By Manuel Kilian
Key Takeaways:
Pairing: Germany as anchor (proprietary factory data, Mittelstand processes, EU standards, high robot density) and India as scale (AI talent, low-cost engineering, Aadhaar/UPI deployment instinct, IndiaAI Mission, a billion-person market).
Four frictions: India’s one-size public rails vs Germany’s bespoke plants; who owns the high-value work (not German IP + Indian outsourcing); EU AI Act/data rules vs Indian localisation; diplomacy in years vs models in weeks. Suggested vehicle: the 150+ German GCCs already in India (~130,000 staff), which move at company speed. The bet is wire the physical world and keep that value—not beat OpenAI or DeepSeek.
Introduction: Is There a Middle-Power Play in AI?
Canadian Prime Minister Mark Carney has lately done more than anyone to make the idea of the middle power fashionable. These are states with real weight that nonetheless fall short of superpower status. At the World Economic Forum’s annual meeting in Switzerland in January 2026, he told attendees that the world had entered an era of rupture rather than an orderly transition, and that “middle powers must act together because if we’re not at the table, we’re on the menu.”[1]
The idea of middle powers is appealing because it offers such states a sense of purpose and a promising role. It is also intellectually convenient, since, in theory, it can be applied to almost anything. However, few fields suit it as well as AI. It is an area where the notion of rupture is truly applicable: the two current superpowers, the US and China, have taken such a leading role that smaller nations are left wondering how they can secure a place at the table.
For all its appeal and hype, the idea has to hold in reality: is there really a ‘middle power play’ for AI? To answer this, it must be put to the test across two propositions. Firstly, that there is room to act outside the American and Chinese duopoly, whose dominance rests on frontier models, and to a lesser extent on compute capacity. Secondly, that middle powers can occupy that terrain by combining complementary strengths, in theory and in practice—as tested here with Germany and India, an established and a rising middle power respectively.
The first proposition, that room exists beyond the American and Chinese duopoly, begins with what that duopoly rests on, namely frontier models—the most advanced general-purpose systems, such as those behind America’s ChatGPT, Claude and Gemini, and China’s DeepSeek and Kimi. The US and China dominate their production: in 2024, American institutions produced 40 notable models, China 15 and Europe three, with few elsewhere.[2] Many middle powers have tried to build their own, but only a few, notably France’s Mistral, have stayed relevant. From this perspective, competing appears to be a losing game.
This reading assumes that the frontier model is where the contest is decided. It may well be at the very top level of complex, agentic tasks; the gap between the leading labs like OpenAI or Anthropic and the rest may even be widening.[3] Winning there, however, might no longer matter.
Models have started being commoditised in the middle of the performance range, which will be sufficient for many tasks. As AI is used to automate a wide range of repetitive, specific assignments, there is pressure to move away from the most advanced closed models, which are still better suited to very challenging work, and towards open models. These are models whose underlying weights are published, so they can be downloaded, run, and adapted freely, unlike closed models reachable only through a provider’s interface.
Such models can potentially run many times faster and at a fraction of the cost,[4] as the following numbers bear out: open-weight models reduced the performance difference with closed ones from 8 percent to just 1.7 percent on some benchmarks within a single year.[5] The inference cost—the price of running a system performing at the level of GPT-3.5—fell more than 280-fold between November 2022 and October 2024.
In short, the US and China continue to dominate the frontier model landscape, and any hype surrounding open models catching up with the absolute cutting-edge can be distracting. The real structural shift, and the precise opportunity for middle powers, lies in growing and channelling demand towards the high-volume, lower-complexity workflows in industry and government that specialised open models can meet. Economically speaking, the further the market moves away from a single dominant model architecture, the better the position of every nation unable to build a frontier model.
This suggests that the value of AI might be moving downstream. If upstream is where frontier models are developed and trained, downstream is where AI is put to use: the unglamorous, complex work of embedding it into existing operations and services, or, in other words, applied AI. For many of these real-world use cases, specialised open models may be sufficient and sometimes the better choice, particularly where cost, control, and integration are the most important factors. Take a narrow, high-volume task such as classifying maintenance records or standardising supplier documents: a small, fine-tuned open model can potentially match a frontier one at a fraction of the cost.
Admittedly, this is an emerging trend, but there is some early evidence to support it. The widely cited 2025 study by the Massachusetts Institute of Technology (MIT) found that around 95 percent of enterprise generative AI pilots had no measurable effect on profit.[6] The figure is sometimes contested, resting on a single study with a narrow definition of failure, but it does not stand alone. A 2025 McKinsey survey of almost 2,000 enterprises found that while 39 percent of respondents could attribute some enterprise-level earnings before interest and taxes (EBIT) impact to AI, only around 6 percent cleared the bar of both material earnings impact and significant value.[7] Both studies point in the same direction: the bottleneck is not model performance, but rather workflow redesign, data readiness, and integration into existing systems—the organisational work that determines whether a deployment can withstand contact with real processes.
Where the Anchor Meets the Scale
With the first proposition established, that room exists for nations beyond the two superpowers, the second follows: that two middle powers, acting together, can take that room and amount to more than the sum of their parts. This claim is best tested on Germany and India.
A concession comes first. The commanding positions in AI’s hardware sit elsewhere, with the most advanced processors being designed in the US and fabricated in Taiwan; the precision gears in a robot’s joints being dominated by a few Japanese firms;[8] and the high-grade magnets in its motors are around 90 percent Chinese-made.[9] Neither Germany nor India plays a decisive role in any of these areas.
Yet while the above can be choke points in global value chains, none of these are where value in applied AI is directly created. The important contest is downstream, requiring four things: proprietary data from a real domain; the ability to embed a model in live operations; a means of deploying and distributing it on a large scale; and the trust and standards that enable it to operate in the real world.
Germany holds one part. It owns valuable industrial processes and the proprietary data associated with them. This data, spanning decades and covering the automotive, chemicals, and machinery industries, lies outside the public internet and cannot be scraped or copied. It has the engineering expertise to determine which problems are worth solving, as well as the German and European certification and standards regime that decides what can operate at all. Its factories are among the most automated in the world, ranking third globally by robot density.[10]
India holds much of the rest. It has a very large AI talent base and was ranked among the first in AI skill penetration in Stanford University’s latest index.[11] It has the engineering scale to build and run applications at low cost—and it has a deployment instinct Germany lacks: a public-technology tradition, proven at scale by Aadhaar, its digital identity system, and the Unified Payments Interface (UPI), its instant payment rails, that knows how to carry a single service to more than a billion people. It is now turning that tradition to AI through the IndiaAI Mission, currently under way.[12] Its vast domestic and rapidly growing market offers a proving ground for applied AI at a scale no single European economy can match.
This is why the pairing can be greater than the sum of its two parts: Germany is the anchor and India is the scale. The anchor comprises the proprietary data, high-value problems, and standards that make applied AI tangible and difficult to replicate. The scale comprises the talent, engineering expertise and reach that transform these concepts into functioning systems and facilitate their adoption by a billion-strong market. Without scale, an anchor is inert; without an anchor, scale is underused. Each supplies what the other lacks.
Four Frictions to Overcome
That is the theory, and it is straightforward. The reality is not. Every middle-power play encounters the same discrepancy between ambition and reality, and this one is no exception. The political momentum is real enough: the same month Carney spoke in Davos, the European Union and India finalised a long-delayed trade agreement, and their leaders shared podiums at successive summits.[13] Moving from words to a functioning AI partnership, however, requires solving concrete, systemic problems. It is not only a question of whether these nations have the political will to work together; it is a question of whether their economies can actually mesh.
Four big challenges stand between the idea and a working partnership. Naming them plainly is the first step to meeting them, and the more squarely each is faced, the better the chance of building something that actually holds.
The first is to fit two different shapes together. India’s success in public technology rests on a high degree of homogeneity. Aadhaar and UPI work because they impose a single, standardised interface on the whole population. In contrast, German industrial strength is based on fragmentation and bespoke solutions, with value locked in the proprietary setups of the Mittelstand, Germany’s dense base of specialised, often family-owned small and mid-sized manufacturers. Specialised chemical plants in Ludwigshafen and premium assembly lines in Stuttgart run on custom legacy systems and idiosyncratic data structures. A population-scale, UPI-style distribution model cannot simply be dropped onto that. The scale engine does not, by itself, fit the anchor’s geometry. The job is thus to build tools that adapt to each setup instead of forcing one template on all of them.
The second is to find a working compromise on the division of labour. The question is where the high-value work in applied AI sits, and how to design the cooperation around it so both sides share in it, and both have reason to make it work. Germany and India come at this from different starting points, so their instincts might naturally pull them apart. Germany would be inclined to treat India as a mere implementation partner; a role India is working to climb out of. If it were left as German IP and Indian delivery, the arrangement would just be outsourcing in a new form, which India has good reason to refuse. Getting the split right is not a detail for later. It is the condition on which the partnership stands or falls.
The third is to agree to a regime for regulatory compliance both sides can live with. Applied AI only works if a domain expert in Germany and a deployment engineer in India can exchange factory data and machine logs, and here the two sides can pull in opposite directions. European rules lean towards caution: data-protection law limits what can cross borders, and the AI Act, the EU’s 2024 law that grades AI systems by risk, adds compliance duties on higher-risk systems. India, meanwhile, brings its own views on sovereignty and localisation. Since the way AI is regulated sets the limits of what can be built from the data, the task is to settle a shared framework up front, one that says what can cross and what must stay.
The fourth is a clash of clocks. State-to-state cooperation, bilateral training schemes, semiconductor memoranda, and joint standards bodies operate within a diplomatic timeframe measured in years. Meanwhile, the underlying technology evolves in weeks. By the time Berlin and Delhi set up a joint framework or certified training pipeline, the problem it was designed to solve is often already outdated. The instruments risk becoming obsolete as soon as they arrive. This is not a failure of will, but of instrument: the classical political cycle, with its summits, communiqués and multi-year programmes, simply does not operate quickly enough for AI. This raises the question of who, if not governments, can establish a partnership at the speed demanded by the field.
Rather than marking a dead end, the four challenges set out exactly what must be targeted, and where institutional creativity is required from both nations.
The answer to who could carry this out is already visible, and it cannot be just another summit. It could fall on so-called global capability centres, in-house offshore units that companies run for their own engineering, and R&D. German companies operate over 150 such centres in India, employing more than 130,000 professionals, with engineering the single largest area of work.[14] German industrial expertise and Indian talent are housed in the same buildings, operating at the speed of business rather than the pace of diplomacy. The anchor and the scale are already present there without the need for a treaty.
It is within these corporate ecosystems—where commercial survival naturally outpaces political scheduling—that the four challenges of shape, division of labour, regulatory compatibility, and clocks can be solved. If these challenges are treated as a practical agenda for co-investment rather than as permanent geopolitical barriers, Germany and India could establish a working model for collective leverage.
For now, the middle-power approach remains a bet, but the direction is clear. Neither nation needs to outrun the superpowers at the frontier; rather, they need to focus on the unglamorous, downstream work of wiring the physical world into code and retaining the value that follows.
Endnotes
Stanford Institute for Human-Centered Artificial Intelligence, “The 2025 AI Index Report” (Stanford, CA: Stanford University, 2025), https://hai.stanford.edu/ai-index/2025-ai-index-report.
Nathan Lambert, “The Next Phase of Open Models,” Interconnects, March 2026, https://www.interconnects.ai/p/the-next-phase-of-open-models.
Stanford Institute for Human-Centered Artificial Intelligence, The 2025 AI Index Report.
Aditya Challapally et al., “The GenAI Divide: State of AI in Business 2025” (Cambridge, MA: MIT Project NANDA, July 2025), https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf.
Press Information Bureau, Government of India, “IndiaAI Mission,” 2025, https://www.pib.gov.in/PressNoteDetails.aspx?NoteId=156786&ModuleId=3®=3&lang=1.
About the author: Manuel Kilian is the co-founder of The Agentic State. His work stands at the intersection of government and technology. He founded and sold the GovTech platform GovMind, built the Global Government Technology Centre Berlin, and advises Germany’s Federal Ministry for Digital Transformation on international affairs and the UAE Prime Minister’s Office on adopting agentic AI at scale.
Source: This article was published by Observer Research Foundation
About Observer Research Foundation
ORF was established on 5 September 1990 as a private, not for profit, ’think tank’ to influence public policy formulation. The Foundation brought together, for the first time, leading Indian economists and policymakers to present An Agenda for Economic Reforms in India. The idea was to help develop a consensus in favour of economic reforms.
View all posts by Observer Research Foundation →











