Summary

OpenAI is positioning ChatGPT as the new front door to the web, not merely a chatbot. With more than 800 million weekly users and API traffic of 6 billion tokens per minute, up from 300 million in 2023, it is embedding Expedia, Zillow, Figma, Target, and Spotify inside conversations. The ambition is explicit: turn ChatGPT into “a new operating system for everything that you might wanna do.”
The platform opportunity carries a privacy risk potentially greater than Facebook’s developer era. Facebook once earned a 30% cut from Credits, with Zynga alone providing 12% of company revenue, before Cambridge Analytica exposed the cost of loose data sharing. ChatGPT may hold therapy transcripts, memories, and intimate advice; OpenAI promises to share the “minimum necessary amount,” but its own Canva warning says conversations and memories may reach developers.
OpenAI’s eventual monetization and ranking rules could determine which internet businesses win distribution. Kevin Roose wonders whether Zillow might be privileged over Redfin because it has an OpenAI integration; Sam Altman says taking inappropriate payment would “clearly destroy that relationship very fast,” while Greg Brockman cautions that determining the best product has “a lot of nuance.” Casey Newton hears an angel saying “be really, really careful” and a devil answering that it is nuanced.
The AMD agreement expands OpenAI’s emerging trillion-dollar GPU empire while giving AMD both a cornerstone customer and help closing its software gap with Nvidia. OpenAI plans to buy 6 gigawatts’ worth of AMD chips, versus 10 gigawatts under its Nvidia deal, beginning with AMD’s newest chip in 2026; it can also earn rights to buy up to 10% of AMD for a penny per share. Joint work on ROCm could make AMD’s hardware more competitive with Nvidia’s CUDA ecosystem.
Financing is now the critical fault line in the AI thesis. OpenAI does not have the roughly $1 trillion it has pledged to spend, so it expects some combination of equity, debt, vendor financing, product revenue, and possibly “new kinds of financial instruments.” The hosts see a path without AGI if adoption keeps compounding, but warn that interlocking commitments could make OpenAI “too big to fail” if models disappoint or scaling stalls.
The compute boom is becoming an economy-wide leveraged infrastructure trade. Jason Furman estimated that data-center and information-processing-software investment accounted for 92% of US GDP growth in the year’s first half, while Bernstein’s Stacy Rasgon wrote that Altman could “crash the global economy for a decade or take us all to the promised land.” The next bottlenecks extend from GPUs to power, cooling, physical sites, gas turbines, and skilled electricians.
Sora 2 demonstrates both genuine creative pull and the fragile economics of AI-generated video. Katie Notopoulos scrapped a prewritten denunciation after finding the app unexpectedly fun, even as copyright disputes, porous guardrails, bullying, and a user base she found overwhelmingly male exposed its liabilities. Katie and Kevin think novelty may fade into group-chat use; Casey argues that better voices, clothing controls, and clips longer than 10 seconds could repeat ChatGPT’s path from toy to 800-million-user habit.

Deep dive

1. DevDay made OpenAI’s platform ambitions measurable

Casey Newton found the venue familiar in an uncanny way: nine years earlier, Facebook had promised Messenger bots would eliminate calls to “1-800-FLOWERS.” The faces and company had changed, he said, “but the promise feels sort of the same.”
Sam Altman reported more than 800 million weekly ChatGPT users and 6 billion API tokens processed per minute, versus 300 million per minute in 2023. Casey’s caveat: startups invent impressive metrics that “the human mind has absolutely no idea how to understand.”
OpenAI also brought Sora 2, GPT-5 Pro, and a smaller voice model to the API. Mattel was shown using Sora 2 to prototype toy designs, while AgentKit offered developers a drag-and-drop interface for assembling agents.

2. Apps turn ChatGPT into a personalized front door

The principal launch was apps inside ChatGPT: users can invoke Expedia, Zillow, Figma, Target, or Spotify without leaving the conversation. Casey tested the concept by having ChatGPT build a playlist directly in his Spotify account.
These are embedded experiences, not merely ChatGPT browsing an external website. In the Zillow demonstration, listings appeared inside ChatGPT after a user described a Pittsburgh move and specified bedrooms, bathrooms, price, and a yard.
Kevin extended the proposition: ChatGPT could inspect prior conversations, infer where someone works and what they enjoy, then recommend suitable neighborhoods. Casey called safe sharing of that accumulated context “the real promise” behind becoming “the new homepage for the web.”
Kevin assumes commerce will eventually monetize the layer: if a user buys from Target inside ChatGPT, OpenAI could take a cut. The opportunity is to sit at the start of users’ web activity and potentially monetize transactions while giving developers access to an audience of 800 million weekly users.

3. Facebook’s platform boom supplies the cautionary precedent

Facebook’s early-2010s platform let developers use profile interests, contact information, friends, and even friends’ contact information. That abundance became “a bonanza for developers” and supported businesses such as FarmVille.
Facebook collected 30% of Facebook Credits revenue, and Zynga alone once represented 12% of Facebook’s total revenue. The arrangement worked commercially before scrutiny of Cambridge Analytica exposed how permissive the underlying privacy regime had been.
Casey stressed that Cambridge Analytica’s ability to swing the 2016 election was “dramatically overstated”; the scandal’s lasting significance was the attention it brought to Facebook’s loose rules. Facebook had actually begun restricting the platform years earlier because it suspected such misuse might someday happen.
ChatGPT may be riskier, Kevin argued, because users disclose therapy conversations, personal worries, and requests for intimate advice. Casey’s comic nightmare was a generated birthday card revealing what he had discussed in therapy: “I would have so much egg on my face.”

4. “Minimum necessary” sharing still asks users for substantial trust

ChatGPT head Nick Turley said apps would receive the “minimum necessary amount of information needed to make the transaction.” A Spotify prompt might therefore transmit only the requested party-playlist instructions rather than a user’s full history.
Casey’s reservation is that “minimum necessary” hand-waves the difficult details, especially once ChatGPT’s memory becomes useful to developers. He believes OpenAI intends to be careful, but its position currently amounts to: “we’re not gonna do anything bad.”
Kevin found the risk explicit when connecting Canva: a warning said attackers may attempt to use ChatGPT to access app data and that conversations and memories might be shared with the developer. Waiting for “privacy-adventurous people” to test the system first, Casey concluded, is entirely rational.

5. Distribution neutrality will collide with platform economics

Kevin asked whether ChatGPT might favor integrated Zillow inventory over Redfin results. The question is commercially decisive: becoming the web’s front door gives OpenAI power not only to answer queries, but to route demand toward selected partners.
Altman replied that user trust is essential: “If we break that or take payment for something we shouldn’t have instead of showing you what we think is best, that would clearly destroy that relationship very fast.”
Brockman introduced the unresolved tension, noting that OpenAI does not always know what the best product is and that serving users contains “a lot of nuance.” Casey translated: an angel says “Don’t be evil,” while the devil says the space is nuanced.
The hosts nevertheless read the launch as an unusually direct statement of intent. OpenAI wants ChatGPT to become an operating system or WeChat-like super app that can “take the rest of the internet and shove it inside what we’re doing.”

6. OpenAI’s hardware tease produced no investable information

Kevin expected the Sam Altman–Jony Ive fireside chat to reveal something about their planned hardware. Instead, the hosts heard unfinished, “GPT-2 level” sentences and learned nothing about the device, its specifications, or how it would be built.
Ive said the clues already existed and that a vision would emerge by remaining “very curious and light on our feet.” Casey’s verdict: “ship it or zip it” — provide specifications or stop staging discussions about the nature of design.
Kevin called the session “the closest thing I’ve ever seen to human-generated slop,” while Casey summarized it as “too many tokens.” For all the surrounding ambition, the hardware program remained rhetoric rather than a disclosed product thesis.

7. AMD gives OpenAI capacity, equity upside, and a second ecosystem

Kevin described AMD as the distant number two in AI GPUs — “the Pepsi of chips” beside Nvidia. OpenAI’s multibillion-dollar agreement targets 6 gigawatts of AMD capacity, a little over half the 10 gigawatts in its Nvidia arrangement.
Deployment is scheduled to begin with AMD’s newest chip in 2026. When purchases reach a gigawatt of compute, OpenAI receives rights to acquire extremely cheap AMD shares, potentially amounting to 10% of the company at a penny per share.
AMD secures future demand; OpenAI receives infrastructure plus what Kevin characterized as a stock rebate. The commitment also tells AMD that if it builds sufficiently capable GPUs, “we will buy them from you,” supporting greater production.
The deeper partnership covers software. Nvidia’s CUDA is considered much stronger than AMD’s ROCm for model training, so OpenAI gains an incentive to improve ROCm and potentially make AMD hardware more compelling to other AI developers.

8. A trillion-dollar buildout depends on unprecedented financing

Combining AMD, Nvidia, and other announced arrangements yields what Kevin called a “trillion dollar GPU empire.” OpenAI’s plan is not the once-discussed $7 trillion chip project, but it does involve roughly $1 trillion of planned spending over several years.
Diversification is partly necessity: OpenAI says it needs more GPUs than any single supplier can provide. Brockman’s striking disclosure was, “You don’t even know the products that we haven’t released because we do not have the compute to power them.”
OpenAI does not have $1 trillion in cash. Its proposed funding mix includes continuing equity raises, debt, vendor financing such as Nvidia’s, and expanding revenue from ChatGPT and other products; Altman also suggested that new financial instruments might be required.
Casey identified “novel financial instruments” as a classic moment to become nervous. Kevin’s concern is circularity: if a model flops, scaling reaches an endpoint, or recession hits, intertwined purchases, investments, and lending could spread failure across the AI industry.

9. The bet can work without AGI, but expectations remain transformational

Casey pushed back on the idea that OpenAI must literally deliver AGI: continued services that enterprises believe make workers more productive could support billions in spending. Kevin conceded that unchanged models plus broader adoption might sustain a revenue flywheel.
Even so, Kevin sees an implicit investor assumption that AI will create trillions of dollars in value and prove more like electricity than an incremental productivity improvement. “Machine-god superintelligence” is unnecessary, but transformational economics make today’s prices easier to defend.
Bernstein analyst Stacy Rasgon’s framing captured the distribution of outcomes: Altman “has the power to crash the global economy for a decade or take us all to the promised land.” Jason Furman estimated data centers and information-processing software accounted for 92% of first-half US GDP growth.
Other developers must strike their own supply agreements in a seller’s market; Anthropic, Google, and xAI are also spending or raising billions. Kevin still sees an OpenAI advantage in Altman’s exceptional fundraising record: on his ability to keep raising money, “I would feel uncomfortable betting against that.”

10. The GPU trade is already becoming an infrastructure trade

Casey remains only partially “chip-pilled”: despite unprecedented numbers, the system is currently working, and he sees “capitalism proceeding apace.” A major data-center disaster, he joked but meant seriously, would make the subject feel more urgent.
Kevin’s next step is becoming “infrastructure-pilled.” The AMD and Nvidia deals total 16 gigawatts. Casey said a gigawatt of energy is about what a nuclear reactor produces and estimated that OpenAI’s infrastructure commitment would require the equivalent energy of 20 nuclear reactors, making energization separate from purchasing GPUs.
The constraints now include data-center sites, cooling systems, electricity, grid fuel, electricians, and specialist labor. Kevin heard traveling data-center electricians compared with fracking-boom workers who went to North Dakota and earned “gobs and gobs of money.”
He expects AI companies to strike deals across natural gas, oil, power supplies, cooling, and the literal energy required to operate the data centers. The investable chain therefore extends well beyond semiconductor vendors into everything required to install, connect, and continuously operate the clusters.

11. Sora 2’s creative pull arrived alongside an immediate rights fight

Sora 2 reached number one among free US App Store downloads, while studios, talent agencies, and creators objected to generations involving copyrighted material. OpenAI changed some outputs, and MrBeast and Casey Neistat raised concerns about the threat to creator revenue.
Neistat embodied the tension: he worried about AI slop’s effect on his livelihood while using Sora creatively in the same video. The tool can divert attention from YouTube, Netflix, or theaters even before it becomes capable of replacing their production quality.
Katie Notopoulos had drafted an article arguing that an AI-video social feed “stinks.” After receiving access, she scrapped it: “This really was the first AI experience where I was like, ‘Oh, I love this. I’m having fun.’”
The conversion came through personalized absurdity, including friends represented with impossible bodies or suffering embarrassing mishaps. Sora was compelling not as anonymous synthetic entertainment, but as a machine for making “videos of my friends doing embarrassing things.”

12. Guardrails are porous, and social use quickly exposes asymmetric harm

Katie’s attempts to reproduce “Stone Cold” Steve Austin’s entrance for Casey initially triggered sexual-content and likeness restrictions. Removing the wrestler’s name while describing the black T-shirt, jorts, breaking glass, guitar music, and beer-chugging produced the intended “Casey 3:16” scene.
The inconsistencies invite adversarial prompting. Katie observed that Sora might reject Hitler while accepting a mustached figure in a World War II uniform speaking with a German accent; Casey joked that Garfield appeared better protected than Hitler after a lasagna interrogation was blocked.
After roughly a week, Katie saw the community becoming a teenage-boy ecosystem filled with homophobic ridicule, including a genre premised on Jake Paul being gay. Kevin pointed to cameo restrictions and user-written boundaries, but expected school bullying to test those controls.
Katie saw “basically no women” on the app. Her explanation was not merely male-skewed early adoption: women more readily understand the potential downside, while Casey noted there is no clear compensating upside.

13. Sora may shrink into group chats — unless the product catches up

Katie’s enthusiasm was already fading: after spending hours generating clips, she felt “kind of over it” and had little interest in synthetic content that did not feature friends. Kevin likewise saw his usage plummet after the novelty wore off.
Kevin’s thesis is that Sora is “a perfect tool for the group chat,” not necessarily a standalone network. Users might generate a private joke, download it, and share it elsewhere, while a broader cultural backlash keeps public AI video from mainstream adoption for now.
Casey agreed adoption was limited but rejected permanence of the limitation. Current shortcomings — inaccurate voices, inability to change outfits, and 10-second clips — are technical or product constraints, not evidence that demand has reached its ceiling.
His analogy was ChatGPT itself: many people tried it once, saw little purpose, then returned as it improved until weekly usage reached 800 million. His closing call was categorical: “Do not sleep on AI-generated video. I truly do think it is here to stay.”

OpenAI's Multi-Billion Deal with AMD & Polymarket, Vercel and Supabase Raise Mega Rounds (Preview) OpenAI Astride the World, Infrastructure Buildouts and Boundless Ambitions, More on Sora and Creation