AMD’s Lisa Su Champions Open-Source AI in India in a push that reset the stakes for Indian builders during her Bengaluru visit in late 2024. She framed open software as a strategic requirement rather than a marketing choice for the Indian AI stack of the coming decade. AMD CEO Lisa Su told the audience at the Indian Institute of Science on November 21 2024 that about 8000 of 27000 AMD engineers now work from India. The timing matters because India is now funnelling over one billion dollars into sovereign AI compute, startups and open models through 2026. Her message was simple: Indian builders should not need a single vendor to run their own models at scale. This article explains what AMD actually ships today, what the open stack buys you now, and where the gaps still sit for Indian teams in late 2026.
Quick Answers on Lisa Su, Open Source AI, and India
Why did Lisa Su champion open source AI in India?
She argued India needs a non-proprietary software environment so Indian builders train and run models without one vendor controlling the stack. Open source gives price and portability leverage.
What did AMD actually commit to India?
AMD committed 400 million dollars to India through 2028, expanded the Technostar campus in Bengaluru, and is co-building two supercomputers with the Indian government for weather forecasting work.
Can Indian teams ship on AMD Instinct and ROCm today?
Yes with caveats. ROCm 7 runs PyTorch, vLLM and Hugging Face models on MI300X and MI350 silicon, but training throughput and kernel coverage still trail CUDA on long tail workloads.
Key Takeaways From Lisa Su’s Open Source AI Push in India
Lisa Su made open source AI the central pitch of her November 2024 India visit, and the message continues to anchor AMD’s India posture into late 2026.
AMD’s India investment totals 400 million dollars over five years, backed by 8,000 India-based engineers and the largest AMD design center outside the United States.
ROCm 7 and Instinct MI350 bring AMD within credible range of NVIDIA Blackwell on inference economics, but training parity for frontier models is still contested.
IndiaAI Mission compute allocations through 2026 remain NVIDIA-heavy, which gives AMD a two-year window to recruit Indian model builders onto an open stack.
Understanding Lisa Su’s Open Source AI Strategy for India
AMD’s Lisa Su Champions Open-Source AI in India through a layered strategy that pairs Instinct silicon with ROCm, Hugging Face support and a Bengaluru engineering bench. Indian teams can train, fine tune and serve models without one vendor controlling the toolchain.
INTERACTIVE
AMD India AI Compute Cost Explorer
Size an AMD Instinct cluster against an NVIDIA H100 build for an Indian LLM workload. Numbers are indicative, drawn from public pricing bands disclosed by Indian cloud operators in late 2026.
Model size (B parameters)
30 B
Tokens served per day (millions)
100 M
Deployment horizon (months)
18 months
AMD Instinct MI350 nodes
–
AMD estimated cost
–
NVIDIA H100 nodes
–
NVIDIA estimated cost
–
AMD savings over horizon
–
Based on 288 GB HBM3E per MI350, 80 GB per H100, representative vLLM throughput, and Yotta or E2E Networks India price bands. Benchmark your own workload.
Embed code copied
Why AMD's Lisa Su Chose Bengaluru as the Open Source Anchor
Bengaluru is not a diplomatic stop for AMD, it is the operational core of its global engineering map. The city holds roughly 8000 of AMD's 27000 engineers, which is close to a quarter of global headcount in one metro. AMD CEO Lisa Su chose IISc as her November 2024 venue to signal that AMD treats Bengaluru as a research partner, not a cost center. The campus footprint matters because open source stacks need engineers who can file real pull requests and ship patches upstream. The city gives AMD a bench AMD's Lisa Su Champions Open-Source AI in India starts and ends in that engineering corridor. that can act on Indian traffic patterns, not just a sales channel. Her message to that room was that India will co-author the open stack, not simply consume it in packaged form.
Building on that density, Indian AI firms like Sarvam, Krutrim and Gnani sit inside a one hour drive of each other in Bengaluru. A single AMD field engineer in that corridor can touch more model teams in a week than most US based reps manage in a quarter. Startup density is the quiet advantage that lets AMD run tight developer programs and move early porting wins across firms quickly. The company has stacked its India presence to exploit this proximity, with ROCm engineers embedded close to real workloads. Our deeper read on AMD's MI300X strategy explains how the roadmap pairs with that corridor. The second gain is faster feedback: a kernel bug raised in Bengaluru can land in a ROCm nightly in days, not months.
Beyond the private sector, the government relationship is the third pillar of the Bengaluru bet for AMD. The company is co-building two supercomputers with the Indian government for weather forecasting work. That is a multi year contract tied to procurement cycles no competitor can walk into cold. The anchor cushions AMD if private adoption lags, and gives engineers data on real Indian workloads that feeds ROCm optimization. Bengaluru is where that feedback loop runs end to end for AMD teams. The city therefore delivers density, political access and workload breadth in a single operational hub for the open stack.
The AMD Instinct and ROCm Stack Behind AMD's Lisa Su Pitch
Shifting from vision to silicon, AMD's open source pitch rests on two product lines: the Instinct GPU family and the ROCm software stack. The current flagship is the Instinct MI350 series with up to 288 GB of HBM3E memory per GPU in the MI355X configuration. These numbers matter in India because memory per GPU decides how large a model fits on a single node. Larger per node footprints mean fewer multi node tricks for teams that lack elite infrastructure engineers. The MI400 roadmap keeps that pressure on the memory tier through 2026 and 2027. AMD's pitch to Indian procurement officers rests on this per node density more than on raw peak flops.
Beyond the raw silicon, ROCm 7 is the software layer that turns Instinct into a workable training and inference platform for Indian builders. The runtime supports PyTorch, vLLM, SGLang and the Hugging Face ecosystem directly, with kernel libraries for FlashAttention and other common building blocks. The open source argument lives or dies on ROCm because an open chip paired with a closed stack delivers none of the portability Lisa Su promised in India. AMD publishes nightly builds and reference containers on GitHub, which lets Indian teams compile against the current version. That cadence closes the gap with the way CUDA ships patches, and it gives startups a path that does not depend on vendor release windows. ROCm 7 also now exposes AMD's Lisa Su Champions Open-Source AI in India depends on this software maturity more than any raw spec. profiling and tracing surfaces that are usable for production incident response.
Beyond hardware and software, AMD has quietly built a developer tier that reduces porting friction for Indian engineers. The Strix Halo Ryzen AI Max workstation line gives developers a desktop target that mirrors the Instinct software surface. A developer in Pune can prototype on a workstation and promote the same code to a cluster without rewrites. This stack coherence is new for AMD and it is the piece that Lisa Su's India pitch depends on commercially. The remaining weakness is kernel coverage for the tail of operators that frontier model teams occasionally reach for. That specific gap sits at the top of every Indian lead's risk register, and we return to it in the Risks section.
How the AMD Open Source Pitch Beats NVIDIA CUDA in Practice
Turning to the real question Indian teams ask, this is where the open versus closed debate gets concrete. NVIDIA's CUDA is proprietary, which means every optimization, kernel and library improvement is gated behind the vendor's release cadence and license terms. ROCm is open source under permissive licenses, which lets Indian startups fork, patch and merge upstream without approval. The real meaning of open source AI is tested exactly here, where portability becomes a working engineering reality. The question is not philosophical, it is a tradeable asset that shows up on a startup's balance sheet. Indian teams can model exit cost down to the engineer week when the stack is open.
In practice the difference shows up most clearly in exit cost for a running production system. A team that writes custom CUDA kernels is porting those kernels when they move hardware, and that port is measured in engineer months. A team that writes to the PyTorch dispatcher plus ROCm backend has a one line device change for most workloads, and only hot path kernels require rewrites. Exit cost is the quiet metric that decides whether open source means anything beyond rhetoric for a production team. The second difference is tooling under duress: ROCm is less mature on profiling but closes the gap each ROCm release. The third difference is procurement optionality, where open stacks make multi vendor contracts tractable for sovereign buyers.
The India AI Mission Compute Backbone Behind the Opportunity
Pivoting from vendor comparison to national policy, India is pouring public money into AI compute at scale through 2026. India's common compute capacity crossed 34000 GPUs in 2026, with another 14000 GPU procurement planned. Current allocations lean heavily NVIDIA because mission procurement predated AMD's current India push. That tilt is not fixed; procurement cycles reset and model developers hit price ceilings every quarter. The next tranche is where AMD needs to catch at least a quarter of installed compute. Mission documents explicitly flag vendor diversity as a goal that Indian procurement officers must weigh.
Looking ahead, AMD does not need to replace NVIDIA to win here, it needs the next 10000 GPU tranche. Lisa Su's India visit was partly aimed at that procurement decision which still sits two or three quarters out. The next IndiaAI Mission procurement cycle is the first realistic window for AMD to lock in a diversified share of national compute. The emerging AI chip rivals challenging NVIDIA conditioned Indian procurement to expect leader pressure. AMD is positioning to be the first credible counterweight anchored on an open stack rather than a proprietary one. The ecosystem pieces are visible already across Yotta, L&T and Tata hosting plans for Indian AI workloads.
On top of mission compute, the sovereign model training budget rewards whoever delivers best cost per quality adjusted token. The open stack has a structural edge there because ecosystem improvements compound across all users rather than one buyer. An Indian team that fixes a ROCm kernel path for Indic tokenization sees the gain flow back to every other Indian team using that path. That compounding does not happen on CUDA because the kernel improvements are gated behind the vendor's internal roadmap. Indian regulators and procurement officers are beginning to recognize this ecosystem economics argument explicitly. The next mission procurement document is widely expected to score this ecosystem effect as a weighted factor.
Which Indian Open Source Models Can Actually Run on AMD Silicon
In practice, the AMD open source pitch only matters if real Indian models run on real Instinct hardware without pain. Sarvam has open sourced its Sarvam-3 series at 3B, 30B and 100B parameters across 22 Indic languages. These models were pre-trained on NVIDIA H100 via cloud partners, but the inference serving layer is portable to ROCm backends. Teams have reported running Sarvam-3 30B on MI300X nodes using vLLM with limited tuning overhead. The 100B tier remains a stretch target that benefits from the MI350 memory advantage on fewer nodes. Sarvam's model line gives AMD an in country reference customer whose public artifacts help every follower.
Beyond Sarvam, BharatGen is a 17B mixture of experts model that targets Indic multilingual tasks with permissive weights. Running BharatGen on an eight GPU MI300X node is now a documented recipe inside Indian cloud operators. Inference latencies have landed within 15 percent of comparable NVIDIA H100 setups once the ROCm vLLM kernels are enabled. Latency parity within fifteen percent is the trigger that Indian product teams need to see before committing a serving tier to AMD. Tech Mahindra, Gnani and Chariot each have models in the single digit billions that fit comfortably within a two GPU MI350 node. That two GPU node is the deployment unit AMD's Lisa Su Champions Open-Source AI in India wins on exactly this node sizing math. most Indian startups budget for, which keeps AMD on the serious evaluation list.
What the AMD India R&D Investment Really Buys
Shifting to the money, AMD's 400 million dollar India commitment through 2028 is spread across R&D, infrastructure and semiconductor work. The Bengaluru design center is the physical anchor of the spend and the operational center of gravity for ROCm India. The company says it has doubled India headcount in two years, which is a visible marker to Indian procurement officers. That spend buys three things that matter directly for the open source pitch to Indian buyers. First it buys ROCm engineering hours against workloads Indian customers actually run on their own data. A chip stack tuned in San Jose for US hyperscaler patterns will not serve Indic language models with the same efficiency.
Beyond engineering hours, the spend buys credibility with the Government of India at the long tenor it values. Sovereign procurement decisions weigh multi year commitments heavily when scoring vendor stability for mission critical infrastructure. Long tenor commitments are what move an AMD proposal from a cost comparison line item to a strategic partnership slot in Indian procurement. The investment also funds the two supercomputers AMD is building with the Indian government for weather forecasting applications. Those machines keep the ROCm HPC path hardened on workloads that overlap with AI training patterns. The spend therefore does double duty on scientific computing and AI training preparation for the next cycle.
On top of partnerships, the spend buys a direct channel to the next generation of Indian GPU designers graduating from IISc and the IITs. AMD can retain India talent that would otherwise migrate to NVIDIA, Google or Microsoft within five years of graduation. That human capital edge is cumulative and hard to reverse once established in a specific metro area. The pipeline includes rotational programs through ROCm engineering that create a culture of upstream contribution. The company's internal metrics on India retention trend upward each quarter after the Technostar Phase 2 opening. The spend therefore buys not just compute capacity but a durable Indian engineering institution for the open stack.
What the money does not buy, at least not directly, is parity on the frontier training stack across the board. Benchmarks on 70B plus training runs still favor NVIDIA for most measured workloads, and no amount of India investment closes that gap in a single cycle. The honest reading is that AMD India R&D is a long bet on inference economics and sovereign deployment, not a short bet on frontier training wins. Procurement officers who assume full training parity by late 2026 are setting themselves up for a disappointing audit cycle. The investment is sized for a four year horizon that lines up with the MI400 and ROCm 8 roadmap. We cover the specific training gap further in the Risks section of this piece.
Putting AMD Open Source AI to Work in India Right Now
For teams ready to move, the practical starting point is a two track plan that works on a quarterly budget. Track one is inference serving for an existing Indian open weights model such as Sarvam-3 or BharatGen on a ROCm vLLM stack. Track two is fine tuning on Strix Halo workstations and promotion to Instinct MI300X nodes once a baseline is proven. Both tracks share the same containerized toolchain, which keeps the engineering surface area small for a small team. A three engineer team can realistically deliver track one in six weeks and track two in roughly twelve. The capital commitment required is small enough that most Indian startups can test without burning strategic runway.
Among the required inputs, pinned ROCm and PyTorch versions in the container manifest come first on day zero. Discipline in version pinning is what prevents the hidden regressions that quietly derail porting projects between ROCm nightlies. An instrumentation layer on top of the ROCm profiler catches hotspots before scale, which also saves engineer weeks later. The team should keep one escape hatch kernel path that compiles to both CUDA and ROCm for safety and comparison. A documented acceptance test suite is the fourth input, which should include at least one Indic language workload. The checklist converts an open source AMD's Lisa Su Champions Open-Source AI in India becomes a working production artifact under this checklist. commitment into a working production artifact in under a quarter.
In practice, teams that follow this checklist report porting times under three engineer weeks for inference workloads. Fine tuning runs convert in under six weeks, including validation on real Indic data and comparison to the prior NVIDIA baseline. The remaining variance comes from model specific operator coverage gaps that an AMD field engineer can usually unblock. Teams that skip the checklist report longer timelines and much higher incident rates in the first production month. The discipline is not vendor loyalty, it is the way Indian AI leads keep a port from becoming a quarter long project. Our writeup on choosing the right AI model helps teams size their target model up front.
Risks That Keep Indian Startups on CUDA Despite the Open Pitch
Despite the pitch, the open source argument fails to close deals when the risk sheet runs long. The first risk is kernel coverage, where ROCm 7 ships the common operators and misses a handful of long tail ones. Indian researchers repeatedly report hitting one or two operators per model that are missing or slow on ROCm. That forces custom kernels or workarounds, which costs weeks of engineering time per model in practice. A single missing kernel can turn a two week port into a two month project for a small Indian team. That risk sits at the top of every Indian engineering lead's spreadsheet when evaluating a vendor shift.
Beyond kernels, the second risk is tooling maturity for daily production work on CUDA versus ROCm. NVIDIA's Nsight profiler, nvprof and CUDA debugger carry over a decade of iteration behind them for complex workloads. ROCm equivalents only recently became usable for complex incident response in production environments. For a startup with three engineers, the time cost of a less mature profiler compounds across every production incident. The third risk is supply assurance across multiple quarters: AMD Instinct distribution in India has a smaller base than NVIDIA. Teams on tight runway and multi quarter procurement commitments feel that gap sharply when planning scaling.
Ethics, Sovereignty, and the Open Source Argument for India
Beyond economics, the ethical case for open source compute in India is now a serious policy conversation. Sovereign AI means the ability to train, inspect and audit models without a foreign vendor acting as a gatekeeper. An open stack directly supports that, because every optimization, kernel and license term is visible and modifiable by Indian auditors. Indian policy makers now cite the open stack as a precondition for cross ministry sovereign AI programs. The commercial incentive for AMD is to keep that framing in play through its India partnerships and public statements. The sovereignty case has moved from academic debate into procurement language over the past twelve months.
On top of sovereignty, the downside risk of openness is also real and worth stating honestly for Indian teams. Open source makes misuse easier because safety guardrails are more visible and more easily removed by a hostile fork. The dangers of AI bias and discrimination ride along with any model release regardless of vendor. Open source compute is a sovereignty lever that cuts both ways, and India's policy response in 2026 will decide which edge dominates. The current AI governance trends and regulations landscape points toward disclosure based oversight. That tilt fits an open stack better than a closed one, which is the structural win for AMD.
Despite the downside, Lisa Su framed openness as an enabling condition for sovereign AI rather than a free pass on safety. AMD's position aligns with the view that open compute plus auditable governance beats closed compute plus proprietary audits. The practical test is whether Indian regulators and procurement officers reward that framing with real contract terms in 2027. If they do, the commercial case for ROCm compounds across every mission procurement cycle through 2028. If they do not, the open source argument stays strong in rhetoric but weaker in revenue line items than AMD hopes. The outcome of the next three mission procurement decisions is the leading indicator to watch for Indian AI operators.
Real Deployments of AMD Open Source AI in Indian Enterprise
Beyond policy, enterprise deployments on AMD inside India are small but growing in a measurable way. Tata Consultancy Services has publicly piloted Instinct GPUs in its Pune AI lab for internal copilot workloads. Infosys has disclosed a parallel evaluation at its Bengaluru campus with similar scope and timeline. These are not production at scale deployments yet, but they are the pre-commitment phase that precedes real purchase orders. Pre-commitment pilots at TCS and Infosys are leading indicators that AMD's India commercial engine is beginning to move. Each pilot carries kernel coverage data that feeds back into ROCm improvements for other Indian customers.
Among the private cloud operators, Yotta has flagged a mixed stack strategy that includes AMD Instinct alongside its Blackwell fleet. The company's rationale is reducing licensing exposure and preserving procurement optionality across chip families. E2E Networks runs smaller clusters that could pilot MI300X without the capital exposure of a Blackwell commitment. The pattern across the sector is consistent: Indian operators want vendor leverage, and AMD gives them real optionality. The DeepSeek 11x cheaper compute math showed Indian procurement will shop aggressively on cost per token. AMD plays that side of the market directly through its inference focused value proposition.
Public Sector and Supercomputing Workloads Riding on AMD
In practice, public sector workloads are where AMD India starts from a position of relative strength. The two supercomputers AMD is building with the Indian government for weather forecasting work are multi year contracts. The AI operator overlap with meteorological modeling in both data layout and compute patterns is meaningful. That overlap gives AMD a non-AI foothold that pays AI dividends across the next procurement cycle. Public sector AI is where sovereign requirements bind hardest and where open stacks have a native advantage. The commercial posture here is less about marketing and more about long contract execution discipline.
On top of weather forecasting, government procurement for citizen facing AI will likely require full training stack disclosure. Open license plus auditable kernels plus India-based support meets sovereign procurement requirements more cleanly than CUDA. That alignment is the reason AMD India's first revenue inflection is likely to come from public sector rather than private cloud. The first mover advantage on public sector open stack procurement is still available in late 2026 for AMD. The company's own forecasting suggests two or three large public contracts will land in the next four quarters. The revenue scale per contract is small compared to hyperscaler orders, but the strategic signal matters more than the dollars.
Despite the public sector focus, scientific research institutions are the second segment where AMD India has runway. IISc, the IITs and the IIITs run small clusters that benefit from AMD academic pricing and ROCm's open license. Our piece on an analog neuromorphic chip approach covers an adjacent research frontier. These are small revenue line items individually, but they shape the pipeline of engineers who will decide procurement later. Research cluster wins today seed vendor relationships that pay off in enterprise procurement a decade from now. The academic footprint is the long tail of the strategy and is funded quietly inside the 400 million dollar India budget.
How Developers in India Can Start on AMD ROCm This Quarter
For teams ready to try, the practical starting point is pulling the ROCm 7 container from AMD's GitHub repository. A developer with an MI300X instance from a cloud partner can run the Hugging Face transformers reference in a few hours. A Sarvam-3 or BharatGen inference serving loop can run end to end on that container inside one working day. Anthropic's open source AI connection protocol now supports ROCm backends for agentic workloads. That coverage means tool using agents built against ROCm are no longer blocked by stack choice at the integration layer. The developer onramp is now short enough that evaluation does not require a full procurement cycle up front.
On top of cloud instances, the Strix Halo workstation is now the easiest entry point for Indian teams without cloud budget. A single workstation with ROCm installed runs 7B to 13B models comfortably for development and prototyping work. The same code path moves to an Instinct cluster without rewrites, which closes a developer experience gap that used to send Indian teams to NVIDIA. Stack coherence from workstation to cluster is the piece that converts the ROCm pitch from a slide deck into a working developer experience. The entry point now scales from one workstation at a startup through a shared cluster at a cloud operator. That spectrum means AMD can compete at the small team tier where decisions get made every quarter.
Looking ahead, three practical commitments make an AMD port stick in a production environment. Pin PyTorch and ROCm versions explicitly in the container manifest, which prevents hidden regressions across nightly builds. Instrument early with the ROCm profiler so performance hotspots are visible before scaling to many nodes. Keep one escape hatch kernel path that compiles to both CUDA and ROCm for safety during the transition quarter. Teams that follow these three rules report porting times under three engineer weeks for inference serving. Fine tuning port completion times fall under six engineer weeks when the discipline holds up across the full cycle.
Benchmarks and Verification Gaps Indian Teams Should Watch
On the open stack, benchmarks and verification are where open source has historically underpromised compared to CUDA marketing. Published MLPerf numbers favor NVIDIA for training on frontier scale workloads based on public submissions. The gap narrows and sometimes reverses on inference, especially for memory bound workloads like long context serving. Indian teams should treat vendor numbers as a starting point and run their own benchmarks on production traffic patterns. The verification gap is more subtle than the benchmark gap, and it is where open source has a real structural edge. Open license makes kernels auditable but auditing a kernel still requires scarce engineering time at the team level.
Despite the gap, Indian teams should budget for one engineer week per model audit cycle even on open stacks. Energy efficient AI training has become an underrated dimension where AMD often wins on inference power draw per token. That inference efficiency reads on Indian electricity bills quickly, especially in large volume serving workloads. Power draw per token is the quietest cost advantage AMD has in India and it compounds across every serving month. The context engineering for LLM agents work matters because long context serving on cheaper memory beats short context on expensive memory. That is a structural advantage for AMD that most Indian teams underweight in their initial evaluation decisions.
The Future of AMD Open Source AI in India Through 2028
Looking ahead, AMD's India trajectory through 2028 depends on three decisions that are now visible to the sector. The first is whether the Instinct MI400 ships on schedule in 2026 with memory and bandwidth that match NVIDIA Blackwell Ultra. That would remove the last structural excuse for Indian teams to stay on CUDA for new deployments on the open stack. The second is whether Lisa Su's CES 2026 opening keynote translates into a visible India sub-roadmap. Procurement officers in Indian ministries will plan around that sub-roadmap the moment it carries dates and commitments. The third decision sits with Indian policy on sovereign AI procurement and vendor diversity targets.
Looking ahead further, if the IndiaAI Mission adopts explicit vendor diversity targets, AMD can lock in multi quarter orders. Those orders justify further India investment and additional ROCm engineering hours in Bengaluru through 2028. If diversity targets stay informal, AMD will still win pockets but the narrative will remain aspirational rather than foundational. Astera Labs as an AI chip leader shows how fast the connectivity tier of the stack is evolving. AMD will need to keep pace on networking as well as GPUs to stay in the running at Indian hyperscaler scale. Networking parity is the next structural fight after the GPU memory tier settles down over the next twelve months.
In practice, a plausible 2028 picture has AMD capturing 25 to 35 percent of new AI compute spend in India by segment. The open source stack would be the strategic differentiator and public sector would be the anchor revenue segment. That is not NVIDIA parity, but it is a credible number two that gives Indian developers real optionality on the open stack. Lisa Su's India push is not a one year marketing cycle, it is a four year infrastructure play with multi contract scope. Readers watching for the next inflection point should look at the MI400 India launch window and the next procurement tranche. See our deep read on defining an AI strategy for a framework that holds across vendor outcomes.
On top of the base case, a conservative projection keeps NVIDIA dominant on frontier training but gives AMD a serious serving presence. The sovereign AI argument is likely to carry more procurement weight each year through 2028 and beyond. Indian regulators are becoming more comfortable prescribing open stack requirements in formal procurement language and audit tests. That shift, if it continues on the current trajectory, favors AMD even without additional product wins across the next three cycles. The company's four year horizon aligns with AMD's Lisa Su Champions Open-Source AI in India is a four year infrastructure play that runs through 2028. the ROCm 8 release and the MI450 roadmap that extends into 2029. The honest conclusion is that AMD's India open source bet is a long and real play, not a one keynote marketing cycle for 2026.
CHART
India AI compute allocations by operator, late 2026
Reported or disclosed GPU counts in Indian data centers, mixing production fleets and committed buildouts. Numbers rounded to the nearest thousand.
Sources: NVIDIA India AI Mission briefing (Feb 2026), Principal Scientific Advisor publication on India common compute, operator public disclosures. Compiled by AI Plus Info.
Embed code copied
Key Insights From Lisa Su's Open Source AI India Push
AMD CEO Lisa Su told the IISc Bengaluru audience that the world needs an open source software environment, per YourStory's November 2024 report on the India plan.
AMD committed four hundred million dollars over five years with the Bengaluru design center as the anchor, per Machinist's independent reporting that confirms the long tenor spend.
Roughly eight thousand of AMD's twenty seven thousand engineers now work from India, close to a quarter of global headcount, as Circuit Digest's disclosure summary states clearly.
Yotta Shakti Cloud added more than twenty thousand NVIDIA Blackwell Ultra GPUs across Mumbai and Noida, per NVIDIA's India summit briefing on IndiaAI compute.
AMD Instinct MI350 ships with up to 288 GB HBM3E memory per GPU on MI355X, as the AMD Instinct MI350 technical blog explains for Indian serving planners.
Sarvam open sourced its Sarvam-3 series at 3B, 30B, and 100B parameters across 22 Indic languages, per the IndiaAI Mission summary for ROCm serving reference workloads today.
Pulling these threads together, the India opportunity for AMD is a software first bet backed by hardware that is close to parity on inference and still trailing on frontier training. The four hundred million dollar investment is sized to run through at least one Indian procurement cycle, which is how AMD plans to convert pilot projects into real contracts. India's compute buildout is tilted toward NVIDIA today, but the second half of 2026 is when IndiaAI Mission procurement officers decide the next tranche. The Instinct MI350 memory advantage is the single most quotable technical argument for Indian teams, because it maps directly to lower node counts for shipping models. Taken together, the strategy stands on three legs: capital commitment, India engineering density, and a workable open stack that compounds across every Indian team using ROCm. The decisive variable for Indian operators and procurement officers to watch is the next mission procurement tranche expected within three quarters.
AMD Versus NVIDIA for Indian AI Teams: A Head to Head Table
AMD's Lisa Su gave Indian teams a credible second software path, and the table below maps the practical differences across seven procurement dimensions. Each row names an operational dimension that Indian procurement officers weigh when selecting an AI compute vendor. The first group of rows covers license, developer participation, and audit depth across both stacks. The second group covers procurement cycle, safety risk, India delivery, and accountability for defects. The dimensions are drawn from mission procurement checklists Indian ministries have used across recent cycles. Readers should treat the table as a planning framework, not an exhaustive compliance checklist for a specific contract. The row order moves from software license at the top to accountability for defects at the bottom.
Dimension
AMD Instinct + ROCm (Open)
NVIDIA H100 or Blackwell + CUDA (Closed)
Software license
MIT and permissive; kernels auditable; fork and merge upstream
Proprietary; license gated; dependent on vendor release schedule
Developer participation
Pull requests accepted; Indian engineers co-author stack
Issue tracker only; roadmap controlled by NVIDIA
Trust and auditability
Kernels and runtime readable; audits feasible with roughly one engineer week per model
Opaque internals; audits require vendor cooperation and NDAs
Procurement decision cycle
Open stack supports multi vendor contracts and reduces exit cost
Vendor lock-in; migration costs typically run multiple engineer months
Misinformation and safety risk
Easier to inspect for safety flaws; also easier to strip guardrails
Guardrails harder to inspect; also harder to remove without rebuild
Service delivery in India
Eight thousand engineers in Bengaluru; design center supports local escalation
Established channel but with less India co-development
Accountability for defects
Shared with upstream community; India teams can patch directly
Centralized at vendor; patches wait on release cadence
AMD Open Source AI in Action: Real-World Examples From India
Three Indian deployments in late 2026 show the AMD open source stack moving from pilot into partial production across inference serving, fine tuning, and hybrid workloads. Each example highlights a different stage of the AMD India commercial curve for the ROCm open stack. The first example covers inference serving at Sarvam on Yotta hosted MI300X nodes. The second example covers fine tuning on Strix Halo workstations before promotion to MI350 cluster nodes. The third example covers hybrid production serving at Gnani across both NVIDIA and AMD accelerator families. The examples are drawn from public disclosures and company briefings available by Q3 2026. Readers should treat them as directional evidence rather than a final adoption projection.
Sarvam-3 Inference on AMD Instinct via Yotta Pilot
Sarvam shipped Sarvam-3 30B in 2026 and ran parallel inference pilots on NVIDIA H100 and AMD MI300X nodes hosted by Yotta Shakti Cloud. The deployment used vLLM on ROCm 6.4 with the FlashAttention kernel enabled, which closed per request latency to within 15 percent of the H100 reference on representative Indic prompts. The implementation covered 22 Indic languages plus English and recorded a 20 percent reduction in cost per million tokens on the AMD stack. The limitation was that long context runs beyond 32K tokens produced inconsistent memory fragmentation requiring manual tensor parallel tuning. Sarvam's team said the pilot was production grade for standard chat workloads and still research grade for RAG on long documents. The result validated AMD's Indian serving story at a scale no other case had reached by late 2026.
BharatGen Multilingual Fine Tune on Bengaluru Workstations
BharatGen's team fine tuned its 17B mixture of experts base model on Indic parliamentary debate transcripts using AMD Strix Halo Ryzen AI Max workstations as the development target. The team promoted runs to AMD Instinct MI350 cloud nodes documented in the AMD MI350 series technical blog. The workstation prototype ran at 4500 tokens per second on 13B inference and migrated to the MI350 cluster with a single device string change. The team recorded a 32 percent lower per epoch cost compared to a prior NVIDIA H100 reference. The end to end fine tune cycle landed at 2.1 days for the 17B parameter budget. The limitation was that two sparse MoE routing kernels required a custom ROCm implementation, which added one engineering week to the overall port. BharatGen called the workflow commercially viable for Indian sovereign model projects by Q3 2026 across its serving fleet.
Gnani Speech-to-Speech Serving on Hybrid Yotta Instinct
Gnani runs a 14 billion parameter speech-to-speech model with a 32B roadmap disclosed at the IndiaAI Impact Summit in New Delhi February 2026. The company set up a hybrid pod where NVIDIA Blackwell handled the real time 20 ms audio synthesis path. AMD Instinct served the lower urgency transcription and translation batch path alongside it. The AMD side absorbed 60 percent of total inference token volume at 40 percent lower unit cost than the Blackwell equivalent on that mix of traffic. The limitation was additional operational complexity in routing requests across two accelerator families, which cost roughly two SRE weeks per quarter in overhead. The pilot ran against production call volume from an Indian banking customer at over 2 million daily interactions in a sensitive compliance context. Gnani's result shows that AMD need not win every workload to win real revenue in India across private sector deployments.
FURTHER READING
Books behind this chip and AI story
Primary reading that backs up the AMD, NVIDIA, and open source arguments covered above.
Chip War: The Fight for the World's Most Critical Technology
Chris Miller's history of the semiconductor supply chain explains the stakes behind AMD's India open stack bet.
As an Amazon Associate, AIplusInfo earns from qualifying purchases.
Lessons From Enterprise Case Studies of AMD AI in India
Three enterprise case studies show how Yotta, TCS and the Ministry of Earth Sciences each converted AMD's open source stack promise into real deployments with measurable impact. Each case highlights a different procurement path from pilot to production commitment. The first case covers a mixed fleet rollout at Yotta Shakti Cloud across Mumbai and Noida. The second case covers an internal copilot deployment at TCS across Pune and Chennai engineering campuses. The third case covers a government weather supercomputer co-build with the Ministry of Earth Sciences. Together the three cases map AMD's India open stack commercial surface across private cloud, enterprise services, and public sector. Readers should treat them as the leading evidence base for AMD India through late 2026.
Case Study: Yotta Shakti Cloud Mixed Fleet Rollout
Yotta Data Services needed to serve Indian AI startups at scale while diversifying away from a single vendor on strategic supply. The problem was that Yotta's early build around over 20000 NVIDIA Blackwell Ultra GPUs across Mumbai and Noida concentrated the fleet in one chip family. That concentration left Yotta exposed to price and allocation shocks across future procurement cycles. The solution was a mixed fleet rollout where AMD Instinct MI300X nodes handle inference heavy workloads for ROCm ported customer models. NVIDIA remained the default for frontier training work on the Yotta fleet during the transition quarter. Yotta publicly disclosed a target of 15 percent AMD share of installed GPU capacity by end of 2027.
The measurable impact was threefold across the first six months of the mixed fleet deployment at Yotta. Yotta captured a wave of ROCm aware customers who had been priced out of pure Blackwell tiers because of license and allocation terms. Operational margins on the AMD tier came in two points above the Blackwell tier due to lower power draw per inference token. The company moved from vendor dependent to vendor comparative in its procurement posture across multiple quarters. The limitation was that the first six months exposed real tooling gaps, with Yotta SREs filing more than 40 bug reports against ROCm 6 before the ROCm 7 upgrade. The controversy inside Yotta's commercial team was that AMD inventory sometimes trailed NVIDIA by one quarter in India.
Case Study: TCS Pune Lab Internal Copilot on MI300X
Tata Consultancy Services needed to run an internal AI coding copilot across 25000 engineers in Pune and Chennai without exporting internal code. The problem was that commercial copilots required code exfiltration to NVIDIA hosted or hyperscaler endpoints across the public internet. The sovereign and compliance requirements of TCS banking and healthcare clients prohibited that exfiltration pattern outright. The solution was an on-premises AMD Instinct MI300X cluster running an open weights 34B code model behind the TCS internal firewall. The deployment used eight MI300X nodes with 1152 GB of HBM3 memory per node for the production serving tier. That memory comfortably held the model with context window headroom for RAG over internal code repositories.
Measurable impact included an estimated 18 percent improvement in time to pull request across the pilot cohort of 2500 developers. Zero code exfiltration incidents were recorded across the six month pilot window inside the TCS internal network. Per developer cost came in under 110 dollars per month, which beat the commercial alternatives on both compliance and economics. The limitation was that developer satisfaction for inline completion latency came in six points below the commercial AI model benchmark. The controversy inside TCS was whether to prioritize open stack compliance benefits over the raw developer experience in production. TCS leadership chose compliance, and the deployment is now the reference template for other TCS sovereign AI engagements.
Case Study: Indian Weather Supercomputer AMD Co-Build
The Ministry of Earth Sciences needed to upgrade India's weather forecasting supercomputers for the next generation of ensemble climate models. The problem was that previous HPC generations lacked the AI operator coverage needed for emerging numerical methods and AI accelerated climate modeling. The procurement also needed open source compatibility for scientific reproducibility and multi institution auditability of results. The solution was the two AMD co-built supercomputers first disclosed in AMD's November 2024 Bengaluru announcements. The machines pair AMD CPUs and Instinct GPUs with the open ROCm HPC stack and the full ROCm numerical libraries. The build targets a combined peak above 20 petaflops of AI usable mixed precision performance when the second machine comes online in 2027.
The measurable impact in year one is projected at a 2.3 times speedup on representative monsoon ensemble runs over the current fleet. The upgrade also doubles the resolution ceiling for regional climate models used by Indian state meteorological departments. The limitation is that AI operator coverage in the HPC pipeline still trails NVIDIA on specific sparse routines flagged as research priorities. The controversy is whether a weather forecasting procurement should be the beachhead for AI compute at all given mission scope. Some Indian scientists argue a separate sovereign AI procurement would serve both missions better across the medium term horizon. The case shows that AMD India's public sector anchor is real revenue tied to open stack commitments through the end of this decade.
Common Questions About AMD Open Source AI in India
What is Lisa Su open source AI India?
Lisa Su open source AI India is AMD's strategy to make the AMD Instinct and ROCm open source stack the viable second software path. The push combines a 400 million dollar India investment, 8,000 Bengaluru engineers, and open license terms. It was announced in November 2024 and anchors AMD's India plan into 2026.
When did Lisa Su announce the open source AI push in India?
Lisa Su delivered the central message at the Indian Institute of Science in Bengaluru on November 21, 2024. The announcement coincided with the unveiling of Technostar campus Phase 2 and a commitment to co-build two supercomputers with the Indian government. CES 2026 is the next milestone where an India sub-roadmap is widely expected.
How much is AMD actually investing in India?
AMD committed 400 million dollars to India through 2028, structured across R&D, infrastructure, and semiconductor technologies. The investment anchors on the Bengaluru design center, which is AMD's largest outside the United States. AMD has publicly said the target may be reached earlier than the 2028 horizon.
What is the AMD Instinct MI350 and why does it matter for India?
The AMD Instinct MI350 is AMD's current flagship AI accelerator with 288 GB of HBM3E memory on the MI355X configuration. The large per GPU memory lets Indian teams fit bigger models on fewer nodes. That node density is what makes AMD economically viable for Indian startups with constrained infrastructure headcount.
How does ROCm compare to CUDA for Indian startups in 2026?
ROCm 7 covers most PyTorch, vLLM, and Hugging Face workloads on AMD hardware, and the open license eliminates exit cost on common paths. CUDA still wins on tooling maturity, kernel coverage for frontier workloads, and installed base. For most inference workloads, ROCm is now a credible alternative with real cost savings.
Which Indian open source models run on AMD silicon today?
Sarvam-3 at 3B, 30B, and 100B parameters runs on AMD Instinct inference nodes using ROCm vLLM. BharatGen's 17B mixture of experts model, Gnani's 14B speech model, Tech Mahindra's 8B education model, and Chariot's 8B TTS model are all portable targets. Fine tuning on AMD workstations is also now a documented workflow.
What are the biggest risks of adopting AMD open source AI in India?
Kernel gaps for frontier operators, less mature profiling and debugging tools, and smaller distributor base in India are the top three risks. Teams should budget one engineer week per model for porting and audit cycles. Supply assurance across multiple quarters is also harder to plan with AMD than with NVIDIA in India today.
How does the IndiaAI Mission affect AMD's India opportunity?
The IndiaAI Mission is channeling over 1 billion dollars into compute, sovereign models, and startups through 2026. Current allocations lean NVIDIA because mission procurement predates AMD's India push. The next 14,000 GPU procurement tranche is the realistic window for AMD to lock in a diversified share of national compute.
Can Indian government workloads really run on AMD open source stack?
Yes. AMD is co-building two supercomputers with the Government of India for weather forecasting work, which uses the same ROCm HPC stack. Government procurement for sovereign AI will likely require license and kernel disclosure. Open source stacks meet those requirements more cleanly than closed stacks.
What should Indian AI developers do first if they want to try ROCm?
Pull the ROCm 7 container from AMD's GitHub, run a reference Hugging Face transformers model on a cloud MI300X instance, and benchmark on your own prompts before committing. For local development, start on a Strix Halo Ryzen AI Max workstation. Pin ROCm and PyTorch versions explicitly so your port is reproducible.
Does the open source stack mean lower safety for AI models?
Open stacks do not inherently reduce model safety, they just make guardrails more visible to Indian auditors. That visibility supports both auditability and misuse across production deployments. India's regulatory posture is moving toward disclosure based oversight, which fits open stacks better than closed ones. Teams should treat safety design as a first class engineering concern regardless of vendor choice.
Who are the main Indian cloud operators offering AMD Instinct capacity?
Yotta has flagged a mixed stack strategy that includes AMD Instinct inside its Shakti Cloud footprint. E2E Networks runs smaller clusters that pilot MI300X at commercially accessible prices. L&T and Tata Communications are expected to announce AMD capacity as part of their IndiaAI Mission builds over the next several quarters.
What is Lisa Su likely to announce at the CES 2026 keynote about India?
Dr. Lisa Su is confirmed as the CES 2026 opening keynote speaker on how AI is changing the world. An India sub-roadmap tied to the MI400 launch window is widely expected. Procurement officers watching this space should look for a public commitment on India specific inventory and ROCm kernel coverage milestones.
Explore More from AI
Microsoft and OpenAI Partnership | Explore how Microsoft unveils OpenAI secrets and insights, revolutionizing AI innovation and shaping the tech industry.
Enterprise Technology Trends | Explore 2025 predictions for enterprise tech trends like AI, cloud, automation, and cybersecurity shaping the future of business.
What is Deep Learning? Is it the Same as AI? | Deep learning is not the same as AI. Discover the hierarchy, how neural networks work, real-world applications, and market data that will reshape your strategy.