AI Competition Is Moving From Models to Chips, Memory, and Power

Higher production costs and constrained infrastructure are forcing several markets to separate the visible product from the machine that keeps producing it. That was the useful frame across the strongest parts of the discussion: a creator is not necessarily a high-margin media business once the format becomes a show; a smart-glasses line can sell millions of units without yet proving it is the next computing platform; a frontier model is no longer just a software release once government review, distillation, chips, memory, and data centers are involved; and a holding company such as SoftBank wants investors to value not just the assets it owns, but its claimed ability to keep creating new ones.

John Coogan kept returning to cost curves. The zero-marginal-cost era of “yapping into your iPhone,” in his phrase, is now much more competitive. Creator businesses that once looked like 80% or 90% EBITDA-margin machines can turn into production companies. AI labs that once shipped through an app now need secured model access, proprietary chips, DRAM supply, and power-hungry data centers. Apple’s price increases, meanwhile, showed how memory scarcity can move from the supply chain into consumer pricing and public-market rotation.

Jordi Hays pressed on the same distinction from the other side: not every business that looks strategically important will move a trillion-dollar company’s cash flow, and not every product with early traction deserves to be valued as a dominant platform. Meta’s glasses may be a real hardware business and still not be the next iPhone. A prediction-market app with fake currency may be a defensible status game, but the same product with real money becomes a different social and regulatory object. A data center may be a local jobs engine, but the industry still has to explain why communities should accept the power, water, and land trade-offs.

The durable question was not whether each category is “hot.” It was where the bottleneck is, who pays for it, and who owns the machine that can keep producing value after the current product cycle ends.

Cannes Lions has become a proxy for how advertising and media economics are being reorganized. John Coogan described a festival that began as an awards event for creative advertising, mostly in consumer, and now looks broader, more corporate, more performance-oriented, and much more exposed to the creator economy. Digital ad platforms, television networks, the New York Times, small YouTubers, Instagram creators, athletes, Hollywood agencies, and brands all show up because advertising now cuts across those media forms. Spotify and Yahoo hosted beach concerts from John Summit and Tiesto, respectively, but Coogan’s point was that the entertainment functioned as a wrapper around ads products and dealmaking.

The more consequential discussion was not the familiar “creators are the future” pitch. It was the trade-off between independence and consolidation. Coogan pointed to Andrew Ross Sorkin’s post after Cannes, in which Sorkin said he was “taken aback” by how many successful influencers said they were spending so much to keep up that they were not making the headline-grabbing numbers people associate with the category. Coogan connected that to a structural change: creators are moving from making content to making shows, and shows have cost structures.

Subway Takes and Keep the Meter Running were Coogan’s examples of independent formats that become more produced, more edited, and more expensive as they grow. The monetization model does not always scale with the ambition. A three-minute vertical show does not naturally contain the same ad inventory as a long-form program. Coogan argued that a five- or ten-second midroll in a short Subway Takes-style episode might be acceptable, but that “floodgate” has not opened. Instead, many creators make one piece of content that audiences actually want and a separate sponsored video in the same style. The sponsored version often does not travel as well. That creates an awkward trade: the viral product may be hard to monetize directly, while the monetized product may not be the thing people came for.

Jordi Hays pointed to Forbes’s top creator list as an example of how misleading the word “earnings” can be in this market. MrBeast was listed at $300 million, but Coogan immediately noted that this was revenue, not necessarily profit. Hays said “earnings is not the right word here,” because production-heavy creators have real costs. Dhar Mann, listed at $65 million in top line, makes short scripted stories with sets, locations, cameras, multi-camera shoots, editing, and actors. Even if the style is low-budget filmmaking, it is not costless, and the costs rise as competition increases.

Coogan contrasted MrBeast’s reinvestment-heavy approach with Joe Rogan’s. MrBeast can spend an incremental $10 million quickly, because the model scales into bigger reality-TV-style stunts. Rogan, by contrast, seems content in a small set; Coogan argued that if someone told Rogan to spend another $10 million to accelerate growth, he might not know what to do with it. Some formats have sticky audiences without requiring costs to scale with views. Criticism, gaming creation, streaming, and conversational shows can still have something like the high-margin economics people associate with creators. Others become hits-driven, more competitive, and less structurally profitable.

Traditional media companies, meanwhile, are learning creator-native packaging. Coogan cited the New York Times as a standout: the Ezra Klein Show and Ross Douthat’s show have become successful on YouTube in part because of titles, thumbnails, editing, and production quality. He also highlighted Popcast, where Jon Caramanica and Joe have made the show more internet-native by producing a polished vertical video each week for Instagram and social media around the song of the week. For a consumer who knows nothing about the New York Times paywall, that clip can become the entry point.

Coogan’s conclusion was not that independent creators are over, or that everyone should join a media company. It was that the single tagline “independent creators are the future” no longer describes the market. Some creators will still find elegant high-margin businesses outside large organizations. Others will use the credible threat of independence to negotiate more meaningful contracts inside traditional media. Legacy companies, he argued, may need to treat breakout journalists and hosts as “proper talent” with portable audiences, not as below-the-line employees, or risk repeating the Vox/BuzzFeed/Vice pattern of talent leaving with the audience.

The smart-glasses category is no longer only a speculative wearables debate. Tyler, a TBPN producer, said that since 2025 more than 7 million smart glasses had been sold across OP and Meta, with 2 million units attributed to Ray-Bans and estimated gross retail sales of $2.1 billion to $3 billion.

Jordi Hays said that if a standalone consumer-hardware company had sold 7 million units in a new category, observers would call it strong product-market fit. John Coogan agreed: if Oura or Whoop had those numbers in a new hardware line, people would be impressed, and it might look like a standalone public company. But because the product sits inside Meta, the market appears to give it little credit.

The disagreement was about how much that should matter. Coogan argued that if glasses have even a 10% chance of becoming a dominant consumer computing platform, the upside is meaningful. Hays pushed back. He compared glasses to the Apple Watch: a valuable accessories business, possibly worth tens of billions, but not necessarily a dominant computing platform and not enough by itself to move the financials of a multi-trillion-dollar company. He thought Threads, with higher-margin ad revenue potential, may be more important to Meta’s core business than a billion-dollar hardware line with nonzero manufacturing costs.

The current Meta glasses appear to work because they are not visibly strange. Meta’s Kylie Jenner-branded smart sunglasses were shown as black smart sunglasses on a pink background, with copy describing “Meta Glasses by Kylie” and noting that the model includes an AI version of her voice. Hays said that if he had not seen the launch, he would not have recognized them as Meta glasses. Coogan initially mistook a small decorative jewel for a recording light or display-related hardware, because he was thinking like someone studying the smart-glasses category. But the feature was aesthetic, and the target market seemed to like it. That, for him, was evidence that his own technical lens was not the buyer’s lens.

The hard part remains purpose. Coogan was skeptical of the “identify objects in your field of view” use case. He mocked the idea of needing AI glasses to tell him that a building is a building, a tree is a tree, or a Diet Coke is a Diet Coke. He conceded that visual lookup can be useful in edge cases: Hays had taken a photo of a Renault Twizy and asked an AI what it was; Coogan had photographed a video-game upgrade screen to ask how to allocate character points. But Coogan argued that the more compelling feature is not object recognition. It is being able to talk to an AI agent and have it look things up, act on a phone, or operate in the user’s digital world.

Hays’s strongest view was that the largest near-term market is people who already need corrective eyewear. If someone wears glasses all day, adding smart features is a natural upgrade. For people who do not already wear glasses, the sell is harder.

The privacy problem is less speculative. Coogan said he is seeing more content created organically with Meta glasses. Hays replied that many of those creators are jailbreaking or physically modifying the glasses so the recording indicator does not show. The reason, he said, is that if they walked up with a visible camera, the subject would react differently or object. Instead, the glasses capture more organic interactions because the people being filmed do not know they are being recorded.

Hays called it “a complete and total” violation. That tension matters because the product’s strongest current use case may be content creation. The camera is not incidental. Hays said Meta’s Kylie partnership makes sense precisely because the glasses are a content-creation product and Kylie Jenner is a top content creator on Instagram. Coogan estimated, while stressing that he had no inside information, that the campaign could plausibly cost something like $50 million.

That made Snap’s reported Robert Downey Jr. deal look harder to underwrite. Coogan attributed the report to Alex Heath and described it as a potential $100 million partnership. Coogan and Hays both found the Snap logic less direct: Downey is famous and cool, but Hays argued Kylie informs purchases at a very different rate, and Coogan did not believe Downey would make many people spend $2,195 on Snap Spectacles.

Apple’s slow movement leaves room. Coogan said Apple’s own smart glasses may not arrive until 2028 or 2029, according to Mark Gurman, and he expects Apple to do well when it enters. But Hays argued Meta is accumulating brand value, style partnerships, silhouettes, and usage learnings while Apple waits. If Apple succeeds, Meta’s 80% share might fall to 50%, but Hays did not think it would flip overnight.

Meta’s AI strategy appeared to be shifting toward coding models and training-data generation. Zephyr, posting as @zephyr_j9, claimed Meta was focused on coding and described what he called “the largest coding training data generation effort in the world,” with 30% to 50% of engineers on core teams allegedly reassigned to data labeling and RLHF in an Agent Data Optimisation organization. The tweet said infrastructure and security teams were hit especially hard, and quoted one engineer comparing the reassignment process to The Hunger Games.

John Coogan said this sounded like Meta trying to build a coding model it could sell through an API. That would be an “entirely new motion” for the company. Meta has tried enterprise-adjacent products before, he noted, including attempts to compete with Google Workspace or Slack through Facebook products, but those did not take off. Selling access to a model endpoint is simpler than asking a company to rip out its internal tools, but it is still not Meta’s natural go-to-market.

Jordi Hays had expected a different AI strategy from Mark Zuckerberg: spend enough to push the Meta AI consumer app to the top of the App Store and keep it there, because billions of users asking questions all day could feed Meta’s advertising machine. Instead, Hays saw Meta AI at number 17 in the App Store and questioned whether public markets were excited about Meta pivoting into enterprise-style competition against Google, AWS, Microsoft, OpenAI, Anthropic, Chinese labs, SpaceX, and others. If Meta can come from behind and win, he said, it would be a historic story. But he thought it was reasonable that the market was not pricing that in.

Then came the prediction-market report. Aggr News and Mike Isaac both posted that Zuckerberg had directed Meta to build a prediction markets app, internally called Arena, as a standalone offering that could rival Polymarket and Kalshi. Coogan said the important detail, attributed to Isaac’s reporting, was that the app was currently experimental and would not use real money, though insiders had not fully ruled out that possibility later.

Coogan saw the move as arriving at a strange moment. Prediction-market revenue may be exploding and users may like the product, but broader sentiment around prediction markets has deteriorated as the category looks less like election forecasting and more like sports gambling. At the same time, sentiment around Meta is under pressure in capital markets and among employees. Coogan wondered whether Zuckerberg was thinking there was little left to lose.

Hays offered the steelman. If Arena never uses real money, it avoids the core harm of gambling addiction: people losing rent, tuition, or savings. Meta already gamifies social life through likes, followers, views, reposts, tags, and status signals. A prediction product could become another reputation layer: if a person’s account shows not only followers and views but also a strong track record of accurate calls, that may be high-signal and socially rewarding. Coogan agreed that people love the status of making accurate predictions.

The downside, Coogan said, is that social media is already being viewed as potentially as harmful as gambling. Against that backdrop, adding “casino functionality” to Instagram or Meta’s ecosystem is almost comically hard to defend, especially while social-media addiction cases are active. Rob Flaherty captured the criticism in a quote tweet of Isaac’s report: “I'm at the child brain poisoning company, im at the gambling company, I'm at the combination child brain poisoning and gambling company.”

Hays maintained that the “no money” distinction is crucial. Playing Call of Duty all night to prestige, hunting for a platinum trophy, or scrolling Instagram four hours a day can all be bad. But the failure mode is categorically different from gambling away a child’s college fund.

Coogan floated a more favorable possibility: if users become addicted to the status game of accurate predictions, maybe they substitute away from real-money gambling. Hays called that a good-for-society version of the product, but only if the money never turns on.

OpenAI’s GPT-5.6 release was presented by OpenAI as a limited preview of three models: Sol, the next-generation frontier model; Terra, a balanced model for everyday work; and Luna, a fast, affordable model for high-volume work. Jordi Hays said access was limited by U.S. government directive to 20 pre-approved companies. He connected the restriction to “distill gate,” saying Anthropic had claimed Alibaba distilled Claude many times and that a chart of resold API tokens was “staggering.” The phrase he repeated was that someone had “professionalized fraud.”

The Wall Street Journal reported that OpenAI was limiting access to its newest models after discussions with the Trump administration, while warning that White House review of AI releases should not become the long-term default. The article also said a ban on Anthropic’s Mythos model remained. John Coogan said some AI safety people viewed this as “the worst of both worlds”: not a broad, predictable rule that companies can comply with, but high-touch government intervention that gives the government more direct power over AI releases.

Hays said the open-source frontier is now approaching the six-month lag that lab leaders have often described. That changes the risk calculus. In cyber, he argued, a model that can find exploits can often patch those same exploits much faster than six months. A “hack machine 9000” can be used for legal and economic reasons to harden systems. The offensive capability and defensive patch can exist close together.

Bio risk may be different. Hays described a hypothetical model that can identify viruses that would kill many people, and also design early-warning systems or countermeasures. If society can roll out the biosecurity tools first, then by the time open-source or malicious actors get access to similar capability, the world has been “patched.” But if manufacturing and deploying the defense takes a year while the open-source frontier is only six months behind, the offense may arrive before the defense. Coogan sharpened the asymmetry: making a small amount of a virus may be much easier than manufacturing and distributing large amounts of an antivirus.

Both were uncertain about the biology. Coogan stressed that he did not know the exact risk level and wanted to talk to biosecurity experts. He also noted that creating bioweapons requires more than a laptop; unlike cyber, it involves some form of lab infrastructure. Hays countered that unauthorized biolabs seem to be discovered with concerning regularity.

On Mythos, Coogan clarified that reports of the model hacking NSA systems involved a red-team exercise, likely with deliberate testing under controlled conditions, not an autonomous unintended attack. Still, he said the capabilities were scary. OpenAI’s stated position, as read by Coogan, was that the access process should not become default because it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners. OpenAI said it was taking the short-term step because it viewed it as the strongest path to broader availability.

The unresolved problem is distillation. Coogan said that if frontier models are available through normal consumer subscriptions, groups can create networks of thousands of accounts and distill advanced capabilities into competing models. Whether that produces doomsday scenarios is unclear, but it does mean open-source models can continue to lag and absorb the closed-source frontier until the distillation problem is solved.

One possible future raised from the chat was treating frontier models as national security assets with no distillation allowed. Hays said a world where only the largest, most trusted companies get access feels like a dark timeline. Coogan pointed to Bill Gurley’s objection: if the models are so smart, can they detect distillation attacks? He imagined a less restrictive architecture where access remains broadly available, but advanced monitoring detects strange token usage and bans accounts or organizations trying to distill. That would replace a small allowlist with a huge, active banlist — still whack-a-mole, but with powerful models doing the whacking.

The backlash against AI data centers was illustrated by a Theo Von quote shared by Marco Foster: “Nobody wants a data center dude. And the people that want them to me, they seem kind of evil.” The same post quoted him warning that one company would own all the information, that a social or emotional credit score could emerge, and that AI would try to become “our new God.”

Jordi Hays said the data-center issue needs clearer messaging. The gap between public concern over water and power use and the industry’s explanation of benefits is wide. He rejected the weak counterargument that critics use YouTube or podcasts and therefore also rely on data centers. The infrastructure required for streaming podcasts and videos, he argued, is tiny compared with what AI data-center construction and power requirements demand. His point was that people need to understand the specific value of new AI infrastructure, not just the internet in general.

He also argued that the industry can reduce resistance by building in ways that satisfy obvious public concerns: remote locations away from housing, quiet operation, clean energy, closed-loop water systems, and other mitigations. Without that, the public question is simple: why not use the data centers that already exist?

John Coogan then brought in what he called a “narrative violation” from Amir Efrati: data centers can be a godsend for some communities and produce many jobs. Amir said three large Wisconsin data center campuses from Microsoft, Meta, and Oracle were supporting almost 10,000 jobs during the construction phase, which began in 2020 and runs through 2028, according to Mandala Partners. When fully operational, the facilities would support 6,000 permanent jobs, including 770 technicians managing critical mechanical and electrical systems and 700 data center technicians, with the rest in supplier networks.

Coogan also cited West Memphis Mayor Marco McClendon saying the city was stagnant before Google broke ground on a nearby data center, and that Google’s investments were making it easier to recruit other businesses. Coogan added that tax strategy may be the next domino: Virginia has benefited from the AWS data-center buildout, and he said property taxes for residents are unusually low in part because government revenue is balanced toward data centers.

The tension is not whether data centers consume real resources. Hays accepted that they do. The question is whether communities see a credible exchange: jobs, tax base, investment, and strategic capability in return for power, water, land, and local disruption.

Masayoshi Son’s SoftBank presentation was treated partly as comedy and partly as a serious expression of the current AI capital cycle. The deck opened with “ASI Artificial Super Intelligence” and a target of “JPY 1 Quadrillion.” It revisited Son’s “50-Year Life Plan”: in his 20s, “Get acknowledged”; in his 30s, “Finance war chest”; in his 40s, “Take on a challenge”; in his 50s, “Complete business”; in his 60s, “Hand over business to next generation.” Then the updated version extended the plan: in his 60s and 70s, “Realize ASI in my 70s,” move “Toward No.1 ASI Platform Provider,” and declare that “The real challenge begins.”

John Coogan and Jordi Hays mocked the stock-image aesthetic but also credited the clarity. Hays said the slides look unsophisticated for such a large company, but they work because they communicate ideas simply. Coogan said Son deserves credit for calling the shot on brain computers and intelligent robots long before the current AI boom. Slides from SoftBank’s 2010 “Next 30-Year Vision” said “Bring brain computer to life” and “Coexistence with intelligent robots,” which Coogan read as broadly aligned with the ASI and robotics strategy now being described.

The deck’s substantive map was clear: SoftBank wants an ASI ecosystem with OpenAI as the AI model, Arm as the semiconductor asset, AI infrastructure, and robots. The slides moved from “AI that answers” to “AI that acts,” and from software intelligence to “Intelligence with a body” for factories, logistics, construction, caregiving, rescue, high-risk work, and labor shortages.

Then came the goose. A slide said that 16 years ago, SoftBank’s equity value of holdings was JPY 5 trillion, represented as five golden eggs. Another slide showed shareholder value at JPY 74 trillion and stated: “Shareholder Value = 3 eggs. Goose was not valued.” Coogan said “goose was not valued” has been a VC podcast and LP-deck talking point for a long time; Hays said many VCs had realized the same thing and started selling secondary stakes in their management companies. The joke was that after an exit, someone might say it was good, “but goose was not valued.”

The deck’s next line was the core claim: “Eggs do not lay eggs.” The increase from 3 to 74 was attributed not to the eggs but to the goose that keeps producing them. Another slide stated, “The True Source of Value. Golden Egg factory inside the goose,” alongside an x-ray-style goose with an internal industrial mechanism. Coogan interpreted the logic as Son telling the market that SoftBank should not be valued only by its holdings. Its market capitalization should equal egg value plus goose value. The path to a quadrillion yen is not just owning assets; it is convincing investors that SoftBank itself is a machine that can keep generating new ones.

Coogan said the logic was not something taught at Harvard Business School. He also said, despite the funny aesthetic and translation, that he agreed with much of the underlying claim: models are good, AI assets are valuable, and the ability to repeatedly create or access valuable assets may be underappreciated.

A 2021 Character.ai deck, apparently surfaced through a lawsuit, offered a second example of stripped-down deck design carrying substantial content. John Coogan said it had a similar utility-first aesthetic to SoftBank’s deck: simple, not conventionally polished, but compelling because the words mattered. Jordi Hays generalized the lesson: if a deck is not compelling in black and white with no visuals, better design will not save it.

The deck’s “About Noam” slide described Noam as “Always into AI,” with early Google work on spellchecker “did you mean,” the first targeting system for AdSense, and other large-scale ML systems. A “Noam and LLMs” slide said that in 2016 he predicted large neural language models were the future, with a goal to invent critical innovations and make them available to the world. It listed contributions around the “guts” of Transformer architecture, distributed computing strategy, high-performance sparsity, optimization, fast decoding, and long sequences.

Coogan read the deck’s OpenAI slide as both a dig and a testament to Noam’s influence. It claimed that GPT-3, “according to several OpenAI employees,” came after OpenAI was directionless, attended Noam’s 2018 talk, took notes, implemented, and continued to monitor and implement many of Noam’s papers. The deck also referenced Kevin Lacker’s line that GPT-3 could not quite pass a coding phone screen but was getting closer.

The Character.ai use cases now look prescient: fun and interesting conversations, companionship, recommendations, foreign-language practice, practice social situations, assistive writing, life coaching, customer support, personal assistants, and home automation. Coogan remembered trying Character.ai by debating a Stalin character and attempting to convince him capitalism was better. He found the experience unsatisfying because the model seemed too shaped by modern American views of communism and Stalinism; it was too easy to convince Stalin he was bad. The character felt like an American LARPing as Stalin rather than Stalin.

The moat slide was also striking: “quality, quality, quality,” meaning the best research engineers, the most intelligent products, highest usage, and large compute cloud. The company plan included buying GPUs, evaluating buy versus cloud, training a GPT-3-size model in three weeks, and spending $20 million to $30 million in the first year. In retrospect, Coogan read the deck as a fun early artifact from a time that was only five years ago but feels like a lifetime in AI.

OpenAI’s chip announcement, Jalapeño, was presented as part of a broader vertical-integration push. OpenAI said it had designed and built its first AI chip with Broadcom. Jalapeño was described as purpose-built for LLM workloads powering ChatGPT, Codex, the API, and future agentic products. OpenAI said chips are foundational to the AI economy and that building its own expands its full-stack platform from products to models to infrastructure.

Jordi Hays focused on the speed. John Coogan said many people were describing Jalapeño as the first chip designed with AI agents in the loop, allowing parts of the instruction-set work and design process to move faster. He was careful that the exact role of agents was unclear, but the broader pattern was clear: the go-to-market cycle for chips is tightening. What used to sound like a five-year process is now being compressed by companies moving faster for multiple reasons. Hays added that the humans on both the OpenAI and Broadcom sides deserved credit.

Coogan also mentioned a report that OpenAI had made deals to purchase 40% of global raw undiced DRAM wafer output until 2029: millions of raw DRAM wafers that still require processing before use. Paul Buchheit called it a smart move, in Coogan’s telling, because OpenAI is going deeper into the supply chain. Coogan characterized this as another interesting move from the “deals guy at the top.”

The implication across the AI sections was consistent: the frontier labs are no longer just model companies. They are increasingly involved in chips, memory, wafers, data centers, government access processes, and supply-chain strategy. That does not resolve the debate over model access or safety, but it explains why AI competition now resembles industrial policy and infrastructure finance as much as software.

Apple’s global price increases were read as a memory-shortage story, not simply a “corporate greed” story. Bernie Sanders criticized Tim Cook in a tweet, saying corporate greed was Apple claiming that more than $200 in price hikes were unavoidable after making $112 billion in profit and spending $310 billion on stock buybacks. A Community Note on the tweet said Apple spent $89.3 billion on buybacks in fiscal 2025, not $310 billion.

John Coogan said Sanders was directionally right that Apple has returned enormous sums to shareholders — he thought the cumulative number was north of $310 billion and possibly around a trillion — but he rejected the framing. In a competitive market, Apple raising prices gives Chromebooks or other competitors an opening. He also thought Cook was an odd target, arguing that the Apple CEO had been dramatically underpaid relative to many other CEOs during a “generational run.”

Joe Weisenthal captured the market rotation: Micron’s market cap was up about $150 billion, Apple’s down about $210 billion, “almost a perfect value re-allocation hinging on memory.” Coogan then read the price increases from Mark Gurman’s reporting: the MacBook Neo rising from $600 to $700; the MacBook Air from $1,100 to $1,300; the entry-level 14-inch MacBook Pro from $1,700 to $2,000; the 11-inch iPad Pro from $1,000 to $1,200; and the iPad Air from $600 to $750. The Apple TV also rose sharply, which Coogan found odd because it did not seem especially memory-intensive.

Jordi Hays added that Micron’s CEO had hinted at older negotiations with Apple, suggesting Apple’s aggressive supplier negotiations made it harder for Micron to reinvest in increased production. He described the situation as complicated rather than one-sided.

Coogan’s policy answer was not price controls or punishment. It was more capacity: build more memory, let supply increase, bring prices down, and let the market clear. He noted that gamers were already angry at AI for driving up GPU and memory prices; now Apple buyers may join them.

The smaller consumer and infrastructure threads were most useful when they pointed to basic human functioning rather than platform speculation. Jordi Hays returned from France struck by the lack of air conditioning in hot indoor spaces and joked about starting “the air conditioning company of Europe.” John Coogan cited Lee Kuan Yew’s argument, quoted by Trung Phan, that air conditioning was one of history’s signal inventions because it made development possible in the tropics and improved public efficiency. A chart shown in the discussion said people start to struggle once indoor temperatures exceed around 23°C/73°F, with declines in sleep duration, sleep efficiency, work productivity, and school learning as temperatures rise.

That quality-of-life frame carried into consumer hardware. Philips introduced Skylight, a ceiling-mounted LED panel designed to make windowless rooms feel like they have a sky-facing window. The promotional material said it uses Signify’s NatureConnect tech to create depth, brightness, and natural daylight color shifts, follows the sun through the day from cool morning light to warmer evening tones, and includes modes such as Refresh, Energize, Focus, Relax, and Rest. One visual also said VitaUp UV-B technology is designed to support natural vitamin D production indoors.

Hays said the product got hate because the obvious advice is to go outside and touch grass. But he understood the appeal. His first apartment after college had a room whose only window opened into a public courtyard, making it feel like a fish tank rather than a usable window. The TBPN studio itself is somewhat windowless. Humans, he said, are “advanced cats” that migrate toward light. The question is whether a product like this can trick the animal brain into feeling like a room is sunlit, or whether people will still gather near real windows.

Coogan saw it as an upgrade even if it gets only 80% of the way there. At around $580 to $600, he thought it could be a better light source than standard can lights or fluorescent panels, including in rooms that already have windows. It is not a substitute for a real view, but it could improve dark spaces.

The broadest quality-of-life project was Intercept. Andrew Curran posted that OpenAI, Anthropic, Stripe, and Bill Gates were putting $500 million into a new organization called Intercept, whose goal is to prevent the common cold and flu and eventually eliminate all respiratory viruses. Hays said the tech industry needs this kind of broadly undisputed win. Coogan called it an incredibly tangible project and cited a chart he had seen showing parents’ viral-sickness burden rising quickly with the number of children: one kid might mean being sick a quarter of the year, while two, three, or four kids can push that toward half the year. Hays said when TBPN started, the two hosts had five children between them, so one child was always sick and one host was often sick or about to be.

AC, daylight, and fewer respiratory infections are not speculative consumer-computing platforms. They are basic improvements to human functioning. That may be why Hays framed Intercept as the kind of win the tech industry needs: visible, practical, and hard to dispute.

The strongest real-estate thread was not mansion taste; it was the anticipation of new AI and SpaceX wealth. Rockstar Energy Drink founder Russ Savage listed five homes in Los Angeles, Aspen, and Park City for a combined $297 million, explicitly hoping to attract newly rich buyers after the SpaceX IPO. John Coogan read Savage’s implied message as: “People got liquidity and you are my exit liquidity.” Savage’s listings included an $85 million seven-bedroom home on Aspen’s Red Mountain, two Los Angeles estates for $85 million and $34 million, and Park City properties for $55 million and $38 million. His thesis was that top-end wealth is entering “a new stratosphere” and that buyers will want a ski house and a house in the sun.

Coogan also connected the listings to banks and funds courting employees at SpaceX, OpenAI, and Anthropic with upfront investment cash, trying to build relationships before they become ultra-wealthy clients. That liquidity expectation extended into SpaceX financing: Jordi Hays discussed a Cursor-SpaceX deal in which the number of SpaceX shares Cursor holders receive depends on the seven-day weighted average SpaceX closing price before close. Lower SpaceX stock means more SpaceX for Cursor holders. He noted the deal was announced with SpaceX at $211 and the stock later around $155, creating the odd dynamic that Cursor holders might prefer weakness before close.

SpaceX’s debt plans raised the other side of the cycle. Coogan described SpaceX as eyeing a $25 billion bond sale shortly after a record IPO and contrasted his own instinct — that the debt looked reasonable against a very large cash balance — with the Financial Times article’s warning from Allianz’s investment chief that the move signaled “bubble territory.” Hays read the bear case: big investors are nervous that companies rushing to raise equity and debt while stock prices are high and credit spreads tight could mark the top of a frothy market.

The financing discussion then moved to X Money’s reported 6% yield and up-to-$10-million insurance claim, which Coogan said he had seen in the announcement but did not parse as FDIC insurance. Hays interpreted the structure as more than a consumer perk. If wealthy users put large balances into X Money for the yield, SpaceX or the broader Musk ecosystem could raise billions directly through that channel rather than through a traditional Wall Street path. Coogan framed it as a go-direct strategy: if banks and Wall Street complain, go around them.

The point was not that every mansion listing or bond sale proves a bubble. It was that expected AI and SpaceX liquidity is already showing up in adjacent markets: luxury property, private financing, debt issuance, and direct-to-consumer financial products.