Quarterly Growth Triples, Hitting a Ceiling: Why Anthropic Is Betting on the Enterprise Market

I viewed this dialogue as a record of how a smooth exponential curve clashes with institutional structures.I decided to trace the moments where curves that connect seamlessly on a graph collide in reality with the “steps” of strained computational resources, workforce restructuring, and the frontiers of war. I will examine why we are betting on Enterprise and the bundle of trust, what we reject at the frontiers of computational resources and war, and what we gradually build up.

The Outline of the Material: Until the Smooth Exponential Curve Touches Reality

The metaphor introduced at the beginning of the discussion is that of a spaceship in special relativity. The sensation of exponential acceleration is explained as the feeling that Earth time has flown by when you wake up after sleeping. This sets the stage for the idea that only when the smooth curve touches reality do the details and colors increase, manifesting as a sense of wonder.

Suppose you were to accelerate away from Earth on a spaceship at relativistic speed. You go to sleep, and when you wake up, two days have passed on Earth.

Translation: Imagine a spaceship accelerating away from Earth at relativistic speed. You fall asleep, wake up, and find that two days have passed on Earth.

The exponential growth discussed in this interview is unique in the way it defies predictions. The interview explains that, on graphs, revenue and valuation growth have long been viewed as a smooth exponential curve. There is a prolonged period where nothing seems to be happening, and then, at a certain point, things suddenly take off. I interpreted this not so much as the world appearing that way, but rather as a sensation where the observer’s sense of time is compressed.

Growth in the first quarter of 2026 is cited as the first sign that this exponential trend is impacting the system. A threefold increase in a single quarter translates to an annualized rate of 80 times.This is where the divergence between the plan’s assumptions and reality becomes visible. The explanation is that, while preparations had been made assuming an annual 10-fold increase in computing resources, a wave of demand equivalent to an 80-fold increase struck locally. Although this figure is described as a temporary surge, this is where the stage is set for the strain.

Over the first quarter of 2026, we saw revenue grow more than threefold on a quarterly basis. Just in a single quarter—not annualized—a threefold increase in the quarter, which, of course, is 80 times over the course of the year when raised to the fourth power.

Translation: In the first quarter of 2026, revenue grew more than threefold on a quarterly basis. This was a threefold increase in a single quarter—not on an annualized basis—and when raised to the fourth power, it equates to an 80-fold increase over the course of the year.

Here, I focused on the gradual nature of the exponential growth and the mismatch with the institutional “staircase.” As indicated by the scaling laws demonstrated by Kaplan et al. in 2020 and the Chinchilla discussion by Hoffmann et al., performance tends to grow regularly in tandem with computational resources and data volume.Assuming a smoothly rising performance curve, the supply side plans to increase computing resources by a factor of 3 to 4 annually. The exponential growth in data center power consumption, as summarized by the IEA in 2024, also serves as a reminder that this expansion requires physical infrastructure to support it.The moment demand surges by a factor of 80, these smooth plans stumble on the stairs. From this point forward, the discussion makes sense when viewed as a matter of how to absorb this stumble through design.

From Asynchrony to a Bundle of Trust: A Design That Encourages Competition

According to the discussion, his upbringing in San Francisco shaped his tolerance for asynchrony. He recounts anecdotes from a time when the Internet revolution was unfolding right before his eyes—how he was engrossed in mathematical doodles, an interest in the universe, and science fiction.The idea that one should pursue a coherent vision even if every expert opposes it is presented as having sprung from that background.

The Methodology of Exploring Dissent

The “nonconformity” discussed in this interview is not about being eccentric for the sake of it. It is described as a method of exploration—seeking out narrow paths that the majority does not take.As illustrated by Turner’s 2006 genealogy of science fiction counterculture, the San Francisco area has historically provided a space where it is permissible to experiment with unconventional ideas. While the probability of missing the mark is high precisely because the vein is narrow, the premise is a long-tail distribution where the payoff is substantial when a hit is struck.The metaphor of a gold vein also comes up in the dialogue. When deciding where to dig, one has no choice but to bet on the coherence of a minority. I understood this approach to betting as a choice that prioritizes logical consistency over reaching consensus.

Variability in Trust and the “Race to the Top”

In the discussion, it is candidly stated that there is significant variation in the reliability of the people building AI. The generalization that “no one trusts each other” is rejected, and the example is given of Demis Hassabis, the creator of Gemini, with whom the speaker has had a 15-year acquaintance and has exchanged ideas on securing computational resources and safety measures.This duality—competing while cooperating—manifests concretely here.

This relationship is held together by the concept of a “race to the top.” It is the idea of creating a situation where trustworthy actors come together and are compelled to adopt the same standards. On the “carrot” side, mutual stimulation leads to the expansion of initiatives such as research on interpretability.On the “implicit stick” side, if a group has already done the right thing, pressure arises for those who do not conform to look inferior by comparison. In the interview, the former is referred to as the “carrot” and the latter as the “implicit stick.” I interpreted this framework as a design that raises standards through incentives for imitation rather than enforcement.It was also noted that the anecdote about the handshake at the AI Summit in India stemmed from confusion in the ceremony’s organization, reflecting a consistent stance against reducing the incident to a narrative of interpersonal conflict.

A Model Compatible with Values

As outlined in the discussion, it is far better to choose a business model that is compatible with one’s values.

When you feel that you can’t trust someone, when you feel that their values aren’t what they claim they are, when you feel they aren’t honest, when you feel they aren’t in it for the reasons they say they are, when you see disturbing patterns of behavior, dishonesty—that makes it very hard to continue working with the company or to continue trusting it.

Translation: When you feel you cannot trust someone, when you feel their values differ from what they claim, when you feel they are not honest, when you feel they are not acting for the reasons they state, and when you observe troubling patterns of behavior or dishonesty, it becomes extremely difficult to continue working for that company or to continue trusting it.

It’s far better to choose a business model that aligns with your values.

Translation: It is far better to choose a business model that aligns with your values.

The reasons for leaving OpenAI are also described not as differences in the details of safety policy, but as a matter of trust. The conclusion is that in a field where difficult decisions with no clear right answer are constant, differences of opinion alone may be a reason to stay, but when combined with a lack of trust in values or integrity, it becomes difficult to continue. He states that once he has left, the market and public opinion will be the judges.

This issue of trust also ties into the initial strategic bet: the decision to focus on coding and enterprise solutions rather than flashy consumer-facing apps.In the discussion, it is noted—as if with the two-sided market framework outlined by Rochet and Tirole in 2003 in mind—that consumer models that generate revenue through advertising tend to create incentives that maximize engagement and addiction.In contrast, they explain that in the enterprise sector, many benefits—such as disease treatment, energy efficiency, education, healthcare in developing countries, and economic growth—take the form of business solutions and are predicated on long-term relationships of trust.The growth of Claude Code and Claude Cowork is cited as an example, accompanied by the observation that by integrating into corporate operations, these services build ongoing relationships rather than relying on short-term attention.I interpreted this choice not as a declaration of right or wrong, but as a matter of incentive design. While conflicts arise in both models, there is nothing unreasonable about the explanation that they chose the side where conflicts can be minimized.

The Ceiling Determined by Computing Resources and the Reconfiguration of Employment

Computing resources represent the most tangible point where the index intersects with institutional frameworks. Throughout the discussion, it is repeatedly emphasized that model quality is paramount and that one should not rely on the inertia created by the effort required to switch models. Winning on quality also serves as the premise for a “race to the top” that consolidates trust. When constraints on computing resources are layered onto this, the discussion of capacity limits and employment issues converge into a single thread.

Computing Resources and the SaaSpocalypse

The supply crunch for NVIDIA H100 chips since 2023 is emblematic of the supply shortages repeatedly mentioned in earnings reports.In the discussion, it was noted that procurement had been supported by market liquidity, based on the forecast that future computing resources would grow three to four times annually. However, the threefold quarterly growth in the first quarter of 2026 significantly exceeded that assumption in a localized sense.When demand equivalent to 80 times the planned growth—which was based on a 10-fold annual increase—suddenly materializes, it is only natural that a sense of scarcity and a loss of confidence in the market would arise. While this shortage is described as temporary, it serves as a prime example of just how abruptly the gap between plans and reality can widen.

Over the first quarter of 2026, we saw revenue grow more than threefold on a quarterly basis. Just in a single quarter—not annualized—a threefold increase, which, of course, when raised to the fourth power, equates to an 80-fold increase over the course of a year.

Translation: In the first quarter of 2026, revenue grew more than threefold on a quarterly basis. This was a threefold increase in a single quarter—not annualized—and if raised to the fourth power, it would equate to an 80-fold increase over the course of the year.

On the SaaS front, the discussion covers the so-called “SaaSpocalypse”—an event in which approximately $285 billion in market capitalization evaporated following the launch of Claude Cowork.As Bessemer’s Cloud Report has long argued, SaaS moats have been supported not only by complex software but also by switching costs such as customer relationships, domain knowledge, and accumulated operational expertise. The discussion concludes that while moats built on complex software are becoming harder to defend, those based on customer relationships, domain knowledge, and new moats will endure.It explains that because the market itself is expanding tenfold, even if existing industries grow by 1.5 times, they will appear relatively small. I interpreted this to mean that the key to adaptation lies in taking stock of which moats will remain.

Reactions to the “Halving Theory”

Remarks regarding employment are reconstructed in light of how they were taken out of context and widely circulated. It is explained that a statement made a year ago—that half of entry-level white-collar jobs would be lost within one to five years—was not a definitive prediction but rather an illustration of the scale of the issue.Even with existing frameworks for employment outlook—such as the BLS Handbook—in mind, the difficulty of forecasting remains, much like the debates on automation during the Industrial Revolution. In the interview, he reflects that although countermeasures such as a token tax, collaboration with companies on redeployment and retraining, and fiscal policy were mentioned from the outset, only a three-second clip went viral.

The idea that this is cheap marketing is itself cheap marketing. This is laziness. This is a failure to engage with serious intellectual work.

Translation: The very notion that this is cheap marketing is itself cheap marketing. This is laziness; it shows a failure to engage with serious intellectual work.

I read this passage as a call to address both the magnitude of the problem and the suite of solutions simultaneously. It is not intended to stoke fears of doom, but rather a judgment that preparations cannot begin without conveying a sense of scale. What is required is not a back-and-forth of “cheap” accusations, but a refusal to shy away from serious consideration.

The Staircase from 90 to 100

The transition described in the discussion is a continuous shift from productivity enhancement to a stage where AI takes on direct responsibility. It is explained that if 90 percent can be automated, we go through a stage where the remaining 10 percent gains 10 times the leverage; as we approach 100 percent, job roles will need to be redefined.Using Anthropic’s internal software engineers as an example, the discussion notes that signs are beginning to emerge of a shift from a stage where AI writes nearly all the code but humans remain productive, to a stage where it is better for AI to perform the tasks directly.

Examples of alternative demand include jobs involving building and constructing in the physical world, roles centered on human relationships, and positions that guide AI. A caveat is added that there are no guarantees for any of these. I interpreted this caveat as a sign of integrity in avoiding definitive claims. This approach—presenting the scale of the issue alongside proposed countermeasures—is consistent throughout the discussion on employment.

The Boundary Between War and Autonomy: Mythos Tests Governance

In the realm of national security, the contrast between the smooth curve of the growth curve and the steep steps of institutional change is most pronounced. The dialogue discusses the history of contracts with the Department of Defense on the classification network under both administrations, as well as the two “lines” drawn in that context. The core concept of this section is the idea of gradually elevating governance in tandem with the growth of capabilities.

Two Red Lines

As outlined in the discussion, two perspectives coexist: the necessity of defense for the democratic camp, and the view that there is no point in defending it at the expense of compromising values. Against the backdrop of Russia’s invasion of Ukraine and the risk of a Taiwan contingency, the company explains that while it supports contracts within the classified network, these are not motivated by financial gain.On the other hand, the company rejects two red lines—mass surveillance and fully autonomous weapons—stating that this decision was made unanimously by all founders. It is reported that the company has not only ceased doing business with governments but has also suggested halting business with all companies that do business with governments. Since there are limits to what can be achieved through contracts, the company expresses a willingness to welcome bipartisan legislation in Congress.

Here, I read this in conjunction with the framework outlined in Executive Order 14110 in October 2023—including its reporting requirements—as well as the principles of distinction and proportionality under international humanitarian law as systematized by the ICRC, and the ongoing discussions on autonomous weapons within the CCW at the UNODA.The idea of not confining the boundaries solely to contracts but gradually elevating them to the level of legislation and international debate is consistent with the overall structure of this dialogue—responding to a smooth exponential curve with a staircase. Moves toward defense adoption, such as the “Replicator” concept being advanced by the DoD since 2023, can also be viewed as attempts to ascend another step on the institutional staircase.

Mythos and Disclosure Restrictions

The latest Mythos model is described as having the capability to autonomously complete all stages of the cyber kill chain. The seven-stage kill chain outlined by Lockheed Martin is a framework that views the entire process—from reconnaissance to weaponization, delivery, exploitation, installation, command and control, and mission accomplishment—as a single, continuous sequence.Mythos has made significant strides from vulnerability discovery to exploitation, as demonstrated by its identification of 271 new vulnerabilities in Firefox.

Some of the early companies we provided this to said things like, “This is a superweapon. You should need a gun license to use it. Please don’t release this.”

Translation: Among the companies to which we initially provided the tool, some said, “This is a superweapon,” “You should have to hold a gun license to use it,” and “Please don’t release this.”

It is said that comments from the companies provided with the tool early on, calling it a “superweapon,” led to the decision to withhold its public release. Classifiers such as Opus 4.7 are cited as countermeasures, but they are considered insufficient because they can be bypassed.It is explained that rushing to make the information public would put it into the hands of attackers, so the priority is to first provide it to the defense side to patch the vulnerabilities. I interpreted this sequence as a form of governance that intentionally delays the disclosure timeline relative to the exponential growth of capabilities—a strategy to buy time while capabilities advance and defenses catch up.

While addressing issues such as the expansion to 1,000 to 5,000 targets per day reported by Bloomberg, the collaboration with Palantir, and the accidental attack on a girls’ school in February, the discussion emphasizes the importance of upholding the principle that humans make the final decisions.The dangers of full autonomy are emphasized, and both deterrence and misidentification are discussed. The possibility that AI warfare could serve as a deterrent is juxtaposed with the risk that unrestricted operations could provoke conflict.

Recursive Improvement and the Theory of Appropriation

Recursive self-improvement is discussed not as a single event occurring at a specific moment, but as a continuous process. Figures are cited showing that AI-driven productivity gains have risen from 10 to 15 percent a year ago to the current 20 to 30 percent. While criticizing the back-and-forth over whether or not to regulate, the discussion explains that countermeasures should be gradually increased in line with the exponential growth of the technology.The figure of a 10 to 25 percent probability of collapse is also discussed as a range resulting from the unpredictability of the technology, accompanied by the clarification that the company’s own actions are actually reducing that risk. Using the analogy of an airline, the report notes the caveat that even if a company is ten times safer than its competitors, crashes cannot be reduced to zero.

In response to arguments for government expropriation, the company points out an asymmetry: while traditional technologies such as nuclear power and the Internet originated with the government, AI is the first technology to have emerged from the private sector. Acknowledging its inherent instability, the company explains its governance structure—the Long-Term Benefit Trust.It is stated that, under the articles of incorporation, the trust can appoint and dismiss a majority of the board of directors, effectively allowing it to remove the CEO. As a check on the government’s power, the involvement of the legislative and judicial branches is called for. I interpreted this dual system of checks as a design intended to prevent power from being concentrated in either branch.The philosophy that governance should be built up step by step—precisely because the index grows smoothly—is consistently applied here as well.

From a Smooth Exponential Curve to a Stage-Based Perspective

What I’ve been tracing so far is a single thread: how to handle the friction that arises between a smooth exponential curve and a stair-step institutional framework. By exploring asynchrony to find narrow veins and bundling variations in trust to raise the standard. When reaching the ceiling of computational resources, we assume market liquidity while reassessing the premises of the plan.We take stock of the SaaS moat and are forced to adapt under the pressure of a market expanding tenfold. At the boundary between war and autonomy, we draw two lines; capabilities like Mythos delay the public rollout of new features to solidify defenses first. We view recursive improvement not as a single moment but as a continuum, and governance is built up continuously as well.I interpreted this series of actions not as a back-and-forth between optimism and panic, but as a stance—like that of a surgeon or military commander—of maintaining composure while never taking one’s eyes off the risks.

The saying that “trust is earned through action” resonates most strongly at the end. Choices that entail losses—such as the decision to withhold the release of Mythos or the blocking of access to China, resulting in hundreds of millions of dollars in losses—are cited as evidence. While the growth curves discussed in the dialogue appear smooth, we can only build our institutions one step at a time.I believe that accepting this gap and building a foundation step by step is the realistic approach to dealing with this technology. The world will be seen in a new light. I feel that what is required of us is not merely to gaze at the exponential curve, but to take the initiative in deciding which steps to take and when.