Open-Weight AI Is Rewriting the Cost of Success. India’s Advantage Is to Build on It Fast.
The gap between the best open-weight model and the best closed model has narrowed to a few points, while the price gap runs one to two orders of magnitude. That compression is the largest opening Indian AI has had recently, and the National Mission’s focus should be to fund the builders who can seize it.
The best open-weight AI model now scores within four points of the best closed one and costs one to two orders of magnitude less to run. On the Artificial Analysis Intelligence Index, Moonshot’s Kimi K3 scores 57 against about 61 for the leading closed model. More to the point for India, Z.ai’s GLM-5.2 reaches 51 under a permissive licence that lets any Indian team run it on hardware it owns.
Recently in Bengaluru, Sarvam made its frontier model achievements public. At its Epoch conference, it announced a trillion-plus parameter model built from scratch, priced its 105-billion parameter model at $0.80 per million tokens against $4.50 for OpenAI’s GPT-5.4 Mini, and launched an inference service hosted in India.
Funding the frontier model ambition was important for the National Mission. It is also insufficient for where the technology is going.
The Frontier Has Opened and Its Price Is Deflating
Capability at the frontier is no longer a scarce input. A year ago, an open model meant a visible quality gap, and at the top that gap has closed. Kimi K3, a 2.8-trillion parameter model whose weights were published in late July, ranks third overall on the Artificial Analysis index, though at 1.56 terabytes it is too large for most buyers to host. The more consequential release is GLM-5.2, which posts 51 on the same index under an MIT licence and runs on infrastructure the buyer controls. DeepSeek’s V4 models score near 44 and cost roughly a tenth of the leading American systems for comparable output. Alibaba’s Qwen has become the most widely adopted open base in the world. On coding, mathematics and reasoning, the strongest Chinese open models trade blows with GPT, Claude and Gemini, while a genuine gap still survives at the absolute frontier.
Demand switching has followed the economics. Enterprises are moving production workloads onto open weights because the quality is close enough, the running cost is an order of magnitude lower, and the model sits on hardware they own rather than behind an interface they do not. A brief episode in June 2026, when export restrictions pulled the top American models offline in several markets, hardened the logic by showing that access to a closed model is a permission that can be withdrawn. For a buyer, the binding question has moved from what a model can do to what it costs to run and who can control it.
The Frontier Labs Are Moving Into Their Customers’ Businesses
The sharper risk in a closed model is that the provider becomes the competitor. Selling tokens is a thin and commoditising business. Selling the finished vertical product the customer was building is not, and every query a customer sends also teaches the provider where the margin sits.
The pattern is now too consistent to read as coincidence. Anthropic has started its own drug-development programmes while pharmaceutical firms such as Novo Nordisk and Bristol Myers Squibb run research on its models. It launched Claude Design into Figma’s category soon after its product head left Figma’s board, which led Figma’s chief executive Dylan Field to say the lab had not been “consistently candid”. It sells Claude for Legal into the workflows that Harvey, a legal-AI firm built on Claude, had turned into a business, and Claude Science into the domain of Benchling, another customer. OpenAI has moved the same way, launching a support agent against Intercom, a clinical product against the hospital-software firms that build on it, and, reportedly, a jobs platform and a code repository that compete with LinkedIn and GitHub, both owned by its largest backer, Microsoft.
Palantir’s Alex Karp has given the mechanism its name. Enterprises are leaking their “alpha”, the proprietary edge that makes them valuable, into models they neither own nor can audit. No promise of restraint is credible against an incentive this strong, which is why the defensive move is architectural rather than contractual. Firms such as Intercom have begun shifting core products onto in-house open-weight models they can inspect and contain.
For India, the implication is direct. A company, or a ministry, that builds a strategic vertical on a foreign closed model is training a future rival and security vulnerability, and cannot control what it is giving away. Open weights remove this exposure at the source.
India Was Not Building to Win the Pre-Training Race, and No Longer Needs To
India was never resourced to build frontier general-purpose models from scratch, and the economics now make that a poor use of the capital it has committed. Sarvam’s 105-billion parameter flagship was trained on IndiaAI Mission compute, a few thousand GPUs for a few months. A model built to compete at the global frontier needs cluster-months of more than ten thousand top-end GPUs and outlays in the hundreds of millions of dollars, and the leading American labs have raised tens of billions more than most other companies to fund precisely that. The deeper problem is that a from-scratch model is a depreciating asset. It is overtaken the month a stronger open model ships, and the open frontier now ships every few weeks, so capital spent replicating pre-training buys an edge with a short half-life.
This is not an argument against sovereign models. A trillion-parameter Indian model trained at home is a strategic asset, and Sarvam’s ambition merits the public support it has drawn, for the reasons of security and self-reliance the company cites.
This is an argument that the flagship cannot be the only strategy. The frontier is now an input India can acquire at lower cost. The value, and the durable sovereignty, sit in what it builds on top.
India’s Highest Return Is in Specialised Models Trained on Proprietary Data
India’s advantage compounds one layer up, where an open model is fine-tuned on proprietary data for a defined domain. A model trained on Indian case law, regional-language clinical notes, curriculum content for a national exam, or the country’s compliance corpus will beat any general frontier system on that task, at a fraction of the cost to build and to run. The domains are large and underserved, spanning education, legal services, healthcare, finance and the Indian languages the global labs do not prioritise. Twenty sovereign model efforts are already funded under the IndiaAI Mission, twelve large language models and eight smaller ones, spanning Sarvam, Gnani’s speech models, BharatGen’s multilingual work and healthcare-specific systems.
A privately-deployed vertical model’s moat is the ability to safely utilise proprietary data, the insulation from alpha leakage, and the closed loop control plane for its utilisation, not just the weights. So this holds even as the base model commoditises underneath it, which the open frontier guarantees it will. Owning the data, the deployment premise, the harness, the control plane, and the fine-tuning loop ensures that the economics are worth the effort. A fine-tune costs a small fraction of a pre-train and earns from a paying market within months.
Public capital that now encourages this mode of AI usage will fund the immediate application of this paradigm with the economy’s largest databases such as the Aadhar stack, the UPI transaction base, the ABHA health records, the Skilling and Employability mission, and more within a trusted domestic environment.
Domestic Inference Is Arriving, and the National AI Mission Must Widen Its Aim
The last missing input, low-cost inference on Indian soil, is now being built at scale. Adani has committed $100 billion (about ₹9.5 lakh crore) by 2035 to AI-ready data centres, taking capacity from 2 GW to 5 GW, with a gigawatt-scale campus in Visakhapatnam alongside Google. Reliance has pledged about ₹10 lakh crore (around $110 billion) over seven years toward gigawatt-scale training and inference. The IndiaAI Mission has crossed 34,000 subsidised GPUs at ₹115 to ₹150 (about $1.20 to $1.60) per GPU-hour and is targeting 100,000 by the end of 2026. Sarvam now sells India-hosted inference commercially.
Inference, not training, is where the recurring revenue and the daily data flows sit, which makes it another layer where sovereignty must be prioritised.
That is why the National AI Mission must widen its aim from funding model development to also running them. We have argued in these pages for a National Inference Mission, and its logic extends from compute to models. The State should keep funding sovereign frontier efforts and fund a far larger base of teams building on open weights, subsidising fine-tuning and inference credits rather than only from-scratch pre-training. It should route this through a dedicated fund, co-invested with the Indian IT majors, and become the first buyer of vertical models across public health, courts, governance, defence, and education.
India may not out-raise or out-build the labs at the frontier, and does not need to. The frontier is now an input available to everyone. The advantage goes to whoever turns it into working products first, and the National Mission’s task must be to fund a hundred such teams rather than a few flagships.
DISCLAIMER
Details of all the funds managed by 3one4 capital can be found here: https://www.3one4capital.com/aif-registration-details
For informational purposes only. This content does not constitute an offer to sell or a solicitation of an offer to buy units of any fund. Investments in AIFs involve market risks.
The views expressed herein are those of the author as of the publication date and are subject to change without notice. Neither the author nor any of the entities under the 3one4 Capital Group have any obligation to update the content. This publications are for informational and educational purposes only and should not be construed as providing any advisory service (including financial, regulatory, or legal). It does not constitute an offer to sell or a solicitation to buy any securities or related financial instruments in any jurisdiction. Readers should perform their own due diligence and consult with relevant advisors before taking any decisions. Any reliance on the information herein is at the reader's own risk, and 3one4 Capital Group assumes no liability for any such reliance.Certain information is based on third-party sources believed to be reliable, but neither the author nor 3one4 Capital Group guarantees its accuracy, recency or completeness. There has been no independent verification of such information or the assumptions on which such information is based, unless expressly mentioned otherwise. References to specific companies, securities, or investment strategies are not endorsements. Unauthorized reproduction, distribution, or use of this document, in whole or in part, is prohibited without prior written consent from the author and/or the 3one4 Capital Group.
