Orply.

Anthropic’s $81.3 Billion IPO Opens a Window Into AI Spending

Lon HarrisJason CalacanisThis Week in StartupsSaturday, November 21, 202616 min read

Anthropic’s $81.3 billion IPO gives investors a way to track enterprise AI spending and the split between subscription and usage-based revenue, Jason Calacanis argues. He says intensive AI use among knowledge workers remains rare, leaving substantial room for growth in demand. Calacanis sees Anthropic as a company worth considering but advises investors to wait rather than rush to buy during the opening-day surge.

Anthropic’s IPO offers a rare view into the AI economy

Lon Harris reported that Anthropic priced its Nasdaq IPO at $83 a share, then opened at $111.50, about 34% higher. The stock briefly approached $115 before falling back toward its offering price. The company enlarged the offering from 825 million shares to about 980 million, raising $81.3 billion—the largest IPO on record. Anthropic reported $33.2 billion in revenue for the first nine months of 2026, up from about $2.5 billion in the same period a year earlier. A CNBC screen shown during the discussion highlighted the $83 offering price and enlarged share sale; a separate finance screen showed the stock back near $83 during trading.

$81.3 billion
raised in Anthropic’s IPO, according to the figures discussed

For Jason Calacanis, the IPO mattered not only as a stock debut but as a new source of information about the AI business. Anthropic’s public reporting can show how much enterprise customers spend and how that spending is divided between subscriptions and usage-based charges. Calacanis said a large share of the company’s revenue comes from consumption and overages rather than fixed subscriptions. He said the disclosures would let observers follow that mix over time.

He compared understanding the market to seeing more cards in a poker hand. Nvidia, Google and Meta have already disclosed pieces of the picture; Anthropic adds another. When OpenAI’s figures become available, he argued, investors will have much fuller visibility into the AI economy and its health. He said the disclosures would help show how enterprise spending is distributed and whether demand is growing enough to support the sector’s infrastructure costs.

On whether he would buy Anthropic, Calacanis called it a company “for the ages” and said owning some shares could make sense, but did not name a target price. His caution was about IPO mechanics: offerings, he said, are now “priced to perfection” rather than deliberately set low to create a first-day pop. An investor using a market order risks paying whatever the market is charging during the initial rush. His advice was to wait a week or two rather than hurry, and to think about an investment over a 10-year horizon.

He was also comfortable with Anthropic’s voting structure. The seven co-founders, through an LLC holding a single Class F share with no economic value, control 50.1% of the vote on general shareholder matters. Calacanis said investors who object can decline to buy the stock. Companies can choose different structures, remain private or go public, and shareholders have accepted founder control at other large technology companies. He acknowledged a broader question about whether a company building highly consequential AI should have that much control, comparing the issue to the influence of social media. He said representatives could take action on such concerns.

The IPO gave Calacanis a reason to make two separate arguments: that Anthropic’s stock deserved consideration, and that its disclosures could improve understanding of AI spending. On the first, he recommended patience rather than buying at the peak of opening-day demand. On the second, he focused on the split between subscription revenue and consumption-based use, which the company’s filings now make possible to track.

The bull case rests on how few people use AI deeply

Calacanis’s central investment thesis was that AI adoption among knowledge workers remains far below what his own experience might suggest. He estimated that only about 1% of Western knowledge workers use AI as intensively as he and Harris do. He framed the estimate around hundreds of millions of knowledge workers in Western markets, which he said have greater spending power than workers elsewhere.

Harris described using AI across most of his workday—roughly nine to nine and a half hours in a ten-hour day. Writing remains a deliberate exception: he may ask Claude for a first draft, but then writes his own version and uses the model’s work as a reference. For other tasks, he said, AI is involved in searching, designing, making things and editing video.

The shift, Harris said, comes when AI stops being a chatbot used for isolated questions and becomes integrated with a person’s tools and context. Instead of spending 45 minutes asking a model something, getting an answer and returning to work, he described an ongoing exchange with multiple agents across the work he is doing. Calacanis added that an agent can keep working in the background while its user is occupied with something else.

Harris described this change as moving from an occasional question-and-answer tool to an ongoing conversation that spans his work and other apps. He said connecting tools such as Plaud to Claude and Grokbot had changed what he could do with a captured conversation. On a commute, he recorded ideas for a blog post. By the time he got home, Claude was outlining it, while Grokbot was finding videos to link. Harris still wrote the post himself; the tools handled preliminary research and tasks such as collecting clips, transcripts and captions. He said the experience let him spend more attention on the writing and creative work.

From that starting point, Calacanis projected a large increase in usage. If the share of workers using AI deeply rises from 1% to 80% or 90%, he argued, the number of intensive users could grow by roughly 80 times. He expects usage per person, the amount of work delegated to agents and their effectiveness to keep increasing as well. He described the eventual scale as potentially hundreds of times current usage, with agents able to take on a dozen or more tasks for each person. His estimates combine the assumptions that more people will adopt AI and that each person will use more of it.

His metaphor was “Professor X with Cerebro”: a worker directing many agents, each capable of taking on different tasks. Calacanis said that could sharply raise output per employee. If AI helps a worker produce much more, he argued, spending more on models, tokens or local computing becomes easier to justify against the cost of that worker’s time. He welcomed Harris moving to a more expensive AI plan, saying the potential productivity gain dwarfs the additional monthly cost.

This is where Calacanis’s optimism becomes a market call. He said investors are underestimating AI’s revenue and efficiency potential, and described current sentiment as “peak cynicism.” He said he was buying technology stocks during the downturn and named Google, Uber, Tesla, Shopify, SpaceX and Anthropic as companies he would consider. His confidence rests on the distance between the small share of workers he believes are using AI intensively now and the much larger share he expects will do so later.

He drew a distinction between large technology companies and newer data-center or “neo-cloud” businesses such as CoreWeave. He called the newer firms potentially attractive but harder to evaluate because their financing can involve layers of loans and bonds. He had not done enough analysis to assess their debt exposure and raised the possibility that a company with too much debt could face a spiral. Their competition with well-capitalized incumbents was another concern.

At the same time, Calacanis argued that the broader buildout has ample potential funding if his usage thesis is right. He pointed to the resources of large, cash-generating technology companies and the capital raised by AI providers, and said the question is whether usage grows enough to justify that investment. He did not treat that outcome as certain; he described it as the condition behind his confidence that infrastructure would be built and demand would follow.

Bitcoin was a separate, sharper version of his market view. Calacanis said he owns some as a hedge, but would move some of it into what he considers more productive AI-related companies. He predicted Bitcoin might never reach its previous high again, or might not do so for a decade, and said he expects it to lag major technology stocks. He treated the prediction as his own investment view, not a claim about what would happen next.

Apple could arrive late and still make a trusted agent

Lon Harris described a Bloomberg report about an Apple job listing for a senior engineering manager to lead work on an agent that can use a computer. The role’s description included using apps, browsing websites, writing commands, writing and debugging code, calling services and handling multi-step tasks. The project was described as private and still early, though the report said Apple had considered a launch as soon as the end of 2027. At the same time, Apple was making it harder for third-party agents to control a Mac.

Jason Calacanis expected Apple to put up barriers and require security steps before software can control a desktop. He said those safeguards are reasonable, but distinguished the Mac from iOS: Apple can tightly control the iPhone and iPad, whereas Mac users expect to open and use a broad range of applications. In his account, agents already encounter prompts to change accessibility settings and confirm access. He saw those hurdles as permissions and security checks, not proof that Apple could—or would—close the desktop to outside agents.

His forecast was that Apple could have one of the world’s three leading agents within 24 months, even if it is late to the market. He argued that Apple’s possible advantage is the combination of its own hardware, local computing capacity, a familiar personal assistant in Siri, and customers’ trust in the company with their personal data. Siri would need to become much better, he said, and Apple might acquire or integrate a strong product rather than build everything itself. He mentioned Perplexity as one possible acquisition, not as a confirmed target.

Calacanis pointed to the value of an agent that can draw on a person’s history rather than respond only to the latest instruction. He described having assistants use Grok bots to examine their past conversations and identify recurring operational problems. One example was travel arrangements that repeatedly failed to account for his need to recline during long car rides. After reviewing the history, the assistants identified that he needed an SUV and noticed other recurring issues, including missing hotel information and lost points.

For Calacanis, that kind of context is central to the promise of an agent: it could identify repeated problems across earlier interactions and use them to improve later decisions. He sees Apple’s access to personal context, hardware and local processing as a favorable combination if the company can improve its software. The job listing offers evidence that Apple is working on computer-use capabilities; the 24-month prediction and the acquisition possibility remain his assessment of what the company could do.

Autonomy needs public trust as well as technical progress

California’s DMV barred Waymo vehicles from state freeways, calling the operation an imminent hazard to public safety. Harris said the DMV cited vehicles entering active construction zones, including three Waymos on US 101 near Burlingame. He also described a November 6 video of a Waymo approaching police vehicles that were slowing traffic on a freeway. The clip shown during the discussion depicted the car nearing a CHP traffic break.

The California restriction was not the same as Waymo’s response elsewhere. Harris said Waymo paused commercial freeway rides in all its markets, including Phoenix, Austin and Miami, while it reviewed performance. He said the California ban would remain until the company showed effective corrective action. The wider pause was the company’s review; the DMV action applied to California freeways.

Calacanis called the decision an underreaction. He argued that freeway driving raises the consequences of a failure: a crash at highway speed can cause a death even when the vehicle is not at fault. He gave the example of a person crossing the highway who is hit by a Waymo, and argued that public reaction could still focus on the fact that a driverless car was involved. In his view, a serious incident could damage the entire industry and give opponents a reason to seek wider restrictions.

His argument was not that Waymo had caused the specific harms he described. It was that the potential political consequences of a fatal crash could be much larger than the details of fault in that individual case. He said companies should be more conservative, including keeping safety drivers in vehicles or limiting speed, even if that slows deployment. He also pointed to other failures and disruptions he said had affected the service, including Waymos stuck in water and a prior connectivity issue in San Francisco.

Calacanis was especially concerned about self-certification—the idea that companies can ask the public and regulators to trust that a system is ready. He said the California requirement to show corrective action reverses that burden: the company has to demonstrate it has addressed the problems. He argued that “we got this” would not be enough when the public is also weighing job losses and the distribution of gains from automation.

Harris agreed that riders need time to become comfortable with autonomous cars. He said his own unease eased after taking several Waymo rides, but warned that trying to move everyone into driverless vehicles at once could provoke a backlash. For him, gradual exposure is part of how the public gets accustomed to the technology.

Calacanis said he remains bullish on autonomy and has invested in several companies in the field. His caution, he argued, was intended to protect the industry’s prospects, not to oppose it. He said he believes ride-sharing could grow substantially and wants autonomous services to succeed. The risk, as he sees it, is that companies move quickly enough to provoke a political response that restricts the whole sector.

He invoked Airbnb as a warning that a useful technology can run into political resistance when its effects are felt locally. In the case of autonomous vehicles, he said, critics may frame driverless cars as taking jobs and moving income from local drivers to technology companies. He described a narrative in which local workers lose income while the money flows to companies and investors elsewhere. He argued that this political framing will matter to the industry’s prospects, alongside its technical performance.

AI regulation reflects a fight over who shares its gains

Harris described recommendations from Governor Gavin Newsom’s AI panel: emergency authority to shut down a model deemed an imminent danger if its developer would not contain it, third-party audits, state safety standards and a possible cap on compute used for recursive self-improvement. A post by Newsom shown during the discussion described the effort as moving toward mandatory safety standards and accountability. Harris stressed that the panel’s recommendations were not law. He said the legislature would return to Sacramento on December 7 and that Governor-elect Xavier Becerra had said he would move quickly to put shutdown authority into law. He also cited Representative Ro Khanna’s call for a special session, saying a framework without action was meaningless.

The question Harris put to Calacanis was whether California could make shutdown authority law and use it to stop a model if officials did not trust it. Calacanis said such an outcome was not far-fetched in California. He compared the possibility to a wealth-tax proposal that, in his account, had come close to passing and would have faced a legal challenge. He expected conflict over the framework, rather than a settled path from recommendations to law.

Calacanis predicted that companies might respond by moving between states or countries, including shifting data centers or models abroad. He named Singapore, Malaysia, the Middle East and North Africa as possible destinations, speculating that firms could seek governments with which they could build closer relationships. He also said President Trump had proposed taxing data centers built overseas, and noted that Trump would be in office for two more years. In Calacanis’s telling, the resulting policy changes could shift again with political power.

He connected the risk of regulation to investor anxiety. Calacanis said fear that restrictive laws might pass was one reason technology stocks were falling. He expected some proposed rules to be performative or ineffective, but treated the prospect of regulation itself as a source of uncertainty for companies and investors. The recommendations were proposals in the source discussion, not enacted rules.

The exchange widened from AI policy to the distribution of AI’s gains. Calacanis argued that people who feel excluded from economic growth may turn to regulation or protest as ways to exert power. He called on wealthy technology leaders to support measures such as a $15 federal minimum wage, broader access to health care and lower taxes for people earning under $100,000. His point was that dismissing public opposition without addressing economic insecurity could make that opposition stronger.

He connected resistance to data centers with a broader sense that ordinary people have few ways to influence powerful institutions. In his account, people who believe the gains are going to a small group while they bear the costs may use local opposition and regulation to get attention. He said technology leaders should ask what would make people feel they share in the benefits of economic growth, rather than dismissing opponents as irrational.

Harris offered a related framing from Republican Governor Ron DeSantis: the claim that many young people have not experienced a genuinely free market because capitalism has been tilted against them. Harris said the formulation—that the task is to make markets work for more people—was politically powerful, regardless of whether DeSantis would follow through on it.

Calacanis agreed with the underlying problem of exclusion, arguing that public anger can grow when people believe gains accrue to a small group while they bear the costs. He described empathy for people under economic pressure as a strength, and said he had resisted pressure from political allies when he disagreed with them about issues including the minimum wage and immigration enforcement. The discussion’s tension was that he remained bullish on AI and free markets while warning that public opposition could intensify if people do not feel included in the gains.

For filmmakers, the dividing line is whether the work succeeds

Harris said critics had accused David Fincher’s new Once Upon a Time in Hollywood sequel of using generative AI imagery, based on alleged visual glitches and scenes shown at early screenings. He described the claims as unconfirmed. Fincher’s response to the Los Angeles Times was that he was not a Luddite, did not fear the unknown, and saw no reason to reject a capability such as changing a person’s eye color simply because it was new.

Calacanis said the audience’s first test should be whether the film tells a compelling story and works visually. He treats AI as one more kind of computer-generated visual effect, comparable in principle to earlier shifts from practical sets to miniatures, virtual sets or CGI. A technique that does not work can still fail the audience: he recalled being impressed by a de-aged Luke Skywalker until the character’s face appeared and said the effect crossed into the uncanny valley. He noted that a fan later improved it and, as he understood it, was hired by Disney to work on effects.

The disagreement is not only aesthetic. Calacanis acknowledged that a sequence made with AI might replace work that would otherwise have gone to a crew. He said the same concern applies when productions replace full-scale sets with miniatures or virtual environments. His prediction is that lower costs will allow more projects to be made, creating different jobs and opportunities even if some existing roles disappear.

Harris drew a distinction between using AI behind the scenes—for example, in storyboarding—and putting visibly poor results into a finished film. But he also questioned how cleanly AI can be separated from conventional effects software. He pointed to Pixar’s work on animating fur as an example of software built to make computer-generated hair move more naturally. At ground level, he said, the difference between that kind of tool and an AI system may be less obvious than critics assume.

For Calacanis, disclosure matters alongside quality. He welcomed restoration work that made old concert footage clearer, provided audiences were not misled into thinking it had been shot with modern equipment. He described seeing a 4K restoration of a Dire Straits concert and being able to make out individual guitar strings. His condition was straightforward: tell the truth about how the image was made.

His ambitions go beyond making films cheaper. Calacanis said he would welcome additional episodes that filled in stories from The Sopranos, or more performances built from old concert footage. Harris suggested there were character histories and gaps a new episode could explore. Calacanis acknowledged that extending familiar works in this way could seem sacrilegious to people involved in them, but argued that new tools could make it possible to produce more entertainment.

The frontier, in your inbox tomorrow at 08:00.

Sign up free. Pick the industry Briefs you want. Tomorrow morning, they land. No credit card.

Sign up free