(Preview) OpenAI’s Enterprise Pivot, The Rise of Agents and Bubble Counterpoints, Nvidia Changes Its Inference Story

Summary

• OpenAI’s reported reset toward coding and business users is less a retreat than a correction toward the market most willing to pay for productivity. Andrew notes that OpenAI has spent the last 3 years pursuing consumer subscriptions. Ben Thompson’s Dropbox analogy carries the logic: an exceptional consumer product eventually had to rebuild for enterprise permissions and authentication because “consumers don’t pay for productivity apps,” while employers readily subscribe when software makes salaried workers more productive.
• The immediate strategic danger is Anthropic, which Ben says is “blowing up in the enterprise.” He cites an apparent run-rate jump “from $14 billion in January to $19 billion” now, while acknowledging that private-company figures are difficult to validate. Ben also says OpenAI has much more compute than Anthropic, with that compute serving consumers; Andrew sees a direct 12-to-24-month opportunity, while Ben warns enterprises could standardize on Anthropic and lock OpenAI out.
• The shift does not mean OpenAI is abandoning consumers, despite the dramatic reactions to its all-hands message about avoiding “side quests.” Sora is an example of a side project that could be deprioritized, but Sam Altman explicitly denied that the hardware effort was shutting down: “Quite the opposite. I think you will love what the team is building.”
• ChatGPT’s enormous consumer reach is simultaneously OpenAI’s strategic advantage and “biggest problem,” because supporting that audience consumes vast compute without a proven matching revenue engine. Advertising could unlock the larger consumer market, but OpenAI would be building an ads system from scratch while Google and Meta already possess mature infrastructure enhanced by AI; enterprise subscriptions offer a clearer near-term path to cash.
• Enterprise AI adoption is partly an organizational-design problem because chatbots require employees to choose to use them, and many workers “are just there to collect a check.” Top-down incentives can produce resentment or superficial compliance: Microsoft’s KPI to integrate Copilot yielded it “everywhere,” including places where Ben says it was unwanted or worked poorly.
• Ben’s counterpoint to the AI-bubble case is that old GPUs appear more economically durable than depreciation skeptics assumed, with their prices actually rising. Major platforms are spending roughly around projected free cash flow rather than becoming broadly overleveraged; debt can rationally match upfront infrastructure costs with returns earned over time, while Oracle is the notable aggressive borrower backed by sharply expanding committed business.
• Apparent Anthropic share gains should be discounted for selection bias, particularly when the evidence comes from Ramp customers. Ramp disproportionately serves startups and highly technical Silicon Valley companies—the same cohort most likely to adopt the “hot Silicon Valley company”—whereas general Fortune 500 companies, Ben says, know OpenAI; after company pushback, he gives OpenAI “the benefit of the doubt.”

Deep dive

1. OpenAI is rediscovering the business model of productivity software

• Andrew frames the reported strategy shift through Fidji Simo’s warning that OpenAI could not afford to be “distracted by side quests.” Leadership was examining what to deprioritize so the company could “nail productivity in general, and particularly productivity on the business front,” with coding and business users becoming the center of gravity. Andrew notes that OpenAI has spent the last 3 years pursuing consumer subscriptions.

• Ben’s historical specimen is Dropbox, once a mind-blowing “USB drive in the sky.” At English schools around Taipei, he used Dropbox, Mac Minis, AppleScript, and nightly jobs to synchronize Keynote curricula across classrooms—an improvised system whose real engine was Dropbox’s effortless promise that a file placed in one folder would appear everywhere.

• Yet Dropbox’s consumer success did not produce the right economics. It spent years rewriting its application around enterprise permissions, authentication, and administration because, outside enthusiasts like Ben, “consumers don’t pay for productivity apps”; enterprises both pay for them and demand the security, support, and updates required for deployment.

• Ben traces the template to Microsoft’s subscription transition under Steve Ballmer around the late 1990s or circa 2000. For a “logical spreadsheet-driven CTO,” recurring payment matches recurring utility: employees remain productive, support stays available, security is maintained, and updates arrive immediately. Consumers, by contrast, usually prefer paying with attention through advertising.

2. Human incentives, not model capability, may cap chatbot adoption

• Ben’s qualification to the productivity thesis is that chatbots require active employee participation. Some workers want to improve their jobs, but many simply “punch in and punch out” and may resent demands to become more productive; Andrew cites finance friends with quarterly meetings where they must explain how they are using AI in their jobs each week.

• The organizational problem is getting hundreds or millions of employees “rowing, broadly speaking, in the same direction.” People optimize toward local KPIs, so Microsoft’s KPI to integrate Copilot predictably produced Copilot everywhere—including places where Ben says it was unwanted and did not work particularly well.

• His model of corporate management is a zigzag: incentives initially move an organization toward its goal, then overshoot as employees maximize the metric, forcing another reorganization and course correction. Ben calls this the “nerd fallacy”: computers are difficult to understand, but once mastered execute even flawed instructions exactly; people and organizational headwinds are much harder to manage.

3. Current AI spending does not yet resemble broad overleverage

• Ben addresses the depreciation argument behind some bubble calls: critics assumed GPUs would be useful for only a couple of years, making reported losses artificially low. “Turns out that doesn’t appear to be the case,” he says, because prices for old GPUs are rising and companies are actually making more money from hardware that has already been depreciated.

• The major platforms’ CapEx broadly tracks projected free cash flow: Meta and Google are near that number, Microsoft is somewhat below it, and Amazon is somewhat above it. Ben’s pushback to alarm over Amazon spending more than free cash flow is straightforward: capital investment consumes cash upfront and earns returns over time, while debt supplies cash upfront and is repaid over time.

• In Ben’s framing, debt is “not a bad thing” but an underused and tax-advantaged tool for technology companies. Oracle is the conspicuous company taking on substantial debt, yet its earnings were “incredible,” and its RPO—committed future business—rose by “like, $50 billion or something like that.” Andrew’s test for a truly bubbly build-out was that all the major companies would be overlevered, not merely spending near internally generated cash.

4. The enterprise focus can coexist with OpenAI’s consumer ambition

• Ben cautions against reading the all-hands report as a wholesale consumer exit. The Sora app was a memorable experiment that “didn’t ultimately matter to the bottom line,” making it a plausible side quest to cut; hardware appears protected after Altman said, “We are not shutting it down. Quite the opposite.”

• Listener Adrian supplies the strongest pushback: OpenAI’s consumer platform could create a larger flywheel as general-purpose models, smartphones, watches, glasses, and other interaction points improve. His internet analogy is that early government and academic adoption eventually gave way to enormous everyday value, with scale enabling targeted advertising that benefits advertisers and users.

• Ben concedes that consumer is generally the larger market and that ChatGPT’s scale should, in theory, support an ad-funded business. The obstacle is execution: “Oh my God, building ads is hard.” Google and Meta already own the machinery and are improving it with AI, whereas OpenAI must build from scratch while supporting the compute-intensive consumer audience. That scale also means, Ben says, OpenAI has much more compute than Anthropic, with that compute serving consumers.

• His counterexample is Microsoft in the 1980s and 1990s: enterprise dominance created the Windows flywheel, effectively delivering the consumer market “for free” and even seeding Microsoft’s later gaming position. Apple showed the opposite risk, nearly failing while charging consumers a premium before enough buyers valued a differentiated computer.

5. Anthropic creates urgency, but its apparent lead is difficult to measure

• The strategic pressure Ben identifies is Anthropic, whose enterprise growth he describes as an exponential curve: “they’re growing, they’re growing, they’re growing, and then holy crap, they’re growing.” He references a roughly $14 billion run rate in January becoming about $19 billion now, while treating the private-company numbers as inherently uncertain.

• Andrew argues OpenAI has direct line of sight to what coding and enterprise products could accomplish over the next 12 to 24 months, especially given Codex’s performance. Ben sees the downside condition clearly: if OpenAI waits, customers could standardize on Anthropic and effectively lock it out of the most reliable subscription market.

• Ben nevertheless retreats from treating Ramp’s spending data as definitive evidence of Anthropic’s lead. Ramp over-indexes toward startups and technical Silicon Valley customers, who also over-index toward Anthropic; general Fortune 500 companies, Ben says, know OpenAI instead. After “very strong pushback” from OpenAI, this is one case where he grants the company the benefit of the doubt.