What is a deep tech startup?

A deep tech startup is built on a substantial scientific or engineering advance rather than on a business model innovation, so it carries technical risk, the thing might not work, stacked on top of the market risk every startup carries. That definition is accurate and it misleads founders, because it implies the defining activity is invention. Most deep tech companies being built now do something different: they take a technology that matured in one industry and point it at an industry that never received it. Heron Power applies four decades of power semiconductor progress to grid transformers. Mariana Minerals applies the same class of optimization algorithms that route ride-share fleets to refinery control. The invention is borrowed. The difficulty is the physical build, the supply chain, the labor pool, and the calendar.

What is the difference between deep tech and high tech?

High tech describes a product that uses advanced technology. Deep tech describes a company whose core bet is that a hard technical or engineering problem can be solved at all, or at least solved at production scale and cost. A SaaS analytics tool is high tech and carries almost no technical risk, because everyone knows it can be built. A solid-state transformer factory or a new mineral refining process is deep tech, because the open question is whether the thing works at rate. The practical difference shows up in the schedule and the balance sheet. High tech companies iterate in weeks on software budgets. Deep tech companies iterate in quarters or years on capital budgets, and their advantage usually comes from being faster at that physical loop than incumbents, not from having better code.

Is AI a deep tech startup category?

Partly. Training frontier models is deep tech by any definition: it is capital intensive, the results are uncertain in advance, and the constraint is physical, namely compute, electricity, and land. Building an application on top of an existing model is not deep tech, because the technical risk was absorbed by someone else and the remaining risk is whether customers want it. The interesting middle is where most founders should look: applying mature AI to an industry that never received it. That inherits a real physical bottleneck, a grid, a plant, a mine, a factory floor, without requiring you to fund a frontier lab, and the defensibility comes from the operational access and data you earn rather than from the model.

How much funding does a deep tech startup need?

More than a software company, and the gap is widening rather than closing. AI has made the software half of an ambitious company genuinely cheap, so founders can now attack problems that used to need a corporate laboratory, but it has done nothing to the cost of steel, land, buildings, or long-lead equipment. The effect is that the non-software share of the budget becomes nearly all of it. Real 2026 reference points: Heron Power raised 38 million dollars in a Series A and 140 million dollars in a Series B before its first factory reaches mass production, and Y Combinator has described funding a small nuclear reactor company expected to need roughly 800 million dollars in a single year. Plan for capital to scale with the ambition of the physical problem, not with headcount.

Deep Tech Startups: The Clock Is the Moat

Search for deep tech startups and you get definitions, directories, and lists of companies. What you do not get is the thing a founder actually needs: what changes about running a company when the bottleneck stops being code and becomes a building, a permit, and a supply chain.

An a16z American Dynamism session put two founders on stage who are living that answer. Turner Caldwell is co-founder and CEO of Mariana Minerals, which mines and refines critical minerals and runs its own sites. Drew Baglino is founder and CEO of Heron Power, which builds solid-state transformers for the electrical grid. Caldwell led Tesla's minerals and metals work. Baglino spent 18 years at Tesla and left in April 2024 as senior vice president of powertrain and energy engineering.

One disclosure, because nobody on page one of this search makes it: a16z hosted the conversation and is an investor in both companies, having co-led Heron's 140 million dollar Series B and led Mariana's Series A, which brought that company's total raised to 85 million dollars.

The session opens by framing the constraint on America's AI future as "in many ways atoms and not algorithms." That is the right diagnosis and slightly the wrong takeaway, because it implies the hard part is the atoms. Listen to what these two actually describe and a different shape appears. Neither invented the core technology his company runs on. Both borrowed a mature one and carried it across an industry line nobody had crossed. Everything hard came after that.

What a deep tech startup actually is

The textbook answer: a deep tech startup is built on a substantial scientific or engineering advance rather than on a business model, which means it carries technical risk, the thing might not work, stacked on top of the market risk every startup carries.

True, and not very useful, because it implies the defining activity is invention. The deep tech companies getting built right now look more like this: a technology that matured somewhere else, pointed at an industry that never received it, run at a clock speed that industry has never run at.

Baglino describes his opening in exactly those terms. Power transistors improved for four decades in parallel with the Moore's law curve in compute, and that improvement reached almost everything: phone chargers, telecommunications, data centers. Almost.

even though there's so much innovation happening at the edge of the grid, on the other side of the wire, there's really been no change

And the other side of that wire is old:

the systems underpinning the grid today are the same largely mechanical systems that were developed over 100 years ago

The consequence he names is the part worth stealing, because it describes a hundred industries and not just this one. Without control and without monitoring, he says, "you end up with an overbuilt system that is fragile." Overbuilt and fragile at the same time is the signature of an industry that solved its reliability problem with steel instead of software, decades before software was an option. Heron's product is that arbitrage in one line. Baglino describes it as building solid-state transformers to "use silicon and software to replace steel, oil, and copper in power conversion."

Caldwell's version is less obvious, because minerals do not sound like a software story. But the technology he is carrying across the line is equally borrowed, and he says so when asked where the talent comes from:

a lot of the underlying optimization algorithms that we're writing for our plants, they look very very similar to the optimization algorithms that are in dog walking apps and Uber ride optimization, underwriting loans, ad optimization

That is the tell. The math that schedules a refinery is the math that schedules a delivery fleet. It was not invented for minerals, it was imported. Mariana staffs roughly a quarter of the company with software and machine learning engineers, and then makes the move that confuses people: "we do not sell software. We are not a SaaS company." It engineers, builds, and operates its own projects. We wrote up why that choice is forced rather than ambitious in AI in mining.

So the useful entry rule for a deep tech startup is not to invent something nobody has. It is to find a mature technology curve, find the industry it never reached, and then be honest that your real competition is whatever kept it from reaching there.

The moat is the clock

Ask what stops new mineral supply in the United States and the standard answer is permitting. Caldwell's answer is more uncomfortable, because it survives the policy fix:

we are too slow at designing, building, and ramping up new minerals capacity, even after we have license to operate

He puts numbers on it. Once you start building, he says, "it can take 5 years to get something built, and then it can take three to five years to get something actually operating at rate." So the clock from approval to full production runs roughly 8 to 10 years, and none of that is regulatory. It is engineering, procurement, construction, and the long grind of getting a plant to hit spec.

That is where the moat is. If the whole industry takes 8 to 10 years from permit to rate, a company that takes four is not 2x better, it is running a different business. It gets twice the shots on goal, compounds its learning twice as fast, and can underwrite projects its competitors cannot finance.

Baglino has done the fast version at least once. He describes building Tesla's Lathrop megafactory with his team in 11 months, from a JCPenney warehouse to first product off the line. And the interesting thing is that Heron is running the same play again rather than treating it as a one-off: its first factory is a repurposed 286,000 square foot distribution warehouse in Morgan Hill, California, targeting mass production in late 2027.

Notice what is being optimized. Not the building, which is bought used. Not the technology, which is a mature semiconductor. The thing being compressed is the calendar, and the way to compress it is to refuse to originate anything you can inherit.

For a founder that generalizes cleanly, and it generalizes past deep tech. In any business where the unit of progress is physical, list the things on your critical path that you plan to create from scratch, then ask of each one whether a used version exists. Mariana did the extreme version of this by buying an already permitted copper operation whose refinery had been producing cathode since 2009, instead of developing one. Permitting is the slowest item on a mine's path, so it bought the item rather than building it.

Labor is not why it costs more here

The standard explanation for why hard things do not get built in the United States is wage cost. Baglino, who has actually built the factories, does not accept the premise. Modern factories are heavily automated, and on a new build he puts the labor gap between China and the United States at "less than 10% of cost of goods sold," possibly under five percent. Treat that as his working estimate from building at scale rather than as a published figure, because it is exactly the kind of number that varies by product.

His claim about what does drive competitiveness is more actionable, and the thing he is describing is China:

everything that you could possibly need to build a car, which has 7,000 parts in it, is within less than a 3-hour drive

Co-location, not wages, and he is explicit that getting to that kind of co-location in the United States would be the unlock. The cost that kills you is logistics time, and logistics time is a function of how far your suppliers are from your line. A three hour radius means a bad part gets replaced the same day. An ocean freight radius means a bad part is a quarter.

This is the least glamorous and most decision-relevant fact in the whole conversation, and it has a direct implication for where you site a deep tech company: pick the location by supplier density first and by incentives second. An incentive package is a one time subsidy. A supplier three hours away is a permanent reduction in your iteration time, which by the previous section is the actual moat.

The honest caveat is that supplier density in the United States is thin in precisely the industries people want to rebuild, which is why both founders spent their one policy ask on it. Caldwell wants the incentive structure that mobilizes private capital behind these projects for long enough that "the rug isn't going to get pulled out from under them." Baglino wants "durable industrial policy that you can plan around," plus federal and state coordination so manufacturing and energy build out in the same places. His most concrete idea is a federal highway trust fund for the grid, which does not exist today, and which he argues is why transmission is a patchwork.

There is no phone tree

The hiring problem in a deep tech startup is not that the people are expensive. It is that the category of person you need may not currently exist in a hireable pool. Baglino:

I can't just go to like a phone tree of power electronics manufacturing engineers or production associates

His answer is to hire from analog industries, and the specifics are better than the principle. Staffing what he describes as Tesla's 4680 manufacturing facility, a 50 gigawatt hour battery plant in Texas, at a moment when there were few battery operations in the country to poach from, he went sideways:

I was hiring people out of high-speed bottling plants and out of syringe manufacturing facilities where they're making billions of syringes

High-speed bottling and syringe manufacturing are not batteries. They are high volume, high precision, continuous-flow production, which is the actual skill. Caldwell runs the same method in a sector that has had 35 years of attrition in its labor pool: he points at oil and gas as the reservoir of operating talent, and at the optimization work happening in consumer apps for the software roles.

The method is worth writing down as a hiring exercise, because it works for any scarce role. Describe the job as a set of physical or mathematical properties rather than as an industry title, then ask which other industry has those properties at volume. Titles do not transfer. Properties do.

There is a second-order point here that most founders get wrong, and Caldwell makes it explicitly. Having good software engineers is not what determines whether software gets adopted in an industrial operation:

what sets the rate of software penetration and technology penetration in these plants and in these mines ultimately is the operating teams

His structural fix is to seat the software engineers next to the operating teams, and he is careful to distinguish it from the model it resembles:

sitting the software engineers right next to the operating teams, but not in like a forward deploy engineer type way where everyone has the same incentives is what's going to yield the best results

That distinction is sharper than it sounds. A forward deployed engineer is embedded with a customer whose incentives are not yours: they want their problem solved, you want a general product. When you own the operation, the embedded engineer and the operator are paid by the same outcome, and the software gets designed for the person who has to use it rather than for the person who bought it.

The jurisdiction is part of your stack

Asked what it takes to build in the United States, Baglino does not talk about cost or code. He talks about whether the people reviewing your project want it to exist:

when you're working with your local jurisdiction, they can use the process for a code compliant project to say no at every step, or they can say yes at every step

Read that twice, because it is a statement about a project that is already compliant. Compliance is not the variable. Disposition is. The same rulebook, applied by a reviewer looking for a path to yes, produces a timeline that a reviewer looking for a reason to say no will never produce.

This is not confined to heavy industry. Y Combinator's request for compute at sea opens by observing that "Artificial intelligence is running out of compute and data centers are running out of electricity and land," and its stated reason for looking offshore is that new data centers take years for approval and can still be killed by local government intervention. When people propose putting server farms in the ocean to escape a permitting posture, local disposition has stopped being a footnote and become a design input.

So treat it as one. Before you sign a lease, talk to the jurisdiction the way you would qualify a customer: who reviews this, what have they approved recently, what did they block and why. The answer is worth more to your schedule than most of the engineering decisions you will agonize over, and the schedule is the moat.

The security argument for building it here

The line from this conversation that a security person notices is Baglino's, about who supplies the equipment the grid runs on:

there's not a lot of suppliers providing that equipment, and most of them are actually headquartered overseas, and that just doesn't seem like a secure position for such critical infrastructure for us to have here in the United States

Strip the geopolitics and what is left is a concentration risk finding of the kind any security review produces. A small number of suppliers, most outside the jurisdiction that depends on them, for a component with a multi-year replacement lead time. That is a single point of failure whose failure mode is measured in years, and no amount of network segmentation downstream compensates for it.

Two working conclusions for founders, and they point in opposite directions.

If you are building the hardware: supply concentration is a real, budgeted buyer pain and not just a talking point. Utilities, data center operators, and defense-adjacent buyers typically have risk functions that already track supplier concentration, which is our read rather than a claim from the session. A second qualified source inside the jurisdiction is a purchase justification you can write down, which matters because it gets you evaluated by a different budget than the one that funds incremental efficiency.

If you are building the software that sits on it: you inherit the security review of the thing you attach to. Anything touching grid control, plant control, or mine operations lands in operational technology, where the segmentation standards are strict, where a cloud call from the control layer is often simply disallowed, and where there is no staging copy of a running line. Decide early where inference runs and what leaves the site, because that is an architecture decision and it stops being cheap once the first deployment exists. The same constraint shows up in AI for manufacturing and across physical AI.

Should you actually start one

The honest case against is capital. Y Combinator's What Actually Makes A Startup Durable makes the point that "what AI is doing is it's enabling founders to attack harder problems," and then follows it to its conclusion: as the problems get more ambitious the capital required rises to match. The example given is a small nuclear reactor company that will need to raise something like 800 million dollars in a year, because "it's really hard to build nuclear reactors without tons and tons of money."

That is the trade. AI has made the software half of an ambitious company genuinely cheap, which is why founders can now credibly attack problems that used to require a corporate laboratory. It has done nothing to the cost of steel, land, or a transformer. So the cheapening of software does not make deep tech cheap, it makes the non-software share of your budget close to all of it.

The second thing to be honest about is duration. Caldwell's description of why autonomy has repeatedly failed in mining is not technical:

folks will give it a shot for a year

A year of effort against a ten year industry clock is a rounding error, and the graveyard of deep tech is full of technically correct projects that were abandoned on schedule. The trait he credits from Tesla is the willingness to keep going when the outcome justifies it, which sounds like a platitude until you price it: it means committing to a problem for longer than one funding cycle, one hype cycle, and one job market.

Baglino adds the version that is specific to startups rather than to Tesla. Many times in Tesla's history, he says, whether the paycheck would clear was bet on the team executing well, and that reality "exists uniquely within startups." It is also the reason legacy industrial companies rarely produce this behavior. A conglomerate selling the same product it sold decades ago has no mechanism that makes a team feel that.

If you are a software founder reading this and the pull is real, there is a smaller version of the bet available. You do not have to buy a mine. You can pick the industry whose technology curve never arrived, spend a quarter embedded with the people who operate it, and build the coordination layer for one narrow part of their work. That is the same thesis at a hundredth of the capital, and it is how most people should test whether they actually want a loop measured in seasons. The wider map of where this sits alongside product, pricing, and team is in our guide to AI for startups.

What to do this week

Find the curve that never arrived. Name one technology that has been improving for a decade in your adjacent industry and is absent in your target one. If you cannot name it, you are proposing invention, which is a different and more expensive business.
Write down your clock. From commitment to a working unit at full rate, how long, and what is the industry norm. If you are not at least twice as fast, you do not yet have a reason to exist.
Audit your critical path for things you plan to originate. Building, permit, supply agreement, certification. For each one, find out whether a used version can be bought. Inheriting beats originating on every line where it is possible.
Pick your site by supplier radius. Map your top ten inputs and measure the drive time. Do this before you evaluate any incentive package, because the radius is permanent and the subsidy is not.
Run the analog-industry hiring exercise. Rewrite your hardest open role as physical or mathematical properties instead of a job title, then list three industries that have those properties at volume. Source from there.
Qualify the jurisdiction like a customer. Before any lease or land commitment, find out who reviews your project, what they have approved in the last two years, and what they blocked.
Decide where inference runs, on paper, now. If your software will touch operational equipment, write down what data leaves the site and what happens on a lost network or a stale model, while it is still a document and not a retrofit.

Working out which layer to enter, what to own, and how fast your loop can actually run is the operating question underneath all of this, and it is what the AI Operating System for Startups course is built around.

Sources

The Founders Who Left Tesla to Rebuild America (a16z), the conversation this article distills, with Turner Caldwell of Mariana Minerals and Drew Baglino of Heron Power.
What Actually Makes A Startup Durable (Y Combinator), for the argument that AI lets founders attack harder problems and that capital needs rise to match.
Compute at Sea (Y Combinator), for the electricity, land, and local-approval constraints on new data centers.
Heron Power raises 140 million dollars (TechCrunch) and the 38 million dollar Series A (TechCrunch), for Heron's funding and investors.
Heron Power selects its first factory location, for the Morgan Hill facility, its size, and the production timeline.
Mariana Minerals restarts Copper One, for the permitted Utah operation Mariana bought rather than developed, and its refinery's production history.
Profiles: Drew Baglino, founder and CEO of Heron Power, and Turner Caldwell, co-founder and CEO of Mariana Minerals. Baglino's Tesla tenure and departure per Wikipedia and Business Insider.