NVIDIA Blackwell Delivers 4x Inference Boost for India’s Sarvam AI Models

Jessie A Ellis
Feb 18, 2026 16:35

NVIDIA’s hardware-software co-design achieves 4x inference speedup for Sarvam AI’s 30B parameter sovereign models, showcasing Blackwell’s NVFP4 capabilities.

NVIDIA’s collaboration with Indian AI startup Sarvam AI has produced a 4x inference performance improvement for sovereign large language models, demonstrating the chipmaker’s full-stack optimization capabilities as it pushes deeper into enterprise AI deployment.

The joint engineering effort, detailed in an NVIDIA developer blog published February 18, 2026, targeted Sarvam AI’s flagship 30B parameter model—a multilingual system supporting 22 Indian languages built for voice-based AI agents with strict latency requirements.

Breaking Down the 4x Speedup

The performance gains came from two distinct optimization phases. First, kernel and scheduling improvements on H100 GPUs delivered a 2x speedup through targeted fixes to bottlenecks in the mixture-of-experts (MoE) routing logic. Engineers achieved a 4.1x improvement in MoE routing alone by fusing operations into single CUDA kernels.

The second 2x gain came from deploying on Blackwell architecture with NVFP4 weight quantization. At higher concurrency points, Blackwell showed even stronger results—2.8x throughput improvement at 100 tokens per second per user compared to optimized H100 performance.

What’s notable: a single Blackwell GPU handled the 30B model more efficiently than multiple H100s running in parallel. The disaggregated serving approach—dedicating separate GPUs to prefill and decode phases—proved optimal for this workload pattern.

The Technical Details That Matter

Sarvam’s models use a heterogeneous MoE architecture with 128 experts and top-6 routing for the 30B variant. The 100B model scales to 32 layers with top-8 routing and implements multi-head latent attention similar to DeepSeek-V3 for aggressive KV cache compression.

Service level agreements drove the optimization targets: sub-1000ms time to first token and under 15ms inter-token latency at the 95th percentile. These aren’t arbitrary benchmarks—they’re requirements for production voice AI applications where latency directly impacts user experience.

The kernel-level work cut transformer layer time by 34%, from 3.4ms to 2.5ms per layer. Fusing query-key normalization with rotary positional embeddings delivered a 7.6x speedup for that specific operation by eliminating redundant memory reads.

Market Context

This announcement follows NVIDIA’s February 12, 2026 disclosure that Blackwell has enabled 10x token cost reductions for certain AI inference workloads through its co-design approach. Meta’s multiyear partnership announced February 17 further validates the strategy of deep integration across GPUs, networking, and software.

NVIDIA stock traded at $182.88 on February 17, down 3.9% amid broader market softness, with market cap holding at $4.66 trillion.

For AI infrastructure buyers, the Sarvam case study provides concrete benchmarks for sovereign AI deployment—particularly relevant as more countries push for locally-controlled model development and data governance. The models were trained using NVIDIA’s Nemotron libraries and NeMo Framework, suggesting a template for similar national AI initiatives.

Image source: Shutterstock