What actually changed
The IndiaAI Mission has now deployed 34,000+ subsidised GPUs at ₹150 per GPU-hour (~£1.40 / ~$1.80 at late-May 2026 rates) through its empanelled cloud partners. The original Mission was sanctioned at ₹10,000 crore (~$1.25 bn) in 2024; the 2026-27 Union Budget added a further ₹10,372 crore, potentially doubling the programme to around ₹20,000 crore. That top-up explicitly moves IndiaAI from what officials describe as a "capex-heavy research phase" to a "population-scale deployment phase".
The empanelled compute pool runs alongside the 12 selected sovereign-LLM partners (Sarvam, BharatGen and others) and feeds the developer demand visible on platforms like Krutrim's cloud, which has reported tens of thousands of developers building on Indian infrastructure in recent operational updates.
Meanwhile in the UK, the £500M UK Sovereign AI Fund has made its first investments — notably Isomorphic Labs on 12 May 2026 — and a £40M frontier AI research lab is being stood up. The UK Sovereign AI Compute roadmap and the existing Isambard-AI deployment together signal a programme that is structurally similar to IndiaAI in intent, though the funding mechanics lean closer to a sovereign investment vehicle than a pure compute subsidy.
- ₹150/hr headline price is genuinely below the May 2026 cloud floor for H200 — but the comparison is incomplete without throughput, queue time and availability SLAs.
- Eligibility is narrow. IndiaAI is startup-only via the MeitY portal and not yet open to foreign companies; UK programme criteria are still being shaped.
- For Indian builders, the unit-economics shift is real — a startup doing 10M inference calls a day on a 70B model can see its compute bill drop materially moving from a hyperscaler H100 to subsidised IndiaAI capacity.
- For UK builders, the window to shape eligibility, allocation and pricing is now — before the programme moves out of research phase.
If you are based in India, file an IndiaAI compute application through the MeitY portal even if you are not GPU-bound yet — approvals are not instant and having a reserved allocation when your traffic spikes is worth the paperwork. If you are based in the UK, send a one-page note to UKRI and DSIT outlining what allocation rules and pricing tiers would actually unlock your roadmap. Programmes harden quickly once allocation patterns set in.
The decision framework: which GPU lane fits your workload?
There is no universally right answer. The right GPU lane depends on five variables:
- Tokens per second required. Real-time chat vs nightly batch is a different problem.
- Data residency. Indian financial-services and healthtech workloads increasingly need to keep data in-country; ditto NHS-adjacent UK use cases.
- Queue tolerance. Subsidised capacity is rationed. Can your product survive a 6-hour delay if a slot is not immediately free?
- Cost cap. A hard monthly compute ceiling pushes you towards reserved or subsidised lanes; a soft one keeps spot and on-demand viable.
- Cold-start vs reserved. Bursty, unpredictable workloads punish anything with a long allocation cycle.
Mapped against those five variables, here is how the main lanes compare in May 2026.
Comparison: five GPU lanes for IN and UK builders
| Lane | Headline price/hr | Throughput band | Eligibility | SLA reality | Best for |
|---|---|---|---|---|---|
| IndiaAI subsidised GPU | ~₹150 (~£1.40 / ~$1.80) | Mixed — newer H100/H200 nodes on some empanelled clouds, older A100 elsewhere | Indian startups, researchers via MeitY portal; not foreign companies | Maturing — queues and driver versions still moving | Predictable batch inference, fine-tuning, sovereign-LLM training |
| Crusoe / specialised cloud (MI300X) | From ~$1.71/hr (see our price-war piece) | High for 70B-class models; AMD ROCm caveats apply | Open commercial | Production SLAs from specialised clouds; commitment terms vary | Cost-optimised inference on ROCm-friendly stacks |
| Hyperscaler on-demand H100 | ~$2-$5/hr on-demand (see H100 pricing guide) | Well-characterised; mature tooling | Open; quota gating applies | Strong production SLAs; spot adds risk but cuts cost | Mainstream production with mixed workloads |
| Hyperscaler on-demand H200 | ~$2.50-$7.00/hr typical; broader range $1.45-$13.78 | 141GB HBM3e; fits 70B+ without sharding | Open; quota gating applies | Strong production SLAs | Long-context inference; memory-bound serving |
| Self-host 8×H100 HGX | Capex $320k-$600k; opex on top | Predictable; you own the throughput | Open — capital required | You set the SLA; you also own the on-call | Steady-state workloads with break-even economics (see our self-host model) |
Two sources to anchor those numbers: the Thunder Compute H200 pricing tracker and the ABHS analysis of the IndiaAI deployment. The single-card H200 sits at $30,000-$45,000 depending on form factor; an 8-GPU HGX H200 server lands between $320,000 and $600,000. Specialised GPU clouds (Thunder Compute, Lambda and others) now sit roughly 3× cheaper than the hyperscalers on the same silicon, even after the most recent round of hyperscaler price cuts.
Subsidised compute is not always cheaper once you factor in the realities. IndiaAI access is restricted to Indian-registered startups and goes through a queue. SLAs are still maturing — expect older driver versions on some empanelled clouds, no spot-instance flexibility, and longer cold-start times than a hyperscaler. If your product is latency-bound or your customer contracts include a five-nines uptime commitment, do not assume the headline price is the right one for you. UK Sovereign AI Compute eligibility is still being defined; do not plan capacity around it as if it were a commercial product.
Worked example: 10M API calls/day on a 70B model
The clearest way to show the impact is to plug an illustrative workload through each lane. The numbers below are deliberately rough — they are an order-of-magnitude tool, not a quote. Treat them as a starting point for your own spreadsheet.
Assumptions: 10M API calls per day; 200 input tokens and 500 output tokens per call on average; a 70B-class open-weight model served at fp8; throughput target of roughly 80 tokens/sec per concurrent stream after batching; a 30-day month. We assume cache-warm sessions for self-host and IndiaAI; we assume on-demand pricing without committed-use discounts for the hyperscaler row. Token economics from our inference-costs analysis are the anchor.
| Lane | Approx GPU-hours / month | Effective rate | Illustrative monthly bill |
|---|---|---|---|
| IndiaAI subsidised (8-GPU node) | ~5,760 | ₹150/hr (~$1.80) | ~₹8.6 lakh (~$10,400 / ~£8,200) |
| Hyperscaler on-demand H100 | ~5,760 | ~$3.50/hr blended | ~$20,160 (~₹16.7 lakh / ~£15,900) |
| Hyperscaler on-demand H200 | ~4,800 (denser memory, larger batches) | ~$4.50/hr blended | ~$21,600 (~₹17.9 lakh / ~£17,000) |
| Self-host 8×H100 (amortised) | ~5,760 | ~$1.20/hr amortised over 3 yrs + ops | ~$10,000-$14,000 once amortised |
| Frontier API (e.g. Claude) | n/a — token-priced | Token-priced | Higher per-call cost, near-zero ops overhead |
Illustrative — see assumptions above. Your mileage will vary with batch size, model quantisation, traffic profile, contractual discounts and queue realities.
The headline takeaway: for steady-state Indian workloads that fit on subsidised capacity, the bill can land roughly 40-60% below a like-for-like hyperscaler line. That is not the 5-8× cliff the headline ₹150 rate might suggest — because you also pay for under-utilised hours, queue waits and the engineering effort to make a less-mature platform behave. But it is a real, defensible saving on the right workload.
Need to hire someone who has shipped on subsidised compute?
AI Tech Connect Verified Builders include engineers who have moved live workloads onto IndiaAI infrastructure, onto Crusoe MI300X, and onto self-hosted 8×H100 rigs. Shortlist up to five — we email you the contact details.
Browse Builders →For Indian builders: what to do next
The IndiaAI compute pool is real, the price is real, and the application process exists. The hard parts are: (1) understanding which empanelled cloud actually has the silicon you need on the date you need it, (2) building your serving stack so it can move between lanes if your subsidy allocation gets capped, and (3) negotiating commit-and-discount terms with a commercial cloud as your fallback. Treat IndiaAI as your primary lane for predictable batch and fine-tuning, and keep a hyperscaler account warm for spikes.
It is also worth tracking the Indian sovereign-silicon roadmap. If domestic accelerators come online over the next 18-24 months, the cost floor on subsidised compute could move again — in your favour.
For UK builders: shape the programme now
The UK programme is at the analogous stage IndiaAI was in roughly 18 months ago. The £500M Sovereign AI Fund has made its first investments — including Isomorphic Labs on 12 May 2026, per the gov.uk announcement — and the broader Sovereign AI Compute roadmap is still being defined. As Sifted's coverage of the frontier-AI lab makes clear, the political appetite is there; the operating model is what is still being drawn.
UK founders should be writing to UKRI and DSIT now. Quantify the workload — tokens/sec, model size, residency requirements, throughput targets — rather than asking generically for "more compute". The teams that shaped the eligibility criteria on the Indian side did so by being early, specific and persistent.
The honest tradeoffs
Subsidised compute is not a free lunch. Three tradeoffs that builders consistently underestimate:
- SLA maturity. A subsidised programme moving from research phase to deployment phase will hit growing pains. Plan around them; do not assume them away.
- Driver and toolchain drift. Empanelled clouds run different stacks. CUDA versions, NCCL configurations and container images vary. Your CI pipeline needs to test across at least two backends if you intend to move workloads.
- Allocation churn. What you get this quarter is not necessarily what you get next quarter. Build with the assumption that your subsidy bucket can be re-sized, and design your serving layer to fall back gracefully.
Bottom line
The ₹150/hr IndiaAI rate is a genuine change in what an Indian AI startup's compute line can look like in 2026 — not a marginal one. The UK is heading down a structurally similar path, with different mechanics and a different funding envelope. The builders who win in both markets will be the ones who treat subsidised compute as one lane in a multi-lane plan, not a silver bullet. Run the numbers on your own workload, file the application, and keep the commercial cloud account warm.
For source detail: the IndiaAI startup financing hub on indiaai.gov.in carries the canonical eligibility and application detail; the ABHS deployment analysis and the Thunder Compute H200 pricing tracker are the data anchors for the cost comparison above.