A Conversation with Dario and Daniela Amodei: Why Is Claude Still Rate-Limited?

At the Code with Claude event in San Francisco on May 6, siblings Dario Amodei and Daniela Amodei took the stage together. This was Anthropic’s second developer conference, held on the same day the company announced it had secured all compute capacity at SpaceX's Colossus 1 data center (over 300 MW, 220,000 NVIDIA GPUs).

The conversation was hosted by Anthropic’s Chief Product Officer Ami Vora (who will succeed Mike Krieger at Labs in January 2026). The discussion began with "the feeling of being on an exponential curve," covering topics such as the developer ecosystem, the next phase of model training logic, Anthropic’s trade-offs in capability release, and what upcoming changes excite Dario most over the next six months.

Below is a transcript of this roughly thirty-minute dialogue, originally from the official Anthropic Code with Claude series.

Original video: https://www.youtube.com/watch?v=7xco5Qd2Oo8

Key Takeaways

• One, Anthropic originally planned its compute infrastructure for “10x annual growth,” but actual growth in Q1 2026 annualized to about 80x, which is the direct reason why Claude has been throttled. Dario openly said he hopes growth slows back down to 10x: “80x is too crazy—we can’t handle it.”

• Two, a year ago at last year’s Code with Claude, Dario told Mike Krieger that the first “one-person, $1 billion valuation” company would emerge in 2026. With seven or eight months left until the end of 2026, the latest update is: there are already two-person startups valued at $1 billion, and solo founders with valuations in the hundreds of millions.

• Three, software engineers are the “leading indicator” of how AI spreads across the economy. How developers use Claude foreshadows how other industries will adopt AI in the future.

• Four, coding ability improves quickly because it’s “verifiable”—you run unit tests to see if it works. The next hard challenge lies in subjective capabilities like security, design quality, and code review—areas that can't be automatically validated by tests. Anthropic is actively training models to tackle these, which will also benefit writing and scientific research.

• Five, “Hold light and shade” is an internal cultural principle at Anthropic. The latest example is their strongest model, Mythos: because it can identify and exploit software vulnerabilities, the company chose not to publicly release it, instead distributing it through Project Glasswing to over 50 organizations to strengthen defensive capabilities.

• Six, what Dario is most excited about in the next six months is organizational-level AI. AI won’t just help one person do the work of many—it will enable repeated execution of tasks within groups and organizations.

[1] What Does 80x Annual Growth Feel Like?

Ami opened with a soul-searching question: You two are living this exponential curve firsthand—what does this kind of growth actually feel like?

Daniela responded with an inside joke from the company. There's a “roller coaster” meme in Anthropic’s Slack channel—a graph where the slope suddenly goes vertical. She said she and Dario are like passengers sitting at the front and back of the cart: “Depending on where you sit, the whiplash feels different.” Then she added a line that made the audience laugh:

We’re not totally sure that the operator of the roller coaster isn't like a 15-year-old who's doing a summer job of questionable mental stability.

Dario’s answer was more “scientific.” He said he and several co-founders wrote this very curve over a decade ago using scaling laws (predictable performance gains as training compute increases), forecasting everything from spending $1,000 per month to eventually billions, and predicting exactly how well models would perform on various tasks. So on paper, what’s happening now was expected.

Note: Dario Amodei first observed the “bigger scale, better performance” pattern during the Deep Speech 2 project at Baidu Research in 2014. In 2020, while at OpenAI, he co-authored the influential scaling laws paper. Multiple members of Anthropic’s seven co-founders contributed to this research. This context explains his statement that they “predicted this curve over a decade ago.”

But he emphasized that writing the curve on paper and seeing it unfold in real life are two entirely different experiences. He used a famous scene from *Interstellar* as an analogy: when the spaceship lands on a planet near a black hole, waves rise 2,000 feet high.

I was a physicist—I know the math, general relativity, how much things can be sheared. But actually seeing it on human scale, there's something deeply strange and unsettling about witnessing it happen.

Dario then grounded the “exponential curve” into three concrete numbers.

First, this year marked the first time in company history that Claude caused an upward inflection point in internal PRs (pull requests). The rate at which Claude writes code now exceeds the rate at which humans contribute.

Second, external growth this year has—for the first time—“surpassed exponential.” Anthropic originally planned compute capacity assuming “10x annual growth,” preparing multiple scenarios from flat to 10x. But in Q1 2026, if you were to annualize the quarterly pace, revenue and usage grew 80x.

Note: Dario used the phrase “if you were to annualize it,” meaning the 80x figure extrapolates a single quarter’s explosive growth over a full year. Actual full-year growth is unlikely to sustain that level, but even discounted, it far exceeds Anthropic’s 10x planning margin.

Third, this is precisely why Anthropic has been throttling access. Dario apologized with a wry tone:

I hope the 80x growth doesn't continue 'cause that's just crazy and it's too hard to handle. I hope for some more normal numbers, a mere 10x.

He then pivoted to another announcement from that day:

As you saw today with the SpaceX compute deal, we're working as quickly as possible to provide more compute than we have in the past.

Note: On May 6, Anthropic simultaneously announced a deal with SpaceX to take over all compute capacity at the Colossus 1 data center (located in Memphis, Tennessee, formerly part of Elon Musk’s xAI). Over 300 MW and more than 220,000 NVIDIA GPUs will come online “within a month.” Other Anthropic compute agreements include up to 5 GW with Amazon (nearly 1 GW online by end of 2026), a 5 GW agreement with Google + Broadcom (coming online starting 2027), and a $30 billion strategic compute partnership with Microsoft + NVIDIA on Azure. Musk had previously criticized Anthropic and Dario publicly, but on May 6 tweeted that after meeting senior Anthropic leadership the previous week, he was “impressed.” This deal itself serves as direct validation of the narrative in this talk—that compute is the real bottleneck.

[2] Why Anthropic Places Developers at the Top of the User Pyramid

Ami then turned the conversation toward the developer community. Nearly everyone in attendance that day was a developer, and she wanted to hear how Dario and Daniela positioned this group.

Daniela was direct: in many ways, developers are Claude’s most important users. There are several reasons:

First, Anthropic itself is largely composed of developers, so they’re especially sensitive to the tools they build.

Second, feedback from the developer community is genuine. Anyone who’s built products knows how rare that is:

You build a product and you're like, I see some numbers—those are nice—but the genuineness with which the developer community engages with us is something that is so special.

Finally, from day one, Anthropic has primarily built for developers and enterprises—a focus Daniela believes is uncommon in the AI space.

She listed domains where Claude has already taken root—medicine, software development, financial services—almost every industry now includes companies centered around developers who are reshaping their businesses with Claude. She described this relationship as both a privilege and a responsibility.

Dario added another perspective. He said technology does not diffuse evenly across the economy; software engineers are always the earliest adopters of new technologies. So the spotlight on programming isn’t accidental—“it’s a miniature preview of how AI will transform the entire economy.”

[3] The “One-Person Billion-Dollar Company” Bet: Seven or Eight Months Left

Dario then extended the thread on “developers” into a specific bet. About a year ago, at Code with Claude 2025, Mike Krieger asked him directly:

In what year will the first company worth $1 billion with only one employee appear?

Dario answered: 2026. Now, with seven or eight months remaining, the audience laughed. Dario added half-jokingly, half-seriously:

That's eternity on the exponential.

He offered a hint: there are already two-person AI-native startups valued at $1 billion, and solo founders with valuations in the hundreds of millions—but strictly speaking, the “one-person, $1 billion” milestone hasn’t been hit yet. In his view, the real significance isn’t about saving labor costs, but that for the first time, a single visionary individual or tiny team could access resource scales previously attainable only after years of accumulation, enabling them to build what they imagine.

We’ve moved from “models helping us write code,” to “models helping us think about software engineering as a task,” to “models helping us think about entire business units, entire economic units, as a task.”

Note: Mike Krieger is co-founder of Instagram, joined Anthropic as Chief Product Officer in 2024, transitioned to the newly formed Anthropic Labs in January 2026 as a technologist focusing on experimental product incubation (including the now-famous Mythos project), succeeded by Ami Vora as CPO. In that prior conversation, Dario estimated a 70%-80% chance the prediction would come true. The bet concludes on December 31, 2026. He did not name the specific “two-person, $1 billion” startup, so this claim cannot currently be independently verified.

[4] From Single Agent to Multi-Agent: The Next Bottleneck Is Verification

Ami followed up by asking Dario how developers’ use of Claude might evolve next. Dario outlined several interlocking trends.

First, a shift from single agent to multi-agent systems. A developer won’t just have one Claude—they’ll have a team of Claudes, possibly arranged hierarchically, with higher-level Claudes delegating subtasks to lower-level ones. Dario used a metaphor he often returns to:

We're gradually making our way to the country of geniuses in the data center. We're starting with a team of smart people in a room or something.

Second, Claude Code is currently focused on boosting productivity for "individuals," but Anthropic is increasingly thinking about efficiency gains across entire teams and organizations—where the combined output of a group of people plus a group of Claudes exceeds the sum of their individual contributions.

Third—and something Dario emphasized repeatedly: you must consider Amdahl's Law. When one part of a system is accelerated to its limit, the bottleneck simply shifts to the unaccelerated portion.

You mentioned PR volume—if you're in an organization where you can now produce three or four times as many PRs as before, you quickly realize there are all these other things holding you back. If only one segment speeds up while the rest doesn’t keep pace, problems will inevitably arise. ("If you're living in a world where you can, within an organization, write three or four times as many PRs as you could previously, you start to understand there are all these other things that are holding you back or that will go wrong if you speed up just that and not everything else.")

He identifies what those “other things” are: security, validation, code review, and design quality. What Anthropic needs to do next isn't further point optimizations, but rather lift the entire cycle of bottlenecks together—so acceleration can be released in a "smooth and reliable" way.

Note: Amdahl's Law originates from a 1967 formula by computer scientist Gene Amdahl on parallel computing. It originally stated that if only part of a program can be parallelized while another part remains serial, the overall speedup is limited by the serial portion. Dario borrows this concept to describe collaboration bottlenecks in engineering organizations—it’s a core analytical framework he returns to throughout the conversation, and which reappears later when discussing products and model training.

[5] The Way We Train Models Must Also Evolve

Ami follows up: Could these trends feed back into how Anthropic trains its models?

Dario offers two layers of response.

The first is already happening: Anthropic is using Claude to accelerate the development of Claude itself.

The second layer is more profound. Dario explains that software engineering has become the fastest-advancing domain for AI due to one key trait: *verifiability*. Given a coding task, the model generates code, which then runs through unit tests—delivering immediate, clear feedback on correctness. This simple, brutal, effective loop makes training highly efficient.

But large parts of software engineering remain unverifiable:

Is this code *really* correct? Can we detect errors? Are there security flaws? These questions aren’t nearly as easy to verify. (“Is this thing really right? Can we find errors? Are there security issues? Not quite as verifiable.”)

The implication is straightforward: training efficiency depends on how easily outcomes can be verified. Code passes tests; correctness is binary—hence rapid progress. Security analysis and design judgment lack such automated verification, so improvement lags. Once Anthropic breaks through on training for these “semi-subjective” tasks, the benefits won’t be confined to software engineering—they’ll extend to writing, scientific research, and beyond.

Dario reframes this with Amdahl’s Law: within software engineering, the “soft,” subjective capabilities—currently the bottleneck—are disproportionately important.

[6] Mission: Walking the Tightrope Between Speed and Responsibility

Ami turns to mission. As Anthropic grows and stakes across the industry rise, what should outsiders understand most about the company?

Daniela outlines two pillars.

One is about doing transformative technology well—ensuring it benefits everyone. Claude is a tool that amplifies human creativity and ambition—that’s the opportunity.

The other acknowledges risks: labor disruption, whether releases are safe, and whether the technology truly helps people.

Daniela says Anthropic aims to treat both sides with equal weight. She introduces a key internal cultural phrase: “Hold light and shade”—embracing both light and shadow.

She cites the recently launched “Mythos and Glasswing” as an example:

With a model capability level like Mythos, the potential applications are enormous. But because of certain security vulnerabilities, we wanted to be slightly more cautious in our release.

She summarizes this tension:

This balance we strike is actually quite delicate. We want to ship fast, build the best products, and release the strongest models—but we also want to act responsibly. Most of our decisions start from calibrating between these two poles.

Note: Claude Mythos Preview was released by Anthropic in April 2026. It demonstrated generational performance in cybersecurity tasks, uncovering numerous zero-day vulnerabilities across major operating systems and browsers. Project Glasswing is a complementary defensive alliance, partnering with dozens of critical infrastructure organizations to use Mythos for scanning and patching vulnerabilities. Due to these security concerns, Mythos was released under strict limitations. “Glassman” in the transcript appears to be a speech recognition error for “Glasswing.”

[7] Product Thinking Under Exponential Curves: Building Products *for* AI vs. *with* AI

On product, Daniela starts with a joke at Ami’s expense: “You just said Dario and I 'leaned in a lot' on product—translated, that means you two meddle in my work every day. Can I please just get some peace?”

But she quickly pivots, acknowledging they’re deeply involved because product is how Anthropic’s vision becomes tangible. She adds a less commonly heard perspective: inside Anthropic, “product” and “research” are two interdependent inputs. Sometimes you think, “We need a better tool,” but more often, “product innovation is driven by emergent model capabilities.”

Her example is programming. Anthropic didn’t set out from day one to build a programming product. At a certain point, the team noticed the model could generate “decent, though imperfect” code, and observed that many power users were developers. That sparked the idea: “Maybe we should build something for this group.” And thus, Claude Code was born.

Dario breaks this down further, distinguishing between building products *for* AI and building products *with* AI.

First, building products *for* AI. He shares several key principles.

One: the technical foundation evolves rapidly in the AI era. In the 2010s, tech stacks advanced steadily, with occasional new frameworks. Today, each leap in model capability suddenly makes previously impossible products viable. So internal experimentation must be continuous—“even if something seems impossible now, come back and try again in a few months.”

He shares a firsthand example:

We actually tried something like Claude Code back in 2022. It was pretty frustrating—the idea was sound, but the models were too weak to extract real value. I’ve been training these models since 2015. They were really dumb. (“If we had tried to do Claude Code in 2022, it wouldn't have worked because the models wouldn't have been strong enough...I've been training these models since 2015. They were really dumb.”)

Two: in the AI era, product saturation is pushed forward by increasingly capable models. Dario notes chatbot interfaces are nearing saturation—there’s still a large market, but marginal gains from smarter models are diminishing. New capabilities today are increasingly manifesting in agentic forms like Claude Code.

Three: the API market will never disappear. New products keep emerging—both internally at Anthropic and externally. Beyond code, in domains like healthcare, law, and finance, each new tier of model capability unlocks fresh application spaces.

Four (and back to Amdahl’s Law): when building products *with* AI, Dario observes a phenomenon internally: release velocity increases by 2x, 4x, even 5x, but then “systemic debt” begins to surface.

Using AI to accelerate shipping lets you achieve throughput that was impossible a year ago—but you also accumulate technical debt at an astonishing rate. Then you’re forced to ask: Can we use AI to pay down that debt, or at least track what it is? And then you realize the team must collaborate in fundamentally different ways. These insights emerge constantly, month after month. (“It's possible to accumulate an extraordinary amount of internal technical debt when you ship that fast. And so then you have to say, well, can we also use the AI models to undo that technical debt or keep track of what it is that we're doing?”)

Thus, the AI era isn’t just about faster release cycles—“the very way you *work* is forced to upgrade at high frequency.”

Ami adds her own insight: the fundamental problems don’t change that quickly—people remain people. But you must maintain “a fresh lens on technology” and accept that “your daily work keeps evolving, because the bottleneck shifts every few months.”

[8] What Excites Dario Most in the Next Six Months

Ami asks Dario to summarize in one sentence: what upcoming model capability excites you most over the next six months?

Dario gives a multidimensional answer: the leap from “personal AI” to “organizational AI.”

What excites me is this idea: AI isn’t just doing the work of many people for one person, but that it performs the work of many people, repeated many times over, within a human organization. (“AI is not just doing the work of many people working for one person, but that it does the work of many people many times over by operating within an organization of humans.”)

He connects this thread to the “$1 billion company with one person” bet: that bet may actually be *understated*. What’s more likely is “a team augmented by AI accomplishing what once required hundreds or thousands of people,” rather than one individual single-handedly running a billion-dollar company.

[9] The Claude Use Cases That Moved Them Most

Finally, Ami turns to Daniela: which user stories touched you the most?

Daniela shares several strikingly diverse examples.

The first is a mobile doctor project in the Global South. In some regions, seeing a real doctor requires walking dozens of miles on dirt roads. Yet health needs persist. Developers used Claude to create diagnostic-style interfaces, delivering vetted medical advice—translating model capability into practical tools for low-resource settings.

She also mentions acceleration in biomedical research, a cause she’s long championed.

The next two are more personal. One developer used Claude to recover wedding photos from a corrupted hard drive. Another used it to monitor the growth of tomatoes in their garden.

Daniela laughs at the tomato story: “I would never have thought of that. But do you have a live camera feed? I’d subscribe.”

On the question of what AI can do, users’ imagination will always outpace product managers’ roadmaps.

Quick Q&A Recap

Q: How fast is Anthropic growing today? Annualized at current quarterly pace: 80x (Dario used the qualifier “if you were to annualize it”—a projection based on short-term explosive growth). Originally prepared for 10x compute capacity, so they’ve been throttling usage.

Q: What did the SpaceX compute deal solve? Within the next month, over 300 MW and more than 220,000 NVIDIA GPUs will come online. Anthropic will rapidly convert this compute into higher usage caps for developers.

Q: Where does the “$1 billion company with one person” bet stand now? There are already cases of two-person $1 billion companies and single individuals achieving hundreds of millions (Dario didn’t name them, so independent verification isn’t possible). His prediction window from Code with Claude 2025 was 2026, with 70%-80% confidence. About seven to eight months remain.

Q: What upcoming model capability excites Dario most in the next six months? Organizational AI. AI no longer just does many people’s work for one person, but repeats that work many times over within a human organization.

Q: How does Anthropic make trade-offs in releasing capabilities? Internally called “holding light and shade.” The Mythos model wasn’t publicly released due to safety risks; instead, it was selectively distributed via Project Glasswing to dozens of institutions for defensive strengthening.

Final Thoughts

The central tension revealed in this conversation is Anthropic’s attempt to balance two extreme positions—a kind of “self-opposing” contradiction.

On one hand, it’s the fastest-growing AI company. An 80x annualized growth rate (even if selectively calculated), the SpaceX compute partnership, and Claude Code driving a visible upward inflection in internal PR volume. Dario admits onstage that 80x is unsustainable and hopes to stabilize at 10x—then, on the same day, closes one of the hardest deals in the industry. This is the strongest evidence yet that they’ve pursued *every available source of compute*.

On the other hand, it's also the most cautious AI company. Models like Mythos are being withheld from release solely due to safety risks, and "holding light and shade" has become a recurring mantra for survival. Faced with such a powerful model, Anthropic has effectively sacrificed speed-to-market.

Balancing these two priorities is far more difficult than what Dario and Daniela described on stage. An 80x growth rate means immense delivery pressure—Dario himself admitted that technical debt is “accumulating at an alarming rate.” Under such intense momentum, applying brakes for safety evaluations and insisting on responsible releases requires more than just principles—it demands daily, hard-nosed trade-offs in resource allocation.

Dario’s repeated references to Amdahl’s Law form the key analytical framework of the entire conversation. It points to a more practical question than “AI makes everything faster”: after acceleration, where does the bottleneck shift? For developers, this question deserves far more attention than yet another announcement about stronger models.

Two signals worth tracking continuously: After Colossus 1 goes live, will usage caps truly be relaxed significantly? Doubling the 5-hour limit while keeping weekly caps unchanged feels like semantics. By year-end, how much of the GW-scale commitments from Amazon, Google, and Microsoft will actually translate into usable compute for users? And when will Mythos exit preview and launch on Glasswing—and under what conditions? The former tests Anthropic’s infrastructure capabilities as a product company; the latter tests how long the “light and shade” principle can endure under commercial pressure.

As for the “$1 billion per person” company bet, there are still seven or eight months until 2026 concludes. Dario has already started reframing it on stage: the real challenge may be “a small group plus AI accomplishing what used to require hundreds.” If that reframing holds, the “one-person unicorn” becomes the least interesting part of the story.

Original video source: Anthropic Code with Claude event in San Francisco, May 6, 2026, “A conversation with Dario Amodei & Daniela Amodei”.