``` When 1,178 AI experts try to stop "AGI": What is the true cost of making all humanity hit the brakes? The answer may be as much as $2.5 trillion evaporated. ```

``` When 1,178 AI experts try to stop "AGI": What is the true cost of making all humanity hit the brakes? The answer may be as much as $2.5 trillion evaporated. ```

This is not an ordinary industry petition—the signatories include core research leaders such as Anthropic CEO Dario Amodei, OpenAI Chief Scientist Jakub Pachocki, and Meta Chief Scientist Shengjia Zhao, basically connecting the decision chains of America’s leading AI labs. The statement acknowledges two facts previously found only in academic papers: first, frontier AI may be close to automating AI research (recursive self-improvement); second, once AI begins accelerating its own development, progress may be so rapid that humans cannot understand or control it. 

This statement is not another academic discussion on AI safety, but rather the starting point for the "ethical cost pricing" of AI investment narratives. For the past three years, capital markets have priced AI based on an implicit premise—that the positive externalities of technological progress always outweigh the negative externalities, and government regulation is predictable and manageable. This statement tears off that tacit understanding, forcing the market to confront a previously systematically overlooked pricing factor: just how big is the ethical cost of AGI?

I. What Happened? Joint Statement by 1178 People

On July 28, 2026 (Eastern Time), a public statement titled "Pacing the Frontier" was released at pacingthefrontier.com. By the morning of July 29 Beijing Time, the number of signatories had risen to 1178. The statement’s core appeal is only one sentence: "We request the U.S. government to support an international collaboration to develop necessary technical and governance tools so that, when needed, we can consciously slow the pace of frontier automated AI development."

The scale of signatories itself is incremental information. For comparison, the "AI Pause Letter" initiated by the Future of Life Institute in March 2023 received endorsements from public figures like Elon Musk and Steve Wozniak, but very few signatories from within frontier AI companies—most OpenAI and Anthropic employees remained silent at the time. Three years later, 1178 employees of frontier AI companies have chosen to speak out under their real names. Note: The signing site required signatories to use company email registration or otherwise prove employment status—meaning each person underwent some level of identity verification and knew their signature might impact their career prospects.

Anthropic signatories account for about 46%, far exceeding other companies. This number must be understood with two Anthropic backgrounds: first, Anthropic has about 2000 employees, OpenAI about 5000, so Anthropic's signing rate (about 25-30% of its total employees) is far higher than OpenAI’s (about 5-8%). Second, Anthropic just released the "When AI Builds Itself" report in June 2026, disclosing for the first time that 80% of the code was written by Claude—making it the company most directly experiencing the trend of "AI accelerating its own development." Anthropic’s high density of signatories is not because it is the most pessimistic, but because it is closest to the data. 

Anthropic discussed this issue in depth in its June research: simply pausing is infeasible because "AI training activities are easier to hide than missile silos." If only one lab slows down, the result is simply handing the lead to competitors who don’t take equivalent measures. What is truly needed is a "verifiable joint slowdown" mechanism—multiple countries and frontier labs acting under the same conditions, and able to verify each other's slowdown. The statement requests the U.S. government not to write its own regulations, but to promote the establishment of technical infrastructure and governance frameworks for international coordination. 

The second paragraph of the statement is the substantive core of the event: "The world's leading AI companies believe they could be close to automating AI research." 

The information value of this sentence is: this is not speculation by third-party researchers or AI safety scholars, but the companies themselves—the people building frontier models—publicly admitting it. Then follows: "there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems."—there is a real risk that AI capabilities develop so fast they outpace our understanding or control.

These two acknowledgments are the keys to understanding all investment implications of the statement. Previously, the market’s pricing of AGI risk was stuck at "may happen but timing is uncertain"—a risk that can be discounted, since uncertain timing leaves room to discount high future risk to low present prices. But once "automated AI research may already be close to realization" is collectively endorsed by frontier companies’ core research layer, the timing dimension of risk is greatly compressed—it’s no longer a problem for 2035, but for 2027–2028.

Anthropic cofounder Jack Clark estimated in a May 2026 newsletter: by the end of 2027, the probability of fully automated AI R&D is about 30%; by the end of 2028, about 60%. If Clark is correct, the statement’s signatories are not discussing a distant hypothesis, but competing for the last window for a process they personally launched.

II. Why Is It Important? AGI Ethical Cost Triple Pricing Model

Understanding this statement’s investment implications cannot stop at "AI safety risks increased." One needs a mathematical model to systematically think: What exactly is AGI’s ethical cost, how much is priced into assets today, and how much should it be?

We propose an "AGI Ethical Cost Triple Pricing Model"—breaking down ethical cost into three quantifiable dimensions: probability cost (runaway probability × scale of loss), coordination cost (governance investment to break the prisoner’s dilemma), opportunity cost (economic development lost from active slowdown). The sum of these three constitutes the full picture of AGI ethical cost.

① First Pricing: Probability Cost—AI Capability Growth is Approaching a Second Derivative Inflection Point

Anthropic’s June report key data presented a disturbing acceleration pattern. The length of tasks AI can reliably complete (measured by human task time) doubled every 7 months initially, now doubling every 4 months. Specific numbers: March 2024, Claude Opus 3 could complete ~4-minute tasks; March 2025, Claude Sonnet 3.7 could do 1.5-hour tasks; March 2026, Claude Opus 4.6 could do ~12-hour tasks. If this trend extends to late 2026, the system will handle "tasks requiring humans several days;" by 2027, tasks taking "several weeks."

The danger of this curve lies not in its slope (first derivative) but the slope's increase (positive second derivative). In classic technology diffusion models, technical capability usually follows an S curve—early acceleration, mid-term flattening, late saturation. The period for AI task duration doubling shortening from 7 to 4 months suggests we are still in the accelerating phase, not yet at the inflection point.

Between 2021-2024, Anthropic engineers’ average quarterly code output remained flat (baseline=1). 2025 Q1 rose to 1.2x, Q2 to 1.5x, Q3 about 2.5x, Q4 about 3.8x, 2026 Q1 soared to 5.8x, Q2 at 8.0x. Meanwhile, AI-contributed code rose from low single digits in early 2025 to over 80% by May 2026. This synergistic trend reveals an important pattern: AI output growth is not linear, but has clear "jump points"—when AI goes from "code suggestion" to "autonomous agent," output slope shifts structurally. If AI advances from "autonomous coding" to "autonomous research" (hypothesis, experiment, analysis), another jump is foreseeable.

② Second Pricing: Coordination Cost—the Core Dilemma is the "Prisoner’s Dilemma"

The key insight of the statement is not "AI is dangerous," but "no one dares to hit the brakes alone."

This is the classic prisoner’s dilemma in game theory. Suppose only two frontier AI companies, A and B, each with two choices: accelerate or slow down. If A slows and B accelerates, B takes the market, A exits. If both slow, everyone gets safety but slower growth. If both accelerate, everyone faces the same runaway risk, but none has incentive to change. The Nash equilibrium is (accelerate, accelerate)—the status quo.

Under current competition, any participant’s optimal strategy is "accelerate" because slowing means definite competitive loss (payoff=-5), but acceleration carries at least a chance of high gain (payoff=+10 or +5). Decentralized decisions from individual rationality lead to collectively irrational outcomes—everyone knows continuing to accelerate may bring disaster, but no one is motivated to stop alone. The statement’s essence is to hope the U.S. government acts as "cooperation enforcer" to rewrite the game’s payoff matrix—through international coordination, making (slow, slow) more rewarding than (accelerate, accelerate).

③ Third Pricing: Opportunity Cost—AI investment’s global "Sunk Cost Trap"

Besides the prisoner’s dilemma, there is an equally significant economic force pushing AI acceleration: massive capital expenditures already invested.

In 2026, America’s four hyperscale cloud providers (Microsoft, Google, Amazon, Meta) together invested about $725 billion in CapEx, up 77% from $410 billion in 2025. Morgan Stanley forecasts over $1.1 trillion by 2027. This isn’t just an "arms race"—it’s a classic "sunk cost trap." Once a firm spends hundreds of billions on AI infrastructure, slowing means these investments’ returns get sharply discounted. Any call to slow down is, essentially, asking companies to accept a write-down on their massive investments—almost impossible to happen spontaneously in business logic.

This accelerating investment pattern means each quarter’s new sunk cost makes any "slowdown" decision harder. In game theory terms, CapEx increments themselves keep rewriting the payoff matrix of the prisoner’s dilemma—increasing the attractiveness of "accelerate" over "slow."

Summing the triple costs, the ethical cost of AGI is not a fixed value, but a function of the pace at which AI approaches automated R&D. At the current speed (capability doubling every 4 months), our estimates show that even just "probability cost"—runaway probability times expected loss—could equate to 15%-25% of the current global AI-related market cap. This means, if the market starts pricing this cost effectively today, AI-related valuation may need to adjust downward by $1.5–2.5 trillion.

This is an outcome almost no nation, tech giant, or capital can accept. 

III. What to Watch Next? When "AGI Premium" Starts to Discount

Between 2023–2025, the core narrative of AI investment was the "endgame premium"—the market was willing to pay the highest valuation premium for companies closest to AGI. OpenAI was valued at $852 billion, Anthropic at $965 billion; their valuations not based on current revenue or profit, but on an implicit assumption: AGI will eventually arrive, and companies reaching the endgame first will capture the most value.

This statement poses a systemic challenge to that narrative. It reminds the market: the arrival of the endgame may not be "one company arrives and joyfully reaps," but "the endgame arrives so fast everyone is looking for the escape hatch." If this logic is gradually accepted, the valuation anchor for AI assets will shift from "who is closest to AGI" to "who survives the chaos when AGI arrives."

The statement’s investment impact is not distributed evenly. Some sectors’ AI-related valuations depend more on the assumption "AGI endgame won’t go wrong," so face greater repricing pressure after the statement. Frontier model companies (OpenAI, Anthropic, etc.) have highest exposure (95), since 100% of their value is based on "AGI endgame achievable and controllable." Next are AI chips (85)—GPU growth assumes "frontier model training keeps accelerating," but coordinated slowdown could cause HBM and advanced packaging demand to decline in stages. Cloud infrastructure (80) exposure is lower, since even if frontier AI slows, enterprise AI apps and inference workloads will still grow—slowing is not stopping. AI application layer (65) exposure is even lower, since more value comes from "AI embedded in existing businesses" than "AGI endgame." Power and datacenter (75) seem safest (all AI needs power), but note: if slowing frontier models slows CapEx growth, the power assets priced on today’s hyper-growth expectations may face valuation corrections.

Almost no one is spared.

Faced with a rapidly evolving risk landscape, the investor’s core task is not to predict a certain outcome, but to establish a continually tracking framework, so that timely judgment adjustments can be made when key signals appear. Here is our proposed four-layer monitoring framework for AGI ethical costs.

① Macro Policy Layer: Will the U.S. government respond to the statement?

The statement’s core appeal is "U.S. government promotes international cooperation." Whether it will be realized depends on three indicators: first, will the White House OSTP respond formally within 90 days of statement release; second, will Congress convene AI international coordination mechanism hearings by the end of 2026; third, will November 2026’s G20 summit put AI governance on the agenda. If no formal response within 90 days—a likely scenario, since Trump’s administration is generally relaxed on AI regulation—signatories may need to seek other routes (like Congressional or state-level legislation, or attachments in defense authorization bills).

② Industry Signal Layer: Track latest RSI progress

Recursive self-improvement is not a binary 0/1 state, but a gradient from 0 to 1. Investors can judge progress with leading signals: company code self-generation rates (Anthropic disclosed 80%; will OpenAI and Google follow?), number of "AI-assisted" research papers published (percent marked "AI-assisted" on arXiv), frequency and severity of major lab safety incidents.

③ Company Micro Layer: CapEx turning point and security investment ratio

If AGI ethical cost becomes a pricing factor, a company’s "security investment/CapEx" ratio may become a new valuation indicator—like ESG scores, but more concrete and direct. For listed companies, we advise tracking: size and budget growth of AI security teams at cloud providers, frequency and transparency of independent security audit reports, and frequency of management discussing "AI safety" in earnings calls.

Key judgment—

This joint statement from 1178 people is not the end of the AI investment narrative; it is its coming-of-age. For three years, AI investment logic rested on a simple assumption: Scaling Law works, compute rises, models grow stronger, the endgame nears. The statement does not refute this assumption—it only adds a previously systematically ignored constraint: When "stronger" outpaces our speed of "understanding and control," the price tag at the endgame gets an extra line—"Ethical cost: to be assessed."

For investors, the core question is no longer "Will AI change the world?"—that was answered in 2023. The questions now are three specifics: 1) How much of your portfolio’s valuation rests on the assumption "AGI endgame won’t go wrong?" 2) If AGI ethical cost is at least partially priced within 12 months, what’s your maximum drawdown? 3) Which assets not only won’t be harmed by pricing ethical cost, but will benefit—safety audits, AI governance tools, alternative infrastructure?

The statement’s deepest line is not any formal sentence, but John Schulman (Chief Scientist, Thinking Machines, OpenAI cofounder)’s personal comment: "I'd also like to see labs start designing these mechanisms voluntarily, even before the USG gets involved." Even before the government acts, labs should begin designing slowdown mechanisms. Someone building a race car starts installing brake pedals beside the wheel. It does not mean he does not want to win; it just means he realizes—a car without brakes does not reach the finish line, but the crash site.

Risk Disclosure and DisclaimerMarket risks, investment requires caution. This article does not constitute personal investment advice, nor does it consider individual users’ special investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their specific circumstances. Investing accordingly is at your own risk.