DEV Community: Vikramaditya Khupse

What India is doing?

Vikramaditya Khupse — Tue, 01 Sep 2026 11:55:11 +0000

Vikramaditya Khupse

Sep 1

Artificial Intelligence in 2026: Where the World Stands, and Where India Fits In

#discuss #ai #llm #news

16 min read

Project is open source! Open for contributions and ideas.

Vikramaditya Khupse — Tue, 01 Sep 2026 11:53:48 +0000

Vikramaditya Khupse

Sep 1

Know Your API Cost and Compute Before You Use.

#ai #productivity #privacy #opensource

4 min read

Artificial Intelligence in 2026: Where the World Stands, and Where India Fits In

Vikramaditya Khupse — Tue, 01 Sep 2026 11:50:40 +0000

Artificial intelligence has moved very quickly in the last few years. What used to be a research topic discussed mainly in universities and technology companies now touches ordinary government offices, farms, courts, and businesses. This article explains, in plain language, what is happening in AI around the world today, who is leading the field, and where India stands. It also looks at what the Indian government is doing about AI, whether AI is actually being used in government work yet (not just talked about), and a bigger question that experts argue about constantly: are we close to building a machine that can think and reason like a human being across every subject - Artificial General Intelligence, or AGI?

Wherever a claim could be uncertain or is still developing, this article says so plainly, instead of presenting every announcement as a finished fact.

Part One: What Is Happening in AI Right Now

The center of attention in AI has shifted. It is no longer just about which company has the smartest model. It is now about what these models can actually do on their own.

AI agents are the big story

Through 2025 and into 2026, AI models themselves have become fairly similar to each other in raw ability. The real competition has moved to "agentic AI" - AI systems that can carry out multi-step tasks by themselves, such as browsing the internet, writing and testing software, filling forms, or managing a workflow from start to finish, with only light supervision from a person. Some of the strongest AI models today can complete tasks that would take a skilled human worker 12 to 24 hours, without needing constant guidance. (Source: Mozilla, State of Open Source AI 2026; 80,000 Hours, "What the hell happened with AGI timelines in 2026?")

Freely available models are catching up to paid ones

Many AI companies keep their best models locked behind a paid service. But open-weight models - where the underlying model is published for anyone to download and use - have grown very strong. By mid-2026, roughly 1 in every 3 AI requests processed through major AI marketplaces was being handled by an open-weight model, and several of these free models now sit close to the very best paid ones in quality. Most of these strong open models currently come out of China. (Source: Mozilla, State of Open Source AI 2026)

Models can now handle text, images, sound, and video together

Nearly every major AI model launched in 2026 can understand and sometimes generate text, pictures, audio, and video, all within one system, rather than needing separate tools for each. These models can also read and remember very long documents at once, sometimes running into the hundreds of thousands of words, in a single conversation.

The cost of using AI keeps falling

Running a large AI model used to be very expensive. That cost has been dropping steadily as companies build more efficient hardware and smarter software techniques - one reason AI is spreading so quickly into ordinary products, apps, and government tools rather than staying limited to large technology firms.

Part Two: The Leading AI Models Today

The table below lists the flagship AI models considered strongest as of September 2026, the company behind each, and the country where that company is based. This field changes quickly, sometimes month to month, so treat this as a snapshot rather than a permanent list.

Model (approximate current version)	Company	Country
GPT-5.6	OpenAI	United States
Claude (Opus 4.8, Sonnet 5, and related versions)	Anthropic	United States
Gemini 3.1 Pro / 3.5	Google DeepMind	United States
Grok 4.5	xAI	United States
Llama (open-weight family)	Meta	United States
DeepSeek V4	DeepSeek	China
Qwen 3.5	Alibaba	China
Kimi K3	Moonshot AI	China
GLM-5.2	Zhipu AI	China
Hy3 and MiMo-V2.5	Tencent and Xiaomi	China
MiniMax M3	MiniMax	China
Mistral Large family	Mistral AI	France

Sources: llm-stats.com model leaderboard, Moonshot AI's Kimi K3 technical paper, and industry tracking sites, accessed August and September 2026.

Quick tally - flagship models by country of origin:

Country	Flagship models listed above
United States	5
China	7
France	1

A clear pattern stands out here. Nearly all of the most talked-about AI companies are either American or Chinese. France, through Mistral, is the only other country with a model in this top tier, and it is generally considered a step behind the leading American and Chinese systems in raw capability, though still respected in Europe.

Part Three: Which Countries Lead in AI

Ranking countries in AI is harder than ranking models, because different research groups measure different things. Two respected sources give somewhat different pictures, and it's worth understanding both rather than picking one.

By research strength and frontier models: the United States leads clearly

The Stanford AI Index, a widely respected yearly report, found that the United States produced far more of the world's most advanced models in 2025 than any other country, and attracted about 23 times more private investment in AI than China. China, however, leads the world in the number of AI research papers published, AI patents filed, and industrial robots installed. South Korea leads in AI patents per person, and small, wealthy nations like Switzerland and Singapore lead in AI researchers per person. (Source: Stanford HAI, 2026 AI Index Report)

Private AI investment, 2025:

Country	Private AI investment
United States	$285.9B
China	$12.4B

China's total is about 1/23rd of the United States' figure.

By a broader measure of AI economy: China moves to the top

A different, broader ranking - the Global AI Economy Index - looks beyond just top models. It also considers computing power, electricity supply, internet infrastructure, how widely AI is actually being used, and how self-sufficient a country is in AI technology. On this broader measure, the order looks quite different. (Source: FNTV Global AI Economy Index)

Global AI Economy Index - ranking:

China
South Korea
Sweden
Singapore
Finland
Germany
Japan
Taiwan
France
United States ...
India

Why this matters: There is no single, agreed answer to "which country is number one in AI." It depends on whether you're measuring who builds the smartest models, or who has built the wider ecosystem of power, computing, and everyday AI use across society. Both the United States and China are, by any measure, far ahead of every other country. The real debate is only about which of the two is ahead, and by how much.

Part Four: Where India Stands

India's position depends on which measuring stick is used, and the picture is genuinely mixed - not simply good or bad.

Metric	India's Rank
Global AI Economy Index (FNTV)	18th
Government AI Readiness Index (Oxford Insights)	27th of 195 countries - but 1st in South & Central Asia
AI researchers & inventors, worldwide (Stanford HAI)	2nd

On the Global AI Economy Index above, India ranks 18th in the world, well behind the leading nations. (Source: FNTV Global AI Economy Index)
On the Oxford Insights Government AI Readiness Index, India ranks 27th out of 195 countries, but 1st in the entire South and Central Asian region. The report specifically credits India's newly released AI governance guidelines and around $1.3 billion (roughly ₹11,000 crore) of new AI infrastructure spending for this improvement.
On talent and skill, India performs unusually well. It ranks 2nd in the world for AI researchers and inventors, and leads the world on a widely used measure of how much AI skill has spread through the general working population. (Source: Stanford HAI 2026; ThePrint, on India's AI talent and brain-drain trend)

Put simply, India has a large and growing pool of people who understand AI, and its government's approach to AI policy is being noticed internationally. But India still lags behind the United States and China badly in the actual computing hardware, funding, and research infrastructure needed to build the most advanced AI systems. India also appears to be losing some of its AI talent to other countries - a brain drain - even while producing that talent at a strong pace.

AI trends within India

India now has one of the largest groups of AI startups in the world, behind only the United States and China, with more than 3,200 active AI startups and roughly $10 billion (about ₹85,000 crore) in AI-related funding in the past year. (Source: Inc42, India AI startup funding coverage, 2026 - treat the exact figures as approximate)
Sarvam AI, a Bengaluru-based company founded by former researchers from IIT Madras, has become India's most prominent home-grown AI lab. It raised around $234 million (about ₹1,990 crore) in mid-2026 and is now valued at $1.5 billion (about ₹12,750 crore), reaching unicorn status. It has built its own AI models, Sarvam-30B and Sarvam-105B, considered competitive with some strong Chinese open models. (Source: TechCrunch, June 2026)
BharatGen, a consortium led by IIT Bombay and 8 other institutions, is building a family of Indian-made models called Param2, with government backing under the IndiaAI Mission.
Not every story has been a success. Krutrim, built by Ola's founder Bhavish Aggarwal and once celebrated as India's first generative AI unicorn, ran into trouble in 2026. It laid off more than 200 employees and withdrew its consumer app, shifting focus toward cloud services instead - a useful reminder that India's AI sector, like the sector everywhere else, has both strong successes and real setbacks. (Source: TechCrunch, May 2026)

Part Five: What the Indian Government Is Doing

The IndiaAI Mission

The central government approved the IndiaAI Mission in March 2024, with a planned budget of about ₹10,372 crore (roughly $1.25 billion), spread over 5 years. The mission rests on 7 pillars: computing infrastructure, home-grown foundation models, useful public datasets, real-world applications, skilling and talent, funding for startups, and "Safe and Trusted AI." Actual yearly spending has been modest compared to the full five-year figure. (Source: Lok Sabha reply, February 2026, reported by Down To Earth)

IndiaAI Mission budget - first year vs. now:

Year	Budget
2024-25	₹173 crore
2026-27 (planned)	₹1,000 crore

Actual mission spending has grown roughly 5.8 times over from its first year to the current year's planned budget.

Building India's own AI models

Under this mission, the government has funded 12 different groups to build India-made, or "sovereign," AI models. Several were formally launched at the India-AI Impact Summit, held in New Delhi in February 2026 - the first time this global summit series was hosted in a Global South country. The models include Sarvam AI's Indus model, BharatGen's Param2 family, and smaller specialized models such as Patram and Shrutam.

Building computing power

India has been steadily adding the computing hardware needed for AI work, long a weak point compared to the United States and China. As of mid-2026, the mission had approved more than 38,000 graphics processing units across more than 200 projects, and a large supercomputer is being installed at a government data centre in Delhi.

Rules and governance

In November 2025, the government released its India AI Governance Guidelines. Importantly, these guidelines choose not to impose strict regulation on AI for now. Instead, they set up a lighter framework, including a new AI Governance and Economic Group, a Technology and Policy Expert Committee, and an AI Safety Institute, to study risks and prepare policy over time rather than restrict AI use immediately. (Source: MeitY / NeGD official press release, November 2025)

A note on what is confirmed and what is not: Government officials have spoken about a possible "next phase" of the IndiaAI Mission with a larger budget, but as of this writing, no such expanded plan has been formally approved. Treat future funding increases as likely, but not yet confirmed.

Part Six: Is AI Actually Being Used in Government Offices?

This is often the most important question for anyone working inside government, since public statements about AI can move far ahead of what's actually working on the ground. The honest answer: some AI tools are genuinely live and used every day, while other well-known projects are still experimental.

Live today:

Kisan e-Mitra - farmer chatbot
NAFIS - fingerprint database
Su-Sahay & SUVAS - court tools
UIDAI-Sarvam - Aadhaar voice tools

Still a pilot:

SUPACE - judicial research aid
Land records - AI digitization
Disaster management - AI (unconfirmed)

Already working and in daily use

Kisan e-Mitra, an AI-based chatbot for farmers under the PM-KISAN scheme, answers questions by voice or text in 11 Indian languages, around the clock. It has already resolved more than 9.5 million (about 95 lakh) farmer queries about scheme status, payments, and complaints, and is active across 28 states and 8 union territories - probably the clearest example so far of AI genuinely working at the scale of ordinary citizens' daily needs. (Source: Social Protection AI case study; Press Information Bureau)
NAFIS, the National Automated Fingerprint Identification System, is a centralized fingerprint database with more than 12.3 million (1.23 crore) records, used by police forces nationwide to identify suspects and match evidence.
The Supreme Court has launched Su-Sahay, a chatbot for checking case status and related services, and continues to use SUVAS, a tool that translates court judgments into 18 Indian languages.
The Unique Identification Authority of India (UIDAI), which runs Aadhaar, has partnered with Sarvam AI to add voice-based interaction, fraud detection, and multilingual support in 10 languages, run on secure government-controlled infrastructure rather than external servers.

Promising, but still at the pilot stage

SUPACE, an AI research assistant built for the Supreme Court to help judges review case material faster, is often mentioned in the news, but the Court itself has confirmed it remains experimental and has not yet been deployed for regular judicial use.
Digitizing land records with AI is happening, but unevenly. The national SVAMITVA scheme uses drones and mapping to survey rural land, and some districts run their own pilots - such as Dantewada in Chhattisgarh digitizing 700,000 (7 lakh) records with blockchain technology, and Karnataka's Kaveri 2.0 cutting registration time from about 8 hours to 50 minutes. These remain state or district-level pilots, not one uniform system, and a national digital property register is still being discussed in Parliament.
No confirmed, large-scale AI system was found for disaster management at the district level. Any claim that such a system is already operational nationwide should be treated with caution until confirmed locally.

For anyone working in district administration, the practical lesson is this: AI is already useful and reliable for citizen-facing helplines and welfare scheme queries, and reasonably mature for identity and fingerprint verification. It's still early days, however, for AI in land records and disaster management, where progress exists but is patchy and depends heavily on which state or district one is in.

Part Seven: How Close Is the World to True Machine Intelligence?

This is the single most debated question in the entire field, and honest experts disagree with each other, sometimes strongly.

Timelines have moved much closer, but nobody agrees on the exact date

Just a few years ago, most researchers surveyed believed a machine with fully general, human-level intelligence (AGI) was roughly 50 years away. By the middle of 2026, forecasting communities that track this question closely had moved their estimates much closer, with a typical middle estimate now falling somewhere around the early 2030s. However, the range of opinion is still very wide, running anywhere from as early as 2027 to well past 2040 - and that range has not narrowed even as the middle estimate has moved closer. (Source: 80,000 Hours, summary of AGI forecasting surveys)

Forecast range for human-level AI (AGI):


80% confidence range	2027 – 2044
Current median estimate	~2031 – 2033
Labs' more optimistic estimates	~2028 – 2030
Skeptics	2035+

The people building AI tend to be the most optimistic

Leaders at the major AI labs, including Anthropic, OpenAI, and Google DeepMind, have generally given nearer-term estimates than outside forecasters, with some suggesting a real possibility of reaching AGI-like capability before the end of this decade. Worth keeping in mind: these companies also benefit commercially from sounding confident about their own progress, so their statements should be read alongside more independent forecasts, not instead of them.

What actually changed in 2026

Researchers who track this field point to a few concrete reasons why confidence grew during 2026: AI agents that can work independently improved faster than expected, AI-related business revenue grew faster than predicted, and at least one major AI company reported that its own systems were measurably speeding up its internal research work. At the same time, AI still struggles with messy, real-world tasks that don't have a clear, checkable right answer, even while it excels at clean, well-defined tasks like solving programming problems. Whether that weakness closes over time or turns out to be a lasting limitation is still an open question. (Source: 80,000 Hours research podcast, August 2026)

Part Eight: Where the World Seems to Be Heading

Countries are regulating AI in very different ways, not converging on one approach

The European Union has taken the strictest approach, with its AI Act now able to fine companies up to 7% of their global revenue for serious violations. In practice, the small team enforcing these rules has struggled to keep pace with fast-moving AI companies.
The United States has no single national AI law. The federal government has generally pushed to reduce regulation, even as more than 100 individual state laws on AI have been passed regardless.
China also has no single AI law, but enforces a set of fast-moving administrative rules with real teeth, since the government can simply order a non-compliant AI service to shut down. China has also proposed a new international body for AI cooperation, positioning itself as a possible leader in global AI governance.
India, as described earlier, has chosen a light-touch approach for now, favoring guidelines and study groups over hard rules, while it builds up its own AI capability.

The debate over jobs is genuine and still unresolved

There are three broad camps among economists and AI leaders on how AI will affect jobs. Some warn of serious and fairly quick job losses, including a leading AI company's own chief executive suggesting AI could eliminate a large share of entry-level office jobs within a few years. Others argue history shows new technology takes decades to fully spread through an economy, so major disruption is further off than feared. A third group expects AI to create more jobs than it destroys, through new businesses and higher productivity. The most careful data available so far, based on actual payroll records in the United States through mid-2026, shows no broad, economy-wide job losses yet, but does show a real and growing effect on very young workers in jobs that are easy for AI to take over - mainly through companies hiring fewer new graduates rather than firing existing staff. (Source: Stanford Digital Economy Lab study, cited by 80,000 Hours, 2026)

Competition between nations is intensifying

The rivalry between the United States and China over AI chips and AI capability continues to shape global politics, with talks between the two governments reportedly planned for later in 2026. Meanwhile, countries like India are trying to build enough of their own AI capability to avoid being completely dependent on either side, though this remains a work in progress rather than a finished achievement.

Closing Summary

AI is advancing quickly, and the center of activity has shifted from simply building smarter models to building AI systems that can act independently on real tasks.
The United States and China lead the world in AI, though which of the two is truly "ahead" depends on how one measures leadership. No other country is close to either of them.
India is a genuine and rising player, especially in AI talent and skilled people, and its government has moved with real speed on policy and funding in the last two years. Even so, India remains well behind the two leaders in computing power and cutting-edge research.
The Indian government has launched a serious and well-funded national AI plan, and a small but growing number of AI tools are already working reliably in real government service delivery, especially for farmers and identity verification. Many other announced uses are still pilots, not finished systems, and should be described honestly as such.
Whether the world is close to building truly general machine intelligence remains a real, open disagreement among serious experts. Confidence has grown that it may arrive sooner than once thought, but no one can say this with certainty, and a wide range of outcomes remains possible.
Around the world, governments are choosing very different paths on how strictly to regulate AI, and the effect of AI on jobs is still being measured rather than fully known. Both questions will likely take a few more years to become clearer.

Compiled from publicly available research, government records, and industry reporting current as of September 2026. Given how quickly this field moves, figures and rankings in this article should be treated as a snapshot in time.

Sources

Know Your API Cost and Compute Before You Use.

Vikramaditya Khupse — Tue, 01 Sep 2026 11:20:47 +0000

NoRefund is an open-source desktop tool that locally calculates tokens, context fit, LLM costs, and GPU VRAM requirements.

I had a 526-page PDF, a stack of personal course notes, that I wanted to feed into an agentic workflow I was building. As an AI engineer, I already know every model has a different context window and a different price per token, that part isn't new.

Before I sent anything anywhere, I wanted three answers: How many tokens is this? Will it actually fit? And how much will it cost me?

Finding all three turned out to be surprisingly annoying.

The problem

I opened one of the usual online token counters. Most of them are built for pasting a paragraph, not processing hundreds of pages, the page froze, truncated my text, or just gave up.

The ones that did handle larger input were optimized for quick text checks rather than large-document and multi-model workflows: often only a couple of models, a rough word-to-token approximation instead of the real tokenizer, and no awareness of pricing tiers, so a document crossing a long-context threshold got priced as if it hadn't. None of them touched memory either, if I wanted to self-host a model, nothing told me whether the weights and KV cache would actually fit my GPU.

And the document itself was real content I didn't want sitting on a server I don't control.

So I built NoRefund.

(Following is the actual result export of my 526 page PDF document)

Model	Provider	Tokens	Context %	Fits	Input $	Output $	Total $
GPT-5.6 Sol	OpenAI	60,771	5.8%	✓	0.2431	0.0205	0.2636
Claude Sonnet 5	Anthropic	~61,322	6.1%	✓	0.1226	0.0102	0.1329
Gemini 3.5 Flash	Google	~61,322	5.8%	✓	0.0920	0.0092	0.1012
DeepSeek V4 Flash	DeepSeek	62,455	6.0%	✓	0.0275	0.0014	0.0288

That comparison, run once, told me more than any single "token count" ever did: which models could even take the file, and what the cheapest option that fit actually was.

So I built NoRefund

NoRefund is a local-first LLM workload analyzer: tokens, context limits, cost, and self-host memory requirements, all computed on your machine.

Document analysis — process files or folders locally and get real token counts.
Context checks — see whether a workload fits a model's context window before you send it.
Cost comparison — the same document, priced across every supported model.
Local inference fit — estimate whether a model's weights and KV cache fit your GPU's VRAM.

Everything runs locally. Your documents are never uploaded to a server, network access is only used when you explicitly download a tokenizer or refresh currency rates.

On tokenizer accuracy, I want to be precise rather than just say "real tokenizers": OpenAI, DeepSeek, Llama, Qwen, and Mistral models run their actual published tokenizer. Anthropic and Google don't publish a downloadable tokenizer for Claude or Gemini, so those fall back to a close approximation, and the app marks every approximate count as such rather than presenting it as exact.

How it works

NoRefund parses the file locally (PDF, DOCX, PPTX, TXT, or Markdown), runs it through the real tokenizer for whichever model you pick, then applies that model's context limit and pricing rules, including tiered long-context rates where a provider has them. You get a token count, a fit verdict, and a cost, without a single network call.

The feature I didn't expect to care about

"Token counter" is easy to picture. "Will this model actually fit on my 24 GB GPU?" is a more interesting engineering problem, and it turned out to be the feature I use most.

Fit Check estimates a model's weight memory, KV cache, and activation overhead against a GPU, Apple Silicon chip, or cloud instance's usable VRAM, before you rent hardware or wire a model into an agent that assumes it'll fit.

Beyond a single document

The same problem shows up worse inside an agentic workflow. Retrieved documents, tool outputs, conversation history, and prior model responses all consume context, a workload that looks cheap at step one can get expensive by step five.

Document:        120k tokens
Tool output:       20k
Conversation:       40k
                  ─────
Total:            180k tokens

That's the kind of number NoRefund is built to surface before the request goes out, not after the bill arrives.

What I learned

The interesting part wasn't that token counting is hard. It's that a raw token count, on its own, doesn't actually answer the question I cared about.

The useful question is: will this workload fit, what will it cost, and can I run it on the hardware I already have? A number without those three answers just gets me back to guessing, only with more confidence than I should have.

That's why I built NoRefund, and why it stays local by default.

Try it

NoRefund is free, open source, and available for Windows, macOS, and Linux.

Try it → GitHub
Download → Releases

19 issues for Open source contributions.

Vikramaditya Khupse — Fri, 28 Aug 2026 06:33:11 +0000

Vikramaditya Khupse

Aug 27

I Got Tired of Guessing How Much My LLM Workflows Would Cost

#ai #opensource #privacy #productivity

4 min read

I Got Tired of Guessing How Much My LLM Workflows Would Cost

Vikramaditya Khupse — Thu, 27 Aug 2026 09:08:03 +0000

NoRefund is an open-source desktop tool that locally calculates tokens, context fit, LLM costs, and GPU VRAM requirements.

Before I sent anything anywhere, I wanted three answers: How many tokens is this? Will it actually fit? And how much will it cost me?

Finding all three turned out to be surprisingly annoying.

The problem

I opened one of the usual online token counters. Most of them are built for pasting a paragraph, not processing hundreds of pages, the page froze, truncated my text, or just gave up.

And the document itself was real content I didn't want sitting on a server I don't control.

So I built NoRefund.

(Following is the actual result export of my 526 page PDF document)

Model	Provider	Tokens	Context %	Fits	Input $	Output $	Total $
GPT-5.6 Sol	OpenAI	60,771	5.8%	✓	0.2431	0.0205	0.2636
Claude Sonnet 5	Anthropic	~61,322	6.1%	✓	0.1226	0.0102	0.1329
Gemini 3.5 Flash	Google	~61,322	5.8%	✓	0.0920	0.0092	0.1012
DeepSeek V4 Flash	DeepSeek	62,455	6.0%	✓	0.0275	0.0014	0.0288

That comparison, run once, told me more than any single "token count" ever did: which models could even take the file, and what the cheapest option that fit actually was.

So I built NoRefund

NoRefund is a local-first LLM workload analyzer: tokens, context limits, cost, and self-host memory requirements, all computed on your machine.

Everything runs locally. Your documents are never uploaded to a server, network access is only used when you explicitly download a tokenizer or refresh currency rates.

How it works

The feature I didn't expect to care about

"Token counter" is easy to picture. "Will this model actually fit on my 24 GB GPU?" is a more interesting engineering problem, and it turned out to be the feature I use most.

Beyond a single document

Document:        120k tokens
Tool output:       20k
Conversation:       40k
                  ─────
Total:            180k tokens

That's the kind of number NoRefund is built to surface before the request goes out, not after the bill arrives.

What I learned

The interesting part wasn't that token counting is hard. It's that a raw token count, on its own, doesn't actually answer the question I cared about.

That's why I built NoRefund, and why it stays local by default.

Try it

NoRefund is free, open source, and available for Windows, macOS, and Linux.

Try it → GitHub
Download → Releases