AI infrastructure development is not the same as AI diffusion
| AUTHOR | Arindam Goswami |
| DATE | October 27, 2026 |
| CATEGORIES | High Tech Geopolitics AI AI Diffusion Research and Development |
India’s AI build-out is almost universally being measured and celebrated in terms of the amount of money being spent. Tracxn data shows that AI infrastructure was the single largest technology segment in terms of funding in the first nine months of 2026 (US$1.2bn). This is in the wake of the over US$200bn commitments from Reliance and Adani at the AI Impact Summit in February. The implicit promise is that this money will be used to take AI beyond the handful of frontier labs. A cross-country analysis of publicly available data reveals that there is no evidence whatsoever that the infrastructure funding to date has achieved this. What does correlate with AI diffusion is installed hosting capacity and the general depth of domestic credit.
The analysis is based on three public data sources. The first is the World Bank’s data on investment in infrastructure with private participation (93 LMICs). The second is Microsoft’s AI user share data (147 economies) which gives an estimate of the share of the working age population using generative AI. The third is the GitHub Innovation Graph which counts developers by language. The analysis focuses on developers working in Jupyter Notebook (the standard format for ML work). The US, China and the UK are all excluded as the key question is about diffusion outside of these three countries.
In all the economies that are comparable in terms of income, internet usage and population, those with more ICT infrastructure funding don’t have higher levels of AI use. They also don’t have more AI developers (either as a percentage of all developers or per head of population). The estimate is so precise that you can rule out a large effect: one standard deviation more ICT funding is associated with at most around 0.7 percentage points of AI use (the average being around 20%). The same is true of energy, transport and water funding. The follow-up analysis of the same set of economies from 2020 to 2025 shows a very small positive association with developer activity, though the estimate is so imprecise that it’s indistinguishable from zero.
Two other variables do matter. The number of secure internet servers per person (the closest public proxy for installed hosting capacity) is the single best predictor of developer activity. The top quarter of economies by server density have a median of 268 Jupyter developers per million, the bottom quarter having 13. The depth of private credit is the predictor of AI tool usage (again even after income is held constant).
This isn’t as surprising as it first appears. The vast majority of generative AI use outside the rich world is powered by models hosted in data centres overseas. A student in Patna needs a phone, a data plan and a model that works in her language. She doesn’t need a GPU cluster in Navi Mumbai. Microsoft’s own Q1 2026 report points to improved local language performance as the key driver of AI uptake across Asia, not any new local compute.
India’s own numbers are consistent with the pattern. Its AI adoption rate increased from 14.2% to 18.5% in the period from H1 2025 to mid-2026. That puts it at 62nd out of 144 economies in the non-frontier group, again broadly in line with what you’d expect based on its level of income and connectivity. It had around 180k Jupyter developers in 2025 (some four times the 2020 figure) and a good deal of that growth predates the major infrastructure announcements. The World Bank has only recorded $1.47bn of private ICT infrastructure commitments for India in 2023-24 (0.06% of GDP). The bulk of the headline commitments have not even reached financial close.
The findings have caveats that must be borne in mind. The World Bank figure only covers projects with private sector participation so the bulk of the balance sheet data centre spend is not included. A significant proportion of the ICT finance recorded is from 2022-24 and the effects observed in 2026 may be too early to be meaningful. The Microsoft data uses shared regional values for the 36 economies which have very thin data. None of the estimates are causal.
Even so the evidence speaks quite clearly to Indian policy. Commitments to investment are not in any way a proxy for AI diffusion. The more pertinent metric is actually who is using it or building with it. The IndiaAI Mission’s subsidised access to compute is in the same ballpark as the data would suggest in terms of capacity, but its actual reach into the small developer community outside of the big firms will need to be measured directly. For actual public use, the amount of money being spent on Indian language model performance and on digital skills is probably more important than the amount of data centre floor space. The massive 2026 pledges may well alter the picture. That will be apparent in the 2027 data. For now the onus is on the argument that investment drives diffusion.