India's $221.4 Billion Software Export Year Rests on Off-Site Services, Not Products

India exported an estimated $221.4 billion of software services in the year to March 2026, up 8.2% from $204.7 billion. Almost all of the dollar increase came from IT services. Software-product development contracted, and off-site delivery supplied more than nine-tenths of the total.

That is evidence of a durable remote-services machine. It is not evidence that India completed a shift to software products, or that artificial intelligence raised output per worker. The Reserve Bank of India's survey measures export value and delivery mode. It does not measure hours, headcount, AI use or labor productivity.

Off-site delivery describes where the work happened. It does not explain how the work was produced.

Problem: the total combines observation and estimation

The RBI contacted 7,569 exporters and received responses from 2,363, a 31.2% response rate. Those respondents included most large companies and reported $196.9 billion of exports. The final estimate added $24.5 billion for smaller nonrespondents, lifting the reported value by 12.4%. Put differently, 11.1% of the headline total was estimated.

This is not a conventional random sample with a published margin of error. The RBI's stated estimation method groups companies by service activity, uses medians because export values are highly skewed, and estimates only the off-site component for nonrespondents. That treatment reflects the observation that major companies account for most on-site exports.

The method is defensible for thousands of small exporters. It also means the off-site share is partly modeled. The release supplies no confidence interval, no top-company concentration table and no respondent-versus-estimate split for each activity. Readers can reproduce the arithmetic, but not the uncertainty around it.

That missing split matters more for small categories than for the headline. The $24.5 billion nonrespondent estimate is almost four times the entire $6.4 billion product-development category. A 10% error across the estimated block would move the total by only about 1.1%. Reallocating a much smaller amount between IT services and products could change the product result materially while leaving the aggregate intact. The published tables do not permit that sensitivity test. This is a disclosure gap, not evidence that the estimate is wrong.

Analysis: IT services carried the increase

The aggregate hides a narrow growth engine. IT services rose to $147.0 billion from $131.3 billion, accounting for $15.7 billion of the total $16.7 billion increase. Engineering services rose to $12.0 billion from $10.8 billion. BPO barely moved, reaching $56.0 billion from $55.8 billion. Software-product development fell to $6.4 billion from $6.8 billion.

Currency translation does not explain the expansion. The RBI's rupee total increased 13.0% to ₹19.56 trillion from ₹17.31 trillion. Its annual-average conversion rates imply roughly ₹88.34 per dollar in FY2025-26 and ₹84.59 a year earlier. A mechanical restatement at the earlier rate would put growth near 13%, not below 8.2%. That is a translation check, not a provider-level constant-currency calculation, but it rules out a weak result concealed by the exchange rate.

The product decline survives the same test. Product-development exports fell 1.9% in rupees and 5.9% in dollars. On-site services present the opposite case: they fell 3.2% in dollars, to $18.4 billion, but rose 1.4% in rupees. Calling that a real contraction would outrun the data.

Off-site exports rose 9.3% to $203.0 billion and their share increased one percentage point to 91.7%. The shift is real in the survey's categories. Its cause remains unobserved.

Company disclosures underline that limit. TCS reported a 2.4% constant-currency revenue decline and 584,519 employees, down from 607,979. Infosys grew 3.1% in constant currency while its headcount rose 1.6% to 328,594. Wipro's IT-services revenue fell 1.6% in constant currency. Three different outcomes do not support one sector-wide automation story.

Buyer concentration is doing more work

The United States bought $119.7 billion, or 54.1% of the total, up from $108.3 billion and 52.9%. US-bound exports grew 10.5%, faster than the aggregate. Europe grew 4.6% to $70.3 billion and lost one percentage point of share. Together the two regions accounted for 85.9% of exports.

Invoice currency is a separate concentration. The dollar's share of invoices increased to 72.6% from 72.0%. A US-dollar invoice does not prove that the buyer is in the US, just as a US buyer does not prove that delivery occurred there. The survey exposes all three dimensions, but they should not be collapsed into one geography claim.

The broader $239.3 billion total adds $17.9 billion of local sales through overseas commercial presence. Those affiliate sales are Mode 3 under international trade-in-services statistics. They are not cross-border exports in the balance-of-payments sense. The $221.4 billion figure instead combines cross-border supply, consumption abroad and the presence of natural persons.

Implications: measure the production function

India's export base passed the resilience test. The stronger claim still lacks evidence.

The next RBI release should publish activity and delivery-mode splits for reported and estimated companies, plus a concentration table that does not identify individual firms. A bridge between survey totals and balance-of-payments software receipts would show revisions and coverage differences. Headcount, billable hours and realized pricing would let readers distinguish more work from higher rates.

Providers can close the AI gap by reporting revenue and effort saved from deployed systems under stable definitions. Annualized AI revenue, trained-employee counts and deal announcements do not reveal whether a client paid for more output with fewer hours.

For now, the defensible conclusion is narrower. India sold substantially more remote IT services, mainly to a more concentrated US buyer base. It did not demonstrate a product pivot. The productivity verdict needs a denominator.