John Goodenough invented the LCO lithium-ion cathode in 1980, Akira Yoshino matched it with a graphite anode and turned it into a practical cell in 1985, Sony commercialized it in 1991, and cars did not carry it at scale until the 2010s. Thirty years, with the core science done in the first two of them. LFP cathode took twenty-eight years to become the majority EV chemistry. Carbon fiber took fifty to become half an airframe. The stock answers are that science is slow, scale-up is hard, and qualification takes what it takes. I have given all three reasons to boards, and they do not account for the number. When someone finally measured where the years went, they were not in the material. They were in the value chains the material had to cross. The first person who showed me that was Clayton Christensen.
Some years ago I found myself seated next to Clayton Christensen on a flight. Most people know him as the professor who wrote The Innovator's Dilemma. Fewer know that before he became a professor he co-founded a company: Ceramics Process Systems, started in 1984 with several MIT professors to make advanced ceramics. He ran it as president and chief executive through the late 1980s and left to do his doctorate. Cabot, where I spent much of my career, was an investor in it and supplied materials for its ceramics under public patents.
He told me about a case that did not work. They had developed a high-performance ceramic for engine components, and a German automaker was evaluating it. The part performed. It was also substantially lighter than the metal component it replaced, and the lightness was the problem. Less reciprocating mass changed the dynamic balance of the engine, and the engine vibrated. Fixing that would have meant redesigning the engine around the part. The program was cancelled.
I have thought about that conversation ever since. The part did not fail. It did exactly what it was designed to do, and the engine program could not use it, because a lighter part changes the balance of the whole engine. Nobody made a mistake. The automaker was right to cancel; it was being asked to redesign an engine to accommodate a component improvement, and no engine program has room for that.
Christensen learned why good technologies fail from his own company. Then he wrote the book.
The other half of the lesson, which was mine
As CTO, I spent part of my Cabot years building a security materials business. Covert taggants, markers made by spray pyrolysis, nanoparticle authentication chemistry: materials that let you tell whether a banknote, a pharmaceutical, or a tanker of fuel is what it claims to be. The technology worked. In July 2014 Cabot sold the business to SICPA for about $20 million.
The press release said value to shareholders and a better owner. The real reason, as I understood it from inside, was fit. The chemistry was good. The fit was not. Cabot is a performance materials company. It sells materials into other people's products, at scale, and its identity, cost structure, and competences are all built around doing that. It does it very well.
A security taggant is not that kind of product. Its value does not sit in the material. It sits in the authentication system around the material: the relationship with a central bank, the track-and-trace platform, the ten-year government excise program, the standing to say that this note is genuine and be believed. Sell the taggant as a material into somebody else's system and you capture a sliver of the value, and you are substitutable.
SICPA owns that system. It protects the majority of the world's banknotes, and it does not sell ink as an isolated input. It sells authentication: the security feature on the note, the excise-stamp program, the fuel marking that tells a ministry whether the diesel in the tanker is the diesel that was taxed. When it bought our business it created a competence center for material-based authentication that fed a system it already owned. Same technology, same people, a different growth path under an owner that already controlled the application.
That is the clearest lesson I have about vertical integration. The same asset is worth different amounts to different owners, and the difference is set by where in the chain the owner stands.
Put the two stories side by side and the whole framework is there. Christensen's ceramic failed because the interface between part and engine was not clean enough for a component improvement to pass through it. Our taggant business found its right owner because the value lived on the far side of an interface Cabot was not going to cross. Both are stories about boundaries: where they sit, what can pass through them, and who should own which ones. Integration is a choice about which boundary is worth owning. The choice turns on where the value sits, not on how frustrating your partners are.
That is the framework this essay builds, first for incumbents and then for entrants, who face the same boundaries with fewer options. The rest of the series puts a measured number on what crossing a boundary you do not own actually costs, and introduces a mechanism I built to change the cost equation.
The five gates, and why they usually run in series
Before anyone scales a new material it has to pass roughly five gates: intrinsic performance; manufacturability under the customer's real process conditions rather than yours; acceptable behavior once it interacts with everything else in the system; reliability and regulatory compliance; and acceptance by the customer's customer. That is as true of a resin for aerospace structures, a barrier coating for packaging, a photoresist, or a refinery catalyst as it is of a battery material.
In a fragmented chain each gate belongs to a different company, and the gates run in sequence rather than in parallel. People read the sequence as inertia. It is not. A cell maker who accepts a cathode supplier's coin-cell data and skips its own pilot-line validation is absorbing performance and yield risk it cannot price. An aerospace integrator who takes a resin supplier's coupon data and skips its own allowables campaign is absorbing safety and certification risk it cannot transfer. Each actor is protecting something real, so each one rationally demands more evidence than the previous one considered sufficient. The calendar is long because each test waits for the result of the one before it. They usually cannot run in parallel.
The decoupling point, and what Musso measured
Christensen did not leave the ceramics problem behind when he became a professor at Harvard. He kept working on it, and one of the people he worked on it with was Chris Musso, who did his doctorate on why materials take so long to commercialize.
In 2004 Christensen, Musso and Scott Anthony published the argument in Research-Technology Management. The sentence that matters is blunt: a company can only successfully commercialize a new technology if it controls all of the activities with which that technology interacts. Attempting otherwise, they wrote, is likely to end in frustration. Read that against the engine. The ceramic interacted with the balance of the engine, and the ceramics company did not control engine design. That sentence describes the failure exactly, and the person who wrote it is the person it happened to.
They also named the thing. The decoupling point is the place in the chain where a technology stops interacting with its surroundings strongly enough to require joint control; below it you can hand off and let someone else optimize independently. Their claim is that strategy depends on finding this point, for each technology, at each moment, because it moves. While performance is not yet good enough, the point of integration sits forward, toward the application. Once the technology is good enough for the market, it shifts backward and specialists take over the pieces.
Eric von Hippel at MIT had already supplied the economics underneath. In a 1994 Management Science paper he defined sticky information as information that is expensive to move to another party in a form that party can actually use, and drew the consequence: when the sticky information a problem needs is held at one site only, the problem gets solved at that site. Problem-solving migrates to wherever the expensive-to-move knowledge sits, the way a smelter gets built near the ore rather than the ore shipped to the smelter. That is the decoupling point restated as a cost. If the evidence a decision needs can only be generated in one place, that is where the work will happen, by design if somebody plans it and by attrition if nobody does.
Musso then went and measured it. His 2005 MIT thesis is called Beating the System: Accelerating Commercialization of New Materials. It opens from a number everyone in materials knows: twentieth-century materials breakthroughs took nearly twenty years from invention to widespread market acceptance, which Tom Eagar at MIT had put on the record a decade earlier in Bringing New Materials to Market, and which Musso calls simply too long for private investment. Then he went through the commercial histories of the major commodity thermoplastics, the materials that eventually went everywhere, and asked case by case what had actually caused the delays.
He found two things: technical deficiencies in the materials, and obstacles in the application value chains, and he sized both. Technical deficiencies were the largest single factor; where present they added over eleven years of insertion delay. Value-chain obstacles dominated the applications that took two to six years to insert, and the slowest cases mostly carried both. The factors everyone assumes matter, material cost, competing materials, plain luck, were much smaller. He found something else as well. The factors that got a plastic inserted into an application were different from the factors that drove its growth afterwards. The skills that win the first adoption are not the skills that win the market.
So there is a named concept, a general rule, and a systematic historical study, all pointing at the same place. In materials the binding constraint sits in the interactions between organizations, not in the material. None of these pieces is mine. What I have not found anywhere is the assembly of them into one working perspective for materials, where interfaces stay interdependent for decades, the parties cannot show each other their processes, and the calendar runs long enough to exhaust the capital before the question is settled. That assembly is what this series is.
The version of this framework I use
For working programs I use a related diagnostic: a validation boundary, where evidence generated by one actor is insufficient for the next actor's decision. It identifies an interface that has not yet become independent enough for a simple handoff. I distinguish it from Christensen and Musso's decoupling point, which describes where such independence becomes possible. Neither is necessarily a company boundary.
Work it through a battery chain. The cathode supplier proves performance in coin cells. The cell maker cannot commit its pilot line on that evidence, so it revalidates. The pack integrator cannot commit thermal and electrical design on the cell maker's data, so it revalidates. The OEM cannot commit abuse tolerance and warranty on the pack integrator's data, so it revalidates again. One material change, four validations. A composite airframe, a semiconductor process, a coating on a packaging line, a catalyst in a refinery: different words, same structure, a validation boundary at every place where evidence has to be regenerated rather than passed along.
The calendar is set by the number of validation boundaries, not by the number of companies. Two companies with three boundaries between them make a slower program than four companies with one. Consolidating suppliers does not help if the interfaces between what they supply are still immature, which is why headcount and org charts predict development speed so badly.
Christensen's architecture theory tells you where to expect these boundaries. While product performance is not yet good enough for the market, every layer's behavior depends on every other layer's, and no interface is clean enough for proof to pass through unaltered. Once performance overshoots what the market needs, interfaces standardize and specialists work behind them independently. A resistor's datasheet is enough for a circuit designer; a cathode supplier's coin cell is not enough for a cell maker. This is why the drop-in claim is so dangerous. Selling a drop-in material is a claim that no validation boundary exists at your interface. If the interface is still interdependent the claim is false, and your customer finds out after committing a line. Christensen's lighter engine part was a drop-in improvement only if the interface was modular, and in an engine it was not.
Reading integration through validation boundaries
With that concept in hand, the famous integration decisions read differently. Integration is an attempt to buy a boundary. That is all it is. You are purchasing the right to run both sides of a validation boundary yourself, so that evidence does not have to be regenerated by a party you cannot direct.
Northvolt bought five and had mastered none of them. Tesla bought the specific ones where knowledge rather than capacity was the barrier, and left the rest alone. BMW bought one for the duration of a learning problem and handed it back when the interface was understood. Cabot sold one, because the value on its far side belonged to a different kind of company. Four different decisions, all correct, from one question: which boundary is worth owning.
Full integration looks like the solution: all five gates under one roof, one P&L, one set of priorities, no negotiation at each handoff. Northvolt is the cleanest test of that logic in the last decade. It set out to own the chain, cathode active material at its flagship plant, cell manufacturing, modules and packs through its Systems division, recycling through Revolt, across multiple countries, inside one company, financed on the promise of a sovereign European battery champion with fewer external dependencies. It never got to operate that chain as a whole. In 2024, under pressure, it halted its cathode-material plans and moved to buy cathode externally, sold sites, and narrowed to cells, shedding the integration link by link on the way down. Chapter 11 in the United States in November 2024. Bankruptcy in Sweden in March 2025.
Notice what did not happen. The gates did not disappear because they were brought in-house. Northvolt still had to prove cathode manufacturability to its own cell line, cell performance to its own pack team, and pack behavior to its own customers. The delays moved from between companies to between internal teams, inside an organization that had not yet mastered any of those interfaces, with investors who had been told the coordination problem was solved. Integration does not remove validation boundaries. It relocates them, and an internal boundary can be worse than an external one because nobody writes it into a plan.
CATL is the instructive contrast, and CATL is not integrated in the way the word usually implies. It is a cell maker with an ecosystem: a majority-owned recycling and precursor subsidiary in Brunp, stakes in lithium and nickel projects, competing cathode and anode suppliers built up around its demand. It owns the one boundary that matters, the cell and its qualification, and contracts for everything else from strength, keeping its suppliers substitutable instead of becoming them. That is integration as a portfolio of options rather than a chain of plants, and it is available only to a company with CATL's volume and downstream pull.
Panasonic is the third position, and the most common one for a specialist supplier. In 2016 it was the largest battery supplier to passenger EVs, with roughly a third of the installed base; CATL had under 6%. Panasonic did everything a good partner does. It co-developed Tesla's cylindrical cells, took a small stake in the company, and moved its production inside Tesla's Nevada factory; by mid-2024, 99% of the automotive cells from its two Osaka plants were shipping to North America. But co-locating across a boundary is not owning it. The qualification, the customer, the format decision and the chemistry decision all stayed on Tesla's side. When Tesla qualified LG and CATL as second sources in 2020, adopted LFP for its standard-range cars, and built its own 4680 cell line, Panasonic's decade of co-specialized assets were pointed at an interface its customer was moving. By the first half of 2026 Panasonic held under 4% of the market and CATL led with 40%. Panasonic still earns on what it kept, a 10.3% operating margin in its automotive battery business in the most recent quarter and the highest of its peer set. It kept its economics and lost the market.
Three companies, three positions on the same chain. Northvolt bought every boundary and could run none of them. Panasonic co-specialized across a boundary it never owned and was moved off it. CATL owns the one boundary that matters and contracts for the rest from strength. The one that is winning is the one that chose which boundary to own.
The question that decides it
If integration is a choice, you need a test for it. One question does most of the work for me.
Where are you knowledge-dependent? Charles Fine and Daniel Whitney at MIT drew the distinction I keep returning to: you can depend on a partner for capacity, or you can depend on them for knowledge. Capacity dependency is uncomfortable, but you can contract around it: second-source it, reserve it, pay for it, eventually build it. Knowledge dependency is the dangerous one. It means you cannot diagnose a problem yourself, cannot judge whether the partner's answer is right, cannot accelerate the step at all. You are not waiting for a machine. You are waiting for somebody else's judgment with no way to evaluate it.
Most materials companies I have worked with have misclassified at least one knowledge dependency as ordinary outsourcing. They treat the converter, the cell maker, the tier-one, the fabricator as a manufacturing channel, when that actor also owns the tacit process windows and the qualification heuristics that determine how fast any evidence moves through the chain. Buy capacity there and nothing changes. Buy the knowledge and everything does.
Tesla gets cited as proof that integration works, and the detail that matters usually gets dropped. Tesla did not integrate and then learn. It acquired specific knowledge first, Maxwell Technologies for dry-electrode process know-how, Hibar for cell-manufacturing equipment, a long research relationship with Jeff Dahn's group for electrochemistry and life prediction, and only then attempted cell production at scale. Tesla bought specific knowledge it could not get fast enough from a supplier. That is a purchase, not a philosophy. Seen from Panasonic's side, those purchases turned Tesla's dependence on Panasonic's knowledge into dependence on Panasonic's capacity, and capacity can be bought elsewhere. Panasonic's customer ran this test, and Panasonic was the answer.
There is a third position between full ownership and arm's-length supply, and BMW used it well. Rather than become a permanent carbon-fiber producer, BMW built a joint venture with SGL Carbon, ran it through the i3 and i8 programs at Moses Lake and in Germany, learned how carbon fiber behaves in automotive body structures at volume, and then sold its stake back to SGL once the interfaces were understood, keeping the supply relationship. Temporary integration to learn a boundary, then a planned exit. Integration with a term, not a change of identity, and far more companies could use it than do. The obstacle is not analytical. Exiting an integration looks like failure to a board even when the exit was the plan.
Why startups integrate anyway
Almost everything written about vertical integration is written about incumbents, and it does not transfer. An incumbent integrates from strength, the way CATL runs its supplier ecosystem. A startup usually integrates to gain access.
The pattern I see over and over runs like this. A startup has a material that works. Nobody will qualify it, because no integrator runs a qualification campaign on an input that has never been in a product, from a company that may not exist in three years, without data from a real line. So the startup builds the next step itself, not because it wants to be a manufacturer but because it needs to create the proof that opens the conversation. It builds a cell to sell a cathode, a module to sell a cell, a product to sell a material. Redwood Materials is the live version. It began in battery recycling, moved into cathode and anode materials, and has moved again into energy-storage systems, most recently for data centers. Each step forward creates a customer for the step behind it.
Sometimes this works. The cases where it works look alike.
Integrate toward the evidence: GE and Tepha
Ceramics fly in engines now. GE Aerospace put ceramic matrix composites into the hot section of the LEAP engine, then into the GE9X on the 777X, and CMCs are central to how it describes jet propulsion for the next several decades. The material class that failed in Christensen's engine in the 1980s is in commercial service, and the company that made that happen is the company that owned the engine.
GE did not start with a material and go looking for an engine. It started with an engine and worked backwards to secure the material. It licensed silicon-carbide fiber technology and committed more than $200 million to two Huntsville factories for fiber and ceramic tape, feeding component manufacture in Asheville. A LEAP engine carries roughly one kilogram of CMC. Two hundred million dollars to put a kilogram of ceramic into an engine is the price of making an interdependent interface work when proof cannot cross it. A fiber datasheet could not establish the finished component's performance, and an independent material supplier could not run the engine program. GE approached the problem from the application and secured the material capability it needed. Christensen's company had faced the same problem from the other direction, with the material and without the engine.
Tepha shows the same pattern in medical polymers. It spent years developing poly-4-hydroxybutyrate, a resorbable polymer that supports soft tissue while it heals and then disappears, and then it did not sell the polymer. It made the implant. GalaFLEX, a bioresorbable scaffold for soft-tissue repair and reinforcement, is Tepha's own device, and the polymer also went into surgical mesh sold by Becton Dickinson. You cannot sell a resorbable polymer to a surgeon. The value is in the resorption profile matched to the tissue and the healing timeline, in the geometry and mechanics of the scaffold, and above all in the clinical evidence, and clinical evidence attaches to a device rather than to a material. A regulator does not clear a polymer. It clears an implant. So Tepha integrated forward, not to capture margin but because the proof it needed could only be generated on a finished device. After more than a decade of partnership, BD bought Tepha in July 2021 and described the purchase as strategic vertical integration of an important supply-chain component. Tepha had developed device-level evidence around its polymer; BD acquired a material capability that had become important to its products. The purpose of ownership changed as the technology matured.
The rule those two cases give you
Integrate toward wherever the evidence has to be generated.
This is von Hippel's rule with the word evidence substituted for information. Proof about a jet engine is generated at the engine. Proof about an implant is generated at the implant, because the clinical trial and the regulatory clearance attach to the device and the material has no independent standing. Whoever holds the site where the evidence lives is the only party who can generate it, which is why GE went backwards to the fiber and Tepha forwards to the device. A material company standing outside the qualified article is permanently dependent on somebody else to validate its own product.
That is a very different reason to integrate than margin capture or supply security, and in my experience it is the reason that most often holds up. It is why a polymer company makes surgical mesh and an engine company grows ceramic fiber. They are not diversifying. They are reaching for the place where the evidence lives.
It explains the disappointing cases too. A battery startup that builds cells to sell cathodes is doing something structurally different, because the qualified article already exists, somebody else already owns it, and that somebody already knows how to certify it. You are not reaching for evidence nobody can generate. You are building a second version of a capability your customer has, while depending on that customer. Northvolt was a startup. Sila and Group14 made the other choice. They stay material suppliers into existing channels and combine production with licensing and partnership, accepting slower value capture in exchange for owning fewer validation gates. It is a legitimate strategy, under-represented in venture pitches because it is harder to tell as a story.
The category boundary
The rule has a limit. What happens when the thing that needs evidence is not a component but an entire category? A new chemistry can be manufacturable, compatible, and cheap on paper and still arrive with an empty record, because the category itself has no established performance, cost, safety, or reliability for a buyer to point at. A product can be drop-in at the process level and still un-buyable at the category level. Somebody has to write the missing record.
Sodium-ion shows how that record gets written, and by whom. Faradion pioneered the chemistry from 2011 and was sold to Reliance. Natron built Prussian-blue sodium cells for years and shut down in 2025. The pioneers existed, and the pioneers did not get to establish the category. Sodium became something the industry plans around when CATL put its name, its scale, and its safety franchise behind it: a first generation in 2021, mass production of Naxtra cells in 2025, the first sodium car on the road. A new chemistry is established not by whoever demonstrates it first but by whoever the ecosystem already trusts to certify a category.
When the product is a new category, the final validation boundary sits past the chain entirely, with the customer broadly defined: the buyers, and behind them the insurers who have to price it, the regulators who have to permit it, the financiers who have to residual-value it, and the resale market that has to believe in it. No actor inside the chain can generate that proof on the market's behalf. The record of a category, what it costs, how it fails, how long it lasts, what it is worth used, can only be written in public, at scale, over time. So the point of integration extends all the way forward, past the product, into whatever the market needs in order to make up its mind.
GM introduced the EV1 in 1996, and Nissan put the Leaf on sale in 2010. Electric cars existed before Tesla became a volume manufacturer. What the wider market still needed was confidence in their performance, charging, service and long-term value, and the incumbents were poorly placed to supply it, because dealers earned on combustion service and automakers' assets ran on engines. Tesla built its business around supplying that confidence. Direct sales gave it control of the customer relationship. The Supercharger network addressed long-distance travel. The Model S made an electric car desirable on its own merits rather than as a compliance product. The connector Tesla built for itself became NACS, the standard its competitors then adopted, which is the prize for the company that establishes a category: it gets to set the reference frame around its own architecture, and that is worth more than margin.
SpaceX faced a harder starting condition, because reusable launch had a pioneer whose record pointed the wrong way. The Space Shuttle had shown that reuse was possible and its cost history gave customers little reason to believe reuse would lower launch costs. SpaceX had to build a different record through repeated flights. NASA's cargo and crew programs lent it credibility the market already trusted, and Starlink later gave it enough internal demand to fly reused boosters at a rate no external customer would have supported. If nobody will buy the category's evidence, buy it yourself.
Four moves for a new category
A company holding a genuinely new category has four moves, and it can make more than one. The first is to integrate to the consumer and write the record yourself, which costs the most and pays the prize of setting the category's reference frame around your own architecture. The second is to arm an unconflicted incumbent and be the knowledge inside its product. The third is to borrow standing from a certifier the market already trusts: a NASA, a UL, a national laboratory, an anchor customer willing to deploy in public. The fourth is to design the product so that every avoidable validation boundary inside the chain is minimized, so that the whole evidence budget is spent at the one boundary that cannot be designed away. For a materials company the fourth is the move you control most directly. Choose chemistries and processes that inherit existing chains, existing anodes, existing plants, existing formats. That is the design logic behind every credible new battery chemistry today.
The open question for any new category is which of the four to make, and that is a more precise question than whether to integrate.
What this leaves
None of these companies is the general answer. The ones that did well bought a specific boundary for a specific reason. The ones that struggled bought a philosophy.
For most companies, most of the time, the honest conclusion is that you will not own the chain. Full integration costs more capital and more organizational competence than the acceleration is worth, unless one deeply coupled step dominates everything, or the application is where your value actually lives and you can own it. Which means you will keep working across boundaries you do not control, with partners who will not show you their process, on a product whose behavior only appears once all the parts are assembled. That is the normal case, not the failure case.
So the question for the rest of this series is what working across a boundary you do not own actually costs, and whether that cost can be changed. A category's record is never written by one company alone; it is written across cell maker, integrator, and operator, by parties who cannot open their books to each other. What that coordination costs, who pays it, and a mechanism I built to change it are where this series goes next.