Dario & Daniela Amodei's Latest Conversation: Why Does Claude Keep Rate-Limiting?Anthropic 兄妹 Dario Amodei 和 Daniela Amodei 最新对话：Claude 为什么一直限速？

At the May 6 Code with Claude San Francisco event, Anthropic siblings Dario Amodei and Daniela Amodei sat down together on stage. This was Anthropic's second developer conference, and on the same day, Anthropic had just announced a deal with SpaceX for the entire compute capacity of the Colossus 1 data center (over 300 MW, more than 220,000 NVIDIA GPUs).

Moderating the conversation was Anthropic's Chief Product Officer Ami Vora (who took over in January 2026 from Mike Krieger, who moved to Labs). The topics ranged from "what it feels like to be on an exponential curve" to the developer ecosystem, the next steps in model training logic, Anthropic's tradeoffs around releasing capabilities, and finally the capability change Dario is most excited about over the next six months.

Below is an edited summary of this roughly half-hour conversation, from the original video in Anthropic's official Code with Claude series.

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

Key takeaways

First, Anthropic originally planned compute around "10x per year," but the actual growth rate in Q1 2026, annualized, came out to about 80x — this is the direct reason Claude has been rate-limited. Dario said plainly he wants growth to return to 10x, "80x is too crazy, we can't handle it."

Second, a year ago at last year's Code with Claude, Dario told Mike Krieger that 2026 would see the first "billion-dollar, one-person company." With seven or eight months left in 2026, the latest development is: there are already AI companies valued at a billion dollars with two people, and cases of single individuals valued at hundreds of millions of dollars.

Third, software engineers are the "leading indicator" for how AI diffuses through the whole economy. How developers use Claude foreshadows how other industries will use it in the future.

Fourth, coding ability has advanced quickly because it's "verifiable" (you know if it works by running unit tests). The next real hard nut to crack is "subjective" abilities like security, design quality, and code review that can't be automatically judged by unit tests. Anthropic is training models to tackle these, which will also feed back into writing and research.

Fifth, "holding 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 didn't release it publicly, instead taking the limited path of Project Glasswing, distributing it to more than 50 institutions to strengthen their defenses.

Sixth, the capability change Dario is most looking forward to over the next six months is organizational-level AI. AI will no longer just be doing the work of many people for one person, but repeating that work many times over within an organization made up of many people.

What does an 80x annualized growth rate feel like

Ami opened with a soul-searching question: you two are the ones who genuinely feel this exponential curve firsthand — what does this growth feel like?

Daniela picked up first with an internal company joke. There's a roller-coaster meme in Anthropic's Slack, the kind where the slope suddenly shoots straight up. She said she and Dario are like sitting at the front and back of the car respectively — "depending on which end you're sitting at, you get a different kind of whiplash." She then added a line that got laughs from the audience:

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 like questionable level of sound mind.

Dario's answer was more "science-minded." He said that he and several co-founders wrote down this curve more than a decade ago through scaling laws (the rule that model capability grows predictably with training compute), predicting "spend $1,000 a month first, then $10,000, then $100,000, all the way up to hundreds of billions, and the model will reach such-and-such level on this task and that task." So on paper, everything happening now was actually within prediction.

Note: Dario Amodei first observed the pattern of "bigger is better" while working on the Deep Speech 2 project at Baidu Research in 2014, and in 2020 co-authored the influential scaling laws paper at OpenAI. Several of Anthropic's seven co-founders participated in that research. This is the background for his statement that "we predicted this curve more than a decade ago."

But he said writing the curve down on paper and actually seeing that curve become reality are two different things. He used the famous scene from Interstellar as an analogy: the spaceship lands on a planet near a black hole, where the waves are 2,000 feet high.

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

Dario then made the "exponential curve" concrete with three numbers.

First, this year is the first time in the company's history that Claude has caused an upward inflection point in the number of internal PRs (pull requests, code merge requests) at Anthropic. Claude's speed at writing code has outpaced the rate at which people are being added.

Second, the company's external growth this year, for the first time, "exceeded the exponential." Anthropic had originally planned compute around "10x per year," preparing multiple scenario plans ranging from "almost no growth" to "10x growth." But in Q1 2026, if you annualize the quarter's pace, revenue and usage came out to 80x.

Note: Dario used the qualifier "if you were to annualize it," meaning the 80x figure extrapolates a single quarter's surge across the full year. The actual full-year growth rate is unlikely to hold at this level, but even discounted, the number still far exceeds the company's 10x planning buffer.

Third, this is why Anthropic has been rate-limiting. Dario used an almost apologetic tone:

80x is too crazy, we genuinely can't handle it — I hope it can come back down to something more normal, like just 10x.

He then connected this to another piece of news from that same 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, and we'll get it to you as fast as our capabilities allow.

Note: The news Anthropic announced simultaneously on May 6 was a deal with SpaceX to use the entire compute capacity of the Colossus 1 data center (located in Memphis, Tennessee, formerly under Elon Musk's xAI), bringing online more than 300 MW and over 220,000 NVIDIA GPUs "within a month." Anthropic's other compute deals include: an agreement with Amazon worth up to 5 GW (with nearly 1 GW coming online before the end of 2026), a 5 GW agreement with Google + Broadcom (starting in 2027), and a $30 billion Azure compute strategic partnership with Microsoft + NVIDIA. Musk had previously publicly criticized Anthropic and Dario multiple times, but on May 6 he tweeted that after meeting with Anthropic leadership the previous week, he came away "with a good impression." This deal announcement itself is the most direct endorsement of the interview's narrative that "compute is a real bottleneck."

Why Anthropic puts developers at the top of the user pyramid

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

Daniela put it very directly: in many senses, developers are Claude's most important users. There are several layers to this reasoning:

First, Anthropic itself is primarily made up of developers internally, and they're the most sensitive to the tools they build.

Second, the feedback from the developer community is genuine. Anyone who's built a product knows how rare that is:

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

Finally, Anthropic has, from day one, built products "primarily for developers and enterprises," which Daniela feels is actually not that common in the AI world.

She listed the fields Claude has already penetrated, including medicine, software development, and financial services — in almost every industry there are developer-centered companies using Claude to reshape their business. She described this relationship as "both a privilege and a responsibility."

Dario added from another angle. He said technology doesn't diffuse evenly through the economy — software engineers are always the group fastest to adopt new technology. So it's no accident that this industry's spotlight is on coding right now: "it's a microcosm preview of how the entire economy is going to be transformed by AI next."

The "one-person billion-dollar company" bet, seven or eight months left

Dario then took the "developer" thread to a specific bet. He said about a year ago, at Code with Claude 2025, Mike Krieger asked him directly:

What year will the first billion-dollar, one-person company appear?

Dario's answer at the time was 2026. Now there are seven or eight months left. The audience laughed. Half-jokingly, half-seriously, Dario added:

That's eternity on the exponential.

He gave a bit of a preview: there are already two-person, billion-dollar companies built on AI, and cases of single individuals valued at hundreds of millions of dollars, but strictly speaking, the "one person, one billion dollars" bet hasn't been fulfilled yet. In his view, the real significance of this isn't "saving on labor costs," but that for the first time, a single individual with an idea, or an extremely small team, can potentially command a scale of resources that used to take years to accumulate, in order to build what they imagine.

We've already gone from "the model is helping us write code," to "the model is helping us think of software engineering as a task," to "the model is helping us think of an entire business unit, an entire economic unit, as a task."

Note: Mike Krieger is a co-founder of Instagram, who joined Anthropic as Chief Product Officer in 2024 and moved in January 2026 to the newly established Anthropic Labs as a technical staff member, focusing on incubating experimental products (the most notable current project being Mythos, mentioned later), with Ami Vora taking over as CPO. In that earlier conversation, Dario gave the probability as "70%-80% likely to happen." The deadline for this bet is December 31, 2026. However, he didn't give a specific named example of the "two-person, billion-dollar company," so this claim currently cannot be independently verified.

From single agent to multi-agent, the next bottleneck is verification

Ami followed up by asking Dario how the way developers use Claude will change going forward. Dario offered several intertwined trends.

First, moving from a single agent to multiple agents. A developer no longer has just one Claude, but a whole group of Claudes, possibly organized hierarchically, where an upper-level Claude subcontracts tasks to lower-level Claudes. Dario used a metaphor he often reaches for:

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 currently mainly boosts productivity for "individuals," but Anthropic is increasingly thinking about boosting productivity for "entire teams and organizations," so that a group of people plus a group of Claudes produces more together than the simple sum.

Third, and something Dario repeatedly emphasizes: look at Amdahl's law. When one segment is accelerated to its limit, the bottleneck jumps to the segment that hasn't been accelerated.

You mentioned the number of PRs — 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 pointed out exactly what these "other things" are: security, verification, code review, design quality. What Anthropic plans to do next isn't to speed up a single point further, but to lift this whole ring of bottlenecks together, so the acceleration can be released "smoothly and reliably."

Note: Amdahl's law comes from a parallel-computing formula proposed by computer scientist Gene Amdahl in 1967. It originally described how, if only part of a program can be accelerated in parallel while another part must remain serial, the overall speed is limited by that serial portion. Dario borrows it to describe collaboration bottlenecks in engineering organizations — it's the core analytical framework he keeps returning to throughout this conversation, and he uses it again later when discussing products and model training.

How model training itself has to change too

Ami followed up: will these trends, in turn, change the way Anthropic trains its models?

Dario's answer had two layers.

The first layer is something already happening: Anthropic is using Claude to accelerate Claude's own development.

The second layer is more interesting. Dario said software engineering has advanced faster in AI than almost any other field because of one special property: verifiability. Give the model a coding task, it writes the code, and running unit tests immediately tells you whether it's right. This feedback loop is simple, blunt, and effective, which makes training especially efficient.

But there's a large part of software engineering that isn't verifiable:

Is this thing really right? Can we find errors? Are there security issues? Not quite as verifiable.

The logic here is straightforward: training efficiency depends on how easy something is to verify. Code can be run against tests, so right and wrong are immediately obvious, and progress is fast; security analysis and design judgment lack this kind of automatic verification mechanism, so progress is slower. Once Anthropic breaks through on training for these "semi-subjective" tasks, the benefits won't be limited to software engineering — writing, scientific research, and other fields will benefit too.

He recast this using Amdahl's law: within software engineering, those "soft, subjective" capabilities, precisely because they are the current bottleneck segment, will become disproportionately important.

Mission: walking a tightrope between shipping fast and shipping responsibly

Ami turned the conversation to mission. As Anthropic grows larger and the stakes for the whole industry keep rising, what does the outside world most need to understand about Anthropic?

Daniela offered two pillars.

One is "how to build this transformative technology well, so that it benefits everyone." Claude is a tool that amplifies the ambition and capability people create — that's the side of opportunity.

The other is acknowledging the risks: risks to the workforce, whether releasing the technology is safe, whether it's genuinely beneficial to people.

Daniela said what Anthropic wants to do is treat both ends with equal weight. She brought up an internal cultural keyword: "Hold light and shade."

She used the recently released "Mythos and Glasswing" as an example:

A model at Mythos's capability level has enormous potential in what it can be used to do. But because there are certain safety vulnerabilities, we wanted to be a bit more careful about the release.

She summed up this tension this way:

The balance we're striking is actually quite delicate. We want to ship things as fast as possible, build the best products, and release the strongest models — but we also want to do it responsibly. Most of our decisions start from calibrating back and forth between these two pillars.

Note: Claude Mythos Preview is a preview model Anthropic released in April 2026 that demonstrated cross-generational capability in cybersecurity tasks, discovering a large number of zero-day vulnerabilities across major operating systems and browsers. Project Glasswing is the accompanying defensive security coalition, partnering with dozens of critical-infrastructure organizations to use Mythos to scan for and patch vulnerabilities. Precisely because of these safety risks, Mythos was released only to a very limited scope. "Glassman" in the transcript appears to be a speech-recognition error for "Glasswing."

A view on products under the exponential curve: building products for AI vs. building products with AI

Turning to products, Daniela first teased Ami a bit. She said, "You just said Dario and I have 'leaned in a lot' on product — translated into plain speech, that means: you two keep sticking your hands into my business, can't you let me just work in peace?"

But she then acknowledged that the two of them really are quite particular about product, because product is the outward expression of what Anthropic is trying to do. She also offered a less commonly heard perspective: inside Anthropic, "product" and "research" are two inputs that pull on each other. Sometimes you think "we should build a better tool," but more often, "product innovation is pushed forward by new capabilities emerging from the model."

Her example was coding: Anthropic didn't set out from day one to build a coding product. At some point, the team noticed the model could already write code that was "pretty good, not perfect," and also observed that many power users were themselves developers, which naturally led to the idea "we should build something for this group" — and that's eventually how Claude Code came about.

Dario then broke this topic down further. He said there are two things to consider separately: building products for AI, and building products with AI.

Starting with the former, he laid out the key rules for building products in the AI era.

First, the defining feature of building products in the AI era is that the technical substrate is changing at breakneck speed. In the 2010s product era, the underlying technology map evolved in an orderly way, with an occasional new framework. In the AI era, each time capability climbs a step, products that were previously completely infeasible suddenly "light up." So internally you need to keep running experiments — "even if something can't be built now, come back and try again in a few months."

He gave a firsthand example:

We actually tried something like Claude Code back in 2022. It was pretty frustrating at the time — the concept was right, but the model was too dumb to extract any real value from. I've been training these models since 2015. They were really dumb.

Second, in the AI era, a product's saturation point is pushed forward by the model becoming too capable. Dario said the chatbot form factor is already close to saturation — the market is still large, but as models keep getting smarter, the marginal gains to the chatbot form factor are no longer significant. Today, each new rung of capability shows up more in agentic forms like Claude Code.

Third, the API market will never disappear. Because new products keep emerging — inside Anthropic, and even more so outside it. Beyond code, developers are building healthcare, legal, and financial applications, and every additional rung of model capability opens up a new batch of application spaces.

Fourth (and this circles back to Amdahl's law), when building products with AI, he's observed a phenomenon inside the company: release velocity gets accelerated 2x, 4x, 5x — but then "systemic debt" starts to surface.

Using AI to accelerate shipping really does let you achieve output that was impossible a year ago; but you also accumulate technical debt at an astonishing rate. Then you're forced to ask: can we also use AI to pay down that debt, or at least help us keep track of what it is? And then you find that the team is forced to collaborate in a completely different way. New realizations like this keep surfacing every month.

That's also why the AI era isn't just about faster release cadence — "even 'how you do things' itself is forced to upgrade at high frequency."

Ami added her own sense of this: the underlying problems themselves won't change that fast — people are still people. But you have to keep "looking at the technology with fresh eyes," and accept that "the content of your daily work is also shifting, because the bottleneck keeps jumping to a new place every so often."

The capability that excites Dario most over the next six months

Ami asked Dario to answer in one sentence: over the next six months, what capability advance in models excites you most?

Dario gave an answer that crossed dimensions: the leap from "individual-level AI" to "organization-level AI."

What excites me is this idea: AI isn't 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 tied this thread back to the bet about "a one-person billion-dollar company": that bet might actually be underestimated. What's more likely to really happen is "a group of people plus AI completing work that used to take hundreds or thousands of people," rather than "one person single-handedly building a billion-dollar startup."

The Claude use cases that moved them most

At the end, Ami turned the topic to Daniela: which user use cases have moved you the most?

Daniela gave a few examples with striking contrasts.

The first was a mobile-doctor project in the Global South. In some regions it's very hard to see an actual doctor — you might have to travel dozens of miles of dirt road to reach the nearest city. But local people still have illnesses and health problems. A developer used Claude to build a "consultation-style" interface that provides vetted medical advice, translating the model's capability into a tool that's actually usable in low-resource settings.

She also mentioned acceleration in biomedical research, an area she's been following closely.

The last two examples were more personal. One developer used Claude to recover wedding photos from a damaged hard drive. Someone else used Claude to track how the tomatoes in their garden were growing.

Daniela was delighted by the tomato example: "I never would have imagined this use case in my life. But — do you have a camera livestream? I'd like to subscribe."

On the question of what AI can be used for, users' imagination always outruns whatever the product managers planned.

Quick Q&A roundup at the end

Q: How fast is Anthropic growing today?

Annualized at Q1's quarterly pace, that's an 80x rate (Dario used the qualifier "if you were to annualize it," which is a number extrapolated from a short burst). Compute had originally been provisioned for 10x growth, so they'd been rate-limiting all along.

Q: What did the SpaceX compute deal solve?

Within the next month, 300 MW and over 220,000 NVIDIA GPUs will come online. Anthropic will convert that compute into higher limits for developers as quickly as possible.

Q: Where does the "$1 billion one-person company" bet stand now?

There are already cases of two-person $1 billion companies and single-person companies worth several hundred million dollars (Dario didn't give specific names, so this can't be independently verified). At Code with Claude 2025, Dario gave a window of 2026, with 70%-80% confidence. There are seven or eight months left before that window closes.

Q: Over the next six months, what capability development excites Dario most?

Organization-level AI. It's no longer just about AI doing the work of many people for one person — it's about repeating that process many times within an organization made up of people.

Q: How does Anthropic weigh trade-offs in releasing capability?

Internally it's called "holding light and shade together." The Mythos model wasn't publicly released due to safety risks; instead, under Project Glasswing, it was distributed in limited quantities to dozens of institutions for defensive hardening work.

Finally

The core takeaway from this conversation is the sense of internal tension as Anthropic tries to hold two extreme positions at once.

On one hand, it's the fastest-growing AI company. An 80x annualized growth rate (even if that number involves some selective calculation), the SpaceX compute partnership, and Claude Code producing an inflection point in internal PR volume. Dario admitted on stage that 80x isn't sustainable and that he wants to get back to 10x — yet that same day, they closed one of the hardest deals to land in the entire industry. That's the strongest evidence of "we've gone after every bit of compute we could find."

On the other hand, it's also the most cautious AI company. A cross-generational model like Mythos was restricted from release purely due to safety risk, and "hold light and shade together" has become the recurring line that keeps coming up as a kind of safeguard. Faced with such a powerful model, Anthropic effectively chose to give up the speed of pushing it straight to market.

Balancing both of these is undoubtedly far harder in reality than what Dario and Daniela described on stage. 80x growth means terrifying delivery pressure — technical debt is "accumulating at an alarming rate," in Dario's own words. Under that kind of intense forward momentum, still hitting the brakes for safety evaluations and insisting on responsible release takes more than a few stated principles — it takes real, day-to-day battles over resource scheduling.

Dario's repeated invocation of Amdahl's Law is the key analytical framework running through the whole conversation. It points to a more practical question than "AI makes everything faster": once you accelerate, where does the bottleneck move to? For developers, that question is worth taking more seriously than "the model got even stronger again."

Two signals worth continuing to track: after Colossus 1 comes online, will limits actually loosen noticeably — doubling the 5-hour limit while leaving the weekly limit unchanged looks more like wordplay — and by year's end, how much of the gigawatt-scale commitments from Amazon, Google, and Microsoft will actually convert into compute users can use; and when, and under what conditions, will Mythos move out of preview from Glasswing. The former tests Anthropic's infrastructure capability as a product company; the latter tests how long the principle of "holding light and shade together" can hold up under commercial pressure.

As for the "$1 billion one-person company" bet, there are seven or eight months left before 2026 ends. Dario is already revising it on stage: the real proposition may actually be "a small group of people plus AI doing what used to take hundreds of people." If that revision is correct, the "solo unicorn" may end up being the relatively uninteresting part of this story.

Source video: Anthropic Code with Claude, San Francisco, May 6, 2026, "A conversation with Dario Amodei & Daniela Amodei."