Four euros a minute

The frontier paradox: capital, industrial power, and who gets to build what comes next.

Siddhartha Yash Kovid · 17 September 2026 · 16 min read

In this essay · 10 sections

I judge a technology venture by whether it is genre-defining. Does it change what the industry does? Does it build something important enough that everyone else has to respond? That is the kind of venture I am interested in building.

When we talk about technological leadership, I want to know which ventures determine what happens next. Listing the companies we already have does not answer that.

ArianeGroup exists. Europe has launch expertise. SpaceX still changed what the rest of the launch industry had to do. Volkswagen has factories, engineers, enormous revenues and decades of experience. That did not put it in charge of the transition to software-defined vehicles. A country can have a company in a sector and still spend years reacting to what another company made possible.

Much of the discussion about innovation stops at counting firms, jobs, investment announcements and programmes. We point to an established industrial base and assume that the next generation of important capabilities will somehow emerge from all this activity. When it does not, we start talking about sovereignty.

Before that conversation begins, someone has to finance uncertain work and give a team access to a real problem. Someone has to pay for solving it. The product needs room to develop through use instead of having to resemble a business that already exists.

We make the uncertain stage unnecessarily difficult, especially for technical ventures whose eventual scope is larger than their first transaction. The final commercial explanation is expected early, with established partners to reassure us and a venture that fits something we already know how to evaluate. Then, after someone else has built the capability, we discover that it is strategically important and find a great deal more money to obtain it.

That is the frontier paradox I care about. We are prepared to pay for the consequence. We are much less comfortable paying for the work that creates it.

What a new capability requires

When I talk about PRISM, my starting point is the technical thesis. Materials development involves work that has to inform other work. Results have to change the next decision. Knowledge has to survive beyond the person who generated it and the particular campaign in which it was generated. I want to build the capability that makes this happen more effectively.

People can call part of that AI for materials, part of it manufacturing, part of it laboratory work or enterprise software. The description changes depending on which part they are looking at. What interests me is how the whole thing works. Recognising one familiar component does not mean the project should be redefined around it.

You cannot fully specify an emerging capability, assign it a permanent commercial category and assume that you have described everything it might become. You learn what it can do by building it and using it. Real requirements expose things you did not know. A customer asks for something difficult. Solving it changes the system. The next customer can benefit from work that did not exist before the first engagement.

The product is being made. Customers can pay for its development through contracts. They have a problem they want solved, and the work is agreed with them. That is a market for the capability while it is being developed.

If an industrial customer wants a new material, they cannot just wake up, click on a website and order it. You negotiate the work with them. What they need, what you will develop and what they will pay are part of that negotiation.

Palantir is the comparison I keep coming back to. Its 2020 filing describes initial pilots at its own expense, continued investment in customer deployments and an acquire–expand–scale model. It describes multi-year relationships and billing that varies between contracts, with engineers working alongside customers and incorporating what they learn into the product. [1]

Contract pricing is how that work gets paid for. The venture keeps developing, and so does what it can sell. Expecting the whole commercial form to be settled at the beginning makes no sense to me.

The ambition is to develop materials faster, before dependence on an incumbent becomes another sovereignty problem. Then the conversation comes back to where the price sheet is.

How about four euros per minute? That sounds like a relatively good number to settle on, if that's what sovereignty is for.

Data use and IP are also negotiated with the customer. Those terms belong to the actual engagement. WIPO's guidance on collaborative innovation describes agreements covering existing knowledge, new results, confidentiality and rights to use the work. [2]

Financing the work

A technical founder does not arrive knowing how to sell everything the venture might eventually develop. You fund the work, then hire people who know how to do that part. That is one of the things the money is for.

Some things get worked out as you go along. The team grows, commercial people join, and the organisation develops around the work. Expecting all of that to be fixed from day one defeats the purpose of supporting a venture's development.

Private capital can be just as conservative when it only funds work comparable to an already successful venture. A venture fund has its own obligations and timetable, which a difficult technical programme may not fit.

The question is how many investors can stay with the work, on what terms and for how long. A difficult experiment can take longer than expected. The financing has to allow the team to continue through that uncertainty.

Operating revenue can give a venture more room. So can founder capital, public development contracts or investors with a longer horizon. The financing has to be compatible with what is actually being built.

Valve released Half-Life in 1998 and Steam in 2003. A game developer went on to build distribution infrastructure that other developers came to rely on. [3]

Developing a larger business out of actual work takes enough time and resources to pursue what the team learns. An early product can finance work whose scope extends well beyond it.

NASA's Commercial Orbital Transportation Services programme shows another way to finance development. NASA's retrospective records $396 million of COTS funding to SpaceX and approximately $454 million financed by SpaceX for that programme. Funding was tied to milestones. Another initial participant, Rocketplane Kistler, had its agreement terminated after it failed to meet financing milestones. [4]

NASA was helping finance the development of a service it intended to use, before the wider commercial market had matured.

Radical AI announced a $55 million seed round in July 2025 for an integrated approach to materials development. The US SBIR database separately records a 2025 award of $1,197,902 for its work on AI-driven materials experimentation. [5, 6]

That is the comparison I am making: what a venture is allowed to attempt, and what resources it gets to attempt it.

The ventures that set the direction

SpaceX changed the expectations placed on launch providers. Reuse became something competitors had to address through their own development plans. Its first landing of an orbital-class booster was on 21 December 2015, after years of work. [7]

ArianeGroup created MaiaSpace in 2022 to develop a competitive reusable launcher. Its own description emphasises combining a startup-style approach with the parent's launch experience. [8]

This is what I mean when I say ArianeGroup exists and SpaceX is still the more important reference point for what the industry is trying to become. The question is who changes the expectations.

Tesla does not have to make the best car for every buyer to influence the industry's direction. Nor does Volkswagen's sales volume settle the question. The issue is what an automotive company is expected to control, develop and update itself, and how software changes the organisation around the vehicle.

Volkswagen's 2024 annual report describes its joint venture with Rivian as building on Rivian's software and electrical architecture. It sets out planned investment of up to $5.8 billion in Rivian and the venture. [9]

Despite its own resources and engineering base, Volkswagen chose to obtain that architecture from a much smaller company.

I would judge a European startup against that kind of consequence. Did somebody have to change their plans because of what it built? Do customers depend on it for something they cannot readily obtain elsewhere? Is it setting terms, or permanently finding a place inside terms set by other organisations?

ASML's 2025 annual report identifies it as the sole manufacturer of EUV lithography systems. That is the degree of industrial importance I am talking about. [10]

How often can the current system do that again? One established success, or a long list of smaller companies, does not establish that it can keep developing ventures whose work becomes difficult for the rest of the world to ignore.

Europe

Europe has excellent people, serious laboratories and organisations capable of difficult technical work. It is a very good place to study and develop scientific depth. That is part of why the commercial problem is so frustrating: the raw capability is present.

The European Investment Bank's 2024 scale-up study followed venture-backed firms that reached the scale-up stage after 2013. By ten years after establishment, the European companies had raised about 50% less capital than their San Francisco peers. The gap persisted across the report's industry, founding-year and business-cycle controls. [11]

That is the gap against which I would judge another announcement that Europe is improving.

Procurement matters just as much. A 2026 study by Kapstein, Ospital and Wolff found the ten largest contractors accounted for 67–90% of defence procurement in the Germany, Poland and UK datasets they examined, compared with less than 40% in the United States. [12]

If most of the purchasing relationship remains with established suppliers, a startup can receive development support and still struggle to become an independent business. It may be allowed to demonstrate a technology, then required to go through an incumbent to sell it. It may win a prize and still have no route to a consequential contract.

A large partner's presence tells us little about a venture's risk until we examine its contribution. Does it provide data that cannot be obtained elsewhere, facilities or engineering knowledge? Does it perform the test, accept the result or commit to buying a deployment?

When a partner contributes none of those things, why does its presence make the venture more credible? Being a large company does not explain what it will actually do on the project.

A partnership can also make the startup's position worse. It can introduce another approval process or put somebody else between the startup and the customer. It can give an incumbent influence over the development of a capability that might eventually threaten its existing business.

The incumbent may want an incremental improvement to its current process. The startup may want to change the process. Requiring the startup to obtain the incumbent's endorsement gives one side considerable influence over what the other is permitted to attempt.

I have very little patience for a system that repeatedly asks the largest companies to certify the relevance of ventures that might replace some part of what they do. If the purpose is to produce another supplier to the existing industry, then that may be a coherent arrangement. If the purpose is to produce something genre-defining, it needs a much more serious explanation.

Public programmes should be assessed through the ventures' customers, revenue and deployments, including whether they retain control of the capability and the customer relationship. Counts of incorporations, events and applications leave all of that out.

By unicorns, I mean ventures with a substantial business behind the valuation. How many has a programme helped develop? How many became independent players, rather than suppliers working underneath an established prime? Compare that with programmes elsewhere and with the money that went into them.

The difficulty is that the costs of refusal are harder to see. An unsuccessful funded project leaves an amount of money that can be audited. A capability that is never developed locally leaves a gap. Years later, the gap appears as imported technology, a lost commercial opportunity or dependence on a foreign supplier. Nobody necessarily connects it to the earlier decision.

The incentives favour avoiding a defensible risk: refusing it leaves fewer visible costs. A system can de-risk activity to death while continuing to describe itself as innovative.

Sovereignty

My expectation is that Europe will continue to recognise some of these capabilities too late. A venture builds something elsewhere. European institutions become uncomfortable with the dependency. The subject becomes a sovereignty priority. A new programme appears, with a much larger budget and a collection of established organisations expected to reproduce the capability.

Why was the capability allowed to become a dependency in the first place? What will the new effort actually deliver, and how long will it take?

There are different ways to respond. An incumbent can create a new subsidiary, as ArianeGroup did with MaiaSpace. An established company can seek an external architecture, as Volkswagen did with Rivian. A group can acquire existing businesses and attempt to combine them. ChapsVision's September 2023 release announced €90 million for European expansion through targeted acquisitions. [13]

Acquisitions cost money. So do the years of development behind the technology being acquired. By the time a capability becomes a sovereignty priority, somebody else has already paid for much of that work.

What bothers me is the willingness to spend heavily after somebody else has established the importance of the work, coupled with the reluctance to tolerate uncertainty earlier. I have called this socialism on Quaaludes. Everything is coordinated, everybody important is included, and the willingness to move arrives once the strategic embarrassment becomes impossible to ignore.

The late programme is competing with years of deployment experience, product development and customer relationships. Reproducing that takes time, even with a large budget.

The later customer also has a different bargaining position. If few alternatives exist, the supplier can negotiate accordingly.

The price of access later will be negotiated with the people who built the capability. They will have customers and investors of their own. Supporting the development earlier would have been a different commercial relationship.

Sovereignty depends on being able to continue operating when a supplier withdraws, change the system when necessary and build an alternative within a useful period. Ownership on paper and the location of a headquarters tell us little about those abilities.

Artificial intelligence

AI is a particularly clear example because people use the same term for very different businesses. A foundation-model company, a chip company and an application that calls somebody else's model can all appear in the same national AI strategy. Their technical contributions and dependencies are completely different.

The Transformer came out of the 2017 paper Attention Is All You Need. NVIDIA introduced CUDA in 2006. The model application that a customer uses today sits on top of years of work in algorithms, computing systems and hardware. [14, 15]

A venture building on those foundations depends on what the upstream provider supplies. Its position changes if that provider changes prices or releases the same feature. The useful question is what the venture has developed that gives it control over its own future.

NVIDIA's position includes the tools, libraries and code built around its platform, along with developers trained to use them. Replacing it involves migrating software and rebuilding operating experience as well as changing hardware.

This is also why a country's list of AI companies tells me very little on its own. Who produces the models people actually want to use? Who controls the computing infrastructure? Who supplies the capital? Who gets the customers? Those questions are much less flattering than counting how many companies have AI in their description.

Stanford's 2025 AI Index counted 40 notable models from US-based institutions in 2024, 15 from China and three from Europe. [16]

Mistral's nationality says little about its infrastructure. Microsoft's February 2024 partnership announcement included Azure supercomputing support for Mistral's training and inference. [17]

For a frontier model, I want to see performance, cost and deployment, as well as what the organisation controls underneath them.

The scale of the American financing is enormous. On 31 March 2026, OpenAI announced $122 billion in committed capital at an $852 billion post-money valuation. [18]

The economics depend on the cost of generating revenue, infrastructure obligations and the resources required for the next generation. Do they improve with scale?

Much of the power sits with whoever can change the terms for everyone else. A national strategy that cannot identify who controls those decisions is mostly describing where the logos are registered.

India

I apply the same standard to India. GDP, market size and the number of successful people of Indian origin are not a substitute for examining what can actually be built in the country.

A large domestic market can sustain ventures that adapt an existing business model. The question is which ventures develop something that the rest of the world then needs.

India approved the IndiaAI Mission in March 2024 with an outlay of ₹10,371.92 crore. The initial compute target included at least 10,000 GPUs. What original work will those resources enable, and what dependencies will remain? [19]

Buying access to compute leaves the hardware dependency in place. A venture adapting somebody else's model for a domestic use case still depends on that model's developer. Those dependencies determine what happens when access becomes expensive or restricted.

National pride becomes an obstacle when it prevents an honest comparison. Can someone with a difficult technical thesis obtain access to experiments, funding and customers without first acquiring the endorsement of the people who already control the industry?

I would rather see a credible answer to that question than another celebration of how many people a country contains or how large its economy will eventually become.

Materials

Materials make this problem much more serious than the usual conversation about software sovereignty suggests. A country can own the design of a rocket or an engine and still depend on external suppliers for what the thing is made from and how it is manufactured.

The dependency can be in the ore, refining, alloy production, powder supply, processing route or qualification data.

USGS estimates that China accounted for 82% of mined tungsten production in 2024. [20]

For a material such as NASA's GRCop-42, a replacement supplier would need the manufacturing process and the data to qualify its output. [21]

This is the question I would ask about sovereignty: if access to a material disappears, how long does it take to replace it?

The timeline depends on the experiments required, the facilities available and the evidence needed to accept the replacement.

NIST's work on additive-manufacturing qualification describes statistical, equivalence-based and model-based approaches. [22]

Materials learning should make that development cycle faster and let each programme use what the previous one established.

The ability to develop a replacement is itself a strategic capability, especially when access to a material or processing route is restricted.

An intermediate capability can determine whether a much larger ambition is possible at all, even if it never becomes a finished consumer product.

Blue Origin's February 2023 account of Blue Alchemist describes terrestrial experiments producing solar cells and transmission wire from lunar-regolith simulants, with work dating to 2021. [23]

What interests me is the investment in the physical prerequisites of sustained activity on the Moon.

The semiconductor problem has the same character. Intel's first-quarter 2026 earnings remarks describe a development partnership with SpaceX, xAI and Tesla on Terafab. [24]

If we only finance the recognisable final product, we can miss the capability that makes the product possible. Then we discover the missing capability when it constrains everything else. By that stage, the problem is expensive and urgent.

My expectation is that governments will become much more interested in the ability to replace constrained inputs once the constraints interrupt important programmes. I would prefer to build the capability before that happens.

What I intend to do

I want to build something groundbreaking. Something that changes the paradigm, rather than carrying the same approach a little further down the road.

Peter Thiel's idea of going from zero to one is close to how I think about this. There are things that no one is building. Those are the things I want to work on. That does not mean every venture has to begin with the invention of a new technology. There can be something new to build with technology that already exists.

The question is what we could be doing that nobody is doing yet. That is where my interest is, and what I mean when I talk about building a genre-defining venture.

Siddhartha Yash Kovid