Bloomberg dedicated the August episode of its flagship show, Bloomberg Tech: Asia, to India’s ambition of becoming a global AI powerhouse, and what it will take to get there. For the investor view, Shery Ahn turned to Hemant Mohapatra, Partner at Lightspeed.

Hemant covered how AI went from 20% of Lightspeed India’s deal flow to nearly 90% in four years, why sovereign models are now being built in every major region, and why AI for India has to be built in India, at Indian price points.

Below is their conversation, lightly edited for clarity.

Shery Ahn: You’ve watched the India startup ecosystem for years. How has the landscape changed over the last few years, and what have you seen in terms of investment opportunities, particularly in AI?

Hemant Mohapatra: Three to four years ago, when AI companies began to show up in our deal flow, about 20% of the companies were AI, and 80% were non-AI. Many of the AI companies then were infrastructure businesses.

In the last year or two, the ratio has flipped. Now about 80-90% of our deal flow is AI. The majority of the founders that come to us and knock on our doors are building something in or around AI. And of that, the majority are building at the application layer, which is where a lot of the momentum, the excitement, and the value has shifted.

The narratives in AI seem to flip fast. Software was all the rage, so much money went into the sector, and suddenly generative AI is threatening to make it obsolete. What is your criteria for picking companies in a sector that changes this fast?

You cannot invest in a category that you don’t understand as an investor. All of us at Lightspeed are AI-native investors. We are using AI tools every single day in our own business. I still write a lot of code. We stay ahead of the research and speak to researchers all the time.

I used to be at AMD and designed the chips, the GPUs and the CPUs, that are powering a lot of this. So we are fairly comfortable as a firm all the way to the bottom of the layers, where power, energy, data centers and chips come in, and all the way to the top, where applications and services sit. You have to stay ahead of the curve to understand what’s happening at the frontier.

Subscribe now

Lightspeed has invested in Anthropic, and you’ve worked at Western companies yourself, so you’ve seen both sides. Has the AI landscape in India developed differently from the West?

We are pretty much everywhere in the world where there’s entrepreneurship. We are investors in Mistral in Europe, which is building a sovereign playbook for Europe, and in Reflection AI in the US, which is building a sovereign playbook for the US. And we are investors in Sarvam, which is building the sovereign playbook for India, but is also going to go more broadly than that.

India has unique challenges. India is a $2,500 to $3,000 per capita GDP country, and it is really hard to serve this country at scale by building software and solutions in a $40,000 to $50,000 per capita GDP country. You have to build something locally to be able to serve something locally.

Price points have to be lower in India, given the GDP per capita. Cursor just launched a $7 per month plan in India, which I think is the lowest in the world, even though India has the highest number of power users anywhere in the world. India has to be serviced at scale at a lower price point, and that requires some very fundamental bottom-up thinking.

Lightspeed was the first institutional investor in Sarvam and came in at Series A for Emergent. The two sit at opposite ends of the shift in deal flow Hemant described, Sarvam at the model layer and Emergent at the application layer.

You were early backers of Sarvam and Emergent. India doesn’t yet have many AI companies at global scale. What does it take to get there?

We’ve been very lucky to partner with them from day one. A good way to separate the two is that Sarvam is much more infrastructure and foundation-model-focused and goes up from there, while Emergent sits at the application layer.

Sarvam is building a very voice-first AI playbook for India, because India speaks dozens of languages. There are hundreds of languages and dialects, but there are 25 that are very, very primary, and they are building a separate tokenizer specifically for those. Those kinds of things need to be thought through bottom-up to service India. They cannot be bolted on top of an existing solution.

Sarvam’s primary goal is to service India at scale, for India-specific problems, at population scale. On the sovereign side, there is a strategy from the central government top-down and state government bottom-up, and Sarvam is going to capture that entire ecosystem for civil and sovereign use cases, including defence.

On the enterprise side, which is where the majority of revenue is right now, we’ve got to compete with the global best. We’ve got to build the best models in performance, cost and usability, and we are working on that. Sarvam is about to launch a one-trillion-parameter model in the next six months, and that will be a very competitive model anywhere globally. And then we go up from there. If the models are really competitive, they can be serviced out of India for any part of the world.

Watch the full conversation on Bloomberg: Lightspeed’s Mohapatra on India’s AI Startups

I write a substack called Chasing Inflection, which is mostly an ode to “why now”: the shifts in technology, customer preference that explain why something which failed repeatedly for a decade suddenly works. Slowly at first, and then all at once.

Consumer AI as a sector will see the most “why now”s. Some of it will be AI tapping into latent needs (think health, wealth) and unlocking markets that are invisible today. Some will be AI building 10x better versions of incumbents by providing something cheaper and hyper-personalized (like commerce, education). But my favourite is when AI unlocks new categories and behaviours that never existed.

And Character.AI is the canonical example of that.

It’s one of the world’s most-loved Consumer AI products, with 20M+ monthly active users who spend close to 2 hours a day on it. Its CEO, Karandeep Anand, spent 14 years at Microsoft building Azure, then 7 at Meta building for 3B people, then ran the business at Brex.

It was a privilege to host Karan and explore what the next set of consumer AI companies would look like.

Some notes from the conversation:

The fourth generation of entertainment is interactive

Karan says that we are in the fourth generation of entertainment.

The first generation was linear: a programming team decides what a billion people watch; Doordarshan, in Karan’s case. Then came on-demand, first cable and then YouTube, which lets anyone make content instead of a studio.

Then hyper-personalized reels, where the ranking algorithm does the matchmaking. And now we’re in a generation where you don’t just consume the content, you shape it.

Gaming is the only place this has existed before, which is part of why it has been one of the largest verticals in the world for decades: you feel in control. Anyone who hates how Game of Thrones ended is a customer for a format where you can re-prompt the ending.

Growth no longer solves everything

The pre-AI playbook was growth at all costs, because growth fixed every other problem eventually. In AI, growth is the problem: more usage means more losses, not more money. Last year there were roughly 700 Character.AI clones in the app store; today there are about 30. They lost because they grew too fast, burned through their tokens and had no runway left.

This failure mode looks different based on which kind of company you are.

If you’re adding AI to something people already pay for, your problem is margin. You ship the feature, absorb a $10-20 COGS hit per user, and the user pays you nothing more than before. Take fitness trackers and watches: people buy the device and pay for the service, and an AI layer on top doesn’t change what they’re willing to pay. So the AI has to be compelling enough to charge for on its own.

If you’re AI-native, the problem runs the other way and it’s harder. Your heaviest users are your most expensive ones, so the cost per unit of usage and monetization have to move at once. Which means at least one number on your dashboard has to measure value capture (subscriptions, in-app purchases, ads) not just sessions and engagement. Otherwise you cannot tell the difference between the users who love you and ones enjoying a subsidy.

There is one India-specific unlock: the market that won’t pay ₹499 a month will happily pay ₹20 per transaction, repeatedly. Gaming figured this out, but most other categories haven’t.

PMF now means asking whether you could re-acquire this user today

In traditional PMF, you acquire a user, look at your D30/D60/D90 cohorts, see retention and scale. That worked when building a competing product was expensive.

Now, if you’ve built Canva for India, within a month there will be 1000 Canvas for India, several of them free. So the test becomes: if this user had a free or cheaper alternative in front of them today, would they still open your app? If you can’t answer that, you don’t have PMF, despite what your cohort chart says.

PMF is no longer a milestone you clear once. It becomes something you re-earn daily, because every day the user has another choice.

Get the category right before you build the feature set

Everyone assumed Character.AI is a companion app, but it isn’t.

It is an entertainment app where you happen to build relationships with characters. Not a product you come to for a relationship and get entertained along the way. The difference sounds pedantic but it determines everything downstream.

For instance, Companion apps need to build features that wake you up and wish you a good morning, check in on how you’re doing and remember your day. Character.AI builds none of that, because companionship was never the job.

Karan starts his day looking at dashboards for 30 minutes, then spends 2 hours in Reddit, Discord and Instagram comments. All 70 employees at the company are also expected to do some version of this.

On one call, their user researcher met a woman who had created Ghost, the lead character from Call of Duty, and was talking to him daily. She isn’t a gamer, her husband is. She was using Ghost to rehearse how he would react to something she wanted to raise with him, then going off and having the real conversation. She wasn’t there for a relationship with Ghost; she was using him to prepare for a real one. That’s roleplay, which is entertainment, and no dashboard or app review would have surfaced that use case.

It is also the cleanest latent-need story we have come across in Consumer AI: not a behaviour that moved online from somewhere else, but one that had nowhere to go before.

Build adjacent to the behaviour you already have, not against your competitors’ roadmaps

The most popular character on Character.AI is Harry Potter’s sister, who does not exist in the books. An entire fan universe wrote her backstory: what it feels like to grow up as the sibling of someone famous.

Every subsequent product came out of that same observed behaviour, first audio dramas, then remixable scenes, then micro-dramas and video. “We don’t build half the features our competitors build, ever,” says Karan.

It’s a discipline he carried over from his WhatsApp years, where the team spent more time deciding what not to add than on copying WeChat or Telegram, on the view that anything past three features costs you a user.

Willingness to pay has almost nothing to do with usage

It’s largely a function of geography instead.

Show a US user 5 ads and they’ll subscribe; show an Indian user 20 and they never will. So Character.AI segments on whether a user is time-rich or wallet-rich, and prices against friction rather than value, because someone is always undercutting them anyway.

Which is also why you should never do flat pricing if your costs are inference tokens: your best users become your worst users economically. Something has to be pay-as-you-go, and Character.AI’s own mistake was introducing usage-based pricing and virtual currency far too late.

Keep every revenue model on the table too. Character.AI now makes more money from ads than from subscriptions; one tailwind Karan credits is that Claude’s pricing experiments taught the market that what’s free today may not be free tomorrow.

Almost every tactical rule in Consumer is being rewritten at the same time: PMF, moats, north stars, network effects, growth-first. What doesn’t change, I’ve gathered from this conversation, is the reason you started and how well you understand the people using your product.

- Priyal Motwani

Subscribe now

When we wired Lightspeed’s first check into Sarvam in mid-2023, the company had two founders, an idea, and a belief. The entire thesis came down to one word: token production. India was about to become the largest consumer of intelligence on the planet and we produced almost none of it. Every token we consumed originated elsewhere, priced elsewhere. Lightspeed became Sarvam’s first institutional investor. We led the seed round and the series A: all of this before the models were built, before the IndiaAI Mission, before sovereign AI was a topic. I wrote the Sarvam Manifesto earlier this year to explain why we believed Sarvam will become one of the most important AI companies in the world. Consider this Part-II.

The world we saw in 2023

Three beliefs, really. (1) Intelligence would follow the internet’s curve in India: from scarce and imported to cheap and everywhere, and whoever produced it locally would compound for decades. (2) Voice would beat text; hundreds of millions of Indians will never type, but everyone talks (and we love talking!) (3) None of this works as a wrapper on someone else’s model; it had to be full-stack, built for our languages at Indian price points, at population scale.

And then the founders. Vivek had built and exited a few startups before and then spent 12 years volunteering inside Aadhaar. Building population-scale infrastructure for India was muscle memory for him. Pratyush left a professorship at IIT Madras behind to do it all over again from zero. And oh, both are PhD scholars. In 2023, pouring tens of millions into two scholars was called brave. Some called it something less polite :)

The stumble, and the grind

I won’t pretend the line from there to Epoch was straight. In May 2025, weeks after the IndiaAI Mission selected Sarvam to build the country’s sovereign model, the team shipped Sarvam-M: a fine-tune on an open base. It was an experiment for us, but when you put something out on the internet, it creates its own memory. And the internet was not kind and for a fortnight the whole country debated whether India could build frontier AI at all. Board calls in those weeks were not fun. What struck me was how unmoved Vivek and Pratyush were. It’s a stepping stone, they kept saying, and then they went dark. Two quarters of hard slog later, they had pre-trained 30B and 105B parameter models from scratch on Indian compute. In February came 14 launches in 14 days. Probably one of the best launch sequences I’ve ever seen. The world noticed Sarvam. Naysayers tuned in. The Prime Minister wore Sarvam’s Kaze glasses on stage. This team powered through. They persevered.

Epoch — the crescendo

Which brings me to two weeks ago. I spent the Sunday before the keynote judging the Epoch buildathon: ~200 builders, 8 hours, everything on the Sarvam stack, top-10 teams earning the main stage. Then two days and twelve launches. As an investor, we are incentivized to sing praises of the companies we invested in, but some of the launches truly blew me away - it felt like we touched global quality many times in these two days. A few of my favorite ones below.

On speech. Every ASR system on earth breaks when two Indians talk over each other… which is roughly whenever two Indians talk :) The standard fix is a three-pass pipeline that chops the audio and loses the overlaps. Saaras V4 Multi-Speaker does diarisation and transcription in a single pass, one model, and the team ran it live against ElevenLabs and Deepgram cleanly pulling apart a male speaker and a female anchor talking over each other in real time. I thought this tech should have immediate PMF in all parliamentary systems!

Bulbul V4 was the other half of the speech models, a single reel that went from theatrical narration to a whisper to something sombre, steered with simple emotion and emphasis tags in the text. There were definitely goosebumpy moments here: when I heard sun lo duryodhan! It immediately took me back to my childhood in Chandigarh when streets would clear out on Sundays because Mahabharata was playing on DD-1. Then another voice declared chattrapati shivaji maharaja cha parakrama! in marathi and I thought of my years spent in mumbai as a young student in college, visiting those crowded ganesh puja pandals, and those many, many afternoons taking trains to and from “CST” (Chhatrapati Shivaji Terminus) to Kanjurmarg! Voice is such an important part of our lives - voices are memories. There is nothing more humanizing and grounding than hearing someone speak your mother tongue; there is a reason the word “mother” sits in there. AI for everyone should speak to everyone as they are. Sarvam’s mission finally felt real, personal.

On inference. One of the most important, long-term bets is inference, from India for India. Sarvam Inference serves frontier open models, including their own 105B alongside GLM and Gemma, entirely from Indian infrastructure, at ~$0.80 per million tokens vs $4.50+ for the comparable US offering. Banks, hospitals, and ministries here increasingly cannot ship data abroad; DPDP made residency a feature. Vivek called it token sovereignty. A country that imports 100% of its tokens is as exposed as one that imports 100% of its crude.

And on the trillion. And finally, announcing an SF research office, with Devendra Chaplot joining as advisor to help build globally competitive models from India. Pratyush said he would start with a 1T+ parameter model, trained from scratch in India, aimed at coding, cybersecurity, simulation and science, targeted to go live in six months.

Within five years, we believe no serious AI workload in India will run without a Sarvam layer somewhere in its loop.

Lightspeed’s journey with Sarvam

Working with this team has been one of the great joys of my investing life. I’ve watched Pratyush grow from a professor to founder-CEO without losing the professor’s precision, and seen Vivek carry his years building Aadhaar into building a team and a company that thinks in populations by default. I am but a small member of a much larger Lightspeed team that has tirelessly worked on anything and everything Sarvam needs. Abha & Sid who have been helping hire a world class team, Anuj & Vidur who have helped Sarvam raise $350M so far, Shashank & Gaurav who have inserted themselves almost as full-time members of this team on special projects, our legal and finance team with Amit, Pavani, and Vinayak tirelessly handling all the complex work Sarvam throws our way, and so many more :)

My favorite poet Jack Gilbert wrote in his poem A Brief for the Defense (which I’d highly recommend) that We must have the stubbornness to accept our gladness in the ruthless furnace of this world. Startups are one of the most intense experiences one could go through in their professional lives - no less than a furnace in its own way. But a furnace is also a forge and it’s been an absolute joy to see Sarvam grow stronger through the years and do what they do best in a way that defines class, quality, and craftsmanship.

Onwards.

^{Note: Video credit for the first video in this note: X user @Adyasha8105}

Subscribe now

On the latest episode of Lunchbreak with Lightspeed, Coral co-founder and CEO Ajay Shrihari described to Lightspeed’s Rohil Bagga what that means in practice:

• A 920-page fax carrying 35 patients, with four versions of the same patient’s file hitting the machine 30 minutes apart because someone upstream edited one field.

• Some of Coral’s customers had 80 or more people whose full-time job was reading faxes and typing them into an EMR.

• 14 to 17 business days to process one referral

Ajay observed the human cost of the broken pipeline firsthand. A friend in New York was stuck between a two-week referral wait and a scan he couldn’t afford to pay for upfront. He flew to India instead, where he got diagnosed and operated on, and he was back in the US in five days. Ajay still couldn’t get him a scan appointment.

Automation that subtracts value

Healthcare is roughly a third of the process-automation revenue pool, and the healthcare providers still hired people rather than buying software because it couldn’t hit the reliability threshold.

Traditional OCR software reads healthcare documents at 55 to 60 percent accuracy. A human reads at about 89. Thus, every document the machine got wrong created rework, denials, and delays that a fully manual process would have avoided. The portal bots built on scripted clicks broke ~40 percent of the time, which meant a patient waiting weeks for the retry.

Ajay’s summary of what a decade of that created:

“RPA is a great band-aid. The problem is the band-aid doesn’t heal. I like to joke that our company is building a very durable band-aid.”

Healthcare providers tried the software, and when it made things worse they went back to hiring people.

Crossing the threshold

Coral brings reliability and intelligence to this market.

The company built its document stack in-house: a cascade of specialized models for splitting, checkbox detection, signature extraction, and clinical reasoning, fine-tuned on millions of de-identified patient documents.

Its browser agents read payer portals visually and replan when a page breaks, which one insurer's does every Monday. Accuracy on the document types that define healthcare's back office now sits at 99.7 percent.

Ajay on why the last stretch matters most:

“You add no value at 90 percent accuracy. At 99 percent, you truly solve the problem.”

A market can look permanently resistant to software, littered with failed vendors and manual workarounds, right up until someone crosses the reliability threshold, at which point the same market looks like it was waiting all along.

One Coral customer, a business with roughly half a billion dollars in revenue, ran referrals through 170 people over four days each. The same work now takes four to five minutes, with 10 to 20 people overseeing it. Most of the rest of the team moved to talking with patients.

Creating a new patient experience

In most of US healthcare, the burden of coordination sits with the patient.

You leave primary care with a referral and a vague window, you call the specialist yourself, and then you wait. Most people around him, Ajay says, describe the system with a version of the same question: do we really have to go?

The experience Coral is building inverts who does the chasing.

“I’m going to tell you who’s the best provider for what you need, and I’m going to follow up with you in six minutes and not break that expectation.”

This only works because of the infrastructure Coral is building. Patients can feel the automation because it is reliable. Ajay’s own test is simple: somewhere a 65-year-old is going through something, probably alone, and someone reached out and said, we’ve got you.

The team you build

Ajay’s biggest learning from 18 months of building is about people. He and Aniket (co-founder of Coral.ai) started out believing that being great at technology meant they could do everything themselves. He now calls that “so far from the truth,” and believes the truth is that the team you build is the company you build.

“The only way we can do what we have to do is if the greatest people in the world choose to work on this specific problem, and want to spend their time and energy solving this one problem.”

That learning is now a recruiting pitch: Coral is looking for people who want to build the best models in healthcare and, in Ajay’s words, earn the right to serve a provider. What he’s recruiting them into hasn’t changed since his first conversation with Lightspeed, a coffee chat scheduled for thirty minutes that ran ninety.

“We are going to build the next great process automation company. I’m going to just keep repeating that until it becomes a reality.”

This write-up is based on Rohil Bagga’s conversation with Ajay Shrihari of Coral on the Lunchbreak with Lightspeed podcast. Watch the full episode below.

Subscribe now

Somewhere between then and now the semiconductor industry ended up scaling 100x in almost all metrics, not just in market-cap but also the one metric that used to give us all nightmares: heat. So I consider this a bit of a personal homecoming that today we are partnering with Discovered Materials to lead their $9M seed round, joined by Y Combinator and angels such as Paul Graham, Gokul Rajaram, and leaders from the AI labs such as OpenAI, Anthropic that are deeply interested in having this problem solved. We met the founders, Advaith Sridhar and Akash Ramdas, earlier this year. This is a unique team that has spent time in Stanford’s materials labs as much as in building verifiers & harnesses at frontier AI startups. Ten minutes into our first meeting we were arguing about junction temperatures and ASTM test rigs and I knew two things: this problem had finally found its team, and I was going to enjoy this diligence more than any in recent memory.

Our thesis can be easily summarized in just a few points and I’ll lay it out up-front and then, try to make Akash & Advaith proud by, as they would, empirically building a case for this investment. 1) Heat is becoming the binding constraint on AI (2) You cannot just plumb your way out of it: air-cooling, liquid-cooling, direct-to-chip cooling are all band-aids that don’t solve the core issue (3) An AI-native materials company that dissipates heat better atomically is the right, and possibly only, answer (4) The materials company will need to be closed-loop across materials discovery, a verifier that filters out unviable materials and an autonomous physical lab that rapidly synthesizes, tests, and invalidates/validates properties predicted in discovery phase.

Let’s break down every part of this thesis and build it back up. This will need a bit of a detour of physics, and asking a lot of whys.

01 Why is Heat the binding constraint on AI?

In 2024, the bottleneck on AI was chip supply. Fast forward to 2026, most are talking about energy as the main bottleneck. There are legends of Elon Musk pushing unpermitted generators next to his datacenters to support demand scale-up on Grok. Space-based data centers are getting attention because of “free” solar power and reduced cooling needs. The real issue isn’t that more heat is getting generated; it’s that it’s getting generated over the same surface area so it’s harder to extract & dissipate it.

Why does the heat keep concentrating? (1) The die size cannot grow (2) Each generation therefore packs far more transistors into that same reticle-sized die to extract more TFLOPs (3) Every transistor draws power. (4) All of that power becomes heat in the same small area. Heat per area = power ÷ area and it’s elementary school math from here on: the numerator keeps growing, the denominator (per-die) stays roughly the same (the package is another story, we’ll get there in a second).

If you are really keen to go deeper in the why-rabbithole, you may want to know why can’t the chips grow any larger — surely, that’d solve much of this issue, right? Let me introduce you to the reticle limit. The reticle limit is the maximum physical area (historically around 800 mm² to 858 mm²) that a photolithography machine can expose on a silicon wafer in a single shot. Because of the physical limits of light diffraction, lenses, and optical alignment, chips cannot be printed larger than this as a single monolithic die. As you can see below, we have been sitting at the reticle limit since early 2020s. Nvidia’s last 3 generations have gone from 54B transistors squeezed into a 826mm2 area to now 104B transistors on a 814mm2 area. And when that wasn’t enough, Blackwell fused two reticle-size dies into one package to reach 208B (~1600mm2), with Rubin Ultra set to stitch four dies together (~3200mm2). The die is pinned at the reticle but the package now grows by tiling dies. Doesn’t that solve the heat problem, you may ask? No — power is growing faster than the tiled silicon. Hopper ran 700W on one die; Blackwell ran ~1.2kW on two; Rubin- and Feynman-era parts are projected to be near 2.3kW and 4.4kW. Every added die is stitched to its neighbors through bonded interfaces where heat gets stuck.

02 Why can’t we just plumb our way out of this heat?

So far we’ve established why more heat is getting generated and why it’s getting concentrated on the same surface area making it harder to dissipate. Let’s ask the next why — why can’t we just plumb our way out of this heat? Let’s first establish just how the power needed to cool scales with generated heat.

At AMD, I lived inside the above chart. Every watt we couldn’t conduct away, we paid for several times over in fans, fins, pumps, and plumbing. To hold the chip’s temperature fixed, the work required to remove additional generated heat doesn’t rise linearly with the heat. It rises with roughly the 3.5th power of heat flux. Double the flux and you need about 11x the cooling work. Now when you put that chart above with the one below, you’d notice just how much more power is getting utilized by chips planned for the next few years. There is an entire order of magnitude difference between power needed by H100 racks vs what is planned for Rubin and Feynman. I don’t need to tell you that when you apply the 3.5^ exponent on an already 10x scale, the numbers get unmanageable very, very quickly.

We know that air is not a good carrier of heat. On a per-rack basis, we crossed the capabilities of air-cooling (~50kW) a few generations ago. Water carries heat roughly 3,500× better than air, which is why everything above the H100 has gone to direct-to-chip liquid cooling. It helps but plumbing only solves the last leg of the journey which is getting heat out of a cold plate and into a fluid. It does nothing about the first and the hardest leg: getting heat out of the silicon and across every interface in its way. Every cooling technique ever demonstrated has a hard ceiling on the heat flux it can absorb and past roughly 2 kW/cm², pushing fluid faster does not help any more. At that point the limit is no longer the pump but rather the stack of materials standing between the transistor and the coolant.

03 Why are new materials the only answer to conduct this heat away?

Heat doesn’t get stuck in the chip. It gets stuck in the gaps between things. Silicon itself conducts heat perfectly well but the problem lies in every boundary the heat must cross on the way out: die-to-lid, lid-to-heatsink, die-to-die in a stack, etc. And it gets dramatically worse the moment you stack chips e.g. 3D integration including memory on logic or die on die. This reduces the area of the GPU and is being done to reduce communication time between memory and compute units. This is the only way to get more performance while working around the current reticle limits till a new way to etch wafers are invented. But as you try to bring things closer together, no matter how close you get them, when seen under a microscope, two “flat” solid surfaces actually touch across only about 10% of their area. The rest is trapped air, and thermally speaking, air is a wall. The core tension here is communication bandwidth vs heat — bringing things closer decreases communication time, but makes the heat problem worse. The chip industry’s roadmap is quite literally burying its heat problem inside the package where no cold plate can ever reach it. Only a better material can.

04 Why we need Discovered Materials

Materials discovery has always been about search. Edison famously burned through some 6,000 candidate filaments over two decades before the lightbulb had a filament worth shipping. A century and a half later, the process at the incumbents such as Dow, Henkel, 3M, Shin-Etsu and others is still recognizably that: brilliant chemists, hand-tuned formulations, quarters-to-years per cycle. Search is the bottleneck. And search is the one thing modern AI has become shockingly good at.

Discovered Material’s engine runs the loop 24/7: an agent proposes ~2000+ candidate crystal structures a day; a separate, proprietary verifier model scores each for stability, heat conduction, dielectric constant, interfacial resistance, synthesizability etc in about ten minutes; physics simulation (DFT) confirms a shortlist over days; and an autonomous wet lab synthesizes and measures what survives. Design → make → measure, all of it compressed from quarters into days.

The materials verifier is the company. Proposing candidates gets cheaper with every frontier-model release and that loop is being commoditized for everyone. Knowing which of the day’s 2000+ candidates deserves a wet-lab run is not easy, and never will be. Discovered’s verifier is trained on a proprietary thermal-conductivity dataset larger than the Materials Project. Tune the sieve too wide and your lab budget explodes; too narrow and you filter out the winner. Getting that balance right is the moat. Is it working? In their YC batch, the team went from a cold start to matching Dow’s commercial TIM performance in about six weeks, beating a Henkel reference along the way.

That said, the TIM market is small at ~$4B today, maybe $8B by 2030, and mostly consumer and automotive at that. TIMs are maybe 10% of the heat problem. They are the beachhead and the material’s discovery engine into higher value categories such as in-package thermal-materials, to atomic-width interfaces in advanced packaging where 3D stacked silicon is headed; and ultimately into materials specified inside the fab process itself, in a semiconductor market headed well past $1.5T with AI driving >50% of that growth.

1,000,000× — the end game

Our brains do their thinking on about 20 watts which is a million-fold efficiency gap between biology and the GPU racks we are building. Adding more registers to the die adds more TFLOPs to the chip. In a way, it helps you grow a bigger, more powerful brain. The opposing force here is heat. As chips get hotter, they automatically trigger thermal throttling i.e. chips actively lowers their clock speeds and voltages to cool down and prevent permanent hardware damage. Every new TFLOP we add to the world is being gatekept by the heat it generates. This is why the work Discovered Materials is doing is so critical. Within three years, we believe no frontier accelerator will ship without an AI-discovered material somewhere in its heat path.

The founders of Discovered Materials have done exactly this work in the past. In January 2025, Akash and his Stanford collaborators published in Science a genuinely surprising result: an ultrathin niobium-phosphide semi-metal whose electrical resistivity falls as the film gets thinner — the exact opposite of copper, whose resistivity blows up in nanoscale wires. At 1.5 nm their films reached ~34 µΩ·cm, several times better than conventional metals at that scale, deposited at 400 °C which is cool enough for real chip manufacturing. It was Stanford Engineering’s most-read story of 2025. Note what that work attacks: the resistive generation of heat inside interconnects. DiscoveredMaterials attacks heat’s dissipation and generation. Twenty years after I left this problem back at AMD the entire AI build-out has circled back to the first problem I ever owned professionally. Backing the team that has published at the atomic scale on both sides of the heat equation before raising a seed round is so rare and is about as close to a homecoming as this job allows.

We’re delighted to partner with Y Combinator and with angels from the labs and firms building the frontier. Welcome to the Lightspeed family, Advaith and Akash.

—

Hemant Mohapatra and Abhiram Tarimala

The room was full of founders less than three years into their journey (in other words, born in the AI era), building services companies across construction/design, semiconductors deign, recruitment, tax advisory, media services and more. Based on this anecdotal data as well as the volume of companies we are seeing in the past 18 months, AI-led services is a huge interest area for founders in India.

In late 2024, we published AI Will Eat Services. Its argument: AI is expanding the market for software from productivity tools (roughly $300B of SaaS spend) to the work itself - the $2T+ spent on outsourced digital work. A large set of services industries are facing existential risks - think law firms, creative agencies, SEO firms, management consulting firms, accounting firms, IT services firms and BPOs. Our view is that AI-led services will be one of the defining business models of this shift and has the potential to create several new venture-backed winners.

Lightspeed has led early-stage financings in several companies in this thesis, including Aqqrue, Bridgetown, Gush, Level3 and Pepper.

Here is what we learnt that could be actionable for founders building AI-led services companies, particularly those building toward top-tier venture funding.

Strategy

Pick which of the three forms you are building.

• A1) Selling and delivering AI-powered services to enterprises: Pepper for GEO/SEO outcomes, Pixis in performance marketing, Bridgetown in commercial diligence for PE firms. Organic buildout.

• A2) Selling and delivering AI-powered services to enterprises. Inorganic buildout or roll-up strategy. Lightspeed partner Isaac Kim has written about services rollups here, here and here.

• B) Selling an AI-native tool to services providers. I’ve seen this work mostly by selling into SMBs (as a scale up strategy) or to larger service providers (as an adjunct growth strategy or channel strategy but not the main one). Large service providers will likely, IMO, try their own hand at building out these technologies/products.

• C) Selling an AI-native tool to enterprises to replace what they used to outsource: Harvey in legal, Sierra and Decagon in support and many others. This is the most heavily financed segment of the market.

Speak the language of outcomes to oust incumbent service providers.

Incumbent agencies (think law firms, creative agencies, SEO firms, management consulting firms, accounting firms, IT services firms, BPOs), some sitting on $50-500mm contracts, all speak the language of inputs: hours times people/seats. They are mostly structurally unable to shift to speaking the language of outcomes, for fear of cannibalizing their own revenues and profit pools. This is a key advantage for new AI-led services startups. Wedge in the way - ask for two lines of business out of those contracts and commit to results instead. Even if you promise a 5x improvement and delivers 3x, the customer stays. They still got a 3x lift, and no other vendor was willing to be measured on results in the first place.

Price against the cost of a human doing the work. This is selling work, not software: sell a 95% improvement rather than a marginal productivity gain. And don’t put all your revenue on pay-for-results terms; pick a deliberate slice. These customers churn faster when targets slip, but the ones who stay become your strongest references.

Go US-first.

Selling to US companies is an ROI conversation. You spend $10, show $13 in results, and the org pays $120 next year for $100 of good work this year. In India, target only the thin tier with real willingness to pay: large banks and financial services companies and IT services firms with US exposure plus the top ~30 FMCG (CPG) companies.

Process Automation/Delivery

Run three tests to measure whether you are genuinely AI-native.

In a pre-AI services company, humans were the production engine and technology the efficiency layer; in an AI-native one, the system produces and humans judge. Ask your own company and fix any “no” before you scale.

1. Does operational leverage sit in systems rather than people?

2. Is revenue decoupled from headcount?

3. Is it the system’s learning that compounds, rather than individual experience?

Turn “learning” into a metric.

Every AI-native services company claims its system learns, so make the claim concrete. The way Pepper’s systems learn about ways of working while serving Unilever makes the P&G client far easier to serve next. Your time-to-value should fall with every new customer you onboard.

Rebuild your delivery pod around agents.

Pepper’s pod went from 8-10 people on a $100-200k account to roughly 3 people and 5 agents by stripping the analyst layer and flattening the hierarchy from 3-4 levels to 1-2.

The key human role that remains is the forward-deployed engineer (FDE): one person who sits with the customer like a consultant and also writes production code, instead of handing requirements down a chain. Companies run this role differently: at Intercom, FDEs push code straight to production; at Outreach, they work mainly as consultants. Decide early how much engineer versus consultant yours will be, and start building the talent pipeline now, because the persona barely exists: ANSR is manufacturing them in-house.

Diagnose churn by segment before you blame the product.

Split your churn by segment (enterprise, mid-markte, SMB) before touching the roadmap.

GTM, Budget, and Economics

Take over an existing budget line.

Don’t ask the customer to create a new budget for you; capture one that already exists. For Rocketlane, their wedge is to land a few seats alongside the incumbent tool, build the connectors, and expand once you reach the data.

Go upmarket (or way downmarket!).

Sell where speed is worth money: there seems to be a barbell here - enterprises and SMBs pay for it, while mid-market tries to in-house the work. On the SMB side, Avoca, EvenUp, Gush have successfully targeted SMBs.

Underwrite ~60-70% gross margins, with a durability plan.

Set the benchmark there: pre-AI services ran 20-30%, and if you’re not near 60-70%, relook at the business. According to Karthik, every industry is susceptible to compression over time so track margins by customer cohort, show them improving as automation rises, and explain why you keep 60%+ when pricing compresses.

Team

While we did not discuss this during the session, I think this is the most important aspect of any startup, including in AI-led services. My view is that a strong AI-pilled technology founder is key in addition to a strong operator/domain expert. In case a roll-up strategy is part of the plan, then it helps to have an M&A expert on the team. It’s tempting to assume that insiders from the large incumbents will be the natural founders for new AI-led services startups - not always the case (but certainly can be).

--

AI-led services will be a fragmented market, not a winner-take-all, and that’s fine: in a market this large, even a non-winner can be enormous. Venture firms have historically funded technology companies more often than services companies, so build to clear a high bar: margins that reach ~70%, revenue that can reach hundreds of millions of dollars within a decade, and growth that is fast and isn’t local and relationship-bound.

Clear it, and the market you are playing for is the one we described in AI Will Eat Services. Not the few hundred billion dollars a year the world spends on software, but the trillions it spends on work itself.

Dev Khare

Subscribe now

When we wrote our first check into Mitti Labs in 2024, this was a carbon company, and the entire thesis came down to one word: trust. A rice carbon credit is worth nothing unless you can prove that a specific farmer, on a specific one-hectare plot, actually stopped flooding the field this season. Everyone else in the category was busy selling the credit but we backed Mitti Labs because they were the only team who focused a lot on ground-truth data — satellite imagery across microwave, SAR and high resolution bands, IoT sensors for soil parameters, time-series data on water usage, plant health and more. It was a lot of hard slog that had to be done outside of the headlines over two years. This team persevered.

So this one is personal. Today Mitti Labs announces a $9.5M Series A led by Aramco Ventures, their first ever investment in India, joined by Godrej Industries Group, Cisco Foundation, Francis Family Fund, and Volta Circle, with Lightspeed doubling down from the seed.

The problem Mitti Labs set out to attack

Nothing about why rice matters has changed since 2024, except we have entered an era of water bankruptcy. India carries 18% of the world’s people on 4% of its freshwater, per-capita water availability has halved since 1970, and on current trends demand runs 50% past supply of freshwater by 2030. Roughly 80–90% of India’s water goes to farming, and rice alone drinks more than 30% of all irrigation water on the planet. No other single decision on Earth moves as much freshwater as how a paddy is irrigated. The same flooded paddy is also a climate machine. Standing water also starves the soil of oxygen, the anaerobic microbes feast, and rice ends up responsible for 10 to 12% of global methane emissions. So the entire world’s agricultural mission boils down to (1) growing more rice (2) on less water (3) in worse weather (4) without torching the climate. Not for the faint of heart.

Why we backed Mitti Labs in 2024 and what two years unlocked

Two years ago, our thesis called out three things as core: (1) own the dMRV, and therefore the data / ground truth in a way that’s remotely measured & monitored (e.g. satellites, AI); without that there is no long-term moat in climate and it becomes a project-management business (2) accelerate farmer-behaviour shift via a top-down demand-first approach where buyers of rice in US/EU are demanding a change and (3) bring all stakeholders — the FPOs, the financiers, the buyers — on one platform to create transparency and trust.

On (1), what began as the world’s best dMRV for rice is now a full GeoAI instrument: time series physical retrievals of soil moisture, field inundation, and residue, generated from a multi-sensor radar stack (C, L, X) that sees structure through monsoon clouds, all fused with deep learning architectures trained on hundreds of thousands of farmers to feed the bio-geochemical engine built by Mitti’s in-house scientists. As a result, Mitti Labs helps farmers optimize Alternate Wetting and Drying (AWD) management practices to save water and decrease GHG emissions. Mitti’s dMRV measures it all remotely how many dry-downs happened, how often, how long, etc. This is revolutionary in the category anywhere in the world. NASA has also given Mitti Labs an $850K grant as a result which is a strong proof point of our technical superiority.

On (2), the demand-first approach has paid off well so far. Mitti Labs now runs the largest AWD rice program in the world: more than 100,000 farmers across six Indian states, growing 3x every year from 8000 farmers in 2024 to 35,000 last year. The top-down buyers such as Ebro Foods (the world’s largest rice company) and Syngenta, alongside global technology and financial-services names have all led to a brand and data flywheel for us that we frankly weren’t expecting to kick in so early.

On (3), the platform play, this is where we got surprised. The data record Mitti Labs built to satisfy a carbon auditor turned out to be something almost no one else on Earth has: a verified, season-by-season account of how 100K+ single-hectare/smallholder farms actually operate. We learned its value when buyers who did not want a single carbon credit began asking for the data itself. Food and agriculture companies wanted the field-level truth for their own supply chains. Others asked to license the platform outright. Mitti Labs now runs a second business selling that data and its insights, entirely separate from carbon.

Surprisingly enough, there is an interesting intersection of Geospatial data and AI that is emerging now. Models that can answer questions about the planet with a chat-like interface, or even predict the state of the earth for scenario modeling. In 2024, NASA and IBM launched Prithvi-2, one of the world’s first geospatial LLMs. Soon after, in 2025, Google DeepMind’s launched AlphaEarth. Yet, the utility of these models is tethered to the quality of their training sets. In the physical world, ground-truth data remains a scarce commodity — a scarcity that deepens across India and Southeast Asia, reaching its peak in the isolated rice belts where almost nothing has been digitized. Mitti Labs bridges this gap by owning the data generation layer, verifying every season from the dirt up, and fusing it with GeoAI and bio-geochemical stacks that transform raw records into actionable truth.

Why we are doubling down

Once you can measure a hectare, you can predict its yield at sub-field resolution, time its irrigation, flag an input switch, and see a bad season coming. And the moment a hectare is measured, you can underwrite it, insure it, lend against it, and trade it. The interesting future customers are as much the food and agriculture companies, governments routing enormous energy and fertilizer subsidies toward resilience, banks, insurers, and commodity desks as they are the carbon buyers.

Moreover, the same field-level water ledger that underpins a credit could one day underwrite the replenishment requirements of AI hyperscalers as they build data centers in these markets. This kind of optionality only appears if you own the measurement layer. Rice alone is a $300B industry serving 150 million farmers that technology has almost entirely skipped. Within five years, we believe no serious model of Asia’s food system, its water, its carbon, or its crop credit, will be trained without Mitti’s ground truth somewhere in its loop.

None of this would be possible without the team and the founders. Devdut Dalal spent years as an operator in food and agriculture. Xavi Laguarta came from the demand side of climate, the corporates and carbon markets we needed cracked. They met in the same section at Harvard and spent a summer in India pressure-testing the idea. Nathan Torbick, who spent a career mapping rice from orbit for NASA and JAXA, drove down to Cambridge and says he was skeptical at first but after a few whiteboard sessions, he was convinced. We’ve been so lucky to see them grow from a 3-member team to now over 150 people across the world!

Welcome to our co-investors in this round, especially our friends at Aramco Ventures who are leading the $9.5M series A. With this capital, we plan to go deeper across India’s rice belt, and expand into the Philippines and Indonesia, further up the GeoAI stack, and toward 2 million credits issued a year by 2028.

—

Hemant Mohapatra and Romit Mehta

A trade professional we spoke to told us something that struck me: his local shopkeepers had begun asking why he no longer visited. He used to spend a meaningful part of every day travelling between stores. Now, most materials come to him.

That is the behaviour shift Fixxly is building around.

For contractors, electricians, plumbers, painters and carpenters, running out of one small item can bring an entire site to a halt. And they run out of these items, often. Lost time and lower job throughput are the costs that these professionals pay. Stack that over almost every single day and this begins to hurt.

On the supply side, the building materials industry is among the largest retail TAMs - $40B. And like most Indian retail, it is fragmented. One shop rarely stocks the SKUs that professionals need. When was the last time your plumber / electrician told you that a part you needed was not available and that they would come back the next day?

Fixxly is consolidating this fragmented industry and meeting users at the point of their demand. Hit the app, find a curated range of building materials, all delivered to the worksite in about 30 minutes.

More interestingly, Fixxly encapsulates the wave of Qcomm 2.0 in the Indian context. The first generation of quick commerce was built on the back of low-value, high-frequency consumer purchases. It was revolutionary and changed expectations permanently: people now assume that products should come to them, precisely when they are needed.

We believe the next generation of commerce companies will apply that behaviour to higher-value transactions.

Building materials are a useful example. Professionals can order 30+ times a month; some we spoke to procure two or three times in a single day. The basket size can exceed INR 2,000 with high double digit margins across multiple categories such as electricals, plumbing and hardware.

That combination - high frequency, high AOV and controlled logistics creates the possibility of a fundamentally more capital-efficient commerce model. Fixxly’s choices will reflect that discipline.

Getting a business off the ground needs a founder to make sharp choices. We instinctively gravitate towards founders who bring that clarity and conviction to a business. In this instance, that clarity comes from Shezan’s two decade long journey in this category - as a buyer at DesignCafe and HomeLane, and on the distribution side at IBO.

I remember our first meeting well. Sitting in a coffee shop at a hotel, he held sway over an hour diving into the economics of the business down to the last rupee. The best commerce founders combine outsized ambition with relentless execution and eyes wide open to the levers of their business. Shezan is that unique combination.

It also explains the team he has assembled including Sachith, his co-founder who brings supply-chain experience from Zepto, Flipkart, Swiggy and Amazon.

As Fixxly scales, it has the potential to revolutionize the $40B industry. Let’s start at the atomic level: with the contractors themselves. They no longer lose half a working day because a fitting, switch or paint accessory is missing. Zoom out and if these components and materials are available on demand, the inventory-driven working capital cycle in this industry could drop significantly., and supply gets an entirely new channel for brands which for decades have relied on fragmented distribution.

We’re delighted to be partnering with Shezan, Sachith and the Fixxly team as they build for that future.

Fixxly is launching early-access hubs in Bengaluru, Hyderabad, and Mumbai. Learn more at fixxly.in.

—

Romit Mehta

Rubrik went public in 2024 in the US and counts some of the largest banks, telcos, etc in the US and globally as their customers. Portkey got acquired by Palo Alto Networks in 2026, three years after founding. Acceldata is a leading data management company with Fortune 50 customers in the US, including some of the largest banks, pharmacos, telcos etc. All are or were Lightspeed portfolio companies, backed by Lightspeed from the seed stage onwards.

Here is what we learnt that could be actionable for founders building in this layer:

Entry Point

• Enter a category that is un-named and/or does not have a young incumbent. Portkey’s category renamed itself FMOps → LLMOps → AgentOps in three years because agents didn’t exist at the start. A category only forms over time, once customers, vendors and analysts show up. You can’t will one into existence and it is hard to know how long it will take for a category to form, think about entering before the space/category has inflected.

• Before writing code, find a non-consensus pain, one that customers feel but the market has written off. When Rubrik started, backup was a 30-year-old category where recovery failed ~50% of the time and customers still paid millions. Acceldata’s test for a real signal was having worked at big data companies like Hortonworks and understanding how customers found it painful to manage their data across data silos.

• The entry point that opens an account is usually small and trivial to adopt. Rubrik got into Best Buy by supporting their VMware setup: they got the vCenter login and discovered everything running above it.

• Sell before you write code. Rubrik’s founder spent the first months pitching prospects rather than building, and built only what kept coming up.

Building

• Lock your architectural non-negotiables on Day One. Rubrik made two bets 12 years ago, before either looked necessary: scale like a hyperscaler, and treat the data as if it must stay secure. When they became a security company, “unhackable backups” wasn’t a retrofit. Customers don’t usually buy the architecture, but the one or two things you refuse to compromise become the moat.

• Sometimes a crisis reveals your best use case. Rubrik built a “time machine” - a rewind of a business’s data to any prior state, down to a single row without knowing its best use. Ashish pointed to Colonial Pipeline as the event that crystallized it: even after the ransom was paid, the attackers’ decryption tool was so slow recovery would have taken ~45 days. Attacked companies don’t want to “recover from backup,” they want to jump to the last clean moment before the malware landed. That reframing turned a backup company into a security company.

• In infra, 99.9% confidence isn’t enough. The product has to work essentially every time. Rubrik puts money behind it: a warranty of up to $10M if they can’t restore a customer’s data within 48 hours of an attack. That promise, plus depth of IP and day-one customers who grew with you, is what makes a company acquirable.

• Build in one place until product-market fit. Splitting engineering across time zones before the first product lands kills iteration speed. Rubrik spun up India-owned product lines only afterward, each owning backend, UI, UX and product plus its P&L. By all means, once there is some early PMF, go set up go-to-market teams in your core market.

Selling

• Start with targeting AI-native companies, not the Fortune 500. AI-native companies (also called frontier companies) ship fast and compress months of product feedback into weeks. Portkey’s early design partners included DoorDash and Shopify. These logos often stay private, so the payoff isn’t only references, it’s a hardened product. When a large enterprise arrives, the feature they ask for already exists.

• Breaking into the US is a network game. Enterprise infra buying often runs through people of Indian origin, so alumni intros beat cold outreach. Explore onboarding a late-career US operator as an advisor for 0.25-1% equity who can open doors. Later on, system integrators can also multiply reach once they can resell you as their own (but this is usually much later).

• Open-source the commoditizing layer. Portkey open-sourced ~80% of its product and kept the enterprise pieces (deployment, governance, auth) closed. Their bet was that as LLM usage goes universal, the routing gateway commoditizes, so they owned that layer in the open. It also killed enterprise self-hosting anxiety, and every customer came inbound or by referral.

—

Dev Khare, Partner, Lightspeed

Subscribe now

Karan Vaidya, co-founder and CTO of Composio, joined Lightspeed partner Hemant Mohapatra and vice president Rohil Bagga on the Lunchbreak with Lightspeed podcast to talk about the gap Composio exists to close: models are far smarter than they were two years ago, but they still can’t act on the apps where work happens.

Hemant frames it with a line he’d read somewhere: intelligence without agency is useless. That position between intelligence and action, he says, is what stood out when Lightspeed partnered with Composio.

Code lives in one place. Work lives everywhere.

Models now offer what Karan calls PhD-plus intelligence. The gap is in putting it to work.

Karan: “You have the best-in-class models, but we provide the hands, the muscles to these models so that they can take action on real-world knowledge work apps. We are the bridging layer, connecting AI with knowledge work.”

The base of that layer is unglamorous: auth, which Composio has solved across 1,000-plus apps.

What sits above auth is the harder part, the part raw intelligence can’t solve on its own.

AI models cracked coding first because everything about code lived inside GitHub. Peer review, tests, deployment, governance, all in one place, every outcome verifiable. Knowledge work has no such place.

Karan: “The problem becomes immensely harder in the case of knowledge work, where it’s not a single app. Salespeople use Salesforce and a lot of other systems. For accounting, people use QuickBooks. It’s divided across so many data silos. There is no single place where I can have verifiability, where I can have governance. That’s where we come in. One of the ways I like to pitch Composio is GitHub for knowledge work.”

Making scattered work reliable

Hemant brings up the benchmark on how agents degrade as workflows grow: 100% accuracy on a single tool doing a single task, another 10 to 20% lost with each tool added, as low as 5% on multi-tool, multi-step workflows. Then the key question: why does plugging in Composio make accuracy on those same workflows sharply rebound?

One reason, Karan explains, is better tools. An agentic pipeline builds and maintains them: it maps API endpoints, dependencies, and edge cases, and when an agent fails in production, the pipeline ships a corrected version. New versions of a toolkit like Gmail’s arrive every other day, and a tool that has absorbed enough real failures ends up ahead of the vendor’s own documentation.

Karan: “We have seen in multiple cases that our tools are better than what the documentation offers. There are edge cases and parameters the documentation has gotten wrong, but because we have seen so much data, we ran into that edge case and corrected it.”

Tool selection is another. Composio surfaces a small, correct set of tools at the moment the agent needs them, what Karan calls tool intelligence. The decay in the benchmark comes from agents facing too many choices per step. Fewer wrong turns on offer, fewer wrong turns taken.

Skill distillation is perhaps the most important reason. Every agent run produces a unique trajectory. Composio records it, distills it into a skill, and serves it back the next time a similar task shows up.

Karan: “The first time, the agent was figuring it out. It was a new task, and it took a zigzag trajectory to reach the outcome. The next time, it gets a pre-recorded skill and achieves the outcome in a straight line. It’s more reliable, token-efficient, faster.”

The skills repository alone lifts agent accuracy by about 15% between a cold hit and a warm one, Karan says, even on a lower-tier model.

The protocol isn’t the moat

Rohil surfaces the million-dollar question: connecting agents to tools is exactly what MCP was supposed to do. Karan takes no issue with the protocol. The moat, he argues, sits underneath it.

Karan: “Protocols will keep on changing. If there are gaps in MCP, people will use CLI. If there are gaps in CLI, people will use skills. I think our customers choose us because of that hard-built infrastructure sitting behind the MCP or the CLI.”

Asked why the largest companies trust a startup with a layer this critical, Karan calls it existential: every enterprise is racing to ship agent products, reliability decides whether those products work, and Composio has the data flywheel.

Self-serve users alone run about 150 million tool executions a month through it, growing 50 to 100% month on month, and those executions are the raw material the tool pipeline and the skills repository learn from.

The flywheel also produces lock-in that runs in Composio’s favor. Credentials, trajectories, and skills all live with Composio, so a customer can switch models without starting over. Skills turn out to be model-specific, and Composio translates them.

Karan: “Skills that work well with Claude don’t work very well with Codex because of model behavior differences. Because of the data we have, we can do the skill translation for them.”

From daily habit to default harness

Rohil asks whether self-serve developer usage converts into enterprise business. Karan says this conversion is core to the go-to-market: engineers and product managers pick up Composio inside their coding agents, and when their companies build agent products of their own, and nearly every company now is, they choose the layer their people already use.

Karan: “We have multi-trillion dollar companies sitting out of Seattle using us at this point as their main tool harness in their agent product.”

Most of what those companies build on it is customer-facing, he adds, because the UX of enterprise software is itself turning agentic.

App makers have started coming inbound too. A bank recently reached out wanting to integrate, because its users said they only use tools available inside Composio.

A baby in one arm

Asked for the hardest moment of building Composio, Karan gives two answers. One was the start itself: his daughter was born around the time the company was, and the early customer calls happened with her in his arms.

The other is hiring. The funnel is strong, the bar is high, and the team still hires very few. Which makes his pitch to engineers worth hearing, because it’s the same as his thesis for the company.

Karan: “There are a lot of open loops in AI right now, which if we close, the impact of AI will tremendously increase. GitHub can close the loop inside the GitHub ecosystem. Linear can do it in theirs. But doing it across apps has tremendous value, because people don’t use just one app. We are sitting at a goldmine, because we are connected to every app, and we have the agent trajectories coming in across apps. It’s a pretty hard problem, and you have to learn from actual trajectories.”

This write-up is based on Hemant Mohapatra and Rohil Bagga’s conversation with Karan Vaidya of Composio on the Lunchbreak with Lightspeed podcast. Watch the full episode below.

The past, present and future of AI infrastructure: the puck is constantly moving

The requirements of AI infrastructure keep changing as the hero actor, the entity the entire stack exists to serve, changes.

In the past, the hero was the LLM. Every team was trying to use models and build on top of them to get to a demo or a good pre-production version. Infrastructure was built to train models, observe and secure usage, and route across models. Gateways like Portkey and OpenRouter, frameworks like LangChain, RAG pipelines, LLM observability and vector databases all came up to serve this hero.

Today the hero is the agent, running in production, and the categories inflecting now all orbit it: memory (Mem0, Supermemory, Letta), tool use (Composio, Arcade), agent evals and observability (LangSmith, Braintrust, Ragas, Maxim), and sandboxes where agents can safely execute (E2B, Modal, Daytona).

The future is agents in the wild. While production is a context you control: known tools, known data, a human nearby, the wild is everything else, and out there the error bars widen in both directions, on the upside of an agent that acts for you and on the cost of one that acts wrongly.

The next frontier of infrastructure narrows those error bars. It starts with agents that improve themselves from their production traces (Raindrop, Judgment Labs, Bento Labs), and models that keep learning after deployment instead of resetting every session. As agents start acting on behalf of their users, identity, security and compliance become the precondition for trusting them with credentials and real actions. Insurance emerges to price the risk that remains. And because the next frontier is physical, the infrastructure to train, run and observe embodied AI inflects alongside all of it.

Skating with the puck

If we follow the puck over time, we can see how infrastructure has changed across a few themes.

Core infra

• Past - infra for LLMs: Gateways, frameworks and RAG. The first act of core infra helped developers talk to models and get the best use out of them. Gateways like Portkey and OpenRouter routed and governed LLM traffic. Frameworks like LangChain and LlamaIndex abstracted away prompts, retrieval and API calls. Vector databases and RAG pipelines rounded out the stack.

• Present - the agent harness: The categories inflecting now give agents the faculties a model alone lacks. Tool use (Composio), memory (Mem0, Supermemory, Letta, Zep), observability and evals (LangSmith, Braintrust, Ragas, Maxim) and sandboxes (E2B, Daytona) have all scaled on the back of making agents effective in production.

• Future - allowing AI to act in the wild: Two related but distinct loops inflect here. The first is self-improvement: the harness closes its own loop, where an agent’s production traces feed evals and the evals rewrite the prompts, tools and plans (Judgment Labs, Raindrop, Bento Labs). The second is continual learning: the model itself keeps updating after deployment, so what it learns in the field compounds in the weights instead of living in an ever-growing context. As both loops mature and agents earn real autonomy, agent identity, compliance and insurance inflect alongside.

Data

• Past - labeled data and RLHF: The first act of the data business was mobilising human experts to label and rank model outputs. It built giants in Scale, Surge, Turing and Mercor.

• Present - RL environments: As the hero became the agent, labs and apps started buying simulated worlds and RL environments where agents run tasks and learn from the results. The data vendors of the past pivoted into this space alongside new ones.

• Future - data for the hard domains: The next leg of data moves into domains where the answer is expensive to get and where the frontier models are building today: robotics, biology, materials science, medicine and law.

Inference and compute

• Past - LLM training and offshoots of serving: Mosaic, Anyscale and others turned scarce GPUs into an API and allowed developers to train models easily.

• Present - the inference inflection: As agents moved into production, the centre of gravity of compute spend shifted from training to inference, giving rise to companies like Fireworks, Together, Fal, Baseten and Modal. All of these companies grew rapidly in revenue. The other battle is cost engineering for multi-turn agent workloads: KV cache reuse, smarter routing and scheduling, heterogeneous compute. From India, Simplismart and Pipeshift have emerged - catering to both sovereign and global demand.

• Future - the edge and new silicon: As AI starts acting in the physical world, data cannot make a round trip to a data centre for every decision, so models and inference moves on-device (Liquid AI, WebAI) and onto new energy-efficient substrates. The next front is routing: with dozens of viable models, sending each request to the one that answers it best, cheapest or with the lowest latency becomes its own category

Security and trust

• Past - LLM guardrails: Prompt injection defence and model scanning inflected early and consolidated fast.

• Present - agent security and identity: An agent holding credentials and taking actions is a new attack surface. The active fronts are agent identity and permissions, automated red-teaming (Gray Swan, Haize Labs), and securing tool calls as agents interact with the real world.

• Future - insurance and certification: When agents take consequential actions, someone has to price the risk. AIUC is tackling this in the US, and companies like Ollive are taking it on from India.

The categories that have scaled into real revenue are the data and inference providers. They all sell to the labs, and the labs scaled first. The layer that serves the application companies, the memory, the gateways, the observability around the agent, is early in its commercialisation journey, because its customers are only now reaching large production volume. Infrastructure is a lagging layer by design and a lot of the timing is arriving now. Token consumption in the enterprise has started to spike, and as it does, the companies that spent the early years building product and distribution ahead of demand are the ones positioned to scale into it.

The puck keeps moving. The teams that win will be the ones that noticed, each time, when the hero changed, and not the ones that picked the right category just once.

India has produced category leaders across generations

We observe that a large percentage of the teams that created a category, held it, and then re-founded themselves into the next one were built by Indian founders, and more of them are now built from India.

1. We’ve seen this across generations, and it fuels future generations: India has had category leaders in devtools - Postman, Hasura, BrowserStack and others before them which created the compounding assets that matter: playbooks for product-led growth and alumni who became founders. Several of today’s AI infra founders come directly out of that lineage

2. The world’s best AI and infra companies already build from India: Glean, Rubrik, Databricks, Snowflake, Nutanix all run large engineering organisations in India that ship core products. Engineers here work on some of the world’s best infrastructure and AI products every day and see the architecture, the gaps and the opportunities up close.

3. Open source made distribution geography-blind: Infrastructure is bought by developers, and developer distribution happens on GitHub, HackerNews and Discord. Mem0, Ragas, Composio and many others all scaled their categories on GitHub.

What we’ve seen from the best infra founders

• Pre-empt the trend: The best infra founders see the trend-line years before it becomes a category, and use the lead time to build product and distribution ahead of demand. By the time a category appears on a market map, the winners have already been building for a few years.

• Constantly skating with the puck: Infra requirements keep changing, especially in AI. Demand in one product evolves into demand for the next. LangChain is the canonical case: the framework got the developers, LangGraph caught the agent orchestration wave, and LangSmith caught the observability wave. Scale and Turing rode labeled data into RLHF into RL environments. Your first product is a wedge into a relationship with developers. The company gets built only when founders keep skating with the puck

• Move closer to your customers: Open source gets the initial traction from anywhere on earth, but sales happen close to the buyer, and for AI infra the buyers are largely concentrated in SF. Being there means hearing about a pain point the minute it emerges. Nearly every Indian infra breakout has converged on the same barbell: a founder in SF within reach of the customer, world-class engineering built from India

• Sell to the companies building at the frontier: The clearest tell of a breakout infra company is who is on the customer list. The top labs and the fastest growing AI-native applications companies are the hardest buyers in the world. They can build anything in house, they evaluate everything, and they churn without sentiment. When they still pick your tool, you are on their critical path. Portkey ran in production at Snorkel. Raindrop monitors agents for Vercel, Replit and Avoca. These logos do two jobs at once: they are the harshest quality bar in the market, and they are a preview of demand, because whatever the frontier companies need today, everyone else needs in eighteen months. If your first ten customers are the companies the rest of the industry is copying, distribution follows.

• Stay plugged in to monetisation: Stars, downloads and Discord members are leading indicators, but a huge share of infra tool usage dies in the POC, and what survives often gets rebuilt in-house once the team understands the problem. The founders who convert know exactly who pays and what they pay for: not the demo, but the production path. When your product handles live traffic, holds state or gates a deployment, ripping you out has a cost, and that cost is your pricing power. The test isn’t “how many developers tried us.” It’s “what breaks if we go down.”

Who we see winning

When we zoom out, we’ve seen certain types of founders shine in infrastructure over and over again:

• People who hit the problem firsthand at the application layer: Working on apps, they feel the pain point before anyone names it and create the solution they wish existed for them. Rohit at Portkey led product and AI at Freshworks and then Pepper, hit the same production problems at both, and built the gateway he wished he had. Abhinav built the first version of Postman while running TeliportMe, because the API tooling he needed did not exist. Prukalpa built the data tooling inside SocialCops that became Atlan.

• People who hit the edge of the problem inside a frontier lab: working at the cutting edge, they run into the deficiencies no one has solved or even seen yet, build the fix for themselves, and that fix becomes a product.

• People who built the previous generation of infrastructure and saw what had to come next: Having shipped the last layer, they know exactly which one is missing. Rohit and Ashwin built data systems at Hortonworks, saw that enterprises needed an independent layer to manage them, and started Acceldata.

• Young, hacky teams who spotted the trend early and built distribution: Open source let them win a category from anywhere, before they had a sales team or even a company. Daniel and Michael Han, two brothers, turned Unsloth from a fine-tuning library into a business with over ten million monthly downloads. Taranjeet and Deshraj are running the same play with Mem0, as are Soham and Karan with Composio.

The next few years will decide who owns the infrastructure layer of AI, and we believe a meaningful share of it will be built from India. If you are building here, at any stage, even pre-product, we would love to meet you. Reach out to us!

-

Dev Khare and Abhiram Tarimala

Note: Acceldata, Composio, Databricks, Exa, Fireworks AI, Glean, Handshake AI, Hasura, Judgment Labs, Langfuse, Luel, Nutanix, Patronus AI, Pepper, Portkey, Raindrop, Rubrik, Snorkel AI, Straiker, Unconventional AI, Unsloth mentioned in the article are or were Lightspeed-backed portfolio companies.

Subscribe now

Lightspeed partner Dev Khare sat down with Aayush Ghosh Choudhury, co-founder and CEO of Scrut Automation, an agentic governance, risk, and compliance (GRC) platform started in late 2021. Lightspeed signed Scrut’s seed term sheet 24 hours after their first meeting, says Dev.

Four years later, the two compared notes on what it takes to build an AI-native security and compliance company when the technology underneath keeps shifting.

A series of epiphanies

A striking number of the companies Lightspeed has backed, Dev notes, were started by people who hit a problem inside another company and left to go solve it. For instance, Portkey’s founders found the AI-gateway gap while building agentic apps at their previous job. Acceldata’s founders came out of Hortonworks with their problem in hand.

Scrut fits but took a detour to get there.

Aayush left McKinsey and joined an early-stage SaaS company just before his promotion. Then he jammed with his co-founder Jayesh to build supply-chain software on top of enterprise ERP systems. While the company didn’t scale, the year they spent on it surfaced a different problem.

Even as a team of under five, selling to large enterprises meant burning serious hours proving SOC2, ISO, and GDPR compliance.

“We had a series of epiphanies that led us to the idea of Scrut, because even as a very small team of less than five people, we had to work very hard towards demonstrating our compliance towards SOC2, ISO, and GDPR, because we were working with very high-quality enterprise customers. That pain led to two things: one, we built a bunch of internal tools that helped us be smarter about how we retrieved those artifacts; and two, we found Kush through our network as a third co-founder.” — Aayush

Small, tech-forward companies were newly carrying this compliance load, and nothing on the market served them well. Scrut’s first version was little more than that cleaned up internal tooling, but it sold fast because so many companies were stuck in the same spot.

That, for Aayush, is the real signal to start a company: the moment a problem you keep poking at starts crowding out your actual job.

A security workforce that doesn’t exist

Scrut sells to small security teams of four or five people, sometimes fewer, who carry compliance and risk workloads far bigger than their headcount.

“It’s not so much a story around, okay, can we go to a company and reduce their headcount by half. Most of the situations that we go into, this workforce doesn’t exist. This capacity doesn’t exist, and people are actually not able to access GRC and security outcomes that they would like to.” — Aayush

The reframe also changes the math on market size. Software licenses, Aayush argues, are maybe 10 to 15% of what a company actually spends to get a compliance outcome. The rest goes to headcount and consultants. Agents that offer expert judgment at a fraction of the cost of human FDEs compete for all of it. He puts the expansion at roughly 10x.

Usage patterns back this up. Since going live last year, about three-quarters of customers have adopted Scrut’s agentic tools. Some have stopped opening the Scrut interface entirely: they hit the platform through its MCP server, ask Claude for a dashboard or summary, or send instructions through Slack. The aim, in Aayush’s words, is to get “closer and closer to infra” and let users pick their own way in.

From dashboards to loops

Aayush has rebuilt Scrut around a changed definition of valuable software. Old SaaS tracked activity: logins, sessions, records created. None of these measure results. Buyers now ask a different question, he says: will this software keep earning its keep, or will its features get copied by others? Tools are easy to build, and the model providers ship many features natively. Scrut’s answer is to design around loops instead of screens.

“When we’re thinking about workflows, we think in terms of loop closures, not in terms of creating a dashboard or generating an answer. What work is getting done? When an agent is generating a certain journey, what context is it feeding on, and is that context reinforcing itself every time the agent completes a loop? That changes how we think about work and how value accrues to the customer: to what extent can the user trust the outcome from the agentic journey, and how does that change over time?” — Aayush

The same thinking changes what the company measures: how much context an agent pulls in, whether jobs complete, and how often a human has to step in and correct something. Cost matters too, enough that the team tracks which parts of a run can move from an Opus-class model down to a Haiku-class one.

The context an account accumulates compounds every time an agent closes a loop and is hard for a competitor to replicate.

“For a user on Scrut in year three, their experience will not be the same as year one, because that context has accumulated over a long time.” — Aayush

“Context” can sound hand-wavy. So, Aayush breaks it into three kinds of traces.

“There are three very specific kinds of context traces.

One is the activities a user carries out on the platform: approving a policy, adding a comment, responding to an audit query, overriding a decision. That explicit action becomes a memory, so that every time an agent invokes it later, it knows what the user’s preferences were.

The second is tests that run in the background, like cron jobs, on a set periodicity, each one generating a trace when it completes.

The third is the reinforcing kind: when an agentic loop invokes part of that context to close reliably, on completion it feeds back into that body of context.” — Aayush

Getting started is easy; breaking out isn’t

Dev has watched startup costs collapse over his career. In 1999 you bought Sun servers and spent millions before a website went live. Today you type a command against a foundation model. Cheaper starts mean more starts, and the pace has gone up with them.

“I remember annual release cycles when I was a product manager 25 years ago. Some people are shipping three times a day now.” — Dev

The bottleneck has moved. Not having an engineering co-founder no longer stops anyone. “There is no longer an excuse to not get started,” Aayush says. A category that used to attract ten startups now attracts a hundred, and the filtering happens later, when quick-and-scrappy has to become production-grade.

“You’ll see many startups get to four or five million in ARR because they built something quick and scrappy. The ones that break out will necessarily have to build the systems that make it reliable.” — Aayush

Whatever wedge you land with, its half-life is now measured in weeks.

“That core wedge you land with commoditizes very quickly. Building the next features isn’t just what customers want; it’s how you stay relevant.” — Dev

Which pushes the advantage back to two unglamorous things: talent and reliability.

At Scrut, the reliability work happens out of customer view.

A team split between the US and India runs reinforcement-learning loops on agent traces. Traces that need better judgment or tone go to working security and GRC practitioners, and their corrections feed back into the models.

That invisible work, more than any feature, is what Scrut is betting the future on.

This write-up is based on Dev Khare’s conversation with Aayush Ghosh Choudhury of Scrut Automation on the Lunchbreak with Lightspeed podcast. Watch the full episode below.

I. The fall of the producer

For most of human history, people were producers. Farmers, blacksmiths, masons, artisans, shopkeepers. Work was personal. Production was distributed. Then came the industrial age, and later, the software age. Factories and software platforms brought efficiencies, but also centralization of supply chains and production. Ownership shifted from individuals to institutions. From millions of tiny businesses to thousands of scaled companies. From producers to employees. From creators to consumers.

In 1900, nearly 50% of Americans were self-employed. By 1977, that number had dropped to just 7%. People stopped running their own shops and started working for Walmart. Or Microsoft. Or Google. This was the era of steady corporate jobs. We traded autonomy for a pension and a gold watch at retirement.

II. The software divide

The internet promised decentralization, but it paradoxically entrenched a few mega-platforms. The new “currency” of production — software — was too complex, too technical, too gated. It required years of training. Those without that training? They simply consumed. Building a business once meant having a product and a storefront. Now it means understanding authentication, cloud infrastructure, security, APIs, database design, and full-stack architecture. You had to be technical, or hire people who were. This was the great shift: not just from local to global, or analog to digital, but from producer to consumer.

Software ate the world. But only a small fraction of humans could write it. Everyone else adapted to using it.

III. AI — when intent meets capability

As people moved away from self-employment to corporate employment, a new implicit trust emerged between employee & employer: I take care of you, you take care of me. When my father started his career in India in the 70s, you often retired from the same job you began. This trust has been slowly eroding over the last half century. When I was starting my career in the 2000s, spending 5–10 years at a company was the norm. Today, people move between companies every few years. Some of the more “evolved” ones are now working 5–6 jobs in parallel :)

This erosion of trust has accelerated since the hyper-digital era of the last 10–15 years, and we are now witnessing a resurgence of self-employment on a global scale. Google searches for “freelancing” and “digital nomads” have skyrocketed since COVID. In the US alone, bootstrapped solo founders are now 39% of all founders, up from just 22% ~10 years ago.

Thanks to the recent advances in AI, we’re now at a rare intersection. Two long arcs of history are converging:

1. Intent: More people than ever want to be independent. They want to freelance, to build, to own, to escape the 9-to-5.

2. Capability: For the first time, AI tools make this possible even if you can’t code.

IV. Why We’re Investing in Emergent

There are plenty of “vibe coding” platforms out there. We’re backing Emergent because it turns ideas into income. Think “Press to Ship” for software. Emergent handles setup, testing, security, compliance, hosting, and scaling. Non-coders go from prompt to live product to revenue without glue work or a bench of engineers.

We believe Emergent has taken some very interesting choices that are showing up in the data.

1. Better token usage = Higher quality code
   One of the most interesting and technically superior choices Emergent has made is to “spend more tokens up front thinking, planning, architecting”. Most AI coding agents aim to minimize compute and response time, often generating code in a single pass with limited validation, but this can sacrifice reliability. Emergent takes the opposite approach: it expends significantly more tokens and agent cycles per task to maximize code quality. Over the last several quarters, Emergent’s code-bug rate has come down from ~30% to <15%. This is much higher than the other tools in the category where >30–40% of the usage tokens are used to fix bugs introduced by the platform. Emergent’s agents perform planning, coding, testing, and even automatic debugging. For example, a QA agent continuously runs tests and patches any failing cases without human prompting. This thorough, multi-agent process yields a much higher reliability score for Emergent’s code output.

In essence, Emergent trades extra tokens (i.e. computing cycles) for higher code reliability and completeness. This is a critical distinction. A conventional AI coding assistant might slap together a web app login page with some placeholder code in a few seconds — impressive, but likely brittle. Emergent’s agentic system might take a bit longer and consume more tolens, but the end result is a fully fleshed-out app that adheres to best practices and actually deploys without falling apart. For a user, that difference is the gap between a throwaway toy and a usable product.

2. Focus on full-stack & production-ready apps vs just prototypes

Most AI coding platforms today sit on the lower-left end of this heatmap: i.e., they minimize compute cost, spitting out code quickly but often with gaps and errors that need manual fixes, which by themselves consume more costs. Emergent deliberately pushes to the high end: it may consume more tokens on a given project, but in return the user gets an application that’s production-ready out of the box. In the long run, this is the winning formula. Users don’t want a pile of AI-generated code they must babysit; they want an app that works. The average code-base on Emergent is ~35K vs ~5K for competing platforms. A typical code-base on Emergent has also complex integrations, backends etc, indicating that a typical Emergent project is far more advanced than a front-end only website, or a prototype.

3. Building an ecosystem for serious builders

One of the clearest signs that Emergent’s ICP is composed of serious builders is the steady increase in the proportion of apps that move from “active” to “production.” Over the past three months, daily deployed apps as a share of daily active apps has risen from roughly 5% to over 20%. This means builders on Emergent aren’t just experimenting or tinkering, they are publishing, shipping, and putting their work into the hands of real users. The platform is becoming less of a sandbox and more of a production backbone, a place where ideas graduate into live businesses at an accelerating rate.

Software creation is no longer the privilege of the few. Just as the iPhone camera turned everyone into a photographer, platforms like Emergent are giving millions the ability to become software producers. The real story isn’t about code, it’s about agency: enabling individuals to move from being passive consumers in the digital economy to active builders of it. Emergent is betting on this shift that the next wave of enduring companies will be built not by large incumbents, but by thousands of serious entrepreneurs who finally have the tools to ship. We are excited and privileged to go on a journey with them together with a world class set of AI angels including the legendary Jeff Dean (Google), alongside Balaji Srinivasan, Devendra Chaplot (Thinking Machines), Edo (Pinecone CEO), Prasanna (ex Rippling founder), and others.

Hemant Mohapatra

Partner, Lightspeed