Your Pricing Unit Decides Your Margin
There are only three units an AI startup can charge in, and they form a ladder: per token, per insight, per outcome. Which one you pick decides whether falling model prices become your margin or your customer's.
I work with founders before there is even an idea.
Co-founder and Managing Partner at Grove Ventures, where I lead first rounds. Taking a vision from zero to product-market fit is a defining craft of our time, and it is the craft I practice, usually from the first conversation.
I write here about where that craft is changing: enterprise AI, infrastructure software, data moats that actually compound, and how company formation works when execution gets cheap and judgment becomes the constraint.
About
I began working with Dov Moran in my twenties and have spent most of my adult life in early-stage investing, at the earliest stages of company building. In 2017 we co-founded Grove Ventures, where I am a Managing Partner. I have incubated multiple companies alongside their founding teams, often starting before the idea existed.
I am drawn to companies that build foundational infrastructure, turn fragmented data into reliable sources of truth, and use that foundation to rethink how decisions are made and how work gets executed. Today that means enterprise AI, infrastructure software, services-as-software, physical AI, and healthcare AI, with a constant curiosity for what comes next.
When I am not at Grove, I am usually exploring the world and its people, creating communities and experiences, parenting (with mixed results), or deep in hands-on AI exploration.
This site is a repository, not a feed. Everything starts as a LinkedIn post; what holds up gets kept here in full, and what keeps coming back gets expanded into essays.
If you are building something that today looks like a point solution and tomorrow becomes a control layer, I would like to hear about it: lotan@grovevc.com
The FDE Paradox
As AI platforms get more advanced, deploying them takes more humans, not fewer. Notes from hosting the executives running Forward Deployed Engineering, from public companies to twenty-person startups.
Stripe Is Buying the Door to AI
Google bought Android for $50 million to own the door to mobile. Stripe bidding $10 billion for OpenRouter is the same move: own the switchboard, then take a cut of every token bill that flows through.
The AI bubble bursts on layer economics
Healthy infrastructure takes a thin slice of the value above it. In AI the slice is bigger than the pie: the model layer collects more than the applications above it earn. The bubble ends the day the software on top can pay for the model underneath and keep a margin.
Service as Software: three company types
Past industrial revolutions made products accessible at the price of standardization. AI delivers personalized service at industrial scale. Three company types are forming; the third builds demand that never existed, and if history repeats, that is where most value gets created.
Show Me the Flywheel
Everyone now agrees data decides the AI era. That consensus is right, and it is only half the picture. Proprietary data is a head start. A moat is data that compounds.
The AI Conversation We're Not Having
A dozen companies own the entire conversation about AI. The most interesting companies being built right now are not on that list. Consensus is priced; hard is the moat.
Diligence your investors
Investors diligence you deeply. You pick a decade-long partner on a few meetings and gut feel. And investors are the hardest thing to replace. We built a tool that flips it: founders running DD on funds, ours included. Information symmetry is healthy for both sides.
Anyone can read the mirror
Now that almost everyone arrives with traction, it no longer marks who is on a venture-scale path. Customers are living in the past. Founders who steer by their feedback are driving by the rearview mirror.
Five ways to build proprietary data
Before a data moat can compound, you have to build the dataset. Five patterns that actually worked: earn it intimately (Navina), go where no one else will (Alice), digitize the physical world (Limitless), structure the public signal (OnFire), create data that never existed (Protai).
Imagine: A Laid-Off Tech Worker Walks Into City Hall
Israel was always the startup nation, world-class at building technology for everyone else. It never became the adoption nation. That gap, which always looked like a weakness, is exactly what makes the layoffs an opportunity.
AI layoffs are a local opportunity
AI driven layoffs in tech are not only a problem. For local economies, they are a real opportunity.
Show me the flywheel
Larry Ellison says data is the resource that decides the AI era.
That is becoming the consensus. It is right, and only half the picture.
Build where SaaS never reached
The markets most at risk are the ones SaaS already conquered, because the bridge to software is built. The opportunity is the huge economy that never went through the wave. Frontier labs are buying the bridge; FDE startups build it themselves. Not competition. Division of labor.
Teramount: constraint before consensus
Data center bandwidth was about to hit a wall and silicon photonics was the unlock, before co-packaged optics was a buzzword. The pattern: identify the constraint before it becomes consensus, back the team that owns the solution.
Lawyers are next: pricing moves to value
When the cost of execution approaches zero, effort-based pricing collapses. Differentiation moves to who creates value and who is willing to be measured on it. It will not stop at lawyers.
Judgment becomes the constraint
AI will not eliminate most roles. It will quietly redefine what good looks like inside them. You will not be measured by hours. You will be measured by the quality of your decisions.
Acquisitions are easy to grasp. Breakthroughs are not.
The market reacts to the acquisition and misses the scientific breakthrough in the same week. Some layers that feel strategic today will commoditize faster than we think. The challenge is not speed. It is the allocation of attention.
The pause as advantage
AI is not another domain. It is a new operating layer, and you cannot truly adopt one while deep in the 24/7. A founder who takes real time between companies for AI bootcamp may return with a different kind of edge.
Four conversations, one insight
It takes time to learn how to save time. The people meant to shape the future are captive to maintaining the present. If something is critical to the future, it needs a recurring slot in the calendar.
Where the SaaS panic is justified, and where it is overdone
Three sources of SaaS advantage, and AI stresses each differently. Panic is justified where advantage rested on momentum. It is overdone for systems of record, network effects, and compounding data. Ask where the advantage comes from, not whether the product is replaceable.
The moment everything clicks
In every ideation process there is a small moment we all wait for, when the picture snaps into focus and the direction forward becomes clear. Accompanying founders to that moment is the heart of the work.
Technical buyers do not buy the way others do
Independent, skeptical, allergic to fluff. The future of GTM is a data layer that replaces opinions with real signals. That thesis is how OnFire was born.
Trust is the key
To turn a service into a product, companies need Palantir's path: product innovation plus organizational and GTM innovation. Every startup here must build a brand an enterprise can trust.
Why We Invested in OnFire
They didn't have a startup yet. They didn't even have a defined idea. What they had was impossible to fake: the kind of team dynamics you only see in teams who have been through real battles together.
Validation when customers live in 2020 and you build for 2030
In the seventies customers wanted better pocket calculators while founders were building Oracle, Microsoft and Apple. Customer feedback matters, and it does not necessarily point forward.
Services Are the Next Software
Turning human-delivered services into software while keeping the human support that lets customers make the transition. The potential dwarfs traditional SaaS, and the barrier is operational, not technological.
Why us. Why now. Future market.
The line runs between companies that are relevant to the AI generation and those that are not. Every founder needs sharp answers to three questions, with proof.
BPOs: the trillion-dollar market waiting for its AI revolution
Companies spend trillions on outsourced services that never went through deep digital transformation. Customers do not want an interface. They want the problem solved.
Four ways to create proprietary data
The pendulum settled on unique data. A data strategy can and should be articulated before the first line of code. Navina's founders refused to write code for a year until theirs was precise.
Building products for the AI era
Keep the familiar interface, put the sophistication under the hood, give users control over how much help they get. And build for a new kind of user: your next customer may not be human.
The Enterprise-AI Startup Playbook
AI is the enabler, not the product. Integration beats technical perfection. Adoption is driven by trust. Proprietary data outlasts the model. The best products create a data flywheel.
The competition moved to the interaction layer
Windsurf and IO in one week. The race is no longer for the smartest model but for the layer through which models get used. Is AI opening technology up, or are we watching fast reconsolidation?
Guardrails: deciding what the product must not do
The model has too many degrees of freedom. Product management flips from defining what to build to deciding what the product must never say, offer, or attempt, and where it stops and says: I do not know.
Laziness is a feature
Every breakthrough started because someone did not want to make the effort. Instead of fighting laziness, listen to it. It points precisely at what needs automation or a paradigm shift.
The end of tribal knowledge
AI can now learn from observing organizational routine, not just from text. Tacit knowledge becomes working infrastructure. The prize is the next giant platform, for every department still run like a tribe instead of a system.
The Enterprise-AI Startup Playbook
Lessons from a decade partnering with AI founders building and selling enterprise solutions: embedding AI where it adds value, driving adoption, and building data strategies that compound.
92% adopted AI. 16% adapted their workflows.
From the Knesset AI subcommittee: juniors are the fastest adopters. Once companies build the right AI workflows, GenAI will not replace juniors. It will turn them into seniors faster.
The six archetypes of SDLC companies
A framework categorizing developer-tooling startups by go-to-market strategy. Identifying your archetype shapes GTM, accelerates growth, and predicts the challenges ahead.
Mapping the Israeli SDLC Landscape
100+ SDLC-focused startups, $5B raised in a decade, and research across dozens of founders and buyers: where budgets actually go, why founders lost faith in PLG, and what sells.
Navina: the ideation journey
Our investment often arrives at the earliest possible stage, before there is a product. The Navina ideation process became the model we have replicated since.
Trust and safety: everything I learned
One accessible resource on the creation and acceleration of the trust and safety industry, written from inside the ActiveFence journey. Kept as a historical market map.
Shift Happens: the state of software infrastructure
Grove's yearly research program on software infrastructure and how AI changes building and delivering software: the 2023 State of SDLC report, the 6 Archetypes framework, and the 2024 study on AI in the development lifecycle.
Everything I write lands here first. If something resonated: