AI Value Is Shifting From Models to Operating-Layer Control

John Gruber framed Apple’s current AI moment as a test of the company’s oldest product instinct: make hard design decisions, then ship the version Apple believes is right, even when the market initially wants something else. The iPhone’s lack of a hardware keyboard remains his clearest example. At launch, people who already used smartphones objected that they could not type all day by “poking at a touch screen.” Apple did not hedge with a second keyboard model. It decided software keyboards were the better use of limited phone real estate, and the market later moved there.

That history matters because Apple’s AI position is now much less clean. Gruber said the coming WWDC is unusually consequential because Apple announced its first serious “Apple Intelligence” push two years earlier, failed to ship the most ambitious parts, and then effectively received a pass last year after postponing them. Now, in his view, “it’s time to show their cards.”

The clearest failure was not merely delay. Gruber called Apple’s advertising problem obvious even to a non-lawyer: the company showed a commercial for a feature that still did not work a year later. “It was false,” he said. That is a different failure mode from Apple’s familiar pattern of waiting until a technology is ready. Tim Cook has often said Apple does not aspire to be first, but to be best. Gruber’s criticism was that Apple behaved as if it had to have something big ready for June 2024, “ready or not,” and it was not.

Gruber described what he expects from Apple’s unusual Google arrangement: Apple would use Gemini technology as an underlying layer for Apple’s own foundation models, without presenting the experience to users as Gemini. If a user does not have or use a Google account and does not want Gemini, Gruber said, Apple’s likely goal is that the user never encounters the word. The model supplier can be Google while the user-facing product remains Apple.

The strategic question is whether Apple can build a good AI system when it does not control the underlying model. John Coogan pushed back that vertical AI companies already build strong products on top of models they did not train, naming Cursor and Harvey as examples. Gruber accepted Cursor as “probably the best example,” but Coogan argued that the harder question is what Apple will permit third-party agents to do inside iOS.

That is the core tension: if ChatGPT, Claude, or Gemini can sit above Siri, can they operate native apps? Can they read Apple Notes, Apple Mail, Safari context, iMessage, calendars, and third-party app data with permission? Or will they remain answer engines with limited surface access?

Gruber argued that Apple needs the deeper version. If third-party AI extensions cannot access the user’s Apple data, then users who want AI to manage their digital lives may move the data elsewhere — into Google Calendar, Google Keep, ChatGPT-controlled workflows, or whatever stack gives the model the best context. Apple’s interest, in this reading, is not only the 30% or 15% subscription revenue it could capture if users subscribe to AI services through iOS. It is keeping Apple Notes, Apple Mail, Safari, and the rest of the Apple ecosystem relevant as the substrate for AI.

Gruber compared Apple’s AI problem to Disney’s brand problem. Disney can own ESPN and participate in sports betting, but a Disney cruise ship still has no casino because gambling is off-brand for the core family experience. Apple faces a similar line-drawing problem. Its own branded AI will likely be conservative: less willing than frontier models to answer certain questions or assist with certain tasks. Third-party extensions can let Apple offer more capability while saying, in effect, that the response comes from ChatGPT, Gemini, or Claude, not Apple Siri.

That structure only works if the extensions are genuinely useful. Apple can “wash their hands” of some content and behavior, Gruber said, but not of the integration problem. On the Mac, hacker-style tools can scrape the screen, drive iMessage, and work around Apple’s restrictions. Coogan raised OpenClaw-style workflows as a possible analogy: software running on a Mac mini, opening iMessage, taking screenshots, and extracting context against Apple’s wishes. Gruber said that kind of workaround is possible only because the Mac is Apple’s open platform. Most Apple users, he estimated, do not have a Mac; they have an iPhone, or an iPhone and iPad. For AI to become normal computing for Apple’s user base, it has to work on iPhone, and “there’s no way to do that without Apple’s help.”

The same issue appears in mobile vibe coding. Jordi Hays asked about apps that let users generate software on iPhone, some of which reportedly were not removed from the App Store but were blocked from shipping updates. Hays laid out the conflict: Apple sees on-the-fly software generation as a threat to the App Store business and a way for users to create apps Apple would not normally allow; developers and users see innovation being stifled.

Gruber said Apple’s answer to this is one of the top things he wants from WWDC. If Apple says nothing about mobile vibe coding, he would treat that as a red flag. In his words, disruption theory is satisfying because the disrupted company does not get to choose whether disruption happens. Entrenched companies often begin by dismissing a new technology, then later say they choose not to be disrupted. “It never works.”

His objection was partly philosophical. iOS is nearly 20 years old, and an iPhone is a computer. An iPad Pro can cost around $2,000 and be as powerful as a MacBook, but users still cannot make apps for the device on the device itself. Gruber said that made sense in the early years of limited hardware and a new App Store model. At this point, “it’s kind of bizarre.”

The possible compromise is already visible in Apple’s own tooling. TestFlight lets developers distribute apps outside the App Store under limits. Gruber suggested a consumer-level version: perhaps AI-generated apps distributed to only 10 devices, for yourself, friends, or family, without full App Store review. The point is not that this exact number is the right policy. It is that Apple needs a story that permits personal software creation while preserving safety, trust, and a business model.

Apple’s broader software-quality problem, in Gruber’s account, comes from measuring the wrong things. Asked whether Apple software is getting better or worse, he said Apple has done well on measurable problems such as crashes. If Notes or Safari crashes, Apple can count that, collect reports, and drive the number down. But confusing FaceTime flows, bad iMessage search, and unclear collaboration interfaces do not show up on a dashboard in the same way.

Hays described having to text Coogan to tell him they were in a FaceTime because the group FaceTime did not reliably ring him. Gruber said that if Steve Jobs had personally experienced that problem, fixing it would have become urgent the next day. Jobs was not only CEO but “the number one user of the products.” Tim Cook, by contrast, has shaped a company more oriented around numbers. Gruber was careful to say many good things came from Cook’s leadership, but the tradeoff is visible: crash rates improve, while “opinionated” judgments about confusing interfaces lose force.

On iMessage search, Gruber said it technically works in the way that two soup cans connected by taut string technically transmit audio. That is not what users mean by search. Users want to find the message where a word appeared a month ago. Apple’s implementation satisfies a technical definition while failing the ordinary user expectation.

His hope was not for a wholesale reversal, but for a course correction under John Ternus: less dashboard-only quality control, more willingness to say “this app is confusing” even when the problem cannot be reduced to a metric.

Corporate AI spending has become a management problem before it has become a settled ROI story. Hays and Coogan treated rising token bills as a real operational issue; Brad Gerstner treated the same evidence as a predictable phase inside a much larger adoption curve.

The first problem is that token usage can become a status game. Hays described “token maxing” dashboards that reportedly encouraged people to leave agents running overnight or generate activity that was not clearly productive because they wanted to rank up on a leaderboard. The issue is not limited to literal leaderboards. A token budget can itself become a target. If a manager gives an employee a $10,000 or $100,000 AI budget, the employee may infer that failing to spend it looks like failing to deploy assigned capital.

Hays summarized the governing tension with two familiar laws. Jevons paradox says that when something becomes more efficient, people often use more of it, not less. Goodhart’s law says that when a measure becomes a target, it ceases to be a good measure. In enterprise AI, cheaper and more capable models can drive more usage, while visible usage metrics can distort behavior.

The mundane examples make the risk legible. Coogan admitted to sometimes using an LLM to get a weather report, though he said that would usually be done through a free chat app, not an expensive coding model. Hays described someone using an agent to scan an iOS contacts list for recently added people after a conference — a task that might be solved by a simple $2 app, but can also become an agentic workflow that emails a synthesis. In a personal context, that tradeoff may be harmless. In an enterprise context, “everything has five extra zeroes behind it, if not nine.”

The deeper question is whether companies are pointing AI at work that matters. Hays worried that AI can pull items from the bottom of a backlog precisely because they are now cheap to attempt, even if those items were previously deprioritized for good reasons. “Maybe the cutting room floor should remain the floor,” he said. If token spend goes into low-leverage work, management will eventually have to explain why revenue, productivity, or cost structure improved.

Still, Hays argued that lumpy short-term ROI can be justified if it makes a workforce AI-native. A board may accept overspending if management can credibly say employees have learned where AI is useful and where it is wasteful — from critical security bug finding to checking the weather — and can now use the tools more judiciously.

Coogan saw a healthy correction dynamic. Companies may have gotten ahead of themselves for a few months, but then stack-ranked usage, cut unnecessary token spend, and kept the valuable workflows. He contrasted high-token-cost vibe-coded projects with little value against builders who use a $200 subscription “like a scalpel not a hammer” and produce something useful.

Gerstner’s answer was that both sides of the AI-spend argument are overstating their case. Bears first said AI revenue would not show up; when it did, they argued it was “all token maxing” with no ROI. AI bulls, by contrast, sometimes act as if every token dollar is already perfectly allocated. Gerstner said the truth is in the middle: millions of actors are making individually rational decisions, experimenting, wasting some money, and finding real returns elsewhere.

He pointed to Altimeter’s own survey of 300 enterprises. The chart shown on screen, attributed to “Altimeter May 2024 Survey,” was titled “Rapid growth continues during optimization” and compared trailing 12-month spend growth with expected raw API growth over the next 12 months across companies already optimizing, planning to optimize, evaluating optimization, or not prioritizing it.

Gerstner interpreted the survey as showing that companies actively optimizing still expected API/token usage to grow by more than 50% over the next 12 months, while companies planning to optimize expected growth around 90%. The point was not that optimization is fake. It was that optimization can coexist with very high growth because penetration is still low. Enterprises are early in coding adoption, barely started in broader knowledge-work AI, and many companies globally are not yet serious AI users.

Gerstner placed Anthropic at the center of the current public-market AI narrative. He said the year began with investors skeptical that AI revenue would show up, and argued that if Anthropic had not delivered the revenue it did, the stock market might be down 10% or 15%. In his account, OpenAI and Google had good numbers, but Anthropic supplied the outperformance that “buoyed the entire AI segment.” He called Anthropic “the fastest growing company in the history of capitalism,” and said its reported high gross margins and possible positive free cash flow in Q2 helped the market recover from being down on the year two months earlier.

He also pointed to hardware and infrastructure results as evidence that the demand is real. Dell, he said, had AI server revenue up 750% year over year and had grown from a $1 billion AI-server business to a $16 billion business. Those were Gerstner’s figures, offered in the context of his AI-infrastructure bull case. He warned that normal pullbacks are still likely: semiconductor stocks could see 10% to 20% consolidations simply after enormous moves.

The investment question is which companies are in “the token flow.” Gerstner argued that software should not be treated as one category. Databricks, Snowflake, and ClickHouse, all Altimeter investments, benefit as token consumption increases because database queries rise with AI use. At Altimeter, he said, database queries are growing faster than token usage. Those companies enable AI workloads and can earn AI multiples.

He contrasted that with front-end application companies that compete more directly with models. Salesforce, he said, may yet get into the token flow, and he praised Marc Benioff, but he viewed Salesforce’s customer-facing solutions as more exposed to model competition than Snowflake’s infrastructure role.

Gerstner’s “SaaSpocalypse” view was more severe than simply saying software multiples had compressed. A chart he prepared was titled “Saasapocalypse? Or Just Back to Market Multiples?” and argued that large-cap application companies now trade roughly in line with the S&P 500, around 22x, while AI-leveraged infrastructure still carries a premium. Gerstner said the correction took software from above-market multiples down to market multiples. That may not be the floor.

If software companies get on the AI train, they can regain above-market multiples. If improving computational intelligence makes a company’s business worse, he said, those companies can trade below market multiples. For Altimeter, most software is therefore in the “too hard” basket.

That also changes early and growth investing. Gerstner said Altimeter is investing in companies in or adjacent to the token flow: compute shortages, semiconductors, data centers, military modernization, and AI infrastructure. In growth-stage software, he said the old Series A-to-Series B revenue pattern no longer guarantees demand. A company going from a few million in revenue at Series A to $20 million at Series B would once have had a line out the door. Today, he said, “you wouldn’t have a single taker” if investors believed the business could be steamrolled by models.

Brad Gerstner rejected the simplest comparison between AI infrastructure and the dot-com fiber buildout. In 1999 and 2000, he said, there were around 35 million people connected to broadband internet. Investors could see what Amazon might become, but adoption could not arrive as quickly as hoped because the distribution layer was small. Today, he said, there are three to four billion people connected, making diffusion radically different.

The limiting factor in AI is not consumer access but physical production: memory wafers, logic wafers, powered shells, and therefore tokens. In the fiber era, companies installed “dark fiber” knowing nobody was yet using it. Gerstner said there is no equivalent today: “There is not a dark GPU in the world” and “there is not a dark token in the world.” Google, Amazon, Microsoft, OpenAI, and Anthropic all reported or indicated token constraints, in his telling. If they had more tokens, he said, they could generate more revenue.

He expects the next nine months to surprise people because compute ramps and algorithmic improvements are happening together. He said OpenAI and Anthropic began the year with a combined three gigawatts of compute, may end the year closer to 10, and end next year closer to 20. He also pointed to competition from Cursor and xAI’s “Macro hard” and “macro harder” compute efforts as part of a broader American frontier-model race.

That optimism sits alongside one political fear: a data-center moratorium. Coogan asked whether a moratorium could be negative for chip suppliers but positive for companies that already have tokens to sell, because scarce capacity would increase pricing power. Gerstner rejected the tradeoff: a moratorium would be “bad for everybody” and “horrific for America.”

His argument was strategic and historical. Activists, he said, had previously shut down supersonic technology and nuclear clean-energy development in the United States. He said China is building many fission reactors while the U.S. has one, and treated that as a warning. A data-center moratorium would, he argued, threaten GDP growth, push the country toward recession and unemployment, and cede the global AI race to China. He framed AI as economic security, job creation, and national security at once.

Yet he did not dismiss local concerns. Returning from his mother’s 90th birthday in rural Indiana, he described communities worried about jobs, their children’s futures, water use, and electricity prices. If activists tell a town that a data center will drain water and raise power bills, he said, residents’ agitation is understandable. His proposed answer is a “sociopolitical bridge” for the next three years: tangible benefits for communities hosting data centers.

Gerstner said he is working with cloud companies, chip companies, offtakers, and the White House on an initiative he was not ready to announce. The goal is to deliver a meaningful dividend to communities where data centers are built. He believes that in three years the benefits of AI will be more obvious: personal assistants in every pocket, enterprise tools that up-level workers, and broader abundance. But he said the next three years require concrete local benefits, not abstract promises.

On Meta, Gerstner said its reported enterprise push makes sense once a company is spending $100 billion a year on capex. He compared it to AWS: Amazon built capacity for Black Friday and Christmas peaks, then rented idle capacity the rest of the year. That became a major business and improved the core retail operation. He described Elon Musk’s analogous effort as “EWS,” or Elon Web Services, turning compute into a sellable service, including through large customers such as Anthropic.

Meta, in Gerstner’s view, faces the same need to monetize infrastructure. If Mark Zuckerberg wants to keep building frontier-level AI capacity, CFO Susan Li and the company have to find ways to monetize it. Moving from a “120% consumer” company into AWS-like infrastructure or enterprise agents is hard, he said, but not impossible. Product-led growth in coding agents has consumer-like adoption dynamics, which may give Meta more relevant DNA than traditional enterprise observers assume. Coogan added that Meta already has relationships with hundreds of thousands of businesses through its ads platform.

Gerstner was more skeptical of services firms building their own AI software stacks. Coogan asked about Kirkland & Ellis reportedly considering a half-billion-dollar investment into internal software. Gerstner’s response was blunt: they have to announce something because competition is coming, but he would not view it as a high-probability bet if he were a partner there. He saw a more likely path in models such as Thrive Holdings buying accounting firms, recruiting strong engineers, partnering deeply with OpenAI, and driving productivity gains inside acquired services businesses. A law firm whose core competence is legal work suddenly writing “killer legal software” to compete with OpenAI and Anthropic struck him as unlikely.

Brad Gerstner described the children’s investment-account project as a four-year effort that passed into law through the Invest in America Act as part of what he called the “big beautiful bill.” He said the app launched the day before the interview and had already reached number three in the U.S. App Store, behind ChatGPT and Claude and ahead of Google. He credited Vlad Tenev and Robinhood, BNY, Joe Gebbia and the National Design Studio, Treasury officials, and others for building the system.

As Gerstner described it, the program gives American children investment accounts tied to the S&P 500. Children born after January 1, 2025 receive $1,000. Children roughly between ages two and 10 receive at least $250. He said there are 35 million children in America under 10, and most will receive $250 from Michael and Susan Dell. Indiana children receive an extra $250 from Gerstner; Connecticut children receive an extra $250 from Ray Dalio; Oklahoma children receive $250 from the state. Beginning in 2027, he said, accounts will be automatic when a child receives a Social Security number.

Gerstner’s larger claim is that the accounts create “universal private ownership.” He repeatedly distinguished them from conventional philanthropy and from 529 accounts. In his view, 529s primarily serve families already able to save. These accounts start every child with ownership, title, and exposure to compounding. “A hundred cents on the dollar goes to the kid,” he said, arguing that the model avoids charitable overhead and abstraction.

He described adopting a school in Durham with 700 students, giving $250 to each child. The school made a spreadsheet, got students signed up, and the principal could QR-code funds into accounts. He said any donor can adopt a school, county, city, or state. The important feature is scalability: a wealthy donor can cover millions of children, while a local group can raise money for one school.

The emotional case is the move from zero to one. Gerstner said he grew up in rural Indiana with “zero,” and that zero is despondent because people do not know how to get to one. Moving from one to two is easier. The account is meant to put every child on the first rung of ownership. He argued that children with assets are more likely to graduate from high school and college, start businesses, and buy homes. He also said if a child starts with $1,000 and saves $50 a month, the account could reach $50,000 by age 18.

Hays extended the thought: if a generation reaches adulthood with $50,000, $100,000, or $200,000, that becomes down-payment capital, which can underwrite more housing demand and potentially support more building. Gerstner agreed and said that over 15 years the program could transfer $3 trillion to $4 trillion of wealth from people who have it to people who otherwise would have zero. He also said President Trump views the program as potentially his biggest legacy and argued it could become more impactful than Social Security because the account is privately owned.

The program is also Gerstner’s answer to the question of how immense Silicon Valley wealth should be redistributed without demonizing success. Michael and Susan Dell’s commitment, which Gerstner described as $6.25 billion — $250 for 25 million kids — was, in his telling, the largest philanthropic gift in history and only a beginning. He called the model “giving pledge 2.0”: instead of promising broad future charity, donors can transfer money directly into accounts for children now.

Coogan said he had originally put the project in the “too hard” bucket, even while respecting Gerstner, because getting such a policy passed and implemented seemed nearly impossible. Gerstner’s reply was simple: “We fucking did it.”

The startup interviews were brief, but they formed a coherent counterpoint to the broader AI-market discussion. Each company was selling into a domain where generic model intelligence is not sufficient: continual learning for AI applications, defense integration, and insurance operations.

Ronak Malde said Trajectory is building “the platform for continual learning,” working with companies such as Harvey, Decagon, Clay, Rogo, and Mercor. The goal is to make agents learn online from the way users actually interact with them. Frontier labs are trying to build a “smart PhD student,” he said, but that student starts every job on day one and can make the same mistake repeatedly. Trajectory wants the opposite: an agent with “10 years of experience on the job.” It does not need frontier math capability. It needs to know a company’s primitives, business context, and rewards.

Malde said that can be achieved with light post-training. He claimed Trajectory has models 10 times smaller than frontier models that can beat them in specific domains, but the more important point is daily improvement: “not only are we beating them today, but we’re improving 1% every day.” Application companies, in his view, believe product craft, domain expertise, and taste will matter for years; Trajectory gives them a research lab “in their back pocket” so they can modify models and harnesses rather than only optimize prompts.

Zane Mountcastle said Picogrid builds technology to integrate mission-critical systems — sensors, drones, robots, and weapon systems — primarily for military applications. Mountcastle described the product as open infrastructure, “connective tissue” or “glue” between platforms. The company has active contracts with the Pentagon, NATO, and allied partners.

Mountcastle said the company’s biggest challenge has been keeping up with demand without stretching too thin. In the prior three or four months, he said, Picogrid had 7x’d its production line. The new funding is aimed at building the team and the infrastructure needed to serve that demand. The company is focused on military customers because the integration bottleneck is most acute there. Autonomous systems are proliferating, but drones, robots, sensors, and weapons must work together. He expects a commercial presence eventually, but dual-use sales motions differ so much that Picogrid is concentrating on defense for now.

Jamie Cuffe said Pace positions itself as an AI operations partner for large insurers and brokers, including Prudential, WTW, and Convex, automating back-office operations so insurers can cover more of the world’s risk.

Cuffe emphasized that insurance is a regulated market where 90% accuracy is not good enough. Many customers require 99.9% or higher accuracy SLAs, and he said Pace hits those “day in day out.” He also said Pace has had a 100% win rate from pilot to production, crediting both the product and forward-deployed engineers working closely with large customers.

Pace uses “agent operating procedures,” natural-language instructions for long-running tasks. As models improve, the product improves. Cuffe singled out computer use as a recent unlock. Many insurance systems depend on desktop applications, legacy web apps, or even green-screen CLIs where APIs are unavailable or expensive to build. On Pace’s evals, he said, computer-use models improved from around 30% accuracy to more than 95%, allowing agents to operate end to end in environments that previously blocked automation.

He tied the product to the insurance “protection gap.” Last year, he said, 60% of the world’s losses were uninsured, representing $9 trillion of risk that should be insured but is not. For established insurers, the opportunity is not merely back-office cost savings. AI-native operations can change the economics enough to offer a 10-person company the kind of service experience once reserved for a 10,000-person company. Pace is seeing, he said, two orders of magnitude less spend than would have been required before agents.

Together, the three fundraises suggested a version of AI adoption that is less about general chatbots and more about embedding learning into specific domains. Trajectory focuses on product feedback loops; Picogrid on physical-system interoperability; Pace on regulated, high-accuracy workflows in legacy environments. In each case, the model is only one layer. The domain harness, operational context, and deployment surface are the product.

Kyle Kuzma described investing as something that began when he entered the NBA, hired a financial advisor, and learned the standard world of stocks and bonds. Kobe Bryant was one of his key mentors and introduced him to venture and technology investing through a fund shortly before Bryant died.

The first lesson from Bryant was to “keep the main thing the main thing.” For Kuzma, the main thing is basketball. No investment, unless perhaps an early SpaceX-like outcome, is likely to produce the return that basketball can produce for an NBA player. But the career is finite. “We can only play basketball for a certain amount of time,” he said, while life after that may last 40 or 50 years.

Kuzma said traditional financial advisors tend to keep athletes conservative, targeting modest annual returns. He wants more, which draws him toward private technology investments with larger possible multiples. Endorsements, equity, and direct investments all have different incentives. Agents and managers often prefer endorsement deals because fees are clear. Equity is harder to justify because “90% of things fail.” Kuzma’s answer is education, curiosity, and surrounding himself with people smarter than he is.

He views athlete access as an underused asset. Everyone wants to meet athletes, be around them, and talk to them. That access can be used for fame and parties, or it can be used to build a platform and enter blue-chip companies. Kuzma said he has mostly invested as an individual so far, joining cap tables, SPVs, or late-stage growth rounds. But he sees the logic of eventually building a fund, as Bryant did, to capture upside from access.

The stage question is changing for him. He has invested in real estate, CPG, and late-stage growth tech where outcomes feel more certain but returns may be 3x, 4x, or 5x. Coogan joked that Kuzma wants the “thousand x,” and Kuzma did not disagree. To reach that, he knows he has to look earlier. But $25,000 to $75,000 checks into pre-seed or SAFE rounds can be awkward if his current focus is larger late-stage opportunities. He spends his own time studying sectors, often in the hours after basketball work, and relies on mentors and networks to help filter opportunities.

On AI in basketball, Kuzma was ambivalent. Players dislike analytics when it feels like a computer micromanaging human judgment. He described being told by a system that he should not take a shot even when he is wide open and knows, as a professional player, that he can make it. He was more positive about technology to speed officiating, such as an out-of-bounds or replay system like Hawk-Eye, because current replay stoppages can consume five minutes of real time.

Kuzma’s American Dynamism thesis was direct. He likes investing in companies aligned with American interests because he loves America and believes investors should consider serious industries such as space and defense. Space, in his view, is becoming a major economy, and control of space has implications for control on Earth. Wars on Earth are fought horizontally; space provides a vertical vantage point. “Whoever really rules space is probably going to rule the world,” he said. Hays called it a strong SpaceX bull case.

His advice to younger investors came from a friend: “Don’t be stupid, and follow the money.” For Kuzma, that means respecting the due diligence and pattern recognition of people who understand a sector better than he does, while still making his own educated judgment.

Blue Origin’s New Glenn rocket explosion during a static-fire test ahead of NG-4 was treated as more than spectacular footage. The hosts showed nighttime and vertical-phone video of a massive fireball, with on-screen text reading “NEW GLENN EXPLODES IN TEST” and “BLUE ORIGIN TEST ENDS IN FIREBALL.” Hays compared the blast to a Christopher Nolan scene or “a nuke,” while Coogan said it looked like a nuclear bomb had gone off.

The business point was the asymmetry of competition. Hays argued that if a company is competing with SpaceX and has “just” a launch business, a pad-damaging setback is especially painful. He connected the point to Gerstner’s earlier public comments as he understood them: launch is a great business, but not enough without Starlink; Starlink is great, but not enough without AI. SpaceX, in this telling, is not merely a launch company. It is a launch, satellite broadband, AI, and compute platform story.

The hosts emphasized that failure is normal in rocketry. Musk replied to Bezos with “Most unfortunate. Rockets are hard,” according to Hays, and Bezos wrote, “Very rough day, we’ll rebuild whatever needs rebuilding and get back to flying, it’s worth it.” Hays said even SpaceX maximalists should want multiple American heavy-lift providers.

Daylight photos shown on screen of Launch Complex 36 looked less totally destroyed than the explosion footage suggested. Coogan said the site looked “way better” than he expected. Hays said he had expected a crater. Even so, Coogan said every nut and bolt on the tower would need inspection for damage or corrosion. Rebuilding is possible, but the work is extensive.

The later SpaceX discussion turned on the opposite side of the same infrastructure problem: what investors are actually buying if SpaceX becomes a public company. Hays read a critique from Ben Thompson, shared by Tae Kim, arguing that a proposed $2 trillion SpaceX valuation was absurd relative to the financial figures Thompson cited: $18.67 billion in revenue, $4.9 billion in losses, and growth slowing from 35% to 33%. The critique also pointed to xAI and X tipping the company from profit into loss through $5.1 billion in AI R&D expense, toward a model described in the post as fifth place and with its founding team recently gone.

Hays did not reject the absurdity of the numbers, but he argued that the critique underrates the new AI computing contracts and the possibility of significant business reacceleration. Elon Musk has been talking about “macro hard” and entering enterprise AI for more than a year, Hays said, so the large AI enterprise software TAM in the S-1 should not be treated as entirely out of the blue.

Gerstner later sharpened that point by saying the SpaceX IPO’s tenor changed because of compute deals, including Cursor and Anthropic. He described Musk as unmatched at “turning electrons into tokens” and said people should expect more Elon data centers on Earth and eventually in space. For Gerstner, SpaceX is no longer only a launch and Starlink story; it is becoming part of the AI infrastructure race.

The pad failure and the valuation debate point in opposite directions but share the same lesson. Rockets fail violently, launchpads break, and rebuilding is slow. At the same time, companies that can combine launch, satellites, power, compute, and AI may become difficult to evaluate as ordinary launch providers. Ben Thompson’s fundamental skepticism, as read through the shared post, stood against Gerstner’s view that compute changes the entire SpaceX story.