--- title: "US$300bn pledged to India's data centres; subsidised GPUs reportedly 20% used" summary: "Data-centre pledges, GPUs onboarded under the IndiaAI Mission and their reported use, from Rhodium Group's 27 October report and government statements, May 2025 to October 2026." url: https://www.thetechbag.com/blog/india-ai-compute-pledges-gpu-use format: Analysis desk: "AI in the enterprise" region: India published: 2026-11-02 facts_checked: 2026-11-02 author: TechBag (The TechBag Journal) tags: ["AI compute","GPUs","IndiaAI Mission","data centres","Rhodium Group","MeitY","Reliance","Adani","Yotta","GPU utilisation","subsidised compute","AI infrastructure"] cover: https://www.thetechbag.com/blog/covers/india-ai-compute-pledges-gpu-use.png --- # US$300bn pledged to India's data centres; subsidised GPUs reportedly 20% used *Data-centre pledges, GPUs onboarded under the IndiaAI Mission and their reported use, from Rhodium Group's 27 October report and government statements, May 2025 to October 2026.* Analysis · AI in the enterprise · By TechBag · Published 2026-11-02 · Facts checked 2026-11-02 · Canonical: https://www.thetechbag.com/blog/india-ai-compute-pledges-gpu-use The short version 1. Rhodium Group counted more than US$300bn of AI data-centre commitments to India in a year, two-thirds of it from Reliance and Adani, and wrote on 27 October that nearly 80% of the GPUs subsidised under the IndiaAI Mission “reportedly sit unused”.[1] 2. The government’s own count was 38,000 GPUs from October 2025 to March 2026 and 45,000+ by June 2026, at ₹65 an hour. It publishes hours used, 93.18 lakh by August 2026 across 237 projects, but no utilisation rate.[6][9][10] 3. For a buyer, the subsidised hours are open to DPIIT-recognised startups and MSMEs, researchers, students and government bodies through the IndiaAI Compute Portal; in the allocations IndiaAI lists, the subsidy was 40% of the bill.[11][12] Everywhere else the shortage in AI compute is of supply. Alphabet said on 28 October that for Google Cloud “there’s huge demand” and that, “given the supply-constraint environment”, it had started renting third-party capacity.[2] Rental prices for the newest chips rose from August to late October, as our [GPU price trend](https://www.thetechbag.com/blog/gpu-rental-prices-turn-up-b200-h100) shows. Rhodium Group’s report on India, published the day before, opens with the opposite problem: “India is building AI compute faster than it can use it.”[1] If you run a company in India that needs GPU time, the question is practical: is the subsidised capacity there for you, and what does the gap between the pledges and the use mean for what you pay? ## The pledges: more than US$300bn, two-thirds from two Indian groups Rhodium’s Table 1 lists twelve commitments, from Reliance’s US$110bn over 2026–2032 and the Adani Group’s US$100bn through a joint venture over 2026–2035, to US$1bn from Airtel’s Nextra.[1] The government’s own framing is narrower: a PIB backgrounder of 14 September said “investments of nearly USD 70 billion are already underway”, with “an additional USD 90 billion in announced projects”, and put installed data-centre capacity at 1.57 GW in August 2026, projected to reach nearly 8 GW by 2030.[3] At the India AI Impact Summit in February, PIB recorded Reliance’s pledge as “USD 110 billion over seven years towards AI-focused infrastructure” and Adani’s as “USD 100 billion by 2035”.[4] **Figure: Reliance and Adani account for US$210bn of the roughly US$316bn of commitments Rhodium lists** *Major AI data-centre investment commitments in India, US$bn, as listed by Rhodium Group, periods from 2025 to 2035* Reliance India, 2026-32 110 US$bn Adani Group (JV with ConneX) India, 2026-35 100 US$bn AirTrunk Australia, 2026-30 30 US$bn Amazon US, 2026-30 21 US$bn Microsoft US, 2025-29 20.5 US$bn Google US, 2026-30 15 US$bn HyperVault (TCS) + TPG India, 2026-32 7 US$bn Yotta India, 2026-27 7 US$bn CtrlS India, 2026-30 2 US$bn NTT Data Japan, 2026-28 1.5 US$bn HCLTech India, 2026-31 1.5 US$bn Airtel (Nextra) India, 2026-30 1 US$bn | | Value (US$bn) | | --- | --- | | Reliance | 110 | | Adani Group (JV with ConneX) | 100 | | AirTrunk | 30 | | Amazon | 21 | | Microsoft | 20.5 | | Google | 15 | | HyperVault (TCS) + TPG | 7 | | Yotta | 7 | | CtrlS | 2 | | NTT Data | 1.5 | | HCLTech | 1.5 | | Airtel (Nextra) | 1 | Source: Rhodium Group, Hyperscaler Heaven?, Table 1, based on company announcements · as of 27 October 2026 · Amazon is listed as 21+. Periods differ by company; totals are commitments, not spending TechBag countHow the total adds up, and why we do not add it to the government's Rhodium’s twelve entries sum to about US$316bn, with Amazon counted at its stated minimum of US$21bn; Reliance and Adani are 66% of it. The periods run from two years (Yotta) to ten (Adani), so the sum is an order of size, not a schedule. The government’s US$70bn underway and US$90bn announced are a different cut of the same boom, and the two cannot be added. The power these sites need is in [our piece on India’s data-centre gigawatts](https://www.thetechbag.com/blog/india-ai-data-centres-power-gigawatts). Rhodium’s account of why the money comes covers cost, power and tax. Mumbai ranks as the second-cheapest major hub in the world to build a data centre, after Shanghai; India installed a record 50 GW of renewable capacity in 2025; and the Union Budget added a 20-year tax holiday for foreign cloud providers. It also notes that India produces 20% of the world’s data but hosts under 3% of global data-centre capacity, and that a typical project can need 30–40 separate approvals.[1] ## The GPUs: 38,000 for five months, then 45,000 The IndiaAI Mission, approved in March 2024 with ₹10,371.92 crore over five years, began with a target of 10,000 GPUs.[6] MeitY said on 30 May 2025 that the common compute capacity had crossed 34,000.[5] From October 2025 PIB put it at 38,000, available at ₹65 an hour, which it called on 10 February “nearly one third of the global average cost”.[6][7] On 17 February, at the AI Impact Summit, the IT minister, Ashwini Vaishnaw, said India would add 20,000 GPUs “in the coming weeks” beyond the existing 38,000.[8] On 25 March PIB still gave “more than 38 thousand”.[9] A PIB factsheet of 13 August gave “45,000+ Graphics Processing Units (GPUs), as of June 2026”.[10] **Figure: The official GPU count held at 38,000 from October to March, and stood at 45,000+ by June** *GPUs onboarded under the IndiaAI Mission as stated by the government, with Rhodium's report on their use, May 2025 to October 2026* | Date | Event | Detail | | --- | --- | --- | | 30 May 2025 | 34,000+ GPUs | MeitY: common compute capacity crosses 34,000 | | 12 Oct 2025 | 38,000 GPUs at ₹65/hour | PIB: up from an initial target of 10,000 | | 17 Feb 2026 | +20,000 promised | Vaishnaw: “in the coming weeks”, beyond 38,000 | | 25 Mar 2026 | 38,000+ onboarded | PIB: 190 projects approved | | 13 Aug 2026 | 45,000+ as of June 2026 | PIB factsheet: 237 projects, 93.18 lakh GPU-hours | | 27 Oct 2026 | ~80% reportedly unused | Rhodium Group; no source given | Source: MeitY (Digital India); PIB releases, backgrounders and factsheets; Rhodium Group · as of 2 November 2026 · Each count is as stated on that date; the government publishes no monthly series Both numbers are the government’s, at different dates: 38,000 in statements from October 2025 to March 2026, and 45,000+ as of June 2026 in the August factsheet. The 38,000 plus the promised 20,000 would be 58,000; the government has not said, in any document we found, how many of the 20,000 had arrived by June. We found no dated official target beyond that February promise. ## The use: one GPU in five, reportedly Rhodium wrote that “nearly 80% of its GPUs reportedly sit unused, despite availability at prices less than 40% below the market rate”. The report names no source for the 80% figure.[1] Its reasons are on the demand side: “only a handful of India’s AI startups actually train models and therefore need dedicated GPU clusters”, and Indian AI startups raised US$1.3bn last year, which it puts at 0.6% of private equity and venture money invested in the world. It also wrote that enterprise and edge data centres used only by Indian entities are “less than 13%” of India’s operational capacity, against roughly 60% leased out by hyperscalers.[1] **Figure: Rhodium reports about one subsidised GPU in five in use, and under 13% of capacity serving only Indian entities** *Shares reported by Rhodium Group: IndiaAI Mission GPUs in use, and operational data-centre capacity in enterprise and edge sites, 2026* 20% IndiaAI Mission GPUs in use (reported, Rhodium) “nearly 80%” reportedly unused; no source given 13% Capacity in enterprise and edge sites used only by Indian entities “less than 13%”; about 60% leased by hyperscalers Source: Rhodium Group, Hyperscaler Heaven? · as of 27 October 2026 · Reported, not measured by TechBag. The 20% is the inverse of Rhodium's 'nearly 80%' unused TechBag countWhat the government's own hours say The August factsheet says 237 projects had used subsidised computing by August 2026, “covering 93.18 lakh GPU hours”.[10] That is the output of 45,000 GPUs running for about 207 hours, under nine days, spread over the time since allocations began in 2025 (our arithmetic). It does not give a utilisation rate, because the GPUs belong to the providers that empanelled them, which can sell the same machines to other customers. Where the capacity goes instead is also in Rhodium’s account. Yotta’s chief executive “estimates that nearly 80% of revenue already comes from foreign clients”, and the company “plans to allocate half of its $7 billion worth of NVIDIA GPUs purchases this year” to NVIDIA’s DGX Cloud Lepton marketplace under a US$1bn, four-year agreement.[1] Rhodium’s reading is that India’s data centres “will primarily be used to scale global cloud ecosystems rather than serve the domestic market”. The government’s reading, in the same months, is of “affordable high-performance computing” that is “lowering entry barriers for Indian researchers, start-ups and innovators”.[1][10] We report both; the hours and the GPU count above are the only measured numbers either side has published. ## What a buyer can do: apply for the subsidised hours The capacity is rented through the IndiaAI Compute Portal. Startups and MSMEs qualify with DPIIT recognition, experience in AI or machine learning, and annual revenue of ₹50 lakh or funding of ₹1 crore; researchers with an h-index of 5 or more, AI publications in journals with an impact factor above 2.5, or 150+ citations; early-stage startups with four years of AI experience and a letter from the IndiaAI CEO or a joint-secretary-rank officer; and students, IndiaAI fellows and government entities on their own criteria.[11] **Figure: Five steps from sign-in to approved GPU hours, with the subsidy requested in the bill** *How to apply for subsidised compute on the IndiaAI Compute Portal, as the portal's own page describes it* Step 1 of 5 1. Sign in Log in through DigiLocker, e-Pramaan or Jan Parichay. Source: IndiaAI Compute Portal · as of 2 November 2026 · Undated portal page. Approval is 'by PMEC as per SLAs'; the portal does not publish the SLAs TechBag countWhat the subsidy was in the listed allocations IndiaAI’s compute-capacity page lists allocations with the GPU type, provider, bill and subsidy. In 14 of the 21 rows with a bill (allocations from April to July 2025), the subsidy was exactly 40% of the bill; the GPUs were mostly H100 and H200.[12] The page shows no later allocations, so we cannot say whether the share has changed. If you want subsidised GPU time in India **Who it is for.** Indian startups, MSMEs, researchers, students and public bodies that meet the portal’s criteria. The portal says it serves “Industry” too and counts 303 end users, but its eligibility list has no category for a large enterprise buying for itself, so ask IndiaAI before you plan on it.[11] **The price.** ₹65 an hour is the government’s figure.[6][7] None of the documents we read names the GPU model the rate applies to or says whether 18% GST is included, so ask the empanelled provider for the bill of materials, with GST shown, before you plan around it. **The rule that shapes the market.** The Union Budget 2026–27 gave eligible foreign cloud service providers a tax holiday until 2047 for “global operations using India-based data-centre infrastructure”, PIB’s September backgrounder says.[3] Rhodium describes it as a holiday for foreign hyperscalers “who sell overseas”, and links it to capacity being sold abroad.[1] Our viewOur view: how to use it If you qualify, apply for the hours you can schedule, and keep a commercial account as the fallback, priced at today’s market rate rather than last year’s. If you do not qualify, the reported spare capacity is a reason to ask Indian GPU providers for a term rate in writing, not a promise of one. How we did this Figures as of 2 November 2026 The pledge chart is Rhodium Group’s Table 1 as published (we did not re-check each company’s announcement); the sum and shares are our arithmetic. GPU counts are quoted from the dated government documents listed, each as of its own date; the government publishes no monthly series, so the timeline shows statements, not a measured line. The 20% in the waffle is the complement of Rhodium’s “nearly 80%” unused, which the report gives without a source. The GPU-hours arithmetic divides 93.18 lakh hours by 45,000 GPUs and is an illustration, not a utilisation rate. The 40% subsidy share is our count of the allocation rows on IndiaAI’s page. Not used: a government factsheet said in coverage to be dated 29 October, which we could not find (the US$70bn and US$90bn figures are from PIB’s 14 September backgrounder), and a 100,000-GPU target for December 2026, for which we found no official statement. Global capacity and prices are in our separate GPU price trend and are not repeated here. Spotted an error? Write to us; corrections are published at the foot of the piece. Sources 1. [Rhodium Group (Ankit Tiwari and Reva Goujon): Hyperscaler Heaven? A Reality Check on India's AI Ambitions](https://rhg.com/research/hyperscaler-heaven-a-reality-check-on-indias-ai-ambitions/) — 27 October 2026 PDF with Table 1: https://rhg.com/wp-content/uploads/2026/10/Hyperscaler-Heaven-A-Reality-Check-on-Indias-AI-Ambition-and-Geopolitical-Alignment.pdf 2. [Alphabet, Q3 2026 earnings call transcript](https://s206.q4cdn.com/479360582/files/doc_events/2026/Oct/28/2026_Q3_Earnings_Transcript.pdf) — 28 October 2026 3. [PIB backgrounder: Data Centres in India, Infrastructure for the Digital Age](https://static.pib.gov.in/WriteReadData/specificdocs/documents/2026/sep/doc_9859_20260914_10550401.pdf) — 14 September 2026 4. [PIB: India AI Impact Summit 2026, Landmark Global Declaration and Major AI Investment Commitments](https://www.pib.gov.in/PressReleasePage.aspx?PRID=2234343&reg=3&lang=1) — 2 March 2026 5. [MeitY, Digital India: India's Common Compute Capacity Crosses 34,000 GPUs, 3 More Startups to Develop and Deploy Indian Foundation Models](https://www.digitalindia.gov.in/?p=10937) — 30 May 2025 6. [PIB backgrounder: Transforming India with AI](https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/oct/doc20251015666901.pdf) — 12 October 2025 7. [PIB backgrounder: Democratising AI in India, Expanding Access to AI Resources, Skills, and Technology Nationwide](https://static.pib.gov.in/WriteReadData/specificdocs/documents/2026/feb/doc2026210784401.pdf) — 10 February 2026 8. [PIB: India to Add 20,000 GPUs Beyond Existing 38,000 to Strengthen National AI Infrastructure](https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2229171&reg=3&lang=2) — 17 February 2026 9. [PIB (MeitY): IndiaAI Mission Expands AI Ecosystem with Affordable Compute and Startup Support](https://www.pib.gov.in/PressReleasePage.aspx?PRID=2245069&reg=3&lang=1) — 25 March 2026 10. [PIB factsheet: Semiconductor and AI Revolution](https://www.pib.gov.in/FactsheetDetails.aspx?id=150860&NoteId=150860&ModuleId=16) — 13 August 2026 11. [IndiaAI Compute Portal: eligibility and how to apply](https://compute.indiaai.gov.in/) — undated page 12. [IndiaAI: IndiaAI Compute Capacity, GPU compute allocation](https://indiaai.gov.in/hub/indiaai-compute-capacity) — undated page; allocations dated April to July 2025 Related on TechBag: [the power India’s AI data centres will need](https://www.thetechbag.com/blog/india-ai-data-centres-power-gigawatts), [GPU rental prices worldwide](https://www.thetechbag.com/blog/gpu-rental-prices-turn-up-b200-h100), and the [Google Cloud page](https://www.thetechbag.com/google/google-cloud). **Disclosure.** TechBag is a software reseller: it earns a margin when a reader buys through it, including from vendors named here. The analysis is TechBag’s own; no vendor paid for or saw this piece before publication. Facts are labelled reported, TechBag count or our view. Go deeper on TechBag - [The power India's AI data centres need](https://www.thetechbag.com/blog/india-ai-data-centres-power-gigawatts) — the electricity side of the same build-out - [GPU rental prices worldwide](https://www.thetechbag.com/blog/gpu-rental-prices-turn-up-b200-h100) — what an hour of H100 and B200 time trades at - [Google Cloud](https://www.thetechbag.com/google/google-cloud) — one of the hyperscalers building in India and renting third-party capacity