Chapter 098　What Did Anthropic Redefine?

Prescription for Incumbents What did Anthropic redefine? Calling it an AI company that puts safety first does not go far enough. Many companies raise a banner. What separates this one is that it built the machinery for enforcing its banner into the design of the legal entity itself, before the company got large. It handed the power to elect a majority of its board to outside trustees. It wrote down, in its own words, the conditions under which it stops training. It gave away the standard it created to a third-party body. The constraints were placed first. The scale came after. This chapter reads that order. One condition has to be stated at the outset: almost every figure in this chapter comes from the company’s own announcements, and none of it has been audited.

1 What makes this enterprise worth a question

Start with the facts. The facts are not the conclusion. The company describes itself this way: “Anthropic is an AI safety and research company. We build reliable, interpretable, and steerable AI systems.” A safety and research company, building AI systems that are reliable, interpretable, and steerable. That is the selfdefinition. The legal form is a Delaware Public Benefit Corporation (PBC). The corporate purpose is set as “the responsible development and maintenance of advanced AI for the long-term benefit of humanity.” Under this form, directors hold the legal authority to balance the financial interests of shareholders against the public benefit. Above that sits the Long-Term Benefit Trust (LTBT), a trust composed of five independent trustees. At the Series C round a new class of stock, Class T, was created, and the LTBT holds it exclusively. Through Class T, the right to elect directors expands in stages as milestones of time and funding are met, reaching the power to elect a majority of the board within four years. The trustees hold the power to appoint and to remove directors, and the right to be notified of actions that would substantially change the company or its business. The company has also explained why the trust exists. AI development generates externalities on an unprecedented scale. So rather than pursuing short-term returns, it places a mechanism designed to respond to extreme outcomes with the interests of humanity in view. On the capital side, the figures are large. On February 12, 2026, the company announced that it had raised $30 billion in a Series G round at a post-money valuation of $380 billion. The lead investors were GIC and Coatue, with D. E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, and MGX as co-leads. The stated use of proceeds was frontier research, product development, and the expansion of infrastructure. The same announcement records runrate revenue of $14 billion, reached less than three years after first revenue, with growth above tenfold in each of the past three years. Customers spending more than $1 million a year number over 500, against roughly 12 two years earlier. Claude Code is stated to have run-rate revenue above $2.5 billion. Here we must record that the evidentiary conditions of this chapter differ from those the series has kept elsewhere. The company is private and carries no disclosure obligation. A companyname search on the U.S. Securities and Exchange Commission’s EDGAR system returns “No matching companies.” There are no filings. There are no audited financial statements and no quarterly disclosure. The previous chapter’s subject, SpaceX, left an audited primary source in the form of a prospectus. This enterprise has left none. Every figure set out above comes from the company itself, and we have no means of verifying it. With that reservation in place, the question stands as follows. What happens to an enterprise that designs its constraints first?

2 Conventional answers and their limits

Three ways of talking about this company circulate. Each is partly right. Each misses the part that matters. The first conventional answer: “an AI company that puts safety first” This is the most widely repeated account. As a matter of fact, the word safety appears in the company’s own self-description. But this account treats safety as a matter of corporate culture or posture. Posture cannot be measured. What can be measured is whether the posture constrains the allocation of resources. On September 19, 2023, the company published its Responsible Scaling Policy (RSP). It is a set of technical and organizational protocols focused on catastrophic risk — cases in which an AI model directly causes large-scale harm, such as the manufacture of biological weapons by terrorists or state actors. The AI Safety Levels (ASL) it sets out have four steps. ASL-1 is a level with no catastrophic risk. ASL-2 shows early indications of dangerous capability. ASL-3 substantially increases the risk of catastrophic misuse or exhibits autonomous capability. ASL-4 and above are left undefined. What matters is that the framework includes the need to pause the training of more powerful models. A pause is the choice not to spend capital that could have been spent. That is not posture. It is a binding condition on capital allocation. The first conventional answer does not see this structure. The second conventional answer: “the rival to OpenAI” The second account sets two companies side by side and argues about which is ahead. This series treated OpenAI in Vol. IX, Ch. 085. For exactly that reason, we decline the comparison in that form. The two should be read as two designs answering the same question. The question is shared. When the capital required to realize a purpose becomes enormous, where do you place the machinery that protects the purpose? OpenAI chose a structure in which a nonprofit parent appoints and removes the directors of a for-profit entity. Anthropic chose a for-profit vessel, the PBC, and then handed the power to elect a majority of directors to a trust of independent trustees. The first places the purpose in a superior legal entity. The second embeds the purpose in the design of a share class. We do not assert which is better. There is no material on which to judge. Neither device has a confirmable record of operating in an actual conflict. Both were designed. Neither has been tested. The third conventional answer: “a company that grew fast in enterprise AI” The third account explains the company through its growth rate and its customer count. Fourteen billion dollars, tenfold growth, 500 customers. The numbers are indeed large. This account has two limits. The first is provenance. As already stated, all of these figures are self-reported and unaudited. They are different in kind from the figures we have used in chapters on listed companies. They must not be handled with the same precision. The second is a distinction the canon states repeatedly. The second dimension of Enterprise Redefinition does not ask what a firm sells. It asks what value the firm delivers. Rising revenue is an outcome, not the redefinition itself. What all three answers miss All three ask how well this company is doing. What should be asked is what this company decided first. Constraints are easy to place while a business is small. What is hard is whether the constraints survive once the business is large. That single point is what makes the case worth a question.

3 What was redefined — an analysis across the five

dimensions We read along the five dimensions of Enterprise Redefinition. The order follows the canon: Purpose, Business, Organization, Capital, and Leadership. What follows is what can be read from public information. It is not an assertion about internal decisions. Purpose — a purpose embedded in the legal form In most enterprises, Purpose sits on the page that carries the corporate philosophy. A philosophy requires no procedure to amend, and rewriting one produces no legal consequence. Here the purpose sits on the side of the legal form. By choosing the PBC, the company placed “the responsible development and maintenance of advanced AI for the long-term benefit of humanity” inside the legal standard by which its directors judge. Directors hold the authority to balance shareholders’ financial interests against the public benefit. Put the other way: pursuing financial interest alone is no longer the self-evident duty. The canon holds that Core Purpose can remain stable while its expression and its means of realization evolve. In this case, the stability of the expression is supported by a documentary structure. Moving Purpose from a poster to a legal obligation. That is the first redefinition. Business — selling trustworthiness as a function The rewriting of the business dimension appears in three words of the self-description. Reliable, interpretable, and steerable. These three are not measures of performance. They are measures of controllability. They ask not what the system can do, but whether it behaves as expected. This design meshes with how the customer decides. The company states that customers spending more than $1 million a year went from roughly 12 to more than 500 in two years. When an organization places a model at the core of its operations, what it judges is not peak performance. It is predictability and explainability. The pursuit of safety and the pursuit of business coincide at this layer. We stop short of asserting more. We hold no independent evidence that this coincidence caused the growth. Organization — binding the inside, releasing the standard to the outside In the organizational dimension, two movements that look opposite occur at once. The first is a constraint turned inward. The RSP and the ASL framework define, in the company’s own documents, the conditions under which the company stops itself. By the company’s account, the pause is not a penalty. It is a design that directly motivates solving the safety problem as the condition for unlocking further scaling. Stopping is what creates the key to proceeding. The second is a release turned outward. On December 9, 2025, the company announced that it would donate the Model Context Protocol (MCP), which it had developed, to the Linux Foundation’s Agentic AI Foundation. That foundation is a directed fund under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI, and supported by Google, Microsoft, AWS, Cloudflare, and Bloomberg. Its purpose is to ensure that agentic AI develops transparently, collaboratively, and in the public interest. As its reason for the donation, the company cited keeping MCP a neutral and open standard, while stating that it would continue to invest in MCP. It co-founded a foundation with OpenAI, a competitor, and released control of the standard it had built. In Vol. IX, Ch. 087 this series examined a structure in which Google supplies computing resources at scale to a rival of its own conversational AI. The same phenomenon is occurring here on the side of standards. As the canon defines it, an organization is not a collection of people. It is a value-creation system made of people, AI, partners, universities, and customers. A structure in which a competitor is simultaneously a co-founder is that definition in its extreme form. Bind the inside, open the outside. The two are the same judgment seen from two sides. Both deliberately narrow the range of what the company can decide alone. Capital — capital without a disclosure obligation The capital dimension is the part of this case to handle most carefully. The figures are large. Thirty billion dollars in the Series G, a $380 billion valuation, $14 billion of run-rate revenue. But we repeat. The company is private and has no filings with the SEC. There are no audited financial statements and no quarterly disclosure. The valuation is not a price set by a market; it is a figure agreed under the terms of a particular transaction. Other chapters in this series stand on a verifiable base of financial statements and SEC documents. This chapter does not. The reader has to know that difference before reading further. With that said, some structure can be read. Of the three stated uses of proceeds, two are research and product. The third is the expansion of infrastructure. On April 6, 2026, the company announced an agreement with Google and Broadcom to bring multiple gigawatts of TPU capacity into use in stages from 2027. Amazon, separately, is an investor: Amazon’s net income of $62.6 billion in the second quarter of 2026 includes a $53.4 billion nonoperating valuation gain related to its investment in Anthropic. The canon’s third equation defines capital as follows. Future Capital = Financial × Human × Learning × Trust × AI × Knowledge × Ecosystem × Purpose This is multiplication. If any single term is zero, the whole product is zero. No term compensates for another. In this case the Ecosystem term is tightly coupled to the Financial term. The computing resource sits outside the company, and that outside includes competitors. Leadership — a design that dilutes control The distinguishing feature of the leadership dimension is not the concentration of authority but its dilution. In the previous chapter, SpaceX showed a structure in which the founder holds approximately 82.4 percent of voting power after listing. That is a design that concentrates control in order to protect a time axis. This enterprise points the other way. It created Class T, handed it to a trust of five independent trustees, and gave that trust the power to elect a majority of the board within four years. Dario Amodei and Daniela Amodei are recorded as directors, but the composition of that board is itself something outside trustees can determine. Founders and investors diluting their own control. This is the reverse of ordinary capital strategy. The canon’s fourth equation defines leadership as follows. Leadership = Purpose × Question Design × Capital Allocation × System Architecture × Trust This equation is multiplicative as well; a zero in any term empties the whole. What stands out is System Architecture. What is being designed is not an individual product decision but the structure of the body that makes decisions. As First Principle 9 states, Leadership Means Designing the Future. But being designed and operating are two different things.

4 Structure — maturity and the value chain

4.1 Where it sits on the Enterprise Redefinition Maturity Model

We hold the case against the Enterprise Redefinition Maturity Model (ERMM). Nothing can be asserted. The company is private, and the means of verifying how it actually operates are limited. What follows is a provisional reading of public information. In the purpose and capital dimensions, high-level behavior can be observed. Placing the purpose in the legal form, and handing governance rights outside at the same moment capital is taken in, is not a reaction to external change. But it is early to call this the level of a Future Value Enterprise. A Future Value Enterprise is the stage at which a firm competes through superior enterprise evolution rather than superior execution. The reproducibility of enterprise evolution becomes observable only across repeated turns. This company’s history is still short. The organization and leadership dimensions are harder still to assess. The RSP is a document the company wrote, and the company can revise it. Whether the state required of a Continuous Redefinition Enterprise — redesign embedded within normal management processes — has been reached cannot be confirmed from outside. Here the canon’s notes apply. Progression across levels is not linear. An organization can hold Level 4 AI capability while remaining Level 2 in leadership. The ERMM evaluates organizational coherence rather than isolated excellence. Maturity is also assessed across all five dimensions in balance: exceptional technological capability with weak leadership redesign does not produce higher maturity, and strong purpose without adaptive organizational systems remains insufficient. Placing this company at a single level is therefore itself a misuse of the model. And Level 5 must not be treated as a target to be reached as fast as possible. The appropriate level differs by industry and environment.

4.2 Where in the Future Value Chain the value was created

The causal order fixed by the canon is as follows. Purpose → Learning → Redefinition → Creation → Enterprise Value Lay that order over this company and the point where value arose can be identified. Value did not arise at Creation. It arose between Purpose and Redefinition. The purpose of responsible development for the longterm benefit of humanity came first, and governance and safety policy were designed to protect it. Those constraints then set the direction of the business: delivering controllability as a function. Creation came after. Enterprise Value comes last. The $380 billion valuation is not a cause but a result. And the valuation of a private company is a function of transaction terms and future expectation, not a settled value. The second equation confirms the same structure. FVCC = Purpose × Learning × Redefinition × AI Integration × Ecosystem × Capital Allocation × Trust Seven terms, multiplied; weakness in any single term weakens the whole. In this company, Purpose, Trust, and Ecosystem move together. Embedding the purpose in the governance structure supports Trust, and that Trust makes possible an Ecosystem co-founded with a competitor. Put the other way, if Purpose falls, two other terms fall with it.

4.3 Constraint and time

The fifth equation places time as an independent term. Future Value = Future Time × Future Capability This is multiplication as well. To hold a stopping condition is to sacrifice growth in Future Capability temporarily in order to extend Future Time. Stop training and the acquisition of capability is delayed. Proceed inside a structure that cannot stop and the possibility remains that the business itself ends. What distinguishes this case is that the trade was written down explicitly.

5 What it looks like in practice — the constraints have

not yet been tested Redefinition is not the act of adding. It is the act of deciding what to protect and what to release. In Vol. X, Ch. 092 this series argued that Purpose begins to function only when it states what is being given up. We set out what this company released, and then set out the conditions under which its advantage breaks. What was released First, the power to elect a majority of the board. Through Class T, a trust of five independent trustees acquires that power in stages. Founders and investors handed over what they would ordinarily keep. Second, the freedom to scale without limit. The ASL framework includes the need to pause the training of more powerful models. A pause means forgoing the revenue that committed capital would have produced. Third, control of a standard the company built. MCP was donated to the Linux Foundation’s Agentic AI Foundation. The co-founders include OpenAI, a competitor. Hold a standard and you control the speed and direction of its adoption. That control was released in exchange for neutrality. What the three have in common is that each closes a short-term option by the company’s own hand. Decisions that create Future Value usually look like this. The conditions under which the advantage breaks Praise is the enemy of analysis. So is condemnation. We set out five conditions. First, there is no record of the machinery operating. The LTBT is a design with strong powers. But within public information, no record can be confirmed of those powers being exercised in an actual conflict. Design and operation are different things. How the heaviest power of all — removing a director — would be handled in a collision with the interests of capital has not yet been observed. Second, the self-binding is not imposed from outside. The RSP is a document the company wrote. It is neither regulation nor contract. The authority to revise it also sits with the company. No external mechanism rules out a revision that loosens the stopping condition at the moment the condition would bite. This is the same tension in the same shape as the one examined in Vol. IX, Ch. 085. The machinery that protects the purpose is administered by the party that declared the purpose. Third, computing resources depend on outsiders. Under the agreement with Google and Broadcom, multiple gigawatts of TPU capacity come into use in stages from 2027. Amazon is an investor. Both are positioned to compete in the AI market. This company stands on the other side of the structure seen in Vol. IX, Ch. 087. If judgment on the supply side changes, its own plans are affected directly. The range within which it can design its own constraints does not extend past supply. Fourth, the relationship between growth and capital intensity is unverified. Run-rate revenue of $14 billion and growth above tenfold in each of three years come from the company’s announcements. Alongside them sit a $30 billion raise and involvement in multi-gigawatt computing infrastructure. Which grows faster, revenue or the capital that supports it, cannot be determined from public information. With no audited cost structure, there is also no material on which to discuss profitability. Fifth, the asymmetry of disclosure is itself a risk. Figures that cannot be verified can be misread favorably as easily as unfavorably. Since trust is one term of a product, the absence of a means of verification caps that term. This is not a matter of the company’s posture. It is a consequence of being private. On comparing two designs Vol. IX, Ch. 085 and this chapter set two answers to one question side by side. When the capital required to realize a purpose becomes enormous, where do you put the machinery that protects the purpose? OpenAI rebuilt the vessel three times, moving the form of the constraint from a cap on returns to governance rights. Anthropic fixed a single vessel at the outset and handed governance rights outside through the design of a share class. The first appears to have chosen adaptation to change. The second appears to have chosen fixity of the initial design. But which is better cannot be judged. Neither device has yet passed through a moment in which the demands of capital and the purpose collide head-on. What can be confirmed is that two designs exist, and that both are untested. To say more would be prediction rather than analysis.

6 What transfers, and questions for the executive

What transfers from this case, and under what conditions? There are four things. First, the idea of moving Purpose from a poster to a legal obligation. In most enterprises the purpose lives on a philosophy page. Rewriting it requires no procedure, and breaking it carries no legal consequence. This company placed its purpose in the legal form itself. That is not the only destination available. Shareholder agreements, board charters, and conditions attached to capital all qualify. Whichever is chosen, a purpose that is not connected to the duty of a decision-maker stays a poster. One condition travels with this. An embedded constraint works precisely because it cannot be removed when it becomes inconvenient. A removable constraint is not a constraint. Second, placing constraints ahead of scale. Constraints are cheap to place while the business is small. Attempt them once the business is large and they collide head-on with existing revenue. But an early constraint is not necessarily a correct one. A constraint set too early closes options before the enterprise has learned. The timing and the granularity of what is constrained have to be designed separately. Third, the option of releasing a standard in order to make it neutral. Keep hold of a mechanism you built and you keep control of its diffusion. Release it and the barrier to adoption falls, opening participation to a wider field, competitors included. The dividing line is whether your revenue arises from the standard itself. If the standard is the source of revenue, releasing it is self-negation. If the standard is a complement, releasing it enlarges the market. Fourth, treating verifiability as one term of value. This company’s figures cannot be verified. That is not a fault, but it does set a ceiling on trust as seen from outside. Read the other way, an enterprise with verifiable disclosure holds a structural advantage there. An enterprise that treats disclosure only as a burden is shrinking that term with its own hands. One caution goes with all of this. This company must not be imported as a model. Neither handing governance rights outside nor a policy of stopping training suits every industry. The appropriate level differs by industry and environment. Finally, three questions. Each can be answered at your next executive meeting. Question 1 — Is your purpose connected to anyone’s duty? If it is not, it is a poster. Can you name the person who carries the responsibility for protecting the purpose? Is that responsibility designed together with the power of appointment and removal? Question 2 — When did you decide the things you have decided not to do? A constraint adopted after the business grew large disappears the moment it collides with revenue. Only constraints set while the business was small survive its growth. Are you placing, now, the constraints you can place now? Question 3 — Which of your claims can an outsider verify? Claims that cannot be verified do not compound the trust term. First Principle 8 states, Trust Compounds Faster Than Capital. Verifiability is the precondition of that compounding. What Anthropic redefined is not the safety of AI. It is the order in which constraints are placed. It placed the constraints first and took the scale afterward. It handed a majority of its board outside, wrote down the conditions for stopping training, and moved its standard to a third-party body. What we are looking at is not results but design. There is Purpose, then Redefinition, then Creation, and Enterprise Value comes last. The figure of $380 billion is only that final term, and it is not even a price set by a market. And this design has not yet been tested. The machinery was built; it has not operated. The day when the demands of capital and the purpose collide head-on will come. Whether the constraints survive that day will decide the answer to this case. Writing that we do not yet hold the answer is the only honesty available to us in this chapter.

In brief

What Anthropic redefined is not the safety of AI. It is the order in which constraints are placed.
What can be read from public information is high-level behavior in the purpose and capital dimensions.
Value did not arise at Creation. It arose between Purpose and Redefinition.
The machinery was designed but has no record of operating, and being private, its figures cannot be verified.

Key concepts

Purpose / Enterprise Redefinition / the Enterprise Redefinition Maturity Model / Future Capital / the Societal Challenge Pattern (→ Vol. VI, Ch. 059)

The chain of ideas

Core Purpose → Redefinition → Trust → Ecosystem → Enterprise Value

Related first principles

Principle 1 — Purpose Precedes Profit. Principle 2 — Future Value Precedes Enterprise Value. Principle 8 — Trust Compounds Faster Than Capital. Principle 9 — Leadership Means Designing the Future.

Related chapters

Vol. IX, Ch. 085 “What Did OpenAI Redefine?” — another design of the vessel, answering the same question
Vol. VI, Ch. 055 “What Does It Mean to Redefine Governance?” — treating the structure that protects a purpose as an object of design
Vol. X, Ch. 092 “What Did TSMC Redefine?” — Purpose functions through what it gives up
Vol. VIII, Ch. 075 “What Is IR in the Age of AI?” — verifiable disclosure sets the ceiling on trust

Papers and companion volumes

Kadowaki, N. (2026). Future Value Theory: A Management Framework for Enterprise, Capital, and Society in the Age of AI. VURA Working Paper. SSRN: https://ssrn.com/abstract=7120980 / Zenodo: https://doi.org/10.5281/zenodo. 21255662
Kadowaki, N. (2026). Enterprise Redefinition: Toward an Enterprise Evolution Theory for the Age of AI. VURA Working Paper. (Published on Zenodo; under review at SSRN)
100 Questions on Management in the Age of AI, #076 “In the Age of AI, Why Does Trust Become Capital?” / #095 “In the Age of AI, Does Purpose Become a Competitive Advantage?”

Chapter 098 What Did Anthropic Redefine?