So you whisper goodbye to ol’ Betty the Boiler, rip her out, and replace the giant hole she left in your factory (and heart) with a central electric motor. Same belt and drive configuration, electricity not steam.

And, nothing changes. At least not measurably.

This is a story that’s been told over and over. With trains, fiber, computers, the internet and now, robotics. Which raises the following questions: when does robotics actually arrive? How do they arrive? When do they deliver productivity gains?

Let’s start with the factory you just gutted.

The factory

Factories in the 1800s were powered by a central steam engine. The boiler sat in the middle, and everything ran off it through belts and line shafts. To get bigger, factories had to go up — which cost money and created increasing risk.

Electricity arrived and the initial approach was to replace the core infrastructure. Take out the steam engine, put in one big electric motor, power everything centrally. It may have been slightly more reliable. But the efficiency gains were not there.

The efficiency only came when the motors were separated and moved to the parts of the factory where the work was, group drive, then unit drive. Electric drive motors let you put power where it’s needed instead of running everything off a central hub.

It took roughly 20 years, 1899 to 1919, for factories to figure out how to use the technology they’d already bought. Warren Devine tracked the shift from shafts to wires; Paul David turned it into the productivity paradox.

What robotics changes about the building

With robots in the loop, a building no longer needs to be constrained and conformed to fit the shape of soft, squishy humans: aisle widths, lighting, reach heights, walking speed, etc. Instead robots can work in closer confines and pack together in shapes and configurations humans can’t. Subject to safety standards, sure, but looser than with a live human.

There is an example of this, running on the electricity timeline. Caged arms in automotive and electronics factories. Unimate started in a GM plant way back in 1961 which led to a wholesale change in how a production line was configured. While one could drop an arm into a factory and see marginal improvement, companies realized that the productivity was unlocked via the redesign, not solely the robot itself. And, researchers were able to see a visible uplift: Economists Graetz and Michaels (Robots at Work) found industrial robots contributed ~0.36 percentage points to annual labor productivity growth across seventeen countries, when examining the period 1993 to 2007. That may not sound like a lot but robots drove ~16% of total productivity gains on only 2% of the capital in the industries that used them. Their gains surpassed another powerhouse of industry — the steam engine — in about a quarter of the time.

This is why robots cluster where they do: warehouses, electronics, automotive, indoor agriculture. Those are the places that can afford to rebuild, or already have. Greenhouses are a good example given their high intensity and low variability. Indoor growers have spent decades reconfiguring so human labor could work faster and more efficiently: pipe rails, controlled lighting, standardized rows and crops bred and trained to fruit at consistent intervals and heights. While Philip K. Dick was dreaming of electric sheep, these designers only had humans in mind. Today’s machines however, reap those benefits and run on rails installed so people could push trolleys.

The GM plant rebuild is an example of unit drive — and it’s why productivity showed up there and essentially nowhere else, yet, in robotics. Robots have already shown what it will take to unlock productivity and it took multiple decades and the capital to reconfigure around the technology.

The humanoid bind

Where do humanoids fit into this?

Start with what’s actually being sold, the pitch is that you don’t have to change anything.

Drop it in as a direct proxy for a human. It walks the stairs that are already there, picks up the tool or vacuum that was already designed for a hand and works the bench that was already built for a person. The building, routines and assembly line stay exactly as they are.

That’s the value proposition and it’s frankly appealing. It is also the reason it can’t deliver what it’s being priced on.

Because the lesson of the boiler is that the gains never came from the new technology alone. They came from what got rebuilt around it. A robot sold on the promise that nothing has to change is a robot sold on the promise of minimal productivity gain.

And therein lies the disconnect. Listen to most humanoid company pitches and they will preach about the utility of their designs in warehouses and auto plants — some of the most redesignable buildings in the economy.

Amazon is the proof. They run roughly 750,000 robots and almost none are human-shaped, because as an owner of the building they get to change the building. They acquired and built around Kiva in 2012 making the shelves come to the picker rather than a machine learning to walk the aisles. The humanoid they did evaluate was relegated to tote recycling at an R&D site.

If you are that buyer and you can redesign the workstation, you redesign the workstation. Purpose-built beats general-purpose in any environment you control. That isn’t a prediction, it’s the last hundred years of industrial automation.

Where it’s actually being sold

And, that is where a sector bull will push-back. That purpose-built doesn’t transfer, it is costly and inefficient to lock-in to a design that can’t easily be adjusted. It’s easy to say one should design a workstation but that can cost billions before manufacturing even begins. In an age of rapidly evolving supply chains, generalized humanoids provide adaptable flexibility and that is a very compelling vision. But Apple ran that experiment. A robot, with hands, built to polish the 2013 Mac Pro couldn't be repointed at the Watch two years later, and the Watch went back to human hands. ( at WattFactory uses the same example to make a related argument about why assembly automation so rarely gets reused across products.)

Note what Apple didn’t do. A company with unlimited resources, some of the best minds and manufacturing partnerships, world class facilities and every incentive to automate, did not double down or redevelop the robot. It returned to those nimble human hands.

And, that’s the bar that humanoids have to clear. It’s not just the purpose-built cell — but the available substitute of the person standing at the station already doing the job. That costs a wage and needs some training, but no capex or certification problem. The humanoid pitch is that a machine can take that slot. Apple ran that experiment and ended up taking the person.

The strongest live counterexample is housed in Spartanburg. Figure 02 spent eleven months in BMW’s body shop, working 1,250 total hours between the two deployed robots and contributing to the assembly of more than 30,000 X3s; Figure 03 is there now, sequencing parts into trolleys. That’s real, it’s paid, and it’s likely the most scrutinized humanoid deployment in manufacturing (with some fun and effective livestreaming videos).

But again, look at what it’s competing against. At a stable, high-volume station the alternative isn’t a worker — it’s a purpose-built cell, and those stations were automated decades ago. What’s left are the stations that never justified a bespoke cell: typically lower volumes with frequent changeovers and, awkward geometry. It’s flexible automation for the residual. That’s a real market and possibly a large one. It’s a capital equipment market, though, not a labor market — and the valuations are written against the second one.

The areas out of sight — the fields, construction sites, mines — will require a different modality, once people figure out how to leverage the new technology. We are already seeing this in agriculture, large purpose-built implements are actuating sprays, electricity, steam, lasers and even hot oil to kill weeds quickly and effectively at a rate significantly more efficient than a crew of humans. Humanoids would not be as efficient in this environment. On the industrial side, the humanoid as pitched is Betty’s replacement motor. Bolted into the space where a person used to stand, with every belt and shaft left exactly where it was.

Where the human shape earns its cost

Where the human shape actually earns its cost is the opposite kind of building. The ones nobody can redesign, at least not on a near-term timeline.

The home. Hotels. Hospitality. Hospitals. Healthcare facilities for the elderly. Basically every building that starts with an H. Human designed for human occupancy as the human shape (Ozempic aside) is likely not to change all that much. This also includes the long tail of small operators who will never fund a redesign because the cash is not there.

That’s the real humanoid market. It’s also the hardest environment in robotics, for three reasons that compound:

• Safety. Industrial robots operate inside a defined envelope by design and leverage various mechanisms for safety like fencing, speed-and-separation monitoring, ISO standards and workforce training by dedicated safety officers. Now compare that to the following:

  • A commercial kitchen has a wet floor, fire and knives, a dinner rush, an aggressive head chef and a sixteen-year-old on their third ever shift.

  • Or, a house with a grabby toddler, a crazy cat, and a staircase.

• The certification regime for an unsupervised bipedal machine operating around untrained people isn’t slow. It doesn’t exist. That’s a regulatory clock, and regulatory clocks run slow (like Flash the DMV worker in Zootopia or AV regulations in California for exhibit A).

• Willingness to pay. A warehouse operator underwrites against a fully loaded labor cost with a payback period attached. A household underwrites against a Roomba, which works because a Roomba can run at 3am and nobody’s waiting for it. The jobs people would actually pay a humanoid to do happen at the same times every day, and the machine spends a chunk of those hours plugged into the wall.

  The warehouse can solve this with spare packs, swap stations, charging docks and dedicated fleet software to schedule the rotation. Every one of those is capital the buyer installs — fine when you own the building, and exactly what the household will never fund. It also stops being a 1:1 replacement of a human somewhere in there.

  • Worth noting that the world's demographics are aging as birth rates slow and labor participation drops across the northern hemisphere, Western Europe and Japan especially. That raises the price of labor and with it the cost ceiling a humanoid has to clear, which is genuine air cover. But scarcity doesn't only raise what a buyer will pay, it raises what they'll change. An operator who cannot hire at any wage has every reason to rebuild the workflow, not just buy a machine that promises they won't have to.

• Variability. A warehouse has a countable number of object geometries and a bounded task list. Racking doesn’t change all that much, is recognizable in various different warehouses and is generally stationary. A house (at least mine) is the opposite. No two floor-plans are the same, with a task list and piles of mess that change daily. It’s the hardest generalization problem in the field, and it’s being held as the fallback market.

  • This is where the foundation model argument lives, where Humanoids can shine and it’s the strongest point those bullish on the sector have. If the model generalizes, the environment doesn’t have to. Allowing a machine the ability to handle floor-plans and tasks it’s never seen or been trained on. That’s hard to do but solves this bullet. It doesn’t touch the other two. The certification regime still doesn’t exist, and the largest market, the home, still isn’t paying warehouse prices for a robot that folds laundry.

  • Worth noting where that capability currently sits. The Wall Street Journal reported this week that General Intuition — the most valuable lab in this space at over $6B — can pilot quadrupeds, wheeled vehicles and drones, but that its approach excludes most bipeds.

So the environments that can be redesigned don’t need a humanoid. And the environments that need one can’t easily be redesigned, can’t be certified in the near-term, and are cash constrained.

The bind

Which produces a bind that has nothing to do with engineering.

The valuations are underwritten on general purpose. One machine, every environment, a TAM that is some share of global labor. If a company narrows that honestly — we just handle totes, or we’re solely an elder care company a decade from now — the TAM contracts. At those numbers a TAM contraction is a down round which brings a particular sort of pain to all involved.

And with the recent rounds and valuations starting to skyrocket, the hype is real and accelerating, which leaves minimal room to shrink the scope of aspirations. And, named partners read as a who’s who in the automotive and electronics sectors: Toyota, BMW, Schaeffler, Foxconn, Mercedes-Benz, Jabil, etc. Not to mention Hyundai recently in-housing Boston Dynamics, Rivian spinning off Mind Robotics and Tesla ramping up Optimus production.

My belief is that the machines arrive roughly on schedule. The cap tables don’t. This is not a technology problem. The technology is real, and what isn't yet is solvable. What's mistimed is the capital against the actual addressable market.

The category as a whole shipped roughly 13,000 humanoids in 2025. Chinese firms accounted for nearly 90% of them — AgiBot at 5,168 units and 39% global share, Unitree 5,215 with UBTECH and Leju taking most of the rest. Set that against $47.4B of venture funding into physical AI in H1 2026, up 80% year over year and more than the entire 2022–2024 period combined. That’s about $3.6M of capital raised for every humanoid that shipped in 2025.

The underlying valuations were all private market driven until Unitree went public. It closed its debut up 460%, at a $51B market value on roughly $250M of revenue. That’s two hundred times trailing revenue, for the only company in the category with a publicly determined price. Lex, in the Financial Times, reached the same verdict this week: Unitree would need a ninefold increase in output, at current unit prices, just to bring its enterprise-value-to-sales multiple in line with Tesla’s — a company not generally accused of trading cheaply. Its humanoids, per the Wall Street Journal, are mainly landing in education and entertainment (if it’s on TV, it’s likely Unitree), or as guinea pigs in robotics labs. The manufacturing application has been sparse so far.

And, adding to the cycle, the forecasters are running ahead of that. Morgan Stanley expects 50,000 humanoid shipments in China this year (~5x prior year and triple what Morgan Stanley had projected in January 2026) and more than 440,000 by 2030. Tesla has told Optimus suppliers to support 8,000 units a month by December. I think it would be foolish to doubt the manufacturing capability — China installed roughly 276,000 industrial robots last year, more than Japan, the US, Germany and South Korea combined. But, building them was never going to be the constraint.

I’ve watched a smaller version of this play out in wave 1 field robotics. Companies raised for scale, discovered they had a market depth problem, and couldn’t reprice honestly without wiping out everyone who had funded the first story. They also ended up pursuing all sorts of failed experiments to manifest the market size that would live up to the capital raised, burning time and capital. Same trap, a few zeros smaller.

When you are carrying a multi-billion dollar valuation and raising hundreds of millions as an early-stage company, you can’t suddenly announce your TAM is much more limited. The capital is priced for — and thus demands — a market that is unlikely to be available. The pivot that would save the technology is exactly the one the cap table can’t absorb.

The capital argument

Robotics is an expensive investment. It takes time, money and specialized vision to implement properly, and the shape of that spend is what Brynjolfsson, Rock and Syverson named the J-curve. You pay up front to install and train with the effect that the company looks less productive at exactly the moment it’s becoming more productive. Then it inverts and everyone who is left looks like geniuses.

Which means the absence of measurable gains tells you nothing. Not whether it’s working, not whether it’s failing. Only that the rebuild isn’t finished.

Now put a term sheet on top of that.

If the redesign is the twenty-year part, then capital underwritten against the redesigned factory is priced against a clock the company doesn’t control. Product-market fit doesn’t catch this. The product can work and the customer can want it, but the money can still be wrong.

That’s a different fit, and it’s the one I keep watching companies fail. Financial-market fit: whether the capital structure matches the commercialization clock. It breaks in two directions.

Bubble: capital running faster than commercialization. Priced on the terminal state, operating in the intermediate one. Humanoids now. AVs in ~2015. Ag, mining and construction Wave 1. Ends in significant write-offs.

Starvation: commercialization running faster than capital. The machine works, value is proven and the customer is ready, but there’s no funding instrument that matches. You see this in certain sectors right now, typically near the trough of disillusionment. Agtech is one of them.

Where I’d be wrong

The first repricing in this category won’t come from a technical failure (unless that’s a robot falling on a squishy human or pet). It’ll come from a Series D that needs the general-purpose story intact to clear the last round, at a company whose only real revenue is shipping into manufacturing pilots. The same companies that will need to clear several compliance and functional safety hurdles before starting sustained trials in the home.

I’m wrong if, by the end of 2028, a humanoid company is shipping thousands of units a year outside automotive and electronics manufacturing at positive gross margin, to buyers (outside of the cap table) purchasing against a labor budget rather than an R&D budget — and those buyers place a second order at list price after the first cohort has run a full payback period. LOIs, frameworks and promises don’t count, nor do sales to research institutions.

There’s a second way I’m wrong, and I believe it’s the more interesting one that shows stronger conviction. If a major manufacturer builds a facility designed around humanoid machines from the ground up — and then keeps buying rather than replacing them with something purpose-built — I’ve misread this badly.

While I am prognosticating, I’ll take a turn arguing the other side. The diffusion I describe below (cheaper components, proliferated talent) when coupled with world models is an argument that this redesign clock runs faster than electricity’s did. The WSJ recently quoted one of General Intuition’s investors who put the field at GPT-2 stage. GPT-2 to GPT-4 took about four years. If world models follow anything like that curve, the generalization problem I just described gets solved on a timeline my falsifier can't outrun. On the cost curve, Unitree’s IPO prospectus highlighted an average sales price dropping ~70% but gross margin increasing ~16 points. Everything upstream got cheaper as they designed and built their own core components and margin benefited accordingly. That cuts against my argument, and I hope it’s right.

Robotics has had its own wipeouts already

None of this is hypothetical. The sector has run the bubble version several times already as the hype arrives before the impact.

Inside the cage, robotics has arrived — the arms in the automotive and electronics plants are doing real work at real scale. Outside it, on your floors, in fields, on the sidewalks and the streets, the machines have been mostly missing in action despite tens if not hundreds of billions going into the sector.

Some of that investment has already seen its wipeout. AVs boomed in the early 2010s on the back of Waymo’s early success and interest from GM, Ford, Uber. It culminated in the wind-down of Cruise, which had taken on ~$10B in capital from GM. Today, Waymo is still here, scaling what might be the most expensive product roll-out in history.

Extend it into other sectors. Ag robotics — hundreds of millions, if not billions, with most companies ending up as failures. Construction robotics. Mining robotics. Rinse and repeat across sectors, and you see a pretty similar story of where the capital went and what it delivered over the past decade.

Short answer: not much.

What gets left behind

None of that makes the money wasted. It might make it early, which is a different thing.

Fiber is the cleanest case. In the 1990s enormous debt was raised on the belief that internet traffic was doubling every 100 days, and the race was on to own the pipe. The bubble popped and took WorldCom, Global Crossing, Williams Communications and 360networks with it. But fiber didn’t disappear, instead it stayed in the ground. Level 3 and others picked it up for pennies and built the infrastructure that made the modern internet and cloud computing possible. Railroads, a century earlier, are the same story with steel instead of glass.

The dot-com boom left something less tangible. Most of what it laid was virtual — know-how, and the exposure of what could come to pass (like cat videos). Those companies mostly failed and failed hard, but they diffused the knowledge, and the wave that followed did more with less because the inputs had gotten cheaper.

In every case the capital chased momentum while the commercialization clock ran slower. The second wave inherited the assets at a cost basis the first wave could never have made work.

So what does robotics leave?

Not the passive assets. There are no significant robotics factories being built that can be easily co-opted by another company. Supply chain centralization and standardization is still in its infancy. What is being laid down is the framework, and the diffusion of knowledge and capability.

I saw this firsthand in AVs. When we were looking to hire a perception engineer, these were incredibly difficult roles to fill — unique, hyper-specialized, and incredibly expensive. We’d interview a stellar candidate only to ask, with some trepidation, where else they were interviewing. Often if Waymo, Cruise, Argo, Uber were mentioned we knew it was a lost cause as their offers were often multiples higher than what Voyage, a relatively well-funded startup, could afford. The scarcity of these key hires slowed our progress and made things significantly more expensive.

Now perception engineers are a dime a dozen. With AI and world models you’re seeing a significant reduction in training and computer vision costs. The inputs and the impediments to scale have been significantly reduced.

This all speeds up time to adoption. And crucially, it cheapens it.

But we do need to keep in mind that diffusion is a real transfer and it is a lossier one than fiber. Fiber doesn’t depreciate in function — fifteen years later and it still works. A dataset has an expiration date.

A purpose-built humanoid manufacturing line has a very thin buyer pool. Components survive and get cheaper while people’s knowledge survives and disperses. Most of the rest evaporates, and it evaporates hardest exactly where the money is going right now.

The part that no one underwrites

From my seat, things are measurably more advanced than even five years ago. Talent has proliferated with AI as a massive accelerant. Financing is becoming more readily available — seven-figure equipment lines are getting done with lenders underwriting the contracts rather than the collateral, betting the machines stay useful and stay in the field. That’s real progress and the most sophisticated capital this sector has ever had. But, one more parable in closing.

The promise of AV was that it enters a human driver’s world, a world where nothing changes. Same roads, same signage, same parking, same everything except look ma no driver.

I modeled that extensively. The machine cost more per mile than a driver in a Crown Vic. A sensor stack, compute, a remote operator and a service depot are a worse cost structure than a person who already owns the car — even with the uptime a human can’t match. AV starts to pencil when it moves away from 1:1 driver replacement and toward something bus-like, with fixed or semi-fixed routes. It pencils when something gets rebuilt around it.

Drop-in replacement for a world that doesn’t need to change, that’s the humanoid pitch with a fifteen-year head start and billions behind it. It still hasn’t penciled.

Most of that rebuild doesn’t run on your books. The buyer pays for it: the fixturing, the retraining, the changed sequence of operations, the reconfiguration. None of it appears in a robotics company’s use of funds. Every cap table above is priced on the machine. Nobody is pricing the rebuild, and the rebuild is where the productivity lives.

Robotics arrives the way electricity did. Not when the machine works, but when enough people have rebuilt around it that you can finally see it in the numbers. Betty came out in an afternoon. The factory took twenty years.

Everyone’s underwriting the afternoon.

Subscribe now

Disclaimer: I work with companies selling into this supply chain.

Daniel Kirstein is the founder of Holdfast Partners, a financial architecture and commercialization consultancy for robotics and hardtech founders. He was head of finance at Voyage, the autonomous vehicle company acquired by Cruise, and later CFO of FarmWise through its commercial launch — firsthand reps in the gap between prototype and scale, including hardware raises from Seed through Series B. Holdfast Partners: hlfst.com.

In June (2026) I wrote about the missing layer in robotics: the tier-one supplier that every mature hardware industry eventually produces, and that this one still hasn’t. Bosch didn’t build cars, he built the part without which no car worked. That part is what let the industry mature.

Three months on, the premise is holding but my prediction isn’t.

Aigen announced Alchemy, a generative world model for the open field - any crop, any weed, any weather, any growth stage - with a claim of reaching a new crop or region in under a week, trained on 99%+ synthetic data. If that holds, the data-collection moat protecting every incumbent in ag robotics just got a lot shorter. Carbon has its Large Plant Model. Bonsai keeps talking about moving from iron to software.

This is the standardization layer, the software version of it at least, and it’s forming.

The catch is that it’s forming inside operating companies that build their own machines, which is the opposite of what made Bosch work. Aigen’s own announcement mentions a Fortune 500 partner already using Alchemy, and that agriculture is just the beginning.

~15 minutes on that, plus the 80/20 problem in a robotics BOM, why collecting lettuce images is worse than it sounds (and why AIGEN’s world model is so interesting), and a well-capitalised industrial manufacturer that’s actively looking at this slot.

Flagging there is an argument I wandered into and didn’t finish: the mid-tier OEM path, and why it might be the most realistic exit left for over-capitalized robotics companies.

The original piece from June 2026.

Spotify link:

Subscribe now

Daniel Kirstein is the founder of Holdfast Partners, a financial architecture and commercialization consultancy for robotics and hardtech founders.

The machines didn’t drive themselves the way the marketing said they would. The dealer subsequently sued in federal court. Monarch Tractor had raised more money than any ag robotics company on record, and this was one of several threads that unraveled on the way to bankruptcy.

Monarch had moved to a sales model before the machine could stand alone.

Call it the Ford Engineer rule: when you drive a car off the lot, would a Ford engineer riding in the back seat — just to monitor things — be a welcome surprise? Or would it tell you something was not quite right with the product you just bought?

For most of us it’s the second one. Hardware you purchase should work, in a range of situations you didn’t describe to the salesperson, without direct oversight from the people who built it.

A number of robotics companies are failing that test right now. They’re shipping machines that still require constant supervision and babysitting, and calling it a sales model. It isn’t. They are aiming for the hardware model but landing on the engineer-required sales model. ERSM or ER for short which feels appropriate because that’s where companies that inadvertently deploy this model end up. This model has the worst economics of both worlds.

Last month I wrote about why RaaS is where most robotics companies start, and why FarmWise eventually had to leave it. This piece is about where a lot of them go next. And, what has to be true before you get there.

A definition, briefly

Hardware sales is the literal exchange of cash for a piece of equipment. Give me that and I’ll give you this. I’d like to believe it’s the oldest model of exchange there is, and it’s the natural evolution point from RaaS.

It has two close cousins — Hardware as a Service and Leasing — both out of scope for today. Not quite the same thing, but rhyming, and they rhyme most on the point that matters here: the customer takes possession, so the machine has to survive without you.

The model itself is simple. It’s the standard buying approach in most CAPEX-heavy industries. It’s understood by farmers and foremen. It’s what happens when you buy a car — test drive, exchange money, drive off the lot (though over 6+ hours and with an upsold tire warranty if you are lucky). The majority of capital equipment on earth gets sold this way. You’re not reinventing anything, and that’s part of the appeal.

What you get

Cash flow. RaaS pays you across the life of the engagement — five-plus years, with churn risk running the entire time. Over the full customer lifetime you likely earn more. It still took five years. A sale recovers in weeks what service recovers in seasons.

Bolt-on monetization. What you sell will need to be serviced, and who better to service it than you. Parts and support form a second revenue channel, occasionally via a durable subscription that’s often financeable. And with software and machine learning models increasingly central to the machine, there’s an opening to monetize the software layer that didn’t really exist a few years ago. You can dial it — lower the upfront and raise the residual, or the reverse — and tune the post-sale offering to what the customer actually values.

Capital intensity. Service means building infrastructure everywhere you go — crews, trucks, warehouses, all of it standing up before the first customer pays. Sales still needs a team and post-sale support, but you can lean on infrastructure that already exists. Dealers, channels, financing partners. The irony is that you’re selling expensive CAPEX, and the selling itself can be asset-light if it’s financed right.

Utilization. An owner-operator gets more out of a machine than a service provider ever could. The service provider has to haul it, get access, coordinate as a third-party vendor, and eat every gap in the schedule. The owner leaves the machine on site without asking anyone’s permission. They know the operation because they built it, so they know exactly where the machine slots in and what’s coming next week. The provider driving two hours each way with a trailer can’t compete with that. Owners also tend to experiment more and push the machine into areas that may not be in the initial remit. Excellent for discovery, not so excellent if the machine isn’t ready.

What has to be true first

The machine stands alone. This is the precondition, not a nice-to-have. Companies take the leap with a machine that still needs three engineers in the field, ongoing guidance, a standing weekly site visit. That machine doesn’t get cheaper as you sell more of it — it gets more expensive, because every unit shipped adds a support obligation you priced at zero. You’ve taken on the capital burden of the sales model and kept the operational burden of the service model. That’s the worst of both.

The unsexy but mandatory stuff. Documentation and user guides. Safety clearances and the right stickers in the right places. Customer Q&As. A support function that answers the phone — or the text, or chat — and tele-operations if the machine needs it. None of this is the exciting stuff (to most people) and all of it has to be in place before the first unit ships, which means it has to be staffed or contracted while you’re still selling. This is a short-term unit economics killer and worth modeling the ramp honestly.

Somebody can fix it. This is the one most robotics companies underestimate. If you sell through a channel, the channel has to be able to service what it sold. Large OEMs’ dealer networks are the public benchmark for what that costs and how long it takes to build — decades of parts depots, trained technicians, and a service reputation that predates most of the equipment sitting on the lot. Whether it’s Caterpillar, John Deere or Toyota, each has invested years and significant sums to ensure the health of their respective networks. You can’t buy that in a quarter. And, if you’re not building it yourself or partnering into it, you’ve just distributed a support obligation across a geography that requires an outsized team and cost.

The second clock

We’ve talked a lot about how one of the biggest benefits of hardware is that one-time large check. What’s often not mentioned is the time to get that check is significantly longer than in RaaS. That second clock is one that surprises most founders, especially those who thought they’d be closing deals within 60 days.

Getting a grower to pay $250 to weed an acre is a fast conversation. Getting the same grower to write a check for a $500,000 machine they’re not yet sure how they’d use is an entirely different one. The deal you thought would close in a month often takes six or more.

The larger the check typically means the greater number of conversations across an organization. You are no longer just dealing with the operators that will use it every day, though they matter and are often your most vocal advocates. Instead you’ll be speaking with their boss, the CFO, the CEO. Each of whom needs a different pitch as they worry about different things and all of which take time. Will it pencil and deep dive on unit economics compared to incumbents → CFO. Operational reliability and interchangeability → Foreman. Ease of use, training and overall utility → On-site operator. All very different people with different goals.

And then, just as you get close, the budget arrives. Is this a CapEx or OpEx expenditure? Where is the customer in their fiscal year? If they budget annually and your machine wasn’t in this year’s plan, you could be waiting nine or ten months for the next window regardless of how well the demo went.

The cash arrives much faster than in service. The first cash, however, arrives much later. Both of those are true at the same time, and the gap between them is where the model either works or quietly kills you.

In closing

Hardware sales can be a game-changing model. It’s tried and true, it’s understood by the customer, and you’re not inventing anything by design.

But it demands a foundation set well before you use it — a machine that works unattended, an organization ready to support it at a distance, and enough capital to survive a sales cycle that runs on the customer’s calendar rather than yours.

Don’t ship the engineer with the car.

Subscribe now

Daniel Kirstein is the founder of Holdfast Partners, a financial architecture and commercialization consultancy for robotics and hardtech founders. He was head of finance at Voyage, the autonomous vehicle company acquired by Cruise, and later CFO of FarmWise through its commercial launch — firsthand reps in the gap between prototype and scale, including hardware raises from Seed through Series B. Holdfast Partners: hlfst.com.

The obvious next move, of course, was to double down. More machines, more crops in more markets. Every instinct in the building said run the play again, but bigger with an eventual international expansion.

I built that model. But the numbers wouldn’t cooperate.

A quick description of the terms first, because they get used loosely. RaaS = robots-as-a-service which means you own the machines, operate them, and charge for the output: the acre weeded or sprayed, the pallet moved, the ride completed. Waymo is a RaaS company. So was FarmWise.

Hardware sales is the older model: sell the machine, collect the check, and then monetize support with parts at a premium. Leasing and hardware-as-a-service sit in between — the customer subscribes to a capability while someone else holds the asset. We’ll come back to these two in a bit.

RaaS is where most robotics companies start, and for good reason. It matches the maturity of the technology. Early machines need babysitting, and the customer doesn’t want to, or can’t, own that problem — they want the outcome. The customer is also likely skeptical of the value proposition, or in other words, if it will even work. You keep the machine close, you learn what actually breaks, the customer gets educated and you get paid while doing it.

But look at what the model demands as you scale. To stand up a new market, you buy the machines, hire and house the crews, rent the warehouse, build the service infrastructure. Often, you end up with more Ford F-150s than a Ford dealership. You’re hundreds of thousands of dollars in the hole — sometimes millions — before the first customer pays you anything. And the payback arrives season by season, job by job — while a small army of account managers work the phones daily, scrapping for acres one grower at a time.

The ongoing cost, both startup CAPEX and ongoing OPEX, is an important aspect of assessing RaaS and for some companies this pencils. However, the deeper constraint is density. A service model lives or dies on concentration and its accompanying metric — utilization. Without enough available demand to keep your robots and operators busy, you end up carrying costs that can quickly shorten your runway. When we mapped the specialty agriculture market honestly, we determined that maybe four or five markets in the country had the density to sustain a profitable service at the scale we needed. Everywhere else ranged between insufficient and manageable but with returns that would never pay back inside the timeline our capital could tolerate.

That last clause is the whole problem. Not “never pay back” — RaaS was (and for many is) a good business. But, never pay back in time.

So we modeled the alternative. Sell the machines and share the risk with the grower. Lean on channels to reach markets we could never justify staffing ourselves. This also allowed for the creation of a post-sale package — support, software, maintenance — that turned a one-time sale into recurring revenue.

And the numbers both reassured and surprised us.

Cash flow transformed. A hefty upfront sale recovered in weeks what the service model recovered in seasons. Expansion became asset-light (a VC favorite term). And the post-sale ARR gave investors and bankers something they knew how to price.

But here’s the part that made it a genuinely hard decision, and the part most founders miss: the RaaS model was projected to earn more money. Run both models over the life of a machine and service revenue wins on Lifetime Value (LTV). The customer relationship runs longer and you capture the full value of every acre.

We didn’t switch because RaaS earned less. We switched because it earned less soon enough.

The switch wasn’t free though. A machine you operate yourself and a machine you hand to a customer are different products. Selling a robot means developing training programs, systems to track spare parts and logistics, and pricing a support tier. And, spending the time to build the often over-looked aspects like manuals and safety stickers. It also means that you possess, or can develop, a machine that is reliable enough to survive without its makers standing next to it (which is no minor feat).

Worth noting that selling the machines themselves required building new muscle that we didn’t possess before. Selling service capacity is very different from selling a big-ticket item in a highly competitive space with multiple substitution options. When you sell a valuable service, and build a reputation, customers often pull you in their direction. When you have to convince someone to sign a (very) large check on new technology — even if it pencils and the demos go well — these conversations can go significantly longer than anticipated and run into months if not years.

But it worked — machines sold and markets opened that we could never have profitably staffed in the time required. The model now fit the capital but that didn’t mean the company was suddenly on easy street. The pivot bought time and options, it didn’t erase the clock.

FarmWise had raised venture capital, and venture capital comes with a clock. A model that maximizes lifetime value while starving the middle years is the wrong model for that capital — no matter what the bottom-right cell of the spreadsheet says. The intuition in the building wasn’t wrong about the business. It was wrong about the money.

And, that’s the thing to keep in mind about pricing models. Often, we treat pricing as a market question — what the market can afford, what your competition is charging and/or what value you are providing to a customer. It is all of that — the art of pricing is never done in a vacuum — but the piece that is often neglected is the capital element. What your balance sheet can actually afford to carry and for how long.

Back to HaaS and Leasing. Back when we were assessing go-to-market options, there were really just the two options — RaaS and Sales — to pick from. Post-sales monetization on the Sales model was one of the only ways to build recurring revenue. But even post-sales monetization was an uphill battle for software and model enhancements. Today, however the environment is changing for the better. Hardware as a Service (HaaS) and Leasing are becoming more prevalent as the financing options to support them, and cover the working capital gap they create, come online. Customers have also now spent almost a decade with robotics in most sectors and understand the intricacies of the hardware and software blend, and that these machines are not iron alone. More options allow for a selection of the business model that best suits your business, state of product readiness and capital stack. It also gives the customer more options ultimately leading to a greater value proposition that best fits their needs.

The discipline of choosing correctly, though, isn’t going anywhere.

Most founders choose the model that earns the most. The survivors choose the model their capital can finish.

Subscribe now

Daniel Kirstein is the founder of Holdfast Partners, a financial architecture and commercialization consultancy for robotics and hardtech founders. He was head of finance at Voyage, the autonomous vehicle company acquired by Cruise, and later CFO of FarmWise through its commercial launch — firsthand reps in the gap between prototype and scale, including hardware raises from Seed through Series B. Holdfast Partners: hlfst.com.

Each of these companies, and the many others of their ilk, had the all-important product-market fit by any reasonable definition. Customers paid, machines worked. They died anyway.

The robots didn’t fail. The money did.

That gap has a name: Financial-Market Fit — the alignment between a company’s capital structure and its commercialization clock.

Product-market fit asks whether the market wants what you built. Financial-market fit asks whether your capital can survive that market’s clock. Whether your go-to-market model, be it Robots as a Service (RaaS), Sales or HaaS/Leasing, can survive the capital constraints inherent with each and the capital clock to the next raise, debt round or profitability.

In software, the two questions collapse into one, because the clock is short — the capital that funds the product lives long enough to see it sell and the ramp speed can be near infinite. In hardware, they diverge. Seasonality, field trials, trust built over years, supply chains and service infrastructure that has to exist before scale does — the market’s clock runs on physics and biology, and capital raised on a software calendar runs out before it. That divergence is where most of the graveyard lives.

At FarmWise we argued for months on the merits of tackling row crops (corn, soy, wheat, cotton) vs the known and high-margin specialty crops (leafy greens, broccoli, celery). The push to accelerate progress, to widen the TAM at the cost of distraction from what was working now. Each of these questions has multiple answers, but all are faced with a capital clock.

A corollary that runs through everything here: capital raises can be a blessing or a curse. They can fuel the next stage of growth, or start a clock that a company races against and often fails to beat. A fundraise is often a narrowing of options, a course set before you know where it goes. Capital that follows proof compounds it. Capital that precedes proof starts a clock.

I write from personal experience within these companies. I was head of finance at Voyage, an autonomous vehicle company, and CFO/Head of Sales at FarmWise, an ag robotics company — through the raises, the commercial launches, pivots and the specific experience of watching a venture timeline try to coexist with a physical one. Today I run Holdfast Partners, working as CFO inside robotics and hardtech companies navigating the same gap.

If you’re new, start with these:

The Robots Didn’t Fail. The Capital and the Commercialization Did. A deep dive of 50+ VC-backed ag robotics companies tracked from 2011 to today — who exited, who died, who’s still building, and the why. This is a keystone article and a good place to start.

Is Agriculture Big Enough for AgTech? A runner up on best place to start. I examine why a trillion-dollar market dissolves into hundreds of micro-markets up close, and over the series, examine the five distinct paths available to companies building in it. It’s a parable about truly understanding your market and your customer.

The Best Go-to-Market Beats the Best Technology. Voyage, and why scoping the problem — the environment, the reachable market, the model that matches your maturity — can decide more than the machine does.

Raising for Robotics: The Greed vs. Fear Equation. A more practical article that highlights the structure and philosophy I use for fundraising. I will continue to refine and update as I support additional rounds in the future.

Subscribe now

The next piece in the commercialization series covers pricing models — why FarmWise walked away from the model that earned the most money. After that: in October, I’m publishing the State of Ag Robotics, a full-outcome analysis built on the dataset behind these essays. Subscribers get it first.

So, if you are building or funding this space, these articles are for you and I hope you find them helpful and share some of your stories with me.

Along the way, they acquired the AV company where I ran finance.

In December 2024, GM pulled the plug on the robotaxi business entirely — more than a decade of work, one of the largest engineering organizations ever pointed at a single problem, folded into driver-assistance features. The technology largely worked. The cars drove themselves through San Francisco every night.

Waymo is the survivor of the once hyper competitive AV sector, and the survival is instructive in a different way. Alphabet has spent almost two decades on the problem, and Waymo has raised more than $27 billion along the way — much of it from Alphabet itself, on top of years of internal spending. The rollout has just reached a dozen cities, arriving years apart. It currently delivers roughly half a million trips a week as of this July, impressive until you compare it to the ~290 million weekly combined passenger and food trips that are completed on Uber. This may be the most expensive, slowest product launch in the history of transportation. It’s also, notably, still not a car you can buy, nor one you can reliably hail.

Neither company had a technology problem.

The company I was at made a different bet. Voyage was a small autonomous vehicle startup — this was about eight years ago, when AV technology was still raw, and Waymo and Cruise were attacking the hardest version of the problem head-on, with bottomless checkbooks, in the densest cities they could find. A problem that didn’t pencil for Cruise and that Waymo is now grappling with. Voyage did something different. We didn’t try to close the technology gap. We made it irrelevant by changing the problem.

Here’s what that means in practice.

An autonomous vehicle loses its sensors at 70 miles per hour on the freeway. It can’t see. It can’t safely stop. That’s a catastrophe.

The same failure on a 25-mph street in an enclosed community is a bad afternoon. You pull over.

That difference — not the technology, the environment — is the entire reason I think the best go-to-market beats the best technology almost every time.

Because here’s the truth about autonomous driving: the base problem — getting a vehicle from A to B on its own — has more or less been solved. Multiple companies have been able to get a vehicle, van, delivery box or truck to drive itself autonomously to its destination. However, as the scope of the environment expands, so does the length of the long tail of adverse outcomes. It’s this long tail — the thousands of rare edge cases that are hard to fully predict but must go right every time a human being is on board — that consumed Cruise and that Waymo is still paying down. It’s where the delays come from, and where the majority of the risk and engineering now lives.

And the long tail shrinks dramatically when you choose where you operate.

Voyage targeted enclosed retirement communities — places like The Villages. Twenty-five miles an hour. Few stoplights, fewer stop signs, predictable road patterns. Mostly golf-carts for company on the road. And the customer was built for it: seniors who could no longer drive, who needed mobility to keep living the way they wanted to. The use case was simpler and the rider was waiting.

That’s not a technological achievement. That’s a go-to-market thesis.

I’ll be honest about how it ended. We were commercially live for roughly two weeks when COVID locked down every retirement community in the country. The gates were closed and so was the market. Voyage never got to fully run the experiment. The company found a soft landing at Cruise in 2021. I’d moved on just before the deal closed.

So I’m not going to tell you the go-to-market won. It didn’t get the chance. What I can tell you, because I was the one running the numbers on whether it could work, is that the scoping was sound. And the logic of that scoping is what’s traveled with me into every sector since.

My job was to make that thesis economical — and this is where it gets interesting.

It’s easy to pull up a list of the top ten retirement communities in America and say “that’s our expansion plan.” But run the numbers and the story falls apart. Every community demands its own operating center, its own staff, its own fixed cost. Standalone, most of them don’t pencil. So we moved to a hub-and-spoke model: anchor on one core community, then bolt on nearby spokes that could share the operational base. Communities that were unviable on their own became prized partners once they fell under an efficient operating footprint. Density paid a second dividend: assets stopped being stranded. A vehicle, a technician, a spare part could rebalance to the next community in an afternoon — try that across three states.

That’s the difference between the market on your slide and the market you can actually win today — between TAM and SOM. Founders love a giant number and a promise to capture all of it. Then they spray across geographies, burn capital on a dozen false starts, and stall. It’s even more punishing in robotics, because you can’t scale a robot without dragging your support functions along behind it. You’re not shipping a software update overnight. Your variable costs follow you into every new market.

Which is exactly why Voyage clustered. And it’s exactly why Waymo still does today — Phoenix first, then San Francisco, then Los Angeles, each city a multi-year commitment before the next one opens. The most capitalized player in the industry, nearly two decades in, runs the same geographic-density playbook a fifty-person startup ran in a retirement community. Not because someone told them to. Because the unit economics leave no other option. Geographic density isn’t a marketing choice. It’s a physics constraint on the business.

There was a second decision underneath the first one: the model had to match where the technology actually was. Voyage was never going to sell cars to seniors. The maturity wasn’t there, and the customer didn’t want to own the problem — they wanted the ride. A service model was the only honest fit. Even Waymo, years and billions later, still isn’t selling vehicles. They’re a RaaS company, because selling would demand a level of technological maturity nobody has delivered yet. The model has to fit where the technology is, not where the pitch deck says it’ll be.

Neither of those decisions was about building a better car. Both were about scoping the problem so the company could actually survive solving it.

Which brings me back to Cruise.

Cruise didn’t die of bad engineering. It died because the version of the problem it chose — dense cities, full generality, everything at once — had a price tag that even General Motors eventually refused to keep paying. The scope was the burn rate. Waymo chose the same scope and survived it, but only because Alphabet is one of the few institutions on earth that can underwrite multiple decades of losses waiting for the long tail to shrink. That’s not a strategy most companies can borrow. It’s a balance sheet most companies can’t.

And Voyage — Voyage got the scoping right and still didn’t make it. Two weeks of commercial life, then a pandemic deleted the customer base. The biggest checkbook in the industry wasn’t sufficient. The smartest scope wasn’t sufficient either. What stopped Voyage wasn’t the scope. The capital structure had no slack for a clock that stopped. Cruise had ten billion dollars aimed at a clock it couldn’t compress. Waymo is still on its clock — but someone else is paying for the time.

Three companies, three answers to the same question — whether the money you raised can survive the clock you’re actually on. Call it Financial-Market Fit. Product-market fit tells you the market wants it. Financial-Market Fit tells you whether you’ll still be there when they’re ready to buy.

Scoping the problem correctly is necessary. It isn’t a guarantee. That’s the uncomfortable part nobody puts in the pitch deck. What it buys you is a real chance — a burn rate you can survive, a market you can actually reach, a model your technology can honestly deliver. Cruise spent ten billion dollars and never bought that chance.

Scoping is the first of a handful of decisions that separate the companies that survive the gap from the ones that die in it. I’ll lay out the rest over the next few weeks.

For now: most founders are pitching a TAM their hardware can’t actually reach. Are you scoped to the market you can win, or the one on the slide?

Subscribe now

Daniel Kirstein is the founder of Holdfast Partners, a financial architecture and commercialization consultancy for robotics and hardtech founders. He was head of finance at Voyage, the autonomous vehicle company acquired by Cruise, and later CFO of an ag robotics company through its commercial launch — firsthand reps in the gap between prototype and scale, including hardware raises from Seed through Series B. Holdfast Partners: hlfst.com.

If you don’t know them: Bending Spoons is a Milan-based serial acquirer. They buy established but stagnating digital businesses — Evernote, AOL, Vimeo, Eventbrite, WeTransfer, Meetup — strip out cost, rebuild the technology, reprice, and run them for cash. Then they use that cash to buy the next one. They started thirteen years ago with roughly $40,000 and small mobile apps, and compounded from there. Revenue was $1.3 billion in 2025, nearly double the year before.

The part that matters most for this piece is the part that gets the least attention: they never sell. This isn’t private equity with a five-year hold and an exit committee. It’s permanent capital. Every acquisition is a cash engine bolted onto the machine forever, funding the next acquisition. The snowball is the business.

Watching Bending Spoons over the years, and in particular their IPO pop during a depressed cycle for Software, I started wondering, could this aggregator model work here, in robotics? There are entire sub-sectors where a dozen companies solved the same problem in slightly different ways — agtech weeding & spraying, autonomous ground vehicles, articulating arms on production lines. The surplus of parallel solutions is exactly the raw material a serial acquirer feeds on.

My first instinct was to write the whole idea off as software-native. Software compounds easily. Low cost of sales, minimal capex, no inventory, no field trucks and frankly easier to re-tool and integrate. Of course the model works there.

Then I went back to the history, and the history disagrees with me.

The model didn’t start in software

The archetype for what Bending Spoons is doing has a name — the serial acquirer, the roll-up, the permanent-capital compounder. Constellation Software is the famous software version: hundreds of small vertical-market software companies, bought cheap, held forever, cash recycled into the next deal.

But one of the original masterclasses was hardware.

In the mid-1980s, two brothers — Steven and Mitchell Rales — controlled a collection of decidedly unglamorous industrial businesses. Mohawk Rubber. A vinyl siding manufacturer. Tool companies. They named the holding company after a Montana creek they fished, applied lean manufacturing discipline to everything they bought, and used the cash flows to keep buying — moving steadily upstream over four decades into instruments, diagnostics, and life sciences.

That company is Danaher. It’s worth north of $135 billion today, and Bending Spoons’ own disclosures nod to it as a model.

So the serial acquirer playbook is not sector-dependent. Danaher proved it works on physical products with supply chains, warranties, and factories. Which sharpens the question considerably: if the model works for hardware in principle, can we anticipate it in robotics?

What has to be true

Working through it, I count five conditions that need to be true for this model to be successful. Robotics currently fails most of them.

The anchor has to cash flow. The whole structure rests on an initial acquisition that generates real, stable free cash — the skeleton everything else hangs on. Danaher had rubber and siding. Bending Spoons had profitable little apps. Robotics has vanishingly few companies that are cash flow positive and stable, and the handful that are, trade at premiums precisely because they’re rare. The raw material for step one barely exists.

The targets have to be cheap or overlooked. Serial acquirers make their money on the buy. Bending Spoons hunts neglected assets nobody else wants. Robotics companies, even struggling ones, tend to carry valuations propped up by the capital they’ve raised and the preferences stacked on the cap table. The distressed ones mostly aren’t cash-flowing, and the cash-flowing ones mostly aren’t distressed. The overlap — the actual hunting ground — is nearly empty right now.

Integration drag has to be low. Danaher didn’t centralize its products; it decentralized operations and centralized the discipline. That works when each business can stand alone without heroic engineering. Buy four robotics companies across four sectors and you inherit four hardware platforms, four bills of materials, four service networks, four sets of field failures at 2am during someone’s harvest. Software has near-zero marginal maintenance. Hardware carries its maintenance with it everywhere it goes.

The playbook has to be repeatable. This is the one people underrate. Danaher had Danaher Business System (a lean manufacturing system) — the same operating system applied to every acquisition. Bending Spoons has the same cost-out, rebuild, reprice sequence run a hundred times. A serial acquirer isn’t a portfolio; it’s a machine that improves what it ingests, the same way, every time. Nobody has demonstrated a repeatable improvement playbook for acquired robotics companies. Yet.

The capital has to be permanent. A fund with a ten-year life can’t compound for forty. Danaher used public equity. Bending Spoons used debt and retained cash flow, and structured the IPO so the founders keep control. Venture and conventional PE timelines break the model before it gets interesting.

Five conditions. Robotics clears maybe one and a half. So the answer to my original question is no — not today.

But “not today” is doing a lot of work in that sentence, because I can see three distinct paths by which this arrives.

Path one: the boring path

Give the industry ten years. A cohort of today’s robotics companies survives, matures, and gets boring — viable, cash-flowing, no longer growing, valuations deflated to something resembling earnings multiples. The technology commoditizes enough that maintaining an acquired platform no longer requires the founding team. At that point a finance operator — a PE-adjacent buyer, or an OEM looking to absorb a fleet — starts rolling them up into one platform.

This will happen. It’s how every capital-intensive industry eventually consolidates. It’s also the least interesting path because anyone can see it coming and there is not much that can accelerate it.

Path two: startups acquiring each other

This one has already started. Bonsai Robotics acquired farm-ng — a tuck-in, a young company buying a peer to expand its product line and customer surface. CropX, on the ag software layer, is explicitly running a roll-up: seven acquisitions and a CEO who calls the company an M&A machine.

As the credit cycle turns and more companies stagnate between rounds, I’d expect the stronger operators in each sub-sector to keep absorbing their peers. But notice the shape of it. When a robotics company acquires, the acquisitions stay tethered to its own thesis — a new crop, a new implement, a capability that feeds the franchise it already has. The acquiring is in service of becoming a bigger version of itself. That’s a strategic consolidator, and it’s bounded by definition: a niche champion cares enormously which sector the cash flow comes from, because the sector is its identity.

A serial acquirer is the opposite. The whole point of the model is that it doesn’t care. Danaher went from rubber to dental to diagnostics to life sciences. Bending Spoons runs note-taking, file transfer, video, and ticketing under one roof. The through-line was never the sector — it was the machine. So the muscle being built in path two is real, and some of these acquirers will discover they’re good at buying. But category consolidation and sector-agnostic compounding are different animals, and the second one won’t grow out of a company whose reason to exist is a single niche.

Path three: the infrastructure buyers

The most interesting path runs through businesses that aren’t robotics companies at all.

Think about an established equipment dealer network. A parts and service business. Companies that already have the unglamorous backbone — repair operations, billing, field technicians, customer relationships built over decades, familiarity with the market and machinery even if not the autonomy stack. These businesses cash flow today. They sit adjacent to robotics, close enough to understand it, stable enough to fund it.

And critically, that backbone is already sector-agnostic. A dealer’s service arm doesn’t know or care whether the box on the lift weeds lettuce, moves pallets, or inspects a bridge — it bills either way. That indifference to what the machine does is exactly the quality a serial acquirer needs and a niche champion can never have. It’s the same reason Danaher could hold rubber and diagnostics under one roof without blinking.

A dealer group that starts acquiring robotics IP and product lines isn’t making the leap a pure-play PE firm would have to make. Day one, they can service what they buy, they can bill for it and they have the customer list. The engineering gap is real — someone has to keep the platforms alive — but it’s a much shorter bridge than the one facing a financial buyer with no sector infrastructure.

This is the inversion that took me a while to see: the first robotics serial acquirer may not be a robotics company that learned to acquire. It’ll be a cash-flowing infrastructure business that learned to hold technology. Which is, when you squint at it, exactly how the Rales brothers did it. They didn’t start with the crown-jewel science. They started with rubber and siding and bought their way upstream.

Where this leaves us

Can the Bending Spoons model work in robotics today? No. The anchor cash flows don’t exist, the targets aren’t cheap, and the integration burden is still too heavy.

Is the model coming? I think yes, and sooner than the ten-year boring path suggests — because the seeds of paths two and three are already in the ground. The startups are learning to acquire. The infrastructure businesses are watching the technology stabilize. And somewhere in the wreckage-to-infrastructure transition this sector is living through, someone is going to buy their first stable, cash-flowing robotics business at a sane price and realize what they’re holding.

The snowball needs a first handful of packed snow. Robotics doesn’t have it yet.

Subscribe now

To raise a round you need to be able to offset the fear and tip the scales towards greed. Too often, founders do a decent job at outlining the opportunity and the stakes, but don’t follow through on de-risking the venture — on reducing fear. And unless you are pitching into a bubble, which creates a separate type of fear that can work in your favor, FOMO, fear trumps greed.

In HardTech, clearing that fear is harder than in SaaS. Every raise fights fear — software has its own real ones: churn, CAC payback, vibe-coding or a moat that won't hold. But hardware stacks more landmines, and the cost of stepping on one is higher. Not only are hardware companies de-risking supply chains and BoMs, but also capital stacks that may not scale with revenue, long pre-sales, and costly post-deployment support. A bad assumption doesn't cost you a quarter — it can cost you a year and a recall.

The market opportunity might be a $100 billion blue ocean, with category-leading technology. The returns to drive greed are there. But the moment a founder waves their hands over the go-to-market strategy, fumbles the unit economics, or fails to explain why their hardware is defensible, fear drops onto the scale. It crushes the opportunity, and the VC passes.

I have helped run multiple Seed through Series B raises for physical tech companies, defending unit economics and go-to-market motions while supporting the company’s vision. In those rooms, my job wasn’t just to help sell the dream. My job was to build the connective narrative that served as the quantitative anchor for the qualitative story.

Through those reps, I watched brilliant founding engineers lose funding because they confused their Technology with their Solution, failing to calm the investor’s operational anxiety.

To raise capital for hardware, robotics, or ag-tech, the standard pitch will fail you. You need a framework built for the physical world that systematically stokes an investor’s greed while dismantling their fear.

Enter the POST² playbook: Problem, Opportunity, Solution, and T² — Technology and Team.

There are many ways to pitch a company, and some of the best pitches can ignore any framework — the founder, the current tech wave, and/or the technology were just that compelling. But for the rest of us, the POST² framework provides the structure to keep a pitch focused while putting equal weight on the parts that matter.

A note on order: the sequence below is the default for a reason — it builds greed before it confronts fear, which is the emotional arc you want a VC to travel. Experienced founders flex it, opening on traction or a jaw-dropping demo when they have one. But if you're not sure, run it in order. The framework is a spine, not a straitjacket.

P for Problem

This is the Greed set-up. It’s important and sets the stage for why a VC should care about your pitch (as a reminder, they operate on Power Law and are looking for the Big Return!).

Most pitches typically get this one right. They start with the problem at hand — what the company is trying to solve, and often the genesis for why the company exists. Information asymmetry. Resource constraints. Time pressure. All real-world problems that are ideally acute and painful.

What some pitches miss is they either go too narrow — not framing a problem big enough to clear a VC's bar, and/or overly focused on cost savings and not revenue uplift. Revenue increases can be harder to prove, but are the same if not more important, as they are unbounded versus cost, which has a natural floor. Or they layer in several problems that all really tie back to the same one problem. You can have multiple problems, but be sure that they are cohesive and support the same narrative.

As an example, I’ve seen decks from automation companies that focused solely on Labor. Labor is a common problem to solve. Labor is scarce, or expensive. A job might be one of the 3 D’s (Dangerous, Dull, Dirty) that dissuades people and is ripe for automation. Often there is a chart that shows cost over time (up and to the right) and the Total Addressable Market in a nice large font. This is a good start. However, the deck then went and expanded the problem set to several adjacent problems like climate change, regulatory hurdles, and chemical costs. All real problems, but diluting the narrative.

The flaw is breadth masquerading as depth. The fix is discipline: one acute pain, one number, one clean line of sight to why it matters.

So P: Set up the Greed with the Problem your company was built to solve. Be sure to anchor in a TAM number to keep VC eyes on the screen.

O for Opportunity

If P: Problem was the set-up for Greed, O: Opportunity is the stabilizer. This is where the problem you have outlined and its accompanying TAM survive first contact with skepticism.

A successful Opportunity section is one that takes the lens of the problem and distills it down to your segment, your space. It also starts the de-risking of Fear. It’s the why now and this. It’s your wedge and your narrative on why it pencils.

At an early stage, it is highly unlikely that you are solving the entire problem you have identified within the next few years. Instead, in the Opportunity section, you are identifying your unique entry point and how it starts to spin the flywheel of inevitability in your favor — how the beachhead you can win now (your Serviceable Obtainable Market) expands into the market you can realistically reach (your Serviceable Addressable Market) and eventually the whole prize (your Total Addressable Market).

Where the Opportunity shines is the identification of an entry point that starts making your story believable — all building up in the pitch to why now is the time for you and your Solution (foreshadowing!).

So O: Take the headline problem, find the entry wedge where you actually win first, and pair it with a credible why-now. That’s how a TAM stops being a number on a slide and starts being a path.

S for Solution

This is the section most focused on de-risking your solution and thus decreasing fear to tolerable levels. It is also where most pitches fail. Too often pitch decks view their unique story as enough, but without a plan you are just an idea, and ideas are worthless. So this is the part where you become credible. And the only way to do that is to have a credible plan.

What plan, you ask? Well, a plan on how you are meeting the opportunity head on and helping to solve the problem you led off with. This is the bridge that takes the pitch from something fantastical to something real and noteworthy. The separation between fantasy and reality. Between Half-Life 3 and Half-Life 2.

A good Solution section is one that grounds in hard data and moves between Technical, Market, and Commercial validation — introducing the technology, the go-to-market plan/customers, pricing, capital stack, and use of funds. Often this section includes a pithy chart showing revenue growth going exponential by year 3–5 and, for bonus points, unit economics. This is the chance to show that the team has done their homework and that this is a real company and opportunity to fund. Don’t skip the numbers — and also pay them heed. It’s good to show growth if it’s reasonable, but pie-in-the-sky numbers reduce your believability.

This is also where the ask typically lives. How much you’re raising, what it buys, and — most importantly — which milestone it gets you to. The ask isn’t a number, it’s a milestone story. “We’re raising $X to hit Y, which de-risks us into the next round” is fundable. “Raising $3–8M for 18 months of runway” is not — a vague ask signals you haven’t done the math, and in hardware the math is the whole game. Tie the number to the milestone, and the milestone to why it justifies the dilution.

So S: This is the rubber-meets-the-road section, where the story gets real and grounded with a quantitative plan. It includes your financial forecast, key customer/partner logos, and product introduction. This is the validation section to reduce Fear and let Greed stand alone.

T² for Technology and Team

Our final section includes the final validation slides. These build on the good work the Solution section has done. They help the investor further understand what your technology is (i.e. your secret sauce) and the qualifications of the people you have enlisted (employees + advisors) to build it.

On technology — the risk here is going deep. You are assuredly proud of your technology. It really is something special. But the risk here (and why it’s at the back of the deck) is for technical founders to spend way too much time on tech at the expense of everything else. Skipping past the Greed build-up, whistling by the Fear reduction, and landing on what they know best — tech.

My advice on tech: keep it short. One or two slides on the best and most differentiated parts of your technology. If you are having to spend 60 minutes explaining to an investor what the tech does, it’s a good sign they aren’t a good fit for your specific business. Additionally, having a separate technology deck that allows the deep dive provides a strong reason for a second, more focused meeting. Treat the initial pitch as setting the hook.

On team — the point here is to show you have the experience and capabilities on hand to realize your dream, or at least progress through the next milestone. For an early-stage company, you will be heavy on engineering talent, with advisors often filling out the commercial/business side. As you grow, this mix balances.

And in hardware, don’t underestimate this slide. Hardware is unforgiving — it breaks, it ships late, it costs real money to fix in the field. A VC who sees battle scars and specific domain expertise on the team is a VC whose fear just dropped a notch. The right team is itself a de-risker, especially in the early days when commercial and product traction is thin.

Also, consider bringing something visual or tangible to the presentation to maximize hardware’s unique strength, it’s real and you can touch it. Videos are excellent to include — it’s hard to get most robots into a conference room, and it helps ground your tech in reality. As a proxy, bringing in a physical sensor or component helps as well.

The team slide (and there is only 1 slide even though your people are wonderful and worthy of 3) is colorful and filled with key faces, brief founder bios, and logos of companies the team has worked at, as well as key partners. Think the cherry on top of the presentation. It’s not carrying the ice cream sundae by itself, but rather the finishing touch.

So T²: These are your final validation slides and help your believability. Keep them higher-level and colorful — they are your closing, the finale you’ve been building towards for the past ~60 minutes.

The Part After the Pitch

You can run a flawless POST² deck and still lose the room in the ten minutes that follow. Because the deck is the part you control — the Q&A is the part the VC controls, and it’s where the real fear lives. VCs may at times try to derail your presentation, jumping ahead or interjecting. This is generally fine if contained, but don’t get knocked off your game because an interruption isn't always just interest — sometimes it's a test.

And, another thing to keep in mind, the hardest questions aren’t attacks. They’re the investor trying to talk themselves into you. Every tough question is a fear they need retired before they can write the check. Your job isn’t to deflect them — it’s to have already done the work so the answer lands cleanly.

In hardware, the fear questions are predictable. Anticipate them and you turn the scariest part of the meeting into your strongest:

“What happens to margins at scale?” — They’re testing whether your unit economics survive contact with reality. Have the cost-down curve ready, and be honest about where it bends. Also be ready to speak to any potential pivots in your revenue model.

“Why won’t a big incumbent just do this?” — They’re probing the moat. “We’re faster” isn’t an answer. A real one names what you have that they can’t easily copy — data, a wedge relationship, a design that’s hard to retrofit.

“How much does the next round cost, and what does it get you?” — They’re checking whether you understand your own capital path, or whether you’ll be back in twelve months diluted and surprised. This is the capital-stack question in disguise, and in hardware it’s the one founders most often flub.

“What breaks first when you scale the fleet?” — They want to know if you’ve actually run the thing in the field or just in the deck. The founders who answer this well have scars. The ones who don’t, hand-wave.

The pattern in those questions: every one of these is a fear wearing a question mark. Walk in having pressure-tested your own deck for them — ideally with someone whose job is to ask the uncomfortable version before the VC does — and the Q&A stops being where raises die and becomes where they close.

In Summary

There is no one way to pitch an investor, and every deck — no matter how well crafted — needs to survive first contact with interested (and sometimes aggressive) questioning that can derail your perfect plans.

So put the framework together, PRACTICE the framework, anticipate the questions that might come up, and PRACTICE again. And if your problem is real (and VC Big), the opportunity is timely, your solution thoughtful and believable, and your team and tech credible — you significantly up your chances for an investment, no matter which order these are presented in.

But either way, remember that, like life, pitches are a dance. You’ll make several false steps. You’ll hear plenty of No’s (or many versions of No). But keep refining, keep practicing, and keep building. Good things typically come to those who do.

Daniel Kirstein is the founder of Holdfast Partners, a financial architecture and commercialization consultancy focused on robotics and hardtech founders. He brings firsthand experience navigating the valley of death between prototype and scale — including supporting multiple hardware raises from Seed through Series B. He writes about hardtech commercialization on LinkedIn and at hlfst.com.

Subscribe now

In 1897, a small Stuttgart workshop became the chief supplier of a truly reliable ignition system for the automobile. Five years later, their chief engineer Gottlob Honold unveiled the high-voltage magneto ignition with spark plug — and Robert Bosch’s courtyard workshop began its transformation into one of the most important industrial companies on earth.

Bosch didn’t build cars. They built the part without which cars as we know them today couldn’t work. And in doing so, they became indispensable to every manufacturer that did.

That’s the missing layer in robotics right now. And it may be the most important business opportunity in the sector that almost nobody is explicitly building toward.

Every Robotics Company is Reinventing the Same Wheel

Across this series, we’ve looked at companies going deep in niches, expanding across crop types, moving into adjacent industries, and leaping entirely beyond agriculture. Each path has merit and each has failure modes. But they all share a common friction that almost nobody talks about publicly.

Every robotics company is rebuilding the same hardware foundation.

Perception systems, compute stacks, wiring harnesses, safety architecture and control systems. The unglamorous infrastructure layer that has to exist before any robot can do anything useful in a field — or a construction site, or a mine. Every company is reinventing the same wheel, taking insufficient off-the-shelf parts, putting finite time and resources into building unique/novel configurations to ultimately solve something that is table stakes and not the differentiation layer. Every robot needs perception, controls, safety but having to build these within a team slows down the ultimate value proposition of a company’s robot.

And, it’s this need to build everything from the ground up that leads to delays, and varying quality levels that don’t scale. I’ve worked with robotics companies that spent months tracking down a random engineer in the Midwest to hand-wind wiring harnesses. While the engineering was good, quality was inconsistent. It was also slow and difficult to replicate. This is not how an industry scales.

The perception stacks that can handle dust, shock and vibration, temperature and full-day operation at the speeds robotics requires — very few exist off the shelf at the quality level the applications demand. Most companies are building their own from scratch, burning months and millions on problems that aren’t their core differentiator.

Every robot needs to see. Every robot needs to move. Every humanoid needs to feel. These are table stakes, not moats. And yet the industry is treating them like competitive advantages worth duplicating at every company independently.

What Bosch Actually Did

When Robert Bosch solved the ignition problem in 1897, the automobile was a curiosity for the pioneers — unreliable, dangerous, temperamental, and dependent on cranking mechanisms that required more physical effort than most people were willing to provide. The ignition system was the unglamorous constraint that had to be solved before the automobile could become a mass-market product.

Bosch didn’t try to build a better car. He solved the constraint that was preventing all cars from working reliably. And because he solved it for the whole industry rather than for one manufacturer, the value compounded across every vehicle made.

That’s the picks-and-shovels insight applied to hardware infrastructure. You don’t have to build the robot that wins. You have to build the part without which no robot can win.

Bosch didn’t wait for BMW or Mercedes to exist before supplying the industry, they didn’t exist yet. Bosch was supplying ignition systems to early automobile manufacturers from 1897 — nearly two decades before BMW was founded in 1916. The tier-one supplier layer didn’t emerge after the platform OEM. It emerged alongside the fragmented early industry and helped determine which OEMs were able to scale at all. The robotics industry doesn’t need its BMW before it can have its Bosch. It needs its Bosch before a BMW becomes possible.

What’s Missing in Robotics Right Now

The gaps are specific and worth naming:

Perception is the most acute. The vision systems required for outdoor robotics — handling direct sunlight, dust, rain, canopy shadow, and the speed constraints of commercial agricultural applications — are being built from off-the-shelf parts that were never designed for these applications and/or rebuilt from scratch at company after company. A standardized, field-hardened integrated perception stack that robotics companies could license and build on top of would compress years of development time across the entire sector. It would also create a unified data ingestion source, where data is being captured by one device and potentially shared across a now unified interface. This allows, for the first time, a common data architecture unifying various actions into one comprehensive dashboard allowing unrivaled insights.

Integrated edge compute is the second gap. The combination of perception, real-time inference, and actuation control running on hardware that can survive the shock, vibration, heat, and intermittent connectivity of field deployment is not a solved problem. It’s being solved repeatedly, independently, at significant cost.

Tactile sensing matters more as the sector moves toward manipulation — harvesting, sorting, handling. The humanoid robotics wave is already running into this wall. An engineer at a large humanoid company told me recently that “the first company to solve tactile sensing is the one that wins.” That’s differentiator language — and they’re right, for now. But the history of enabling technologies suggests a different endpoint. Lidar was a differentiator until it wasn’t. The first AV companies to deploy it had a meaningful edge, then it became table stakes before becoming a commodity. The company that wins the tactile sensing race won’t stay a differentiator forever — it may eventually become the supplier that every manipulation robot depends on. That’s the Bosch moment for tactile sensing. And it’s closer than most people in the humanoid space realize.

Wiring, safety and integration is the unglamorous one. The physical infrastructure connecting sensors, compute, actuators, safety and power systems in a robot that has to survive years of field deployment is being figured out one robot at a time. It also leads to varying capabilities from a safety perspective and may delay global roll-outs. It is not how an industry scales.

A software layer does exist. ROS — the Robot Operating System, now in its second generation — has become a de facto open-source middleware standard that handles communication between robot components. It’s genuinely useful and widely adopted. But ROS, while enabling and an accelerant is middleware, not infrastructure. It helps components talk to each other once you’ve built them. It doesn’t replace the perception system, the compute stack, the safety architecture, or the wiring harness that ROS will eventually connect.

In addition, some companies are beginning to fill that gap from the software side. Carbon Robotics has been deliberate about the value of its Large Plant Model — a computer vision system trained on millions of acres of commercial laser weeding data that increasingly looks like a licensable asset as much as a product feature. Bonsai Robotics has framed its entire thesis around shifting “from iron to software” — the autonomy stack as the durable asset, the hardware as the delivery mechanism. Both are signaling that the most defensible value in their business may not be the machine. It may be what the machine learned.

That’s the software version of the tier-one supplier thesis — and it’s converging with the hardware opportunity from the other side. The autonomous vehicle industry offers the clearest precedent. Velodyne, Luminar, Ouster — a generation of LiDAR suppliers emerged to serve the AV wave, providing a standardized sensing layer that AV companies could integrate and build on rather than develop independently. That supplier layer compressed the development timeline for the whole industry and created significant independent value. Velodyne went public. Luminar went public. The picks-and-shovels play worked.

Robotics doesn’t have that layer yet — in hardware or software. But the conditions that produce it are forming from both directions simultaneously.

How It Starts

There are a few plausible paths to the first robotics tier-one supplier, here are two:

The most straightforward is organic — a company that starts by solving one specific infrastructure problem exceptionally well and expands from there. The way Bosch started with ignition and gradually became the electrical nervous system of the automobile. The way Velodyne started with spinning LiDAR and became the perception standard for an entire industry. Potentially with Bonsai and Carbon solving the autonomy and computer vision respectively. One solved problem, done better than anyone else, deployed across many customers.

The more opportunistic path is IP harvesting. The first wave of robotics has left behind significant technology — perception models trained on millions of acres of field data, autonomy stacks validated in demanding outdoor conditions, hardware designs refined through years of commercial deployment. Most of it is sitting in companies that ran out of runway before they could capture its value. Some of it was acquired by OEMs who may not be the natural home for it, some sits with companies that would welcome licensing it rather than letting it depreciate. Some of it is genuinely in the dustbin of history.

The company that can surface, aggregate, standardize, and productize that IP — building a library of field-hardened components that the next wave of robotics companies can license rather than rebuild — is doing something genuinely valuable and genuinely scarce. It is also a bet on the overall trajectory of the Robotics industry.

It’s a different business than building a robot. It requires a different kind of founder, a different kind of patience, and a different relationship with the companies it serves. But the economic logic is compelling: you’re not betting on one robot winning. You’re supplying the picks and shovels to everyone who’s in the field.

The Series Conclusion: Five Paths, One Question

Across four parts and five distinct paths, this series has been circling the same question: is agriculture big enough for the technology companies trying to serve it?

The answer, it turns out, depends entirely on which company you’re asking about, the capital they’ve taken on and which path they’ve chosen.

Go deep and stay there — own a niche with the right margin or volume, align the capital to match, and build a real business without mistaking it for a platform play. Harvest CROO and TRIC Robotics are doing this right. The ceiling is real but so is the business.

Expand the niche — let the software travel as far as the hardware allows, plan for the ceiling before you hit it, and build the expansion into the capital structure from the start. FarmWise, Burro, and Lumo each found their version of this path. The rule holds: the software travels. The hardware finds its home.

Ag-adjacent expansion — move into sectors that share agriculture’s operating characteristics without fully leaving the farm. Verdant’s expansion into sod, grass seed, and golf turf is the clearest current example. The unit economics need to be rebuilt from scratch in every new market. Adjacency is not transferability.

Agriculture as a stepping stone — treat the proving ground as the credential, take the autonomy stack into defense, construction, or mining, and build a company whose ceiling isn’t defined by crop types. Bluewhite showed where this ends. Bonsai is showing where it could go. The companies that make this crossing successfully will be the ones who planned for it before they needed it.

Become the platform — stop building robots and start building what robots need. The Bosch of field robotics doesn’t exist yet. The conditions that will produce it are forming. The company that gets there first won’t need to win a single crop type or a single customer vertical. It will supply the infrastructure layer that every robotics company — in agriculture, construction, mining, and beyond — is currently rebuilding independently.

None of these paths is wrong. All of them have failure modes. The companies that get lost aren’t the ones that choose the wrong path — they’re the ones that raise capital for one path while quietly executing another, and discover the mismatch when the runway or market gets tight.

Agriculture is demanding enough to build genuinely capable field robotics. It may not always be large enough to reward the capital required to build them. The founders and investors who understand that distinction — early, clearly, and without the filter of a pitch deck — are the ones most likely to find their way through.

The trillion-dollar number was always real. The trap was never that agriculture was too small.

The trap was thinking agriculture was the ceiling.

For the companies building it right — it’s the floor.

This concludes the “Is Agriculture Big Enough for AgTech?” series. To the now 100 (!) subscribers, everyone who commented, and everyone who DM’d — thank you. This series was better for your engagement.

Next: the same capital structure and commercialization dynamics playing out beyond the farm — in autonomous vehicles, construction robotics, and defense. The playbook turns out to be more familiar than it looks.

Subscribe now

For most ag robotics companies, there comes a point where the math stops working the way it did in the pitch deck.

The niche is proven and the tech works. The grower relationships are real, real value is discernable and revenue is growing. And then someone runs the numbers on what it would take to reach the next crop type, the next geography, the next market or even just the next customer — and realizes the hardware ceiling is closer than it looked from the fundraise.

What happens next separates the companies with a plan from the ones that discover they needed one and while we are still early in this transition outside from pure Ag plays, I believe that as the first wave of automation plays continue maturing we will see more frequent IP transfer between disparate sectors.

Some will double down, expanding within ag. A few may become platforms. And a growing number are asking a different question entirely: what if the most valuable thing we built isn’t an ag robot at all?

Path 3: Ag-Adjacent — Same Hardware, Different Grass

The initial move outside agriculture is usually incremental, a step into the somewhat unknown.

The companies taking this path aren’t abandoning agriculture — they’re extending technology proven in ag into adjacent managed environments. Turf & Sod production. Grass seed. Greenhouses. Managed landscapes like golf courses. The problems look similar on the surface: outdoor autonomous operation, plant-level precision, unstructured environments that change seasonally, need for resilient architecture with high uptime. There are however several differences; customers are different, competitive dynamics and unit economics look different. And the market, in some cases, is significantly larger.

Verdant Robotics is the most interesting early example of this move. They spent years building and refining the SharpShooter — a precision application platform that uses “aim-then-apply” computer vision to deliver millimeter-accurate inputs at the plant level across specialty vegetables. The core insight driving their expansion is that the underlying capability — identifying a biological target in an outdoor environment and delivering a precise intervention — is not crop-specific. It’s target-specific. And those targets exist well beyond specialty vegetable fields.

Their Aim & Apply platform — the AI vision system, the individually controlled turrets, the prediction engine — is designed to be product-agnostic. CEO Gabe Sibley, in a recent AgFunder article (May 2026) describes it as “a category unto itself”: the same core technology that guides a shot on a tractor-pulled implement today could direct applications from a swarm robot, a center pivot, or platforms that don’t exist yet. SharpShooter is the first product built on the platform and likely won’t be the last. That framing pushes Verdant closer to the platform path than a pure ag-adjacent expansion story — worth keeping tabs on as the product line evolves.

In April 2026, Verdant officially took that next step and expanded the SharpShooter into grass seed, sod production, and golf turf management. The move is more technically challenging than it sounds. In specialty vegetable fields, the vision problem is relatively clean and a level 1 boss challenge: green plants against brown soil. In grass seed and sod production, the system has to distinguish between grass crop plants and grass weed species — targets that look nearly identical. Verdant’s own description of the challenge captures it well: these are sectors “where crop and weed can look nearly identical, making precision weed control one of the hardest ID problems in agriculture.”

The market they’re entering is significant. The US sod market is worth roughly $2 billion annually, with 2.5 billion square feet produced across roughly 400,000 acres. The global grass seed market adds another $5 billion. The customers — sod farms supplying landscape contractors, homebuilders, sports fields, and restoration projects — face the same structural pressures as specialty crop growers: labor scarcity, rising input costs, and increasing regulatory pressure on chemical use. There is no incumbent robotic solution in this market, which currently relies heavily on manual labor and blanket boom spraying. Verdant is one of the first to have made this crossing.

But the unit economics question is worth asking directly — because the answer is different here than in specialty vegetables.

In specialty crops, the SharpShooter replaces expensive, scarce hand labor — a clear and compelling ROI. In sod and grass seed, the alternative isn’t an increasingly scarce hand-weeding crew. Those boom sprayers, often self-propelled or tractor pulled, cover 30-60 acres per hour at a fraction of the cost. The SharpShooter runs at 5-7 acres per hour. That speed gap is significant, and Verdant knows it — their CCO, Curtis Garner has publicly named 5 mph as the target speed (AgFunder, May 2026), roughly triple where they started. This is a similar issue facing companies that attempt the leap from specialty to lower-margin row crops: the speed required to make unit economics work at scale typically breaks the current hardware form factor. The comparative unit economics story, however, isnt solely about speed, but also chemical input costs. By targeting selectively rather than broadcasting, Verdant cuts chemical usage by up to 95% — reducing both input costs and the logistical overhead of managing chemical supply in the field.

The economic case in sod, however, doesn’t rest solely on labor and chemical savings. It also rests on quality protection. A 3-5% presence of foreign grass in a sod field leads to full pallet rejection, with remediation costs running $1,500-$2,000 per acre. In a market where contamination risk is the primary financial threat, the ability to identify and eliminate weeds within a grass crop — at the 2mm seedling stage, before they can spread — has real value regardless of coverage rate. The system is already in commercial use in grass seed and demoing with major sod producers in Texas, Florida, and Georgia.

Whether the full value proposition justifies the price point and speed differential at scale is still an open question. The economics will become clear over the next season or two. But it illustrates something important about ag-adjacent expansion: the same technology can carry a completely different value proposition in a new market, and the unit economics need to be rebuilt from scratch rather than assumed to transfer.

Verdant is one of the first Wave 1 ag robotics companies to make this move publicly and deliberately. The expansion into grass seed, sod, and golf turf is a bet that the SharpShooter’s precision architecture — and the Aim & Apply platform underneath it — travels further than the specialty vegetable fields it was originally designed for. The ceiling in specialty vegetables may be real. The ceiling on precision application as a category is much higher — if the unit economics can be made to work.

Path 4: The Full Leap — Agriculture as Proving Ground

The second move is harder to make, nascent and harder to explain.

It requires a founder to look at the technology they’ve spent years building for agriculture and conclude that the most important thing about it isn’t that it works in fields — it’s that it works outdoors, in unstructured environments, under conditions that would break lesser systems.

Heat and cold, shock and vibration, irregular terrain, intermittent connectivity, unpredictable obstacles and changing weather conditions. Variable lighting from predawn darkness to midday glare. Irrigation water, mud, dust and everything else that comes with operating close to the ground.

Agriculture is, in a specific and underappreciated sense, one of the most demanding proving grounds for outdoor autonomous systems on earth. The sector is missing some things — maritime waterproofing, deep subsurface pressure differentials, the radiation environment of space. But within the envelope of what field robotics actually requires — autonomous navigation, real-time perception and decision-making, continuous operation in environments that are never quite the same twice — agriculture tests it while demanding high uptime and performance.

A thesis is quietly forming among a small number of investors and founders: once you’ve made it in agriculture, you can make it almost anywhere.

The financial data validates the gravitational pull. Median Series A post-money valuations for defense robotics companies reached $105M in 2025, compared to $50M for non-defense — a gap that has widened every year since 2021. Defense procurement cycles, while long, are predictable and often come with large contracts that have high renewal rates. The customer has both the budget and the urgency to deploy at scale. For a company that has spent years hardening an autonomy stack in some of the most demanding outdoor environments on earth, that valuation gap is a powerful force.

Bluewhite is the clearest public illustration of where this thesis leads. They built an autonomous driving platform for agricultural machinery — GPS navigation, obstacle detection, path planning in unstructured outdoor environments. The commercial wrapper was agriculture. Elbit Systems, one of Israel’s largest defense contractors, saw something different. In May 2026, they acquired 100% of Bluewhite’s shares.

The backstory matters here. Bluewhite was founded in 2017 by Ben Alfi — a former Israeli Air Force combat pilot and Head of Unmanned Systems R&D — along with co-founders Yair Shahar and Aviram Shmueli. Alfi said he was inspired to bring autonomous technology to agriculture after 25 years of military service. The platform was designed to be OEM-agnostic from day one — able to retrofit any ground vehicle regardless of manufacturer.

That origin story reframes the acquisition entirely. This wasn’t an ag company that pivoted to defense. It was a military autonomy veteran who used agriculture as the proving ground he always intended to move beyond. The field was never just a farm — not even on day one.

Elbit didn’t buy a tractor company. They bought a field robotics platform — outdoor autonomy validated in one of the most demanding unstructured environments on earth — that happened to have been built and proven in agriculture first. The defense applications for that capability are significant and funded in ways that agricultural companies are not.

Bonsai Robotics is navigating a version of this path with more deliberate intention — and with a team assembled specifically for the journey beyond agriculture, whether that intent survives first contact with non-ag procurement timelines is a different question.

Tyler Niday, co-founder and CEO, spent years as an engineering leader at Blue River Technology and John Deere before founding Bonsai with a specific thesis: that vision-based autonomy developed for agriculture is the foundation for outdoor autonomy everywhere. His co-founder and CTO, Ugur Oezdemir, brings the same background — vision-based autonomy across Airbus, Blue River, and John Deere. Their COO, John Teeple, launched Deere Labs and led precision ag strategy. This is not a team that assembled to build a better tractor. It’s a team that understands exactly what agricultural autonomy produces and where it leads.

The company built vision-first autonomy software for agricultural machinery, then acquired farm-ng in July 2025 to gain the Amiga hardware platform — giving them both the software stack and now a next-generation machine to iterate and then deploy it on. Integrations like this take time to fully realize, but we are now starting to see the practical benefits of their platform being utilized like with the Rootline Robotics team from Cornell University leveraging the Amiga to win the recent UCANR Robotics challenge. Their stated ambition is explicit: “We’re shifting the industry from iron to software,” Niday has said, and the “software” he’s describing is designed to run on any machine operating beyond the factory floor.

The $15M Series A, announced in January 2025, was framed not as an ag-specific bet but as a platform play — bringing AV-grade autonomy off-road. Niday’s own words at the time: “With our use case now proven, Bonsai Robotics can move forward with expansion into other areas of agriculture and beyond to other off-road environments.” The agricultural deployments across vineyards, orchards, and bedded crops are feeling more like the proof of concept and the off-road environments that follow are the thesis.

What makes Bonsai’s trajectory interesting is that the team is building the kind of platform that solves for the variability that defines agriculture — terrain, crop architecture, weather, task type — and in doing so builds a system capable of scaling across use cases that share those same variables. Construction, mining and defense being the potential targets where the problems are dangerous, dirty, and repetitive — and they’re agnostic of which industry produces them. A vision-based autonomy stack trained on the unpredictability of an orchard doesn’t need to be rebuilt for a quarry. It just needs to be redeployed.

A third signal is emerging from within my own client base, with several clients building, from inception dual use products. These products are being designed intentionally for path that forks: agriculture first to prove the use-case and generate value, then defense. Not as a pivot when agriculture hits its ceiling, but as a deliberate product roadmap with two branches sharing a common autonomy foundation. Agriculture is the proving ground. The defense application is the plan, not the contingency. I can’t name them — they’re in stealth — but the structure of that bet is worth naming because it represents something genuinely new: a company that treats agriculture not as the sole market, to be pivoted away from when things get hard, but as a viable and attractive entry point to a more broadly focused business that falls more inline with John Deere and Caterpillar who focus on various sectors simultaneously (construction, mining, ag).

And, what makes this signal all the more interesting, is that these companies are choosing to not move within agriculture, to not double down on row crops, but rather see greater opportunities outside of agriculture.

The counter is likely true as well, and one could argue that products currently in the defense, mining and construction space may look at agriculture and its massive market and make the leap across. This is feasible and viable, but given how segmented agriculture is, will be a longer path than agriculture leaping outside to these verticals. It however, will be fascinating to watch unfold over the years ahead.

The Crossing Is Harder Than It Looks

The thesis is sound. The financial logic is compelling. But the graveyard between “proven ag technology” and “successfully deployed in a new sector” is real — and most companies never even reach the water to even attempt the crossing.

Most ag robotics companies run out of runway inside agriculture before the option to expand becomes real and this is the first route to failure. Advanced Farm had working technology and $34M raised. Farmwise had paying customers and years of field data. Neither made it to the crossing — not because the expansion thesis was wrong, but because agriculture’s capital structure and commercialization timeline didn’t give them enough runway to find out and there was enough energy expended on staying afloat in Ag to have enough left to thoughtfully plan a crossing.

The second failure mode is attempting the crossing from the wrong position. The leap requires something most ag robotics companies don’t have when they’re finally ready to make it: a profitable or near-profitable core business generating enough cash — or enough investor confidence — to fund expansion R&D while the new market develops its own revenue stream.

Defense procurement cycles are long. Construction adoption is slow. Mining requires certifications and partnerships that take years to build. A company arriving at the door of a new sector with six months of runway and a half-finished product isn’t making the leap. It’s falling.

The companies most likely to successfully make the crossing share a few characteristics. Their core ag business generates real recurring revenue — not just pilots, not just letters of intent, but commercial deployments that cover operating costs. Their autonomy stack is genuinely sector-agnostic, not just claimed to be. Their hardware was designed for outdoor resilience from the beginning, not retrofitted for it. And critically — they’ve made the decision to expand before they need to, not because they’ve hit a ceiling and have no other option.

A company expanding from strength moves differently than a company expanding from desperation. Investors in a new sector can tell the difference immediately. So can the customers.

Bluewhite didn’t make it across under its own power. The technology was good enough that CNH and eventually Elbit saw value in it — but the company itself ran out of runway before it could capture that value directly. The IP survived. The company didn’t. That’s a version of success for the investors who backed the technology. It’s a different story for the founders who built it.

Why This Transition Is Coming

The first wave of ag robotics is maturing. The companies that survived it — the ones with real deployments, real revenue, real field data — are sitting on something more valuable than they may realize: years of autonomy development in one of the most demanding outdoor environments on earth, at a moment when the rest of the robotics world is looking for exactly that. The current second wave will learn from this and will be more intentional about their roadmaps.

The timing is not accidental. Several forces are converging simultaneously.

Wave 1 companies that successfully navigated the capital structure and commercialization gauntlet are now hitting the natural ceiling of their agricultural addressable markets. The hardware form factor finds its home. The software keeps traveling. The question becomes where it travels next.

The defense and construction autonomy markets are growing faster than agricultural robotics and are better capitalized. The problems they need solved — autonomous navigation in unstructured outdoor environments, continuous operation under variable conditions, resilient hardware that doesn’t require a controlled environment — are problems ag robotics has been solving for years. The sector has been running an inadvertent R&D program for field robotics at large, funded by agricultural applications and validated by agricultural conditions.

The investment community is beginning to connect those dots. The thesis that agriculture is a proving ground for field robotics more broadly is moving from fringe observation to emerging consensus. When Elbit acquires an ag autonomy stack, when a BofA research report frames agriculture as the entry point for physical AI at scale, when YCombinator includes agriculture robotics as a featured category in their Summer 2026 Requests for Startups — with Garry Tan writing the section personally, when founders begin designing ag-first roadmaps with defense branches built in from inception — the direction of travel becomes clearer.

The companies that will make this transition successfully won’t be the ones scrambling for a new market when agriculture stops growing. They’ll be the ones who understood from the beginning that agriculture was where you earned the right to go somewhere else — and built accordingly.

What This Means for the Sector

It would be easy to read the companies leaving agriculture — or planning to — as a signal that the sector has failed them. It isn’t.

Agriculture is producing field robotics companies capable of operating in some of the most demanding outdoor environments on earth. That’s a genuine industrial asset, even if the exit value doesn’t always accrue to the agricultural market directly. The knowledge, the IP, the autonomy stacks, the computer vision models trained on millions of acres — these don’t disappear when a company expands its addressable market, they make for a more sustainable, resource rich company that is better able to manage the cyclical seasonality and invest what it learns from other sectors back into agriculture.

A sector that produces technology capable of operating in construction, mining, and defense is worth investing in — not just for agricultural returns, but for the option value embedded in what gets built along the way. And, I would argue that more attention to these companies and their inherent value is a net positive for the agricultural sector, that we will see better funding, more ambitious builders and more novel solutions.

The trillion-dollar number in Part 1 was always real. The trap was never that agriculture was too small. The trap was thinking agriculture was the ceiling.

For the companies building it right, it’s the floor.

The Five Paths

Across this series, five distinct paths have emerged for ag robotics and hardtech companies navigating the fundamental tension between what the market rewards and what the technology can reach:

1. Go deep and stay there: own a niche with the right margin or volume, align the capital structure to match, and build a real business without mistaking it for a platform play.

2. Expand the niche: let the software travel as far as the hardware allows, extend the addressable market incrementally, and plan for the ceiling before you hit it.

3. Ag-adjacent expansion: move into sectors that are outside pure play agriculture, but share similar characteristics.

4. Leave agriculture behind: treat the proving ground as the credential, take the autonomy stack into adjacent or entirely different industries, and build a company whose ceiling isn’t defined by crop types or geographies.

A fifth path — becoming the platform others build on — deserves its own treatment and is the subject of Part 4.

None of these paths is wrong and all have their opportunities and associated risks. The companies that get lost aren’t the ones that choose the wrong path — they’re the ones that raise capital for one path while quietly executing another, and discover the mismatch when the runway is gone.

Agriculture is demanding enough to build genuinely capable field robotics. It may not always be large enough, in its multiple sub-markets, to reward the capital required to build them. The founders and investors who understand that distinction — early, clearly, and without the filter of a pitch deck — are the ones most likely to find their way through.

Next: Part 4 explores the fifth path in detail: becoming the platform others build on, and what it would take for ag robotics to produce its first tier-one supplier.

Then our next series looks the same capital structure and commercialization dynamics playing out beyond agriculture — in autonomous vehicles, construction robotics, and defense. The playbook turns out to be more familiar than it looks.

Subscribe now

There is no single way to build an ag robotics company. There are, however, a handful of distinct bets — and most founders don’t realize which one they’ve made until the market tells them.

The bet isn’t just about technology. It’s about the shape of your market, the patience of your capital, and how far your core hardware can actually travel. Get that alignment right and a relatively small company can be a very good business. Get it wrong and you end up with a technology that works and a company that doesn’t.

Here are the paths I’ve been watching companies navigate.

“Get that alignment right and a relatively small company can be a very good business. Get it wrong and you end up with a technology that works and a company that doesn’t.”

Path 1: Go Deep and Stay There

The simplest bet in ag robotics is also the one most investors struggle to underwrite: find a niche, dominate it, and be honest that dominating it is enough.

For this to work, three things need to be true. The margin in the niche has to be high enough to support a capital-intensive robotics business. Or, the volume has to be large enough to compensate for thin margins. And — critically — the expectations of everyone around the table have to be calibrated to a business that may never be worth a billion dollars, but can be a very healthy and important one.

That last condition is where most companies quietly fail.

Harvest CROO is the cleanest example I know of a company that got this alignment right. They’ve been building strawberry harvesting robots since 2013. Not blueberry robots. No lettuce robots. Strawberry robots. Everything about the company — the computer vision, the actuation, the field operations, the data platform — is oriented around making strawberry harvest better.

What makes Harvest CROO structurally interesting isn’t just the focus. It’s who’s behind them. Roughly two-thirds of the US strawberry industry is represented on their cap table. Growers, not VCs, are the primary investors. The people writing the checks are the same people who need the technology to work — which means the pressure to grow into a different market, or pivot to chase a bigger TAM, is structurally absent. The capital and the mission are pointing at the same thing.

Strawberries also earn this kind of focus. They’re one of the highest-labor, highest-margin specialty crops, and the economics of automating even part of the harvest are compelling in ways that row crops simply aren’t.

TRIC Robotics is a younger version of the same thesis, though with a different mechanism and a slightly different capital profile. Their robots — called Luna — use UV-C light and bug vacuums to eliminate pests and fungal pathogens on strawberry fields without chemicals, with the ability to run overnight so growers wake up to treated fields. They’ve raised $5.5M in seed funding, deployed across more than a thousand acres on California’s Central Coast, and are expanding their fleet rather than their crop scope — staying narrow by design, becoming an integral part of the farming operation rather than chasing adjacencies that would dilute their focus.

The niche logic is almost identical to Harvest CROO’s. Strawberries face up to eight+ pesticide applications per crop cycle — expensive, increasingly regulated, and facing resistance. A focused solution in a high-value crop, with capital structured around proving the model rather than racing to scale.

SwarmFarm is worth noting here as the volume-play version of this path — and as a company that inverted the usual trajectory. Most ag robotics companies start in high-margin specialty crops and wonder whether they can eventually reach broadacre. SwarmFarm, an Australian company founded by farmers Andrew and Jocie Bate, started in broadacre — grain, cotton, dryland cropping — and built their way outward from there. Their lightweight autonomous SwarmBots have logged over two million hectares of operation across grain, cotton, horticulture, turf, and orchards in Australia, and they recently raised $30M to expand into North America.

What makes SwarmFarm interesting isn't just the reverse trajectory. It's the capital structure logic underneath it. Broadacre doesn't have specialty crop margins, but it has volume — and if you can build a machine simple and cheap enough to work economically at broadacre scale, the TAM is enormous. That math only works with patient capital behind it. SwarmFarm was largely self-funded by its farmer founders, raised mostly from ag-focused investors, and their lead Series B investor — Edaphon, a Belgian evergreen fund — has no fixed exit horizon, giving the company room to let broadacre unit economics mature rather than racing to a liquidity event. Their open SwarmConnect platform, which functions as an app store allowing third parties to build crop-specific applications on top of their robots, is also worth noting — because a machine built for volume turns out to travel across crop types just as effectively as one built for margin. That's a Path 2 story, and we'll come back to it.

The cautionary tale for this path is also its clearest illustration of where the bet breaks.

Advanced Farm Technologies started in strawberries. They built a working robotic harvester, commercialized it in California, and raised $34M — including backing from Kubota and Yamaha. They had real growers, real machines, real fruit picked. But they never reached a scale that justified the capital raised, and by the time of their 2023 round, which included a minority investment from CNH Industrial, the apple harvester had become the story they were selling.

It wasn’t enough. In early 2025 they shut down, citing a funding shortfall despite the apple technology being close to commercialization. The team was laid off. The 2025 harvest season was cancelled.

The hard question isn’t whether the apple harvester would eventually have worked. It’s whether the market they were moving into had the margin profile to justify the R&D cost of getting there. Apples are not strawberries. The harvest window is different, the orchard geometry is different, and the mechanics of picking fruit from a vertical trellis don’t transfer cleanly from in-soil row beds. When the next check didn’t come, there was no profitable strawberry business underneath to catch them. The technology was good enough that CNH ultimately acquired the IP. The company just ran out of runway before it could capture the value it had created.

Modest traction in the original niche, an expansion bet that burned the runway, no floor beneath it. That’s the failure mode of the expansion bet. But what happens when the machine you built doesn’t need to be rebuilt to capture the next market?

Path 2: Expand the Niche

The second path starts in the same place — a high-value crop, a focused technology — but the bet is different. The question isn’t whether you can own a niche. It’s whether your core hardware can travel to enough adjacent niches to expand the addressable market without rebuilding the machine.

FarmWise is the company I know best here, having served as CFO through their commercial launch and transition from Robotics as a Service (RaaS) to a Hardware sales model. The starting point was lettuce — specifically the Salinas Valley, one of the most concentrated leafy green production regions in the world. Lettuce was chosen deliberately: high margin, high density, uniform row configuration, and a geography where you could build relationships with enough growers to get real field time without crossing a dozen different agronomic environments.

What FarmWise learned is that the mechanical weeding problem in lettuce is, in a specific and limited sense, the same problem in broccoli. And in celery. And in artichokes. And eventually in sweet potato, herbs and tomatoes. Most specialty crops in California are grown on similar bed widths which means the physical form factor of the machine doesn’t need to change. What changes is the vision system’s ability to distinguish the target plant from a weed in a new environment.

Software, it turns out, is a lot easier to retrain than hardware is to rebuild.

So the expansion happened incrementally: romaine to other lettuce varieties, then radicchio, then broccoli, then crops requiring progressively more sophisticated perception — tomatoes, sweet potatoes, crops where the plant structure looks fundamentally different from a leafy green. Each new crop extended the TAM without requiring a new machine.

But the ceiling was real. FarmWise couldn’t go into spinach, where the density makes mechanical weeding physically impossible. Couldn’t go into commodity corn, where the margins are too thin to justify the cost. Couldn’t go into orchards or vineyards, where the speed required to operate economically wasn’t there. The hardware form factor eventually constrains you — not immediately, but eventually.

Burro is executing a version of this path from a different starting point and with a different product. Rather than a task-specific robot — weeding, harvesting — Burro built an autonomous mobility platform: a robot that follows workers, tows loads, and navigates field environments so humans don’t have to carry things. They started in table grapes and nurseries, and have expanded into blueberries, blackberries, raspberries, stone fruit, and beyond. The insight driving their expansion is similar to FarmWise’s: the core capability — outdoor autonomous mobility across row-based environments — transfers across crops more cleanly than task-specific hardware does. Rows are rows, more or less, whether the crop is grapes or blueberries.

Burro has also started positioning themselves as a platform — describing their robot as “a physical API for a growing list of technology partners.” That framing matters. It suggests they see their long-term value not just as a mobility robot but as the infrastructure layer on top of which other applications get built. SwarmFarm is making a similar bet with SwarmConnect. Both are pointing toward something that doesn’t fully exist yet in ag robotics but will.

SwarmFarm belongs here too. The same platform that works economically across broadacre grain and cotton in Australia has since expanded into horticulture, turf, and orchards — not by rebuilding the machine, but by extending SwarmConnect to let third-party developers build crop-specific applications on top of it. The hardware travels because it was designed to be simple, lightweight, and crop-agnostic from day one. The software ecosystem does the rest. It's the volume-play version of the same thesis FarmWise and Burro are executing from a margin-first starting point — and it suggests the expand-the-niche path may be more accessible than it looks when the underlying platform is open by design.

Lumo illustrates a subtler version of the same thesis — and one that points toward where the real value in ag automation may ultimately compound.

Lumo started in wine grapes, one of the highest-value and most irrigation-dependent crops in California. Their hardware — precision irrigation valves installed at the vine level — gives growers the ability to manage water remotely, with real-time visibility into what each block of vines is receiving and when. The immediate value proposition is straightforward: save water, save labor, improve yield consistency.

But the more interesting story is what the hardware enables over time. Every valve installed is a data collection point. Every season adds to a picture of how water interacts with yield, with weather, with soil type, with varietal differences. The hardware is the entry point. The data is the moat.

Lumo has since expanded into berries, citrus, fruit orchards — apples and cherries in Washington have been a particularly strong market — and other irrigated crops, following the same expand-the-niche logic as FarmWise and Burro. But with one important structural difference. Irrigation infrastructure, once installed, is deeply embedded. Growers don’t pull out a drip system the way they might switch a software subscription. The switching cost is physical, which means Lumo’s retention economics look more like infrastructure than SaaS even though the revenue model increasingly resembles the latter.

Devon Wright, Lumo’s CEO, puts it this way: “The real compounding value isn’t just in the valves — it’s in the longitudinal data they generate season after season. As AI matures, this data is becoming exponentially more valuable, enabling massively better insights and optimizations tailored to each grower’s specific blocks, soils, and varieties. Once a grower has Lumo infrastructure in the ground, they’re getting meaningfully better recommendations every year. That data moat, combined with the physical switching costs, creates a very different retention profile than software alone.”

The ceiling exists here too — dryland farming and flood-irrigated regions are out of reach without a fundamentally different product. But within the irrigated specialty crop universe, the hardware-enabled SaaS model may be the most capital-efficient path through the expand-the-niche thesis. You’re not rebuilding the machine for every new crop. You’re extending a platform that gets more valuable the more seasons it runs.

The rule that emerges across all three: the software travels. The hardware finds its home.

“The hardware is the entry point. The data is the moat.”

The Platform Question

There’s a path adjacent to both of these that hasn’t fully materialized yet but I expect to become significant over the next five years: the transition from product company to platform.

The companies that execute the go-deep or expand-the-niche strategy well tend to accumulate something more valuable than revenue. They accumulate proprietary data — field images, plant health models, yield correlations, actuation feedback loops — and hardware IP that took years and real money to develop. At some point, that IP becomes potentially more valuable licensed to others than operated in the field.

The automotive industry has a Bosch. It has a Delphi. A tier-one supplier layer that doesn’t sell cars but is present in almost every car made. That layer doesn’t exist yet in ag robotics. We’re still waiting for the Bosch of ag robotics. Or the Android. The platform that doesn’t grow the crop but is present in almost every operation that does. The sector is too young and too fragmented. But I’d expect it to form — and the companies that survive Wave 2 with proven technology and deep field data are the most likely candidates to become it. SwarmFarm’s SwarmConnect and Burro’s platform framing are early signals of what that could look like.

Whether any current company successfully makes that full transition remains to be seen. It requires a different organizational identity than building and operating machines. But the economic logic is there, and the industry will eventually produce it.

“The software travels. The hardware finds its home.”

Four paths. Two covered here — go deep and stay there, or expand the niche until the hardware finds its ceiling. Two more to come, and Part 3 looks at the companies making more radical bets — those expanding into ag adjacent industries, and those making the full leap: that agriculture was the proving ground, not the destination.

Subscribe now

Share

The partner leaned back and said: “This is great. But when are you going to address row crops?”

I’ve thought about that question for years.

Because it revealed something about how investors — and founders — misread agriculture. The assumption embedded in that question is that agriculture is one market, that success in specialty crops is a stepping stone to the real prize, and that the real prize is corn, cotton, wheat, and soy — the hundreds of millions of acres that dwarf anything happening in berries, vines, orchards or leafy greens.

The assumption is wrong. And it’s killed more ag robotics companies than bad technology ever did.

Agriculture is enormous. Depending on how you measure it, the global food and agriculture economy touches somewhere between $8 and $10 trillion annually. Machine sales alone — tractors, harvesters, sprayers, planters — run close to $150-200 billion a year globally. In the United States, farm gate receipts regularly exceed $500 billion. The industry is recession-resistant in a way almost nothing else is. People need to eat in recessions. Crops go in the ground regardless of what the Fed does with interest rates. To add to the opportunity, agriculture is undergoing several existential shifts: labor is disappearing and aging, wages are increasing substantially, weeds are becoming more resistant (while new chemicals lag), water challenges and general accelerating input costs. By any measure, this is a category worth building in, with real problems to be solved.

It’s also, however, one of the most fragmented markets on earth.

Fragmentation in agriculture isn’t just about crop types, though the crop type problem alone is severe. A strawberry harvesting platform is not a lettuce harvesting platform. A precision sprayer calibrated for onions doesn’t transfer to tree crops. A weeding robot built for the flat, 80” uniform rows of a California leafy greens operation looks nothing like what you’d need in an irregular specialty crop environment in the Southeast, or the muck in Florida, or flood irrigated Arizona. The technology doesn’t transfer. The agronomic knowledge doesn’t transfer. The customer relationships — which in agriculture run deep, run local, and on decades of trust — definitely don’t transfer.

Fragmentation also runs along geography, growing season, farm size, labor model, equipment compatibility, and regulatory environment. What works in the Central Valley doesn’t work in the Midwest. What works for a 2,000-acre operation doesn’t work for a 200-acre family farm. What growers in one region consider an acceptable return on technology investment is completely foreign to growers in another.

The result is that agriculture’s trillion-dollar headline number dissolves, when you get close to it, into hundreds of micro-markets — some worth $500 million, some worth $50 million, most requiring a fundamentally different product, go-to-market, and customer relationship to access.

Back to that investor meeting.

Row crops — corn, cotton, wheat, soy — sound like the obvious answer to the fragmentation problem. Hundreds of millions of acres centralized in a few core geos with expansive farms. Mechanically intensive operations that already spend heavily on equipment. A buyer profile that’s often more financially sophisticated than small specialty crop growers. On paper, it’s compelling.

In practice, row crops are among the most mechanically efficient agriculture segments on earth. Decades of investment from John Deere, CNH, AGCO, and their suppliers have driven per-acre economics to a place where the margin available for new technology is thin. The operations are large, but they’re also heavily consolidated and extremely price-sensitive. The problems that remain unsolved are real, but they’re genuinely hard — and the incumbents aren’t sleeping.

Specialty crops are the opposite. Labor-intensive, high-value and largely untouched by the mechanization wave that swept row crops. A working robot in strawberries or lettuce is going into a market that has almost no mechanical competition. The acres are fewer but the margin available for technology is significant.

We knew this at FarmWise. The investors who passed because we weren’t chasing row crops were, in a specific and measurable sense, wrong about where the near-term opportunity was. But they weren’t wrong about the underlying constraint: specialty crops, however high-margin, have a ceiling. And VC fund structures aren’t built for ceilings.

This is the trap.

Agriculture is big enough to build real businesses. It may not be big enough — in the right configuration — to produce the kind of exits that justify large VC fund economics. And the companies that raise as though it is, without resolving that tension in their capital structure, are the ones that end up in the gap: too big to be acquired cheaply, too small to IPO, too specialized to expand.

The question isn’t whether agriculture is a good industry. It is. The question is whether the way most agtech companies have capitalized themselves matches the actual structure of the market they’re operating in.

The answer, for most of the first wave of AgTech, was no.

In the pieces that follow, we’ll look at how a handful of companies are navigating this differently — and what their choices reveal about the four or five distinct paths available to ag robotics companies right now.

Some are doubling down on niche ownership, building deep in a single crop type with capital structures designed for patience rather than scale. Some are attempting platform expansion — moving across crop types or into adjacent sectors — betting that their core technology is more portable than it appears. Some are making a more radical bet: that ag robotics isn’t really ag at all, that what they’ve built is hardened field robotics capable of operating in mining, construction, and defense, and that agriculture was the proving ground rather than the destination.

Each path has merit. Each has failure modes. And most companies pursuing them don’t realize, at the moment they raise, which path they’ve actually chosen.

That misalignment — between the bet a company is implicitly making and the capital structure it’s raised to support — is where most agtech companies are lost. Not in the field. Not in technology. In the gap between what they told investors and what the market actually rewards.

Next week I look at two companies that chose different paths. One went deeper into the niche. The other bet that what they built wasn’t really ag at all.

The trillion-dollar number is real. The trap is real too.

Subscribe now

Part 3: Where the exits are forming and what founders building today should take from it

AGCO didn’t exist until 1990. It was assembled through 21 acquisitions in its first fifteen years, growing from a few hundred million in revenue to over $5B. CNH Industrial is a merger of Case and New Holland, themselves the products of decades of consolidation. The entire Big 4 was manufactured through exactly this dynamic — fragmented early innovation, a brutal shake-out, a handful of survivors becoming acquirers, and those acquirers eventually becoming the incumbent infrastructure.

We are in the beginning of the shake-out. The dataset proves it.

The exit map is wider than it looks

Today’s exit market runs almost entirely through one buyer type: the legacy OEM. John Deere has acquired Blue River, Bear Flag, Mineral’s IP, and GUSS Automation. Kubota has been an active strategic investor through its Innovation Center, taking equity in Agtonomy and acquiring Bloomfield Robotics. Yamaha Motor acquired Robotics Plus and The Yield in February 2025 and built an entirely new ag division around it. SDF Group acquired VitiBot. CNH and AGCO are taking minority stakes and forming partnerships but haven’t pulled the trigger on full acquisitions yet. That’s a handful of potential buyers — with John Deere as essentially the only one operating at meaningful capacity.

However, the map is expanding in four ways that aren’t obvious yet.

Champions are forming on the startup side. Carbon at $100M revenue with M&A board experience is beginning to look like an acquirer, not just a target. Their March 2026 hire of Kevan Krysler as CFO — formerly CFO of Everpure, with senior finance roles at VMware and KPMG behind him — reads as deliberate positioning. That’s a public-company finance profile being applied to a hardware company. Either the economics now justify it, or the company is signaling what it wants the next chapter to look like. Bonsai Robotics acquired farm-ng in July 2025 — relatively small, startup-to-startup, but the pattern is forming. On the software and data layer, CropX is already explicitly executing a roll-up: seven acquisitions, conversations with over 200 targets, what their CEO calls “an M&A machine.” The robotics side doesn’t have a CropX yet. It probably will within five years.

Specialist capital is building structurally different vehicles. Rolling funds, evergreen models, non-dilutive offerings, milestone-based deployment, on-farm validation. Reservoir’s incubator model in Salinas is one of the more visible examples. Farmhand Ventures’ redeemable equity is another. These funds aren’t trying to force ag robotics into a software-shaped return profile. They’re building return profiles that match the sector’s actual cycles.

End-users are becoming acquirers. Taylor Farms bought FarmWise not because they wanted to be in the robotics business but because the technology was too embedded in their operations to let die. Oishii bought Tortuga’s IP for the same reason. This is vertical integration by necessity, not strategy. It’s a thin lane — you essentially have to become indispensable to one large customer before they acquire — but it’s real.

Private equity is the lane likely soon to activate. Profitable single-crop robotics companies — strawberries, wine grapes, organic lettuce — assembled into portfolios at patient return expectations is a different kind of exit entirely. It hasn’t happened at scale. It will.

The sector isn’t broken. It’s early. And early looks like wreckage before it looks like infrastructure.

If you’re building today

The question isn’t VC or bootstrap. The question is whether your problem requires scale to prove or depth to prove.

Laser weeding across 100 crop types globally requires manufacturing scale, international distribution, and years of dataset building before the unit economics work. You cannot bootstrap that expansion. It’s a real VC-scale problem, and Carbon solved it the right way with the right founder. TRIC Robotics — deploying UV-C light and bug vacuums to eliminate mites, mildew, and mold in strawberry fields, rooted in a decade of USDA research, expanding from nine to fifty units on a $5.5M seed round — is a different kind of problem. Narrow, deep, capital-efficient, building toward profitability in a defensible niche before expanding.

Most problems in ag robotics right now look more like TRIC than Carbon. They don’t need $50M to prove. They need one crop, one geography, a few deeply trusted grower relationships, and the patience to get the machine right before expanding.

Get the depth-vs-scale question wrong and no amount of capital — or lack of it — saves you.

Four patterns I watched play out at FarmWise and across the dataset. They show up regardless of technology or funding level.

The pilot is not the business. Guardian Agriculture had FAA approval, commercial traction, and a solid team. Monarch raised over $220M, a record Series C, and a manufacturing partner who sold the factory to build AI data centers. Same pattern in both cases: the story scaled faster than the economics. If your growers won’t pay for the machine today, investor capital doesn’t fix that. It delays the reckoning and makes it more expensive.

The prototype is not the GTM. Winning the technical sale to the agronomist is not the same as winning the economic sale to the farm owner. The robot working in the field is not the same as the robot working in the business. Getting to a signed multi-year contract that survives the first breakdown, the first bad season, and the first competitor offering a cheaper demo — that’s the GTM motion. Most companies never got there. Most confused grower enthusiasm for commercial traction and raised their next round on that confusion.

Plan to the exit map. Given VC fund dynamics, a $25M raise implicitly promises an acquisition of $200M+. The size of the raise sets the floor on what success has to look like. A profitable, capital-efficient company that exits for $40M is a great outcome on $5M raised and a disaster on $50M raised. In this sector, the buyer pool that can absorb a $200M+ exit is small, and most of those buyers aren’t actively acquiring yet. Build OEM relationships early — not just as a channel but as a future buyer. Let them touch the technology before they need to. Let them see grower traction before you pitch them. You’re not selling today. You’re pre-selling to the buyer your investors will eventually need.

The season is not a metaphor. I watched this one up close. Every hardware revision that takes three months longer than expected, every pilot extension, every fundraise that slips — these don’t cost you three months in ag. They cost you a season plus three months. The founders who survived understood it going in and built their entire operating cadence around it. The ones who didn’t kept getting surprised by the same thing, then blamed the sector for being slow.

Closing

The first wave of AgTech produced a lot of wreckage, a few winners, and the conditions for something better. The capital structure was wrong for most of it. The exit market was too narrow for almost all of it. And the founders who optimized for growth over profitability, for story over economics, mostly didn’t make it.

The current wave gets to learn from that. Capital-efficient founders building to profitability in defensible niches. Champions emerging with the scale to become acquirers. OEMs waking up to what John Deere has quietly been doing for ten years. End-users embedding the technology so deeply they’ll buy the company rather than lose the machine.

The investors who stayed in the sector through the capital drought are saying quietly that the turn is coming and will come fast. The founders who survived to meet it will have an advantage no amount of late capital can buy.

Early looks like wreckage before it looks like infrastructure.

Daniel Kirstein is the founder of Holdfast Partners, a financial architecture and commercialization consultancy focused on robotics and hardtech founders. He has worked with companies including FarmWise and brings firsthand experience navigating the valley of death between prototype and scale. He writes about hardtech commercialization on LinkedIn and Substack.

Part 2 is about why the capital was wrong — and what the two companies that survived did differently.

Why the capital was wrong

Venture capital was designed for a specific shape of company. Fast product cycles. Near-zero marginal cost at scale. Time-to-revenue measured in months. The fund lifecycle — typically ten years with a three-year deployment window — was calibrated for software companies that can grow from zero to meaningful revenue in eighteen months.

Ag robotics hands you the opposite of almost every one of those assumptions.

If you miss a planting season, you’re waiting six to nine months for the next commercial window. Hardware revisions that take three months longer than expected don’t cost you three months — they cost you a season plus three months. Getting a robot genuinely integrated with a production operation takes years of trust-building, not a pilot. And the capital that’s supposed to be patient has investors who need to mark their portfolios and raise their next fund.

The mismatch isn’t a failure of investor judgment. The VCs who backed these companies weren’t irrational — they were using the right tool for the wrong job. And some of the best-intentioned founders in the sector raised the most money and built the most ambitious plans, and those are exactly the companies that blew up the hardest.

But capital alone doesn’t explain the graveyard. The other half of the failure was commercialization — specifically the gap between a robot that works in a pilot and a robot that works in a production operation at scale. Almost every company in this dataset died with successful technology that was providing real value to customers. FarmWise had Taylor Farms as a customer for years before Taylor Farms became the acquirer. Guardian had $100M in letters of intent. Monarch had enterprise partnerships and had sold units to large reputable customers. The technology worked in controlled conditions. What didn’t work was the path from controlled conditions to a signed multi-year contract with a grower who’s putting their crop at risk and needs local service infrastructure when something breaks at 2am during harvest. VC capital accelerated into that gap rather than waiting for founders to solve it. The companies that survived used capital to scale what was already working. The ones that didn’t used capital to buy time, and time ran out.

The patterns repeat across the dataset: a service model that transfers all capital intensity to the startup while delivering thin margins; a pricing architecture that doesn’t survive the first renewal conversation; a capital stack borrowed from software applied to hardware with completely different working capital cycles. These aren’t random failures. They’re the same traps, company after company.

No company illustrates the compounding nature of both failures better than Monarch. They raised $133M in a single Series C round in July 2024 — bringing total capital raised to over $220M. The narrative was perfect — first fully electric driver-optional tractor, record round, global expansion ahead. What followed was a case study in compounding hardware risk. Foxconn, Monarch’s contract manufacturer, sold its Ohio facility to a SoftBank affiliate to build AI data centers, abruptly abandoning its electric vehicle manufacturing commitments and leaving Monarch without a production partner. Simultaneously, an Idaho dealership called Burks Tractor filed a federal lawsuit after ten tractors purchased for nearly $800,000 failed to operate autonomously as advertised — breach of contract, misrepresentation of the “driver-optional” software. Manufacturing partner gone, dealer lawsuits mounting, demand collapsing in the California wine market. Monarch entered an assignment for the benefit of creditors and in April 2026 Caterpillar acquired its assets — the software-defined vehicle platform, perception stack, and electrification systems — to deploy primarily in construction and industrial machinery. The ag use case is no longer central. The problem wasn't just the tractor. It was what $220M in capital demands — aggressive expansion into hardware and market risks that no amount of funding could eliminate.

The winners and what made them different

Two companies from the first wave of Ag Robotics have the profile of genuine long-term winners. They took different paths. What they share tells you more than either story alone.

Carbon Robotics just crossed $100M in revenue. Its founder Paul Mikesell previously built Isilon Systems which went public and was later sold to EMC for $2.5B. He understood VC cycles, knew how to sequence a raise, and brought a go-to-market playbook the sector had never seen. Carbon’s pre-order model — structured like a consumer launch dropped into a B2B industry still running bespoke one-grower pilots — generated hype, a waitlist, and a story. Carbon is now appointing board members with M&A expertise. That’s not just a governance move. That’s positioning and a potential sign of things to come.

The LaserWeeder wasn’t perfect when it launched. What was perfect was everything around it.

Carbon priced on value, not cost. Weeding in high-density crops where hand labor runs $1,000+ per acre gave them room most founders never take. Eyebrows went up when Carbon crossed the $1M machine price. In retrospect it was the right call — it captured the economics the technology actually created rather than leaving them on the table.

But the more important move was operational. Carbon was intentionally in the field daily. Not quarterly check-ins, not remote monitoring — physically present, treating early customers like design partners. The kind of hypercare that costs margin and doesn’t scale, but buys you iteration cycles and grower loyalty that no amount of capital can purchase after the fact. When something wasn’t right, they knew before the customer did. That’s how you iterate through imperfection without losing the relationships that matter.

Demand momentum on the front end. Obsessive customer presence on the back end. That combination is what $100M in revenue was built on.

Ecorobotix took thirteen years to get there. Founded in Switzerland in 2011, they’ve raised $230M in total — including $150M across Series C and D rounds in 2024 and 2025 — and now operate in over 20 countries. Their precision spraying technology reduces pesticide use by up to 95%. Thirteen years of patient capital, mostly from European investors and early grants, with return horizons a US VC fund wouldn’t typically tolerate.

Where Carbon moved fast and loud, Eco was more methodical. European farm equipment tends to be lighter and less over-engineered than its American equivalent — and meaningfully cheaper. Eco entered markets as a fast, precise spray machine that growers could actually afford to try. Besides price and accuracy their key differentiation was their speed at up to 10 acres per hour. The speed made precision application viable at scale for the first time in broad-acre crops like onions, opening a market that hadn’t existed before.

Additionally, Eco’s approach to market expansion utilized a different playbook. In the US, Eco didn’t initially build a sales team. They worked with domain experts, who gave them early market penetration and proof of fit before they committed resources.

Carbon and Eco don’t look like the same company. But underneath the different tempos is the same discipline: they priced with intent, differentiated on something real, and didn’t expand until they’d earned the right to. Everything else — the pre-order list, the partner GTM, the hypercare, the thirteen years — follows from that.

In Closing

Part 3 next week closes the series: where the exits are forming, how the sector consolidates from here, and what founders building today should take from the first wave's lessons. If any of this is hitting close to home, DM me — always up for the conversation.

Everyone in ag robotics knows the narrative. High failure rates. Capital-intensive hardware. Companies burning through runway before they ever see a commercial season. The technology is promising with a compelling market need, but the graveyard is growing

I spent years inside an ag robotics company - through a significant raise, through commercial launch, through the specific experience of watching what happens when a venture capital timeline tries to coexist with an agricultural one. What I saw wasn’t technology failing. It was a capital structure that was never designed for this sector getting applied to it anyway, running into a commercialization pathway that required patience and the predictable damage that followed.

The robots mostly worked. The money mostly didn’t.

What the data actually shows

I tracked nearly 50 VC-backed ag robotics companies founded between 2011 and 2023 - every company I could find with meaningful funding in the sector. The scorecard is not what the “high failure rate” narrative suggests, but it’s not good either.

I focused this dataset on AgTech because agriculture is the most unforgiving crucible for robotics commercialization. It offers a massive market with a dense concentration of startups, but with a unique brutality: the financial traps that take three years to kill a warehouse robotics company will kill an AgTech company in six months. If a capital structure can survive the agricultural timeline, it can survive anywhere.

Five companies achieved clean strategic exits at real multiples. Blue River Technology and Bear Flag Robotics were both acquired by John Deere - in 2017 for $305M and in 2021 for $250M respectively. GUSS Automation followed in August 2025, a largely bootstrapped company that had deployed 250+ machines across 2.6 million acres before Deere finalized the deal. Yamaha Motor acquired New Zealand-based Robotics Plus in February 2025, forming an entirely new agricultural division around it. SDF Group — the European manufacturer behind Same Deutz-Fahr and Lamborghini Tractori — acquired French vineyard robotics specialist VitiBot in 2022. All five exits went to OEMs. Not a single financial exit. That tells you something about who actually values this technology.

Seven companies are actively scaling: Carbon Robotics, Ecorobotix, Niqo Robotics, Burro, Bonsai, Aigen, and SwarmFarm. Another cluster — Four Growers, Agtonomy, Dogtooth, Saga Robotics, TRIC — are building real commercial traction in narrow niches with capital-efficient models. Four Growers is notable: John Deere is already a participant in their business through the Startup Collaborator program — the same relationship Deere built with Bear Flag before acquiring it for $250M. These are the companies Wave 2 is being built on.

Then there are the dead. Abundant Robotics shut down in 2021, IP salvaged for parts. Guardian Agriculture had FAA approval, Time magazine recognition, and a paying customer when it closed in August 2025. Mineral — Alphabet’s ag robotics moonshot with essentially unlimited resources — wound down in 2024, licensing its IP to Driscoll’s and John Deere. Small Robot Company closed in February 2024 with a signed term sheet in hand. The investment failed to materialize before runway ran out. Monarch Tractor requires its own sentence: they raised $220M, the largest ag robotics round on record, lost their contract manufacturer when Foxconn sold the Ohio facility to build AI data centers, faced federal dealer lawsuits over autonomy claims that didn’t match reality, and in April 2026 had its assets acquired by Caterpillar for deployment primarily in construction and industrial use.

Then there are the ones that don’t fit cleanly into any column. FarmWise raised $65M and was acquired by its own customer, Taylor Farms. Tortuga AgTech raised $27.7M, went four years without a raise, and quietly sold its IP to Oishii. Naio Technologies raised $62M, watched revenue fall from €3.96M in 2021 to €2.4M in 2024, and entered judicial receivership before being acquired by a new owner. Root AI was acquired by AppHarvest for $60M — and then AppHarvest itself filed for bankruptcy with $341M in debts, liquidating everything with it. Traptic was acquired by Bowery Farming — and then Bowery went bankrupt. Two separate cascading chains where the technology outlasted neither the startup that built it nor the company that bought it.

Harvest CROO offers a separate pattern worth naming — a company that spent over a decade operating outside the VC fundraise cycle on under $6M in traditional capital, backed substantially by grower-aligned financing. In April 2025, after twelve years of focused development on one of the hardest automation problems in agriculture, the company declared commercial viability with field trials demonstrating performance on par with human strawberry harvesting. That trajectory sits outside the VC-driven narrative entirely — and it’s a pattern worth studying separately.

And there are the companies, still showing up at trade shows, still posting occasional content, but not raising and not scaling. Stout developed real commercial traction but has drifted without leadership (though a new CEO was recently brought onboard) or a raise. Sabanto deployed 100+ autonomous tractors and still can’t raise. Fieldwork Robotics has Driscoll’s as a partner and funded through convertible notes at 20% interest with crowdfunding language showing four months of runway at minimum. This category is where this sector’s particular cruelty lives — companies with real technology, real customers, and limited viable paths to the capital they need to finish the job.

Five clean exits. Nine dead or bankrupt. Five distressed. Several grinding it out. The rest building quietly in niches the market hasn’t fully found yet.

That’s not a failure rate story. That’s a structural story.

Next week, in Part 2, I break down the exact financial architecture and operational pivots that allowed the few Wave 1 winners to survive and scale. If you are actively raising a round right now and want to read Part 2 before I publish it next Thursday, shoot me a DM and I’ll send you the draft.