contracts formed the basis of such trust, and other institutional developments followed. However, in close-knit and kinship-based communities such as India, the trust quotient is still high within but low without. That inhibits scale. The low-trust quotient also gives rise to elaborate verification, compliance and reporting requirements. Further, by and large, hierarchical societies are not made for disruption, change and innovation but for maintaining the status quo. Even in such societies, in places and industries where this pattern is broken, innovation, competitiveness and scale emerge. The information technology sector and the startup ecosystem that emerged in Bengaluru in the nineties are examples.

But, ‘get out of the way’ and trust people, we must, for we have no other choice. ‘Business as usual’ carries a high risk of economic growth stagnation, if not economic stagnation. Yes, trust is a two-way street and the non-government actors in the economy have to vindicate the trust reposed. In fact, quite a significant chunk of the complicated compliance requirements stem from the efforts of businesses wanting to keep out domestic and foreign competition to the detriment of other industries and the economy. Nonetheless, wiping out the trust deficit in the country is imperative and government agencies have to set the agenda in this regard. Then, it is a good bet that the Indian public will overcome the challenges and turn them into opportunities on the way to Viksit Bharat by 2047.

Two, the personal income tax cut garnered a lot of attention. It provides some much-needed relief to the salaried class. However, a point worth noting is that through this rebate, the government is moving away from its stated position of implementing a simple tax code that reduces exemptions and rebates. Were there any alternatives to increase consumption? In a Business Standard article written before the budget, Dr R Kavita Rao suggested a unilateral reduction in CGST and petroleum cess. I found this section of the article insightful:

Assuming that the Union government has accepted the idea that a stimulus in the form of taxes is a necessity in the present context, two other tax initiatives can be explored through changes in the indirect taxes systems. First, consider petroleum taxes, especially on diesel. Diesel is an input into a number of activities in the economy. Depreciation of the rupee will increase the cost of crude, which in turn will translate into higher diesel prices, further getting transmitted into higher inflation. A downward adjustment of the cess of diesel could be a useful tool to explore. There are fewer challenges in transmitting lower taxes into lower prices for this commodity.

Another alternative to consider is a reduction in goods and services tax (GST), particularly in Central GST. In popular perception, GST appears to be a regime with little flexibility for individual actors. While competition among states regarding the structure of the GST regime may invoke thoughts of a race to the bottom, the same challenge does not exist for changes in the CGST structure. In principle, it should be possible to reduce the CGST component of GST while leaving the State GST rates unchanged. Since the proposed policy interventions are being viewed as a stimulus to be provided by the Union government, a short-term reduction in the CGST rate could also be a possible policy choice. Given that GST is paid by a broader spectrum of people across the country, a stimulus through this channel could have a larger impact. This would be akin to the excise duty stimulus offered during the global financial crisis and can be reversed to signal an end to the stimulus as well.

Three, like every year, I analysed the defence budget, which was more of the same. Given the geopolitical uncertainty induced by Trump and Xi, my colleague Nitin Pai had earlier argued for a surge in defence spending. However, the government’s view seems that the China challenge must be managed through non-military means. Perhaps the government considers other tools of statecraft — diplomatic, economic, or non-conventional — more suitable for the purpose. We can only hope it is busy sharpening these tools even as defence spending continues to decline in relative terms. Nevertheless, here’s a chart capturing defence expenditure trends over the last few years.

Four, here’s an update on India’s semiconductor mission. Not much has changed since our detailed coverage in July 2024. Here are the highlights.

The government’s plan to revive the state-owned fabrication unit, SCL Mohali, saw no progress this year. As against the budgeted sum of ₹900 cr this year, only ₹11 cr was spent. Nevertheless, the government has planned another attempt costing ₹400 crore this year.

OSAT (assembly and test) plants have seen the most action. Total spending incurred by the government thus far has been ₹3144 crore, while another ₹3900 cr has been budgeted for this year.

Construction at India’s sole CMOS fabrication unit at Dholera has begun. That accounted for ₹1200 crore this year, with another ₹2500 crore budgeted for the upcoming financial year.

The sad news concerns a sector where India’s comparative advantage actually lies, i.e., fabless firms. The Design-linked incentive promised to support 100 start-ups in their go-to-market strategy, but the budgetary allocations show how far off the target the scheme is. As against the budgeted expenditure of ₹400 crore over the last two years, only ₹135 crore has been spent. Another ₹200 crore expenditure is planned this year. The reasons for tardiness include strict provisions that put firms raising substantial foreign money out of contention, confusion on the government’s right to the company’s intellectual property, and entrusting a government company, which is also a player in this domain, as the nodal regulator of this scheme. We have outlined many of these problems here.

The display fab scheme hasn’t attracted any interest, and the government hasn’t budgeted any amount for the upcoming year. I have long argued that display fabs are not strategic, and spending taxpayer money on them to reduce imports from China is a wasteful exercise. Good riddance if the government stops wasting time, money, and effort here.

---

HomeWork

Reading and listening recommendations on public policy matters

1. [Podcast] This Ideas of India episode featuring economist Anant Sudarshan is packed with TILs for anyone interested in Indian economic policy

2. [Paper] Don’t miss reading this paper on Technological Sovereignty as Ability, not Autarky

3. [Article] Nitin Pai bats for a Cabinet Committee on Science & Technology on the lines of a Cabinet Committee on Security (CCS)

---