Google Cloud Backlog Grows as Search Ads Face an AI Monetization Problem

John Coogan described Alphabet’s earnings as a rare miss, with the stock down 12.5%. Revenue was $127.6 billion against consensus of $128 billion, a shortfall of about a third of a percent. Earnings per share came in at $2.12, versus expectations of $3.01.

Part of the EPS gap came from an $8.7 billion loss on equity investments tied to SpaceX stock, which Coogan said reduced EPS by 55 cents. Excluding that effect would put EPS at $2.67—still below expectations. The write-down explained part of the miss, not all of it.

The strongest operating story was Google Cloud. Revenue doubled to $30.7 billion, operating profit tripled to $11.1 billion, and operating margin reached 36.2%. Coogan said Cloud had become a substantial business in its own right, a striking reversal from reporting that Larry Page considered pulling the plug on it in 2017, when Google was struggling with internal competition and the prospect of taking on AWS.

The $561 billion Cloud backlog is a large number, but its timing matters. Coogan noted that some contracts extend years into the future; he cited contracts running to 2033 or 2036. Still, he said, 55% of the backlog was expected to be recognized as revenue within two years—roughly $275 billion. For a business that had generated $30.7 billion in quarterly revenue, that expected conversion rate was one reason he thought Cloud should support Alphabet’s valuation. The hosts also pointed to demand for Google’s cloud services and TPUs, including OpenAI’s purchases.

Search presents a less settled picture. Advertising growth slowed from 17% the previous quarter to 12%. Coogan said AI was accelerating query volume and that people were using AI Mode, sometimes because it appears in place of the default search experience. But AI Mode currently monetizes less effectively than traditional search.

The challenge, in his account, is not simply getting people to ask questions. It is earning from the answer. Coogan compared a search results page to a screen whose available pixels Google wants to monetize. Traditional results have accumulated sponsored placements over time: he said that for some queries, users can scroll through much of the first page before finding an organic result. An AI response is not yet covered in the same way. Google will have to work out how to place paid links, offers, and brand messages into responses while earning from the space it gives users.

That problem connects to a broader criticism of Search. Coogan cited a post by Max Anderson, who said he had spent more than $500,000 a month on Google Ads and called Search revenue growth “artificial and extremely unhealthy” for the business. Anderson argued that declining legacy search volume was being offset by higher charges for lower-quality clicks, including clicks advertisers did not want. Coogan said the post’s arguments had been challenged, but its underlying concern—that Search may already be heavily monetized—was familiar.

He used mattress searches as an example: paid shopping placements, several sponsored brand results, and review sites that he said had been purchased by mattress companies can leave users with few clearly independent destinations. Yet the business remains sticky, he said. People keep searching through familiar flows, so the concerns have not translated into a collapse in usage. That persistence gives Google time to adapt, but the shift to AI responses creates a different advertising surface to monetize.

Alphabet is making a large bet on infrastructure at the same time. Its 2026 capital-spending guidance was $195 billion to $205 billion, and Coogan said the company signaled a significant increase in 2027. Third-quarter spending was $51.4 billion. Microsoft’s capex was $38.37 billion against an estimate of $45.9 billion; Google’s was $51.36 billion against $55.7 billion. Jordi Hays said the lower-than-expected spending reflected delays, including power and other infrastructure constraints, rather than a lack of willingness to invest.

Coogan pointed to the difficulty of getting GPUs installed and data centers through permitting. Demand may be enormous, but deployment still runs into the physical world: equipment has to be racked, power made available, and sites brought online. For companies selling compute, he added, supply constraints can also mean that margins remain healthy even when the build-out is delayed. Spending less than forecast in a quarter does not necessarily mean a company has changed its ambitions; it may mean the infrastructure is not ready to absorb the planned investment.

Google reported that Gemini had passed one billion monthly active users and that its first-party AI model APIs were processing 32 billion tokens a minute, up 40% quarter over quarter. Coogan cautioned that token volume is hard to interpret across increasingly efficient models; what matters is whether systems deliver useful answers. Google announced Gemini 4 Carbon two days before earnings, Coogan said, helping reassure investors that the company remained on the frontier.

For Coogan, internal model capability has strategic value even if Google’s models do not show up prominently in third-party usage data. The company can fine-tune models for its own infrastructure and products: Search needs fast responses, while YouTube could use models to work across its video transcripts. Jordi added that access to frontier intelligence at a cost below market prices could matter strategically. The optimistic case is not only that Gemini attracts users directly, but that Google can use its models throughout Search, YouTube, and internal workloads. The counterweight is that Search monetization is slowing just as the infrastructure bill climbs.

Jordi also noted that Berkshire’s recent investment in Alphabet had an approximate blended share price of $350, compared with a share price of about $280 at the time of their discussion. He described the drop as bringing back an older Google bear case. That tension sat alongside the earnings report: Cloud is growing rapidly, but Search remains a central business whose advertising growth and future role in an AI-first experience are less straightforward.

A separate AI question is whether training in games and simulated environments can produce skills that transfer to difficult work outside those environments. John Coogan pointed to a post describing rumors that OpenAI’s internal model had made a reasoning jump through more aggressive reinforcement learning, including training in games and simulated environments. The post claimed that strategies learned in those settings could transfer to research mathematics, citing Navier–Stokes and ETH results. It described Anthropic as pursuing a similar direction, training its internal model on long, difficult tasks in demanding environments and claiming transfer to other fields.

The post also pointed to hiring and job descriptions as clues: it cited OpenAI hiring around synthetic reinforcement learning, self-play, simulators, and synthetic feedback, and said Elon Musk’s Digital Optimus role called for measuring transfer from games into general computer use. Ashlee Vance responded by asking whether the labs were returning to Demis Hassabis’s original vision for reaching AGI after a relatively brief period of brute-force scaling.

Coogan connected the possibility to Google’s continued work on Genie. He contrasted Google’s ongoing interest with Anthropic’s reported focus on coding and OpenAI’s past “side quests.” Some, such as Sora, may have been costly in compute, he said; others, such as computer use, could prove more valuable over time. He recalled OpenAI’s Operator as an early attempt to have a system click through websites in a virtual computer environment. At the time, he said, it was “pretty janky”; he later described computer use as “basically solved.”

Genie’s value remains an open question in Coogan’s account. It might be an expensive entertainment product, or it might become a useful training asset that gives models another route to improvement. The central issue is whether skills learned in simulated worlds transfer to reasoning and practical computer use. The posts and examples raised that possibility, but did not settle it.

Jordi Hays read from what he described as a Times report alleging that Alice Huang, a Chinese citizen and executive at the Singapore-based data-center company Mega Speed, was a Chinese intelligence asset. As Jordi recounted the report, U.S. officials had been scrutinizing a top executive at a Singapore-based data-center company over the possibility that advanced Nvidia chips were being steered to China. He said the report described surveillance of Huang revealing that she had developed a close relationship with Nvidia CEO Jensen Huang, and said President Trump had been briefed roughly a year earlier and took no action. These were allegations and reported intelligence findings as relayed by Jordi, not details the hosts independently established.

The hosts said they had discussed Alice Huang and Mega Speed on an earlier show, including her access to Jensen Huang and her background as a reporter for Chinese state media. Jordi said they had not explicitly suggested she was a spy because they lacked details, though he had found the situation suspicious. In his account, the report added an allegation that officials had concluded she was an intelligence asset.

Coogan focused on why access to an Nvidia executive might matter in a chip-control dispute. He suggested that moving more Nvidia chips into Malaysia could increase the computing capacity available for China. He also described Malaysia’s expanding compute capacity as a potential warning sign if it seemed disproportionate to the local business and economy. In his analogy, an unusual measurement can indicate a problem before it identifies the cause.

The hosts discussed routes around restrictions as examples of the incentives at stake. Coogan mentioned reports of people removing labels from servers to move them across borders, then described an account of people buying DJI drones in a Western country and reselling them at much higher prices near a conflict zone. His point was that restrictions and price differences can create strong incentives for arbitrage. He then speculated that someone seeking access to an executive might try to understand how chip-diversion controls work and how they could be circumvented.

Jordi and Coogan noted that chip-control legislation could be expanded or revised. They also emphasized the difficulty of enforcing restrictions when potential profits are high. Coogan’s broader argument was that diversion is not only a matter of intercepting shipments: people have reason to learn how controls operate, find gaps, and exploit them. The alleged relationship described in Jordi’s account of the report sits against that larger contest over access to advanced chips and the routes by which they might reach China.

A video framed as “Losing our house because of day trader” appeared to show a family embracing outside their home. Jordi described the clip as designed to provoke an emotional reaction; the phrase itself—“because of day trader”—struck him as odd. Looking further into the account, he said, led to follow-up videos about winning the money back, then to a link in the creator’s bio and a sequence of trading products: a free course and paid courses priced at $49 and $100.

The hosts did not know whether the family had actually lost its house. Jordi said the video could be part of a marketing funnel for an information product, and the creator could even be doing well financially without trading. Coogan called the pitch counterintuitive: present yourself as having failed badly, then invite viewers to follow your recovery. They agreed that the apparent loss was upsetting, while noting how an alarming post can hold attention and lead viewers toward further videos and purchases.

The hosts then discussed a different day-trading model: paid simulated trading challenges. In the model they described, customers trade in a simulated market tied to real data, and the company may promise to provide capital if they perform well. The pitch, as Jordi summarized it, is that customers prove their skill and may be “staked,” becoming small independent traders. Jordi said the hosts had found one company in this market that had collected hundreds of millions of dollars in fees, and added that there were many such firms.

Customers may pay around $100 a month for a simulated account, and, according to the hosts, some take out multiple subscriptions. Jordi said some firms had been exposed as Ponzi schemes. He described a possible mechanism in which a firm might use other customers’ money to make payouts to simulated traders it judged successful. In his account, a reward or an “attaboy” can reinforce the prospect that a customer could eventually receive capital to trade; the hosts did not establish that this applied uniformly across firms.

Jordi described the pitch as a win-win: customers pay to demonstrate their ability, and a firm may then provide capital. Coogan said he had initially imagined people in viral trading videos as trading against established hedge funds such as Citadel or Jane Street. He said the arrangement they were discussing could be different: customers may be paying a subscription to a firm running simulated trades rather than trading against a hedge fund. They are not necessarily putting their own money into the market. Instead, some may be carrying recurring charges on credit cards in the hope of qualifying for capital.

That distinction changes the picture suggested by the phrase “day trader.” A viewer might assume someone has put savings into the market and lost them. In a simulated challenge, the customer may instead be paying a firm for repeated chances to demonstrate performance, with the possible reward being access to capital. Jordi contrasted that arrangement with ordinary investing: investors do not typically receive an “attaboy” and an offer of capital for each successful trade. The prospect of being backed, alongside the feedback and rewards the hosts described, can make repeated subscriptions appealing—even when customers might otherwise invest the same money.

The discussion tied two online pitches together without establishing that they were connected: the emotional story of a family losing a home through trading, and the business model some firms use to sell simulated trading with the possibility of future funding. The first may draw viewers into a course funnel; the second charges recurring fees for a chance to prove trading skill. Both can keep the possibility of a comeback—or a funded account—in view.