AI Data-Center Forecast Rises to 50 Gigawatts as Demand Economics Remain Unsettled

Jordi Hays said SemiAnalysis had raised its 2027 data-center capacity forecast from 43 to 50 gigawatts. That increase sharpened a mismatch in the current debate: questions about whether AI companies can sustain revenue growth sit alongside evidence, in the hosts’ account, that a large amount of infrastructure is already on its way.

John Coogan said he trusted SemiAnalysis’s supply-side read because it considered satellite imagery and deals already signed. A market selloff, he argued, would not quickly stop projects already planned and committed to. If conditions worsened for months or rates rose, some projects might be delayed or deals unwound. But a brief downturn would not make construction halt at once.

Demand was less settled. Coogan expects personal AI agents to use substantial compute, and said he had been using Dot to find follow-ups in his email that he might otherwise miss. He also likes using an LLM to research the web: instead of opening a couple dozen tabs, handling cookie prompts and logging into separate services, he can get a synthesized answer with quotations and links. He described that as a better consumer experience, but was less certain how providers would make enough money from it. Would consumers pay for inference, or would revenue come from advertising or transaction fees? He mentioned a bearish estimate putting the agent-commerce market at about $40 billion and questioned whether that would justify the investment.

Hays argued that automating isolated tasks may not be enough. The larger opportunity, he said, would be automating roles or extending what people can do. A viral account of someone working a “26 hour day” to stay available to their agents illustrated the strain that can accompany trying to keep those tools productive. The hosts joked about the premise, but their economic question was serious: could agents change what a worker or business can accomplish enough to support much higher spending?

Coogan saw a possible case for large AI budgets where returns are clear. He said it made sense that a software engineer earning $300,000 might receive a similar-sized budget for tokens if the tools generated sufficient value. He was less sure what an equivalent budget would do for a salesperson or marketer. If a sales representative spent $50,000 a month on inference, he asked, could the best models reliably produce $100,000 in value the representative would not otherwise have generated?

That question matters to the infrastructure argument because the hosts did not expect consumer subscriptions alone to pay for the build-out. Hays said consumers would not fund all the capital expenditure. Coogan agreed that enterprise and business-to-business spending would need to rise well beyond subscriptions costing $20, $100 or $200 a month. He offered a hypothetical: a $10 million business now spending $100,000 a year on inference might need to spend $1 million, perhaps while growing to $15 million in revenue. The example described the scale of change the investment thesis would require, not a forecast for a particular company.

The optimism behind that thesis appears in a survey of 2030 GDP expectations discussed by the hosts. The chart shown on screen put the median prediction at 17.1% among Anthropic staff, 12.5% among site visitors and 5.4% among US adults; a second chart added a 2.0% long-run median for the Federal Open Market Committee. Coogan asked whether the survey meant real GDP growth or included inflation, then proceeded on the assumption it meant real growth. He contrasted the Anthropic staff figure with the roughly 2% growth he said the US typically sees, calling it a prediction of a huge economic boom.

When Hays asked for Coogan’s own number, he answered “200% GDP growth,” joking that he wanted to triple every year and was manifesting it. The hosts also read a social-media response calling the Anthropic figure “borderline mental illness.” Hays and Coogan said the optimism was not unique to AI labs: Coogan pointed to the 5.4% median among US adults as bullish too.

Coogan challenged the idea that AI’s effects were still invisible in ordinary commerce. In a discussion about when a shopper might encounter an AI-made product in Walmart, he said AI-generated goods were already being sold through online shopping and social platforms. He cited beach towels and a pool float shaped like a tank as examples of what he called an “AI slop economy.” Those examples were evidence, in his account, of products reaching buyers—not proof of a specific effect on GDP.

Epic Games founder Tim Sweeney offered a broader production-side prediction: games developers had wanted to make, but could not afford to produce, would become possible, creating a “new golden age of gaming.” Coogan welcomed Sweeney’s optimism and said he expected him to adapt as he had through earlier technological shifts. A reply to Sweeney asked whether this meant Epic would no longer be necessary; the hosts did not resolve that question.

The hosts’ skepticism about AI forecasts echoes their criticism of the way You Can See Everything presents Elizabeth Holmes and Theranos. John Coogan said some viewers seemed to come away more sympathetic to Holmes, perhaps seeing her as manipulated or as someone with more to her story than the fraud. Jordi Hays put one version of that reaction plainly: a viewer might think she was bad, but would probably have been nice to them in person.

Coogan’s objection was that sympathy should not be mistaken for evidence of sound execution. He described Theranos’s account of its invention—a box that could test blood for many things—as making routine medical-device development sound like magic. He said more than 60 companies in the diagnostic-device category had sold for over $100 million since Theranos began. His point was not that building a diagnostic instrument is easy. It was that the work belongs to an established industry, where companies build and sell testing devices; it is not an unprecedented feat only Holmes could accomplish.

Coogan connected Holmes’s rise to the “Ted Talk entrepreneur” era. He cited a post by Nikita Bier about the culture that grew around child-prodigy narratives, college accelerators and young-founder organizations after The Social Network. Coogan said he had seen hard-tech founders present fraudulent products and studies with fudged numbers at conferences, while adults excused the conduct as “fake it till you make it.” In his view, Holmes rode that wave especially well, then continued using the same playbook after the culture had started to change.

Medical devices also face constraints on what companies can claim publicly. Coogan said a founder can build a device without FDA authorization, but marketing or selling it requires authorization. He added that founders generally cannot publicly claim that an unapproved product works; they have to describe it as a work in progress. That regulatory context, as he presented it, contrasts with the sweeping claims he associated with Theranos.

His counterexample was GeneWeave, whose co-founder Diego Rey was Coogan’s visiting partner at Y Combinator in 2018. Coogan described the company’s device as a lab-testing machine made small enough to put in a hospital, where it could test for antibiotic-resistant bacteria. The underlying chemistry was hard science, he said, but not “future sci-fi magic.” A result in an hour rather than two days could help hospital work move more smoothly and make people healthier. Coogan said GeneWeave raised $25 million and sold for $425 million; he said Roche later obtained FDA authorization for its cobas vivo Dx MRSA test in 2019. He noted that this was not the kind of product likely to appear on a TED stage, though he thought it should.

For Coogan, a credible Holmes redemption would need to rest on specific knowledge, not an abstract comeback story. He said he wanted her to make a strong call about biotech, explain how a market might evolve, or offer a market map showing that she understood what companies such as Roche were doing. Hays agreed that the film did not show Holmes particularly engaged with science. If she made a specific, testable argument about biotech and later proved right, Coogan said, that would give people a different reason to take her seriously.

The hosts also discussed the film’s treatment of money and responsibility without resolving the legal questions. Coogan said the film showed Holmes’s partner, Billy Evans, negotiating a share of the movie’s back-end profits and Holmes later backing out of the arrangement. He wondered whether money paid to Evans could support their family without going directly into an account that might be subject to restitution claims. He said he could not establish whether the two were legally married. Hays speculated that if Holmes were involved in routing earnings to Evans, a judge might still treat the money as hers. They left both the arrangement’s legal effect and its implications for restitution open.

The questions of control and incentives also surfaced in the hosts’ discussion of Holmes and Sunny Balwani’s outcomes. Jordi Hays said the film left him with the impression that Holmes had received more prison time than Balwani, Theranos’s other co-founder. John Coogan corrected that impression: he said Balwani was sentenced to 12 years and 11 months, while Holmes’s sentence had been reduced from 11 years and three months to 10 years. They also said neither had sold secondary shares. Hays called the fraud “purely for the love of the game”; Coogan’s point was that, as they understood it, neither had cashed out before the company collapsed.

Coogan then joked about a way founders might sell secondary without making it look like a vote of no confidence: take on a large personal loan first, sell shares and use the proceeds to pay off the debt. Investors might hear “paying off my debt” and imagine ordinary personal borrowing, he said, rather than an equity-backed line of credit. He presented it as a possible maneuver, not as evidence that any particular founder had used it.

The Poppi sale offered a less hypothetical example of how headline values can depend on deal terms. The hosts described PepsiCo’s acquisition of the beverage brand as worth $2.2 billion: $1.9 billion paid at closing and $300 million contingent on later performance milestones. Pepsi announced that the earnout would be $12 million rather than $300 million.

The hosts pushed back on reactions focused only on the missing earnout. They said the founders had already received $1.9 billion—around 90% of the stated deal value—and called that an excellent outcome. The economics would have been very different, they argued, if $300 million had been paid upfront and $1.9 billion had depended on milestones.

The risk is that an earnout depends on a company the sellers no longer control. The buyer controls budgets, distribution and other decisions that can affect performance. One post shown during the discussion called an earnout “a future bet on a business you no longer control.” A participant cited a rule of thumb that sellers should expect perhaps a quarter of an earnout and write the rest down to zero. The point was not that every earnout pays out at the same rate, but that contingent value should not be treated as guaranteed.

Hark’s attempted $1 million pizza giveaway made the gap between an appealing offer and the permissions behind it unusually visible. Jordi Hays said users were asked to download the app and order through it, but the promotion appeared to have problems. A screenshot showed a Google authorization page requesting access to Drive files, contacts, spreadsheets, calendars, email and an organization’s workplace directory, including permissions to edit or delete some data. The hosts did not establish whether every permission shown was required to participate.

John Coogan joked that a free pizza seemed to require letting the app edit a tax return. Hays described one user trying to get help through the app. When offered the choice of reporting the problem to Hark or filing with the Federal Trade Commission, the user chose the federal complaint. Another screenshot showed a draft complaint about the promotion, including the claim that users were asked to connect email. The hosts treated the escalation as comic: a giveaway that had been flaky for a few minutes on launch night had prompted a formal complaint. They did not present it as evidence of a broader privacy outcome.

They also wondered why Hark had been set up separately from Brett Adcock’s company Figure. Coogan suggested that a separate company could raise fresh funding and offer a new co-founder equity stake, with a possible later acquisition into a larger business. He compared the structure to xAI, while Hays noted that Elon Musk had already built car and rocket businesses. They left Hark’s rationale uncertain.