India's software exports hit 5.2% of GDP as AI moves the work upmarket
The AI-displacement story priced into India's IT industry for three years does not show up in the export data yet — the counter-evidence is in a new ING report. The same week, an open-source project published head-to-head numbers against the closed model everyone is arguing about.
India's software services exports have climbed to about 5.2% of GDP, from 3.3% before the pandemic, and export growth was still running at roughly 12% year-on-year in mid-2026 — slightly above its pre-pandemic rate. That is the finding in an ING research note published Wednesday by Deepali Bhargava, the bank's head of Asia-Pacific research, who set out to test whether generative AI is actually eating the outsourcing market it was supposed to eat. The sharper number sits underneath: business services exports — finance and accounting, risk management, legal and compliance support, engineering design, R&D, analytics, consulting — have more than doubled their share of GDP to 3.3% from 1.6% before Covid. Those are the tasks hardest to automate, and they are the ones India is winning.
The aggregate picture holds against the global market too. India's exports of digitally delivered services are up 45% since 2022, against 32% worldwide, which means the country gained share while China and Japan lost ground. India already accounts for more than half of the global outsourcing industry, roughly $205 billion in annual software services revenue and about 5.8 million workers — not a small-sample result.
The labour market is where the disruption shows, and it looks different from the headline fear. India's top IT firms added just 17 net employees in the first nine months of FY2026, down sharply from the previous year, with entry- and mid-level roles under the most pressure — productivity rising, labour intensity falling. NASSCOM still expects the technology workforce to reach around 6 million in FY26, up 2.3%, which is what reshaping rather than replacing looks like in a payroll file. The same ING note finds the Philippines, the other outsourcing powerhouse, with combined tech and business services exports at 7.1% of GDP, up from 6.3% before the pandemic, and IT and professional services employment up about 4.5% over the past year. If AI were destroying outsourcing demand, 2026 is the year it should be visible in the export line. What is visible instead is the mix moving upmarket — a bet on skills, not headcount.
The open-source answer to Jev shipped, with a benchmark table aimed straight at TypeSafe's closed model. Laya, from Convai Innovations, is a non-autoregressive "System 1" decision engine: it answers typed questions — a choice from a fixed label set, a score on an ordinal rubric, or a calibrated true/false probability — in a single forward pass, with no text generation and therefore nothing to parse or hallucinate. It is Apache 2.0, installable from PyPI, and ships three checkpoints behind a router that picks one per request by script and language: a 421M English model, a 322M multilingual model covering 100-plus languages, and a fine-tuned typed-decisions checkpoint. The author measures 32.8 ms for a single question on a T4 and 7.2 ms per question batched.
The comparison is the point. Against TypeSafe's Jev — the $40 million lab whose pitch is the same architecture — Laya's published table claims 0.766 against Jev's 0.727 on 2,000 typed decisions, an expected calibration error of 0.081 against 0.246, and a median latency 7.8 times lower. The Jev figures come from third-party benchmark runs and TypeSafe itself, not from Laya's author, and the 0.766 checkpoint was fine-tuned on that benchmark's own training split — so read it as a floor on what an open model can be tuned to, not a clean win. The author also publishes his own limits, which is rarer than it should be: the base checkpoints are near chance zero-shot on typed decisions, the English checkpoint collapses outside Latin script while staying confident, and choice questions degrade past roughly 20 options. We covered the credit fight over this architecture two days ago — The Jev idea was published a year ago — open, and unnoticed — and this release is the part that argument was missing: something anyone can download and test.
What to watch: whether anyone independently reproduces Laya's numbers now that it is pip-installable, and whether India's IT hiring line stays flat while exports keep compounding.
If AI is automating the routine work but the export numbers keep climbing, what is actually being displaced — tasks, or junior headcount? Tell us in the comments.
Sources: Laya (GitHub) · AbdelStark's Jev benchmarks (GitHub)
