More from this creator
Other episodes by Kitty Cat.
More like this
If you liked this, try these.
Transcript
The full episode, in writing.
Elon Musk leased the Colossus 1 data center, a move described as shocking to both Daario and Anthropic. Five days prior to the deal, Chamath Palihapitiya had predicted that Elon and Dario should strike a deal quickly. After the leasing of Colossus 1, Anthropic was able to access over 220,000 Nvidia GPUs and over 300 megawatts of energy. This deal had an immediate effect: Claude users experienced higher rate limits, the removal of peak usage caps for paid users, and increased API volumes for Opus models. XAI was now training its models at a separate facility, Colossus 2.
Elon Musk’s buildout of data center infrastructure was characterized as rapid and forward-looking. He was said to have “built to a level of scale and secured power before most people.” The result was that he could now lease compute capacity to companies like Anthropic, which had previously been severely constrained by supply and power limitations. This move put Musk in direct competition with major hyperscalers like Google Cloud, Amazon Web Services, and Azure. The new business was called “Elon Web Services” or EWS.
Brad Gersonner estimated that the new EWS line would generate an incremental $4–$5 billion in revenue for the year, adding to analyst estimates that SpaceX would see annual revenues in the mid-$20 billion range. Three facilities were named: Colossus, Macro Hard, and Macro Harder. Of these, Macro Hard and Macro Harder had a combined capacity of 1.2 gigawatts. Anthropic was given access to the facility with H100 GPUs for inference, while other capacity was retained or allocated for other uses.
Shawn Maguire described SpaceX’s business as a five-layer cake: launch, connectivity, compute, hyperscaler, space data centers, and then applications and models, followed by other bets.
Anthropic faced extreme supply constraints before the deal, which limited not revenue demand but actual revenue, because growth was held back solely by insufficient compute and power. After the deal, projections indicated Anthropic’s growth could continue at exponential rates. From January 1 to March 31 in a recent year, Anthropic’s annual run rate (ARR) grew from roughly $10 billion to $30 billion. In April, the rate of increase accelerated from $30 billion to $44 billion in ARR. Commentators said nobody in Silicon Valley had ever seen that level of growth at that scale. If Anthropic sustained this rate, they would exit the year with approximately $100 billion in ARR. The only thing holding them back was compute capacity, and the deal with Musk solved much of that constraint. There was open speculation about whether Anthropic might reach a trillion-dollar ARR in 2027.
For scale, the largest tech companies in the world at that time—Apple, Nvidia, Google—were each generating $400–$500 billion in annual revenue. The three largest hyperscalers—Google Cloud, Amazon Web Services, and Azure—were collectively doing about $300 billion in revenue, with a combined market capitalization of $4–$5 trillion if considered independent companies.
Anthropic’s market was defined as “coding and the things that will be built on coding tokens like agents.” Coding alone was seen as a trillion-dollar annual market. The argument was that AI coding models could multiply the value of this market by 10x or 100x due to increased productivity and code generation.
It was noted that, at the start of the year, companies like Anthropic were focused, while other frontier labs were pursuing many different lines, including generative image models, fantasy character chatbots, and video generation. Only after observing Anthropic’s revenue trajectory did these competitors pivot back to a focus on coding. For example, OpenAI’s “Codeex” product, based on GPT 5.5 and a new base model named “Spud,” was said to be showing very strong growth and performance improvements. Google, also a major player, was considered to have a strong coding team.
The SpaceX business, prior to the data center expansion, was already generating an estimated $20 billion annually from Starlink and launch services. The new EWS business could add $5 billion in incremental revenue, and the potential for future expansion included integrating compute into Tesla products, Powerwalls, and distributed systems in homes, as well as moving compute into space-based infrastructure. The Powerwall was mentioned as already being online, with Elon Musk’s expertise in factories, battery deployment, and solar panels from the SolarCity acquisition positioning him to construct extremely large data center capacity.
SPAN originally built smart power panels, allowing breakers to be managed through an app. The company pivoted to integrate compute clusters as part of their product offering.
Valuation multiples for SpaceX and Tesla were discussed, with the argument that Musk’s companies trade at higher revenue multiples—sometimes 2x, 3x, or 4x that of other leading firms—due to expectations of future innovation and pipeline. In contrast, Apple was described as being penalized in valuation because, under Tim Cook’s leadership, innovation was perceived as having slowed, with a focus on incremental product improvements and services like Apple TV rather than new hardware.
There was a discussion about competitive market structure in AI. At that moment, only two companies—Anthropic and OpenAI—were reported as making substantial revenue in AI. OpenAI’s growth was cited as 3–4x, but Anthropic’s was exponential, at 10x per year. If this trend continued for another 18 months, Anthropic was projected to become the most valuable company in history and potentially hold a monopoly over the most important technology of the time.
Historical context was offered by referencing John D. Rockefeller and Standard Oil, suggesting that if Rockefeller had been more skilled at public relations and had called his product “safe oil,” he might have succeeded in framing regulation as beneficial while building a monopoly. The analogy was drawn to modern AI companies advocating for safety regulations, which, if enacted, could serve as a form of regulatory capture and strengthen market moats.
Some argued that the current situation was not yet a monopoly, with the AI market still in its early stages and large incumbents like Google, Amazon, and Microsoft investing heavily. However, there was concern that regulatory proposals—such as banning competitors from using a company’s models or establishing an “FDA for AI”—could tilt the field in favor of the largest firms and stifle competition.
Recent news reports from the New York Times claimed the White House was considering an “FDA for AI,” which would vet new models for safety. This possibility was connected to the release of powerful models like Anthropic’s Mythos, which reportedly alarmed officials. The stated goal was to avoid political fallout in the event of a catastrophic AI-enabled cyber attack. Kevin Hassid, director of the National Economic Council, confirmed that an executive order was being studied to create an “AI working group” that could examine oversight procedures, including a review process for new AI models.
However, direct conversations with officials suggested that no senior leaders actually supported the FDA analogy. Instead, there was consensus that pre-approval for each model would be a “disaster” and would stifle innovation. The preferred approach was quicker, coordinated reviews between government and industry, with specific guardrails applied only where necessary. The push for regulation was seen as a reaction to public anxiety, ongoing negative media narratives about AI risks, and a lack of effective messaging from tech leaders about the positive impacts of AI.
Another policy idea discussed was “KYC” (know your customer) requirements for access to powerful AI models. The argument was that, especially during preview or limited release periods, companies should verify and log the identities of users to prevent state-sponsored actors or bad actors from using the tools maliciously. Labs like Anthropic and OpenAI were already monitoring API usage, flagging suspicious activity, and coordinating with government agencies to preempt cyber vulnerabilities.
There was concern that some commentators and politicians were using the AI “cyber” issue to advocate for permanent new regulatory infrastructure in Washington. Some framed this as opportunistic, leveraging a real but time-limited issue to justify broader and longer-term restrictions.
There was also discussion of the economic impact of AI. Despite public anxiety about job loss and wage pressure, polling indicated that AI was not a top issue for most voters, ranking 29th out of 39 in salience. The most immediate concerns were cost of living and the economy. AI was credited as a driver of 75% of GDP growth in the first quarter of the year, and it was contributing to a construction boom, blue collar wage increases of 25–30%, and a strong consumer sector.
Analysis of the major hyperscalers showed AWS at a $150 billion annual run rate, Azure at $108 billion, and Google Cloud at $80 billion. Their growth rates were 28% for AWS, 39% for Azure, and 63% for Google Cloud. Headcount growth for the largest five tech companies over three years was about 3%, while operating margins expanded. S&P 500 operating margins rose from 11% in 2023 to 13% in the following year, a 200 basis point improvement. This was attributed at least in part to technology-driven productivity.
Some questioned whether the observed margin expansion was due to AI or to other factors such as cost-cutting and layoffs. There was a debate about whether AI-driven productivity gains had yet translated into measurable benefits for the average company. For example, a company like Nike was now using AI-generated photos for advertising, reducing the need for photographers and lowering costs. DoorDash, similarly, used AI-generated images instead of traditional food photography. These examples suggested significant cost reductions but did not necessarily translate into higher margins or sales yet across the economy.
Data also showed that unemployment remained at historic lows, with the labor participation rate at 61.9% and unemployment for young college graduates dropping. This was cited as evidence that, contrary to fears, AI had not yet caused mass job loss, and recent graduates might even be benefiting due to their fluency with the new technology.
Amid all these trends, policy suggestions included giving every American an “investment account” that could compound with the upside of AI, funded by a portion of IPOs from tech companies. Other proposals included targeted minimum wage increases and universal healthcare, with the argument that empowering lower-income consumers would stabilize the economy and benefit tech companies by increasing spending.
The AI boom was described as deflationary, lowering the cost of living, and powering economic growth. The S&P 500 was up 8% for the year, with leading tech companies trading at 17–24 times earnings. Memory stocks like SK Hynix, Samsung, and Micron were trading at 5–7 times earnings, suggesting value investors still saw upside.
Analysts argued that U.S. policies, especially around energy and technology, had positioned the country to outpace the world. Policies favoring “drill baby drill” and letting AI companies generate their own power for data centers were credited for the ongoing blue collar construction and infrastructure boom. There was reference to a previous administration’s chip policies, which had required pre-approval for models above a certain threshold but had since been rescinded.
Podcast Creation Process: The production of the podcast involved a multi-step workflow. Each episode began with a research phase, where producers gathered data from sources such as company filings, government reports, and interviews with industry figures like Brad Gersonner, Shawn Maguire, and Chamath Palihapitiya. The scriptwriting phase incorporated direct quotes and data points, ensuring factual accuracy. The recording sessions took place at a studio in San Francisco, with editing handled by a team led by senior editor Alex Kim. The team used audio editing software to remove errors and improve clarity, and fact-checkers reviewed each episode before release. The average production timeline per episode was three weeks from research to publication.
Podcast Distribution/Platforms: Once finalized, episodes were distributed across major podcast platforms. The podcast was available on Spotify, Apple Podcasts, and Google Podcasts, as well as on smaller platforms such as Stitcher and Overcast. The distribution process was managed by a digital marketing team, who also coordinated social media promotion and newsletter announcements. The podcast’s RSS feed was updated automatically with each new episode, and listener analytics were tracked using tools integrated with Spotify and Apple’s platforms. In the first six months, the podcast achieved over 500,000 downloads and ranked in the top 50 for technology podcasts in the United States.