I like @deedydas's work but but this take misses context Sarvam-M isn’t a vanity fine-tune; it’s India’s first open-weights 24 B Indic-centric LLM built under brutal GPU & data scarcity. Judging it by few hours of HuggingFace stats badly misses the point. Most people outside India don't appreciate that compute is quite the invisible ceiling - H100 clusters are still not commercially stocked in India - US export caps tightening next week will squeeze supply even further - Indian teams literally queue for hours of A100/H100 time that US & CN labs get on tap Data is also the long-tail problem Indic languages form <0.01 % of CommonCrawl. You read that right—two orders of magnitude less than Chinese or Spanish. Any local lab must build its corpus first, then train. That’s months of ETL before the first gradient step. Synthetic data is GPU-constrained. Talent pipeline is still forming HPC RLHF compiler-level optimisation is new ground in India; Sarvam’s run has already up-skilled dozens of engineers who now know how to wrangle 10 k GPU-hours, FP8 PTQ & GRPO reward engines. Their detailed blog post democratizes a lot of this learning. You can’t AWS-credit your way to that muscle memory. What Sarvam actually shipped - 3.7 M high-diversity Indic prompts, deduped & quality-scored - Two-phase non-think/think alignment that adds 2 pp on IndicGen - GRPO RL with partial-credit rewards—LiveCodeBench jump 0.23→0.44 - FP8 look-ahead decoding: 2× tokens/s, ½ $/M tok on H100 That means a 🇮🇳-hosted midsize model now matches Gemma-3 27 B and Llama-3.3 70 B on Indic reasoning while costing a fraction to serve. That’s some engineering leverage & definitely not hype. Model adoption is anyways a long-tail - one needs to ship multiple models of non-frontier quality to eventually be able to get to the one that's truly at the frontier (at least along dimensions that we care about). Plus, there's a whole host of Indic-language use-cases where this sovereign model would work much better compared to using any other open-weights model. Look at (LiveCodeBench 0.23→0.44, 2× tokens/s) If you ask for stats, you'll learn that some of their conversational AI platform reaches out to about 50M people in just a week. What's next possibly? - Maybe we all recognize the data problem & do Nation-scale data-collection drives (something like CommonCrawl-IN) - Public RL-as-a-service clusters so smaller labs can replicate GRPO - For devs wanting to push the Indic NLP forward, consider forking Sarvam-M, fine-tuning on your domain corpus, benchmarking on Indic-Eval, contributing back patches. Each derivative model widens the knowledge base & closes the English–Indic gap. In summary, celebrating Sarvam's work (I'm not an investor) isn't nationalism, it's recognizing an innovation feat under constraints - India can't out-GPU Mountain View today but there's technical merit on display here, regardless of the metrics. 👏 @pratykumar, @AashaySachdeva, @HarveenChadha & other friends from @SarvamAI Here's to more AI in 🇮🇳

.@saranormous - i think there’s a nuance Coding is not merely a large market for frontier labs, it is a keystone domain of outsized strategic importance. It combines extreme inference token consumption, high-quality verifiable reasoning traces for post-training, developer mindshare (the builders of future AI systems) & the clearest near-term path to agentic self-improvement. Labs therefore have intense incentives to capture or heavily influence this layer. Yet “capturing it alone” is structurally difficult. Their core competency (scaling general intelligence) does not translate directly into superior IDE integration, enterprise SDLC workflows, or closed-loop telemetry. @cursor_ai’s decision to train its own Composer-series models (including via massive SpaceXAI partnerships) illustrates that the workflow layer can move upstream into model training while retaining product advantages. The durable moat is less raw data volume and more asymmetric access to outcome-verified, context-rich signals plus distribution velocity in the actual development loop.

The "Scaling Laws" paper is probably the least understood paper in AI I read this as empirical confirmation of what scaling laws have long predicted: once models reach sufficient competence at the “perspiration” layers of R&D - code synthesis, debugging, experiment execution & local iteration - the effective rate of progress becomes gated by data, compute & the remaining cognitive bottlenecks rather than raw human headcount The reported task-horizon doubling every ~4 months is consistent w/ the smooth, continuous capability curves we observed in earlier scaling work, it is not discontinuous “intelligence explosion” yet, but it is measurable compression of the outer loop Once models cross competence thresholds on execution tasks, they begin producing better data & iteration signals for themselves, compressing R&D loops in a manner fully consistent w/ smooth scaling-law extrapolation once post-training, test-time & agentic compute are included The fundamental insight & bet that Anthropic appears to have made internally is that once a model can reliably perform the constituent subtasks (code synthesis, debugging, experiment execution, local iteration), the horizon over which it can chain those subtasks without constant human intervention lengthens predictably anthropic.com/institute/recu…

The true answer is layered I think: (1) Hardware economics diverge sharply. DeepSeek V4 is explicitly optimized for & runs on Huawei Ascend 950/950PR clusters with native CANN framework support and close co-design. This bypasses (or dramatically reduces) NVIDIA’s high ASPs and ~60-75% gross margins on GPUs. US providers remain heavily NVIDIA-dependent for peak performance and ecosystem maturity; even with custom silicon progress (Google TPU, Amazon Inferentia/Trainium, Microsoft Maia), most frontier API players carry that premium. Huawei Ascend is not yet universally superior in perf/watt or absolute throughput, but the effective cost per FLOP for DeepSeek’s workloads is structurally lower - partly domestic supply chain, partly national strategic pricing (2) Architectural & systems-level inference superiority. DeepSeek’s MoE designs with extreme sparsity (reports of 27:1 activation ratios in discussion), Multi-head Latent Attention (MLA) slashing KV cache ~93%, speculative decoding, FP8, and custom redundant-expert deployment create a much lower theoretical and realized floor per token. These are not marginal tweaks; they are co-designed with the serving stack. US labs have strong caching and quantization, but fewer are pushing sparsity and KV-cache innovations at the same intensity across the board. (3) Pricing as strategic investment, not pure margin play. DeepSeek (ecosystem-aligned) appears to treat API pricing partly as customer acquisition and data/feedback flywheel fuel - classic land-grab economics. US labs balance margin signaling to investors/LPs, R&D recovery, and willingness-to-pay segmentation. Developer discourse confirms the pricing feels “too good,” raising questions of short-term cash burn vs long-term moat building It is both hardware/stack asymmetry & genuine optimization excellence, enabled by China’s full-stack sovereignty push. US players face real cost pressures and ecosystem lock-in; DeepSeek exploits the opening with talent density in systems programming and a willingness to price for adoption. This is not sustainable purely on subsidies forever, but it is durable enough to bifurcate the market.

For a while, @aakrit & I've quietly done this at Activate: spend time with exceptional operators, CTOs & senior builders before there’s a deck, company, or fully formed idea. Now opening it up Mixture of Experts - a small private dinner for people at the edge of founding First edition tomorrow in Bengaluru Request below on the Luma to join or DM if someone exceptional should be in the room. luma.com/1vmm7hq9

Several interesting ways to read this: (1) In the short-term, it’s unlikely that Grok will be a frontier model & @SpaceX is behind on the pure model race (2) They do seem to have a new business model of being the go-to neocloud for labs that need burst capacity (both Cursor & Anthropic) (3) Elon wants to take down OpenAI at any costs Given compute is the binding constraint in the AI race & Colossus 1 is one of the world’s largest clusters, handing it wholesale to Anthropic signals xAI has either (a) excess capacity while Colossus 2 (already scaling toward 550k GPUs/~1 GW) absorbs Grok 5 training (6–10T parameter variants) or (b) chosen revenue and platform validation over pure model supremacy