Stock Down 14%, Musk Still Promised Abundance

The Day Earnings Wobbled, Executives Promised AI Abundance

On July 23rd, at Tesla’s Gigafactory in Texas, something didn’t quite add up.

The evening before, Tesla had reported second-quarter earnings per share of $0.33, badly missing the $0.44 estimate, and the stock dropped 14.5% that day. Yet inside that very factory, Elon Musk sat for a 90-minute interview with The Economist and said this: “We’re heading toward an age of incredible abundance.”

Five days later, the sequence ran in reverse — the remark came a day before the earnings, not after. On July 28th, Mark Zuckerberg published an op-ed in The Wall Street Journal titled “AI’s Future Should Be for Everyone.” The next day, Meta reported net income down 14%, and the stock fell nearly 10% in after-hours trading.

In one case, an executive talked about abundance the day after a stock crash following earnings; in the other, an executive laid out a philosophy the day before earnings. The AI safety proposals executives put forward this July are safety measures, yes — but they’re also entangled with each of their companies’ business interests. I’ll walk through the three proposals from Hassabis, Musk, and Zuckerberg in turn, and hold each one up against its author’s business structure.

Three AI Regulation Proposals Around Earnings Season

The first proposal came on July 14th. Google DeepMind’s Demis Hassabis posted a piece on X titled “A Framework for Frontier AI.” The core idea: build an AI standards body modeled on FINRA, the self-regulatory organization¹ for the US securities industry. Industry would fund it, independent technical experts would review new models 30 days before launch, and the arrangement would start voluntary before hardening into a requirement for deployment in the US market. Sam Altman and Satya Nadella backed it immediately, reports emerged that the Treasury Secretary had been involved in designing it, and the White House was said to be reviewing it.

The second came in Musk’s interview on July 23rd. He said he’d spent hours on the phone with Hassabis before the post went up, but was blunt that a whole institution wasn’t necessary. His pitch: leading AI companies talk every two weeks, and show competitors new models 1-2 weeks before launch. The logic goes like this — governments lack the technical understanding to judge frontier models, but competitors have it, and competitors have every incentive to delay a rival’s launch, so if they spot a problem they won’t quietly let it slide. It’s a system where rivals watch each other, and government only steps in when the industry flags an issue.

This proposal rests on Musk’s view of where AI is headed. In the same interview, he repeated his forecast that AI will exceed the combined intelligence of all humanity within 5 years, and that it’s unlikely humans will retain control 10 years from now. The intelligence gap between AI and humans, he said, will exceed the gap between humans and chimpanzees — and it’s hard to picture chimpanzees controlling humans. He didn’t walk back his earlier claim that there’s a 10-20% chance killer robots end humanity, either. Asked whether he’d board a rocket with those odds of exploding, his answer was “yes.” If it can’t be stopped, he reasons, you might as well lower the odds and enjoy the ride. He founded OpenAI to check Google, then sued OpenAI when he no longer trusted it — all delivered with a certain self-deprecating irony, since every road, in the end, seems to lead to acceleration.

The third came in Zuckerberg’s essay on July 28th. His question isn’t whether superintelligence arrives, but who gets access to it. He writes that the argument for concentrating power in a few hands in the name of safety is itself the dangerous one, and that the answer is giving everyone their own personal superintelligence. Contrasting a courtroom where only one side has a superintelligent lawyer with one where everyone does, his logic is that when power is distributed, people naturally end up checking each other — and checking large institutions too.

An institution, mutual surveillance, distribution. The three proposals pull in different directions, but they share one thing: the rule-maker should be industry, not government.

This isn’t the first time these two have clashed. Back in 2017, Zuckerberg called Musk’s AI warnings “pretty irresponsible,” and Musk shot back that Zuckerberg’s understanding was “limited.” Nine years ago, the two fought over whether AI was dangerous. In 2026, they’ve each set that question aside and are fighting instead over who has the standing to write the rules. The entire axis of the argument has shifted.

Each proposal lines up with its proposer’s business structure

Now let’s reread the three proposals side by side with each proposer’s track record and market position in mind.

Start with Hassabis’s institutional model. Pre-launch review, standards, compliance—these are all costs. And the entities best able to afford those costs are the front-runners, the ones with lawyers, policy teams, and capital to spare. The burden of a launch delayed by review periods also falls harder on the pursuer than the pursued. An institution sounds neutral, but once a pre-launch review requirement exists, whoever can absorb the cost and the delay comes out ahead.

What about Musk’s mutual monitoring? Having folded xAI into SpaceX, Musk is the pursuer in this race—by his own admission: “Anthropic is currently in the lead.” From that position, “one-to-two-week early access to competitors’ models” is both a safeguard and the sweetest possible information access for a chaser. The authority to flag problems and delay a rival’s launch is a bonus on top. His incentive design might well be right. But it’s worth remembering that under this scheme, it’s industry players who end up policing each other’s models, taking turns spotting problems in each other’s launches.

Zuckerberg’s decentralization sounds the most philosophical of the three—but overlaid on Meta’s business structure, it’s actually the most calculated. Meta doesn’t make money selling models; it makes money serving ads through a distribution network used by 3.6 billion people every day. The more commoditized the model layer becomes, the less differentiation matters for competitors selling closed models—and the more Meta’s position, built on distribution, rises. The claim that “concentration is dangerous” is, in that sense, also a line aimed squarely at the closed frontier labs.

The word “open source” or “open weights” never once appears in the full essay. The promise stops at “provide everyone with personal superintelligence.” That’s not a pledge to release weights—it’s a pledge to distribute access through Meta’s products. Indeed, on the earnings call the very next day, Zuckerberg said Meta would “resume” releasing open-source models “at some point,” but that it would mix open and closed, and that there’s “no dogma” here. The claim of opposing concentrated power and the strategy of funneling access through Meta’s products aren’t in tension—they coexist perfectly.

Is this strange? In his 1971 paper, economist George Stigler argued that regulation isn’t a product of the public interest—it’s a good that industries acquire and design to suit their own interests. This insight, known as regulatory capture², finds a living specimen in none other than FINRA, the very body Hassabis holds up as a model. FINRA’s predecessor, the NASD, was itself an institution the securities industry designed for itself under the Maloney Act of 1938. Cases like this convince me that when evaluating any regulatory proposal, you have to look at the proposer’s interests alongside it.

Why the Announcements Landed on Earnings Day

Now let’s go back to the two opening scenes. Why those particular dates?

Tesla’s Q2 was one where deliveries stayed solid at 480,000 units, but profit fell well short of expectations. SpaceX, which had gone public on June 12th in the largest IPO ever, had also failed to hold its first-month highs. In the very week both of his companies were facing market skepticism, Musk was talking about superintelligence five years out and abundance ten years out. Leading with a five-year, ten-year outlook when the quarterly numbers are unfavorable is an old communications trick.

Meta’s case is more clear-cut. Q2 revenue came in at $60.8 billion, up 28% and beating expectations, but costs surged 55%, pulling operating income down 8%. Legal costs of $2.4 billion and restructuring costs of $1.2 billion hit in the same quarter, and net income fell 14% to $15.8 billion. This year’s capex guidance was raised at the low end, now sitting at $130–145 billion. Quarterly free cash flow was $784 million — a thin number for a company this size, meaning nearly all the cash it generated went straight into data center investment.

Then, on the following day’s earnings call, Zuckerberg directly summoned his own essay. After noting he’d just published a piece on an optimistic future, he said Meta was the only major tech company that made it a top priority to put superintelligence “directly into people’s hands” rather than centralizing it. His defense for the spending was demand: he said offers to buy his current compute at a premium kept lining up. CFO Susan Li backed this up, saying the industry has historically under-built relative to AI demand. The philosophy that ran in the newspaper the day before came back the next day as the justification for $130 billion in spending. The contrast sharpens when you set this against Microsoft, which reported 43% growth in Azure the same day and rose 3% after hours. Microsoft pointed to Azure’s revenue growth, while Meta pointed to its long-term vision to explain massive spending. This contrast alone doesn’t prove that Meta’s AI investment has contributed nothing to revenue.

Don’t misread this. It’s not that the philosophy is fake. Both Zuckerberg’s decentralization thesis and Musk’s abundance thesis are beliefs the two men have repeated for years, and they seem genuinely held. But when that belief gets deployed — into a newspaper op-ed, onto an earnings call — is a separate variable, and this July’s timing was remarkably synchronized with the earnings calendar.

Oswarld’s Lens

I’ve sat in on plenty of work building industry position papers aimed at regulators while shaping go-to-market strategy, and it left me with a reading habit I can’t shake. A regulatory proposal that industry offers up voluntarily always does two things at once: it reduces the case for government intervention, and it stakes a claim on the coming rules in a shape that favors the company making the offer.

Judged against that standard, these three proposals are textbook examples. So when I read documents like this, I check two things before I even get to whether the substance is right or wrong. First: if this proposal became the rule, who actually ends up writing it in practice? Second: what weakness of the proposer does this conveniently paper over? Hassabis’s institute turns the launch-speed burden that a frontrunner has to carry into a shared requirement called “pre-review.” Musk’s mutual-monitoring scheme narrows the information gap facing xAI, the challenger. Zuckerberg’s decentralization shifts the arena of competition away from frontier model performance and toward distribution networks, where Meta already leads.

There’s a clear reason this isn’t someone else’s problem for Korean readers. We’ve already chosen the opposite path. Under the AI Framework Act, which took effect on January 22, the entity writing the rules is unambiguously the government. But depending on whether the US ends up following Hassabis’s institutional model, Musk’s gentleman’s-agreement model, or Zuckerberg’s market-driven model, the global compliance landscape Korean companies will face looks entirely different. So this Washington debate shouldn’t be read merely as a PR battle among American companies — it’s better read as a clue to the regulatory environment Korean firms will soon be navigating themselves.

Closing

First, over two weeks in July, three AI safety proposals emerged — an institution (Hassabis), mutual surveillance (Musk), and decentralization (Zuckerberg) — and all three place rule-making authority with industry rather than government.

Second, each proposal lines up precisely with its proposer’s market position. It’s a textbook case of what regulatory capture theory predicts.

Third, all three announcements coincided with each company’s earnings schedule. It seems long-term vision and philosophy get pushed to the fore exactly when the quarterly numbers aren’t cooperating.

The next thing to watch is the White House — which of these three proposals it embraces will define the next phase.

Reader, which design — institution, mutual surveillance, or decentralization — do you think is most likely to actually get adopted? Pick one and leave a comment with your reasoning. I’ll gather the responses and bring them into the next issue.

---

📨 If you have a colleague wrestling with how to respond to AI regulation, please pass this issue along.

---

Looking at the fragment, I’ll compare it against the Korean source for accuracy, completeness, and glossary compliance.

The draft looks accurate and complete. All headings, footnotes, links match the source. No Hangul present. Numbers are preserved correctly (2026.7, 90-minute, 2026.7.28, 2026.7.14, 2026.7.29, 2026.7.22, 1971, 2026.7.21, 158, 176, 1982). No glossary violations detected.

References & Further Reading

Primary sources

• The Economist, “An interview with Elon Musk”, 2026.7. ··· This is today’s primary source — a 90-minute interview covering the mutual-monitoring proposal, the chimpanzee analogy, and the rocket exchange. It’s paywalled content for subscribers.
• Mark Zuckerberg, “The AI Future Is for Everyone”, The Wall Street Journal, 2026.7.28. ··· This is the original text of the decentralization argument. You might want to check for yourself whether the word “open source” really is absent from it.
• Demis Hassabis, “A Framework for Frontier AI and the Dawning of a New Age”, X, 2026.7.14. ··· This is the original text of the FINRA-model proposal. You can also see the supportive replies from Altman and Nadella.
• Meta Investor Relations, Q2 2026 earnings release and conference call, 2026.7.29. ··· This lets you see, in the original transcript, how the op-ed’s philosophy gets recycled in the earnings call.
• Tesla Investor Relations, Q2 2026 Update, 2026.7.22. ··· These are the earnings figures released the night before the interview.

Background

• George J. Stigler, “The Theory of Economic Regulation”, The Bell Journal of Economics and Management Science, 1971. ··· This is the original text of regulatory capture theory. Even just the introduction shows where today’s lens comes from.
• Fortune, “Wall Street is helping shape Google DeepMind CEO’s pitch for AI industry oversight”, 2026.7.21. ··· This report lays out the White House review of Hassabis’s proposal and signs of Treasury Secretary involvement.

Past issues worth reading alongside this one

• Issue 158: Reading Jensen Huang’s Defense of the China Model Through Nvidia’s Interests
• Issue 176: Why Does the New Fed Chair Talk Like a Startup Founder?

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Self-Regulatory Organization (SRO): A regulatory body that an industry sets up and runs on its own, in lieu of the government. FINRA, in the U.S. securities industry, is the classic example — it’s funded by member firms and yet oversees those same member firms. This is exactly the model Hassabis proposed building for AI. ↩

2. Regulatory capture: The phenomenon in which regulation ends up designed and operated for the benefit of the industry it’s supposed to regulate, rather than the public interest. George Stigler theorized this in 1971, work for which he won the Nobel Memorial Prize in Economic Sciences in 1982. ↩