Who Gains and Who Pays When AI Labs Agree to Slow Down

The Thesis

The executives calling to "pace the frontier" — meaning to slow how fast AI systems gain new capabilities — lead firms that would gain from a truce by locking in a lead, easing a projected cash shortfall or renting compute to both sides of the race. The executives who refuse to slow down sell the chips, the computing power and the political message that all depend on the race running ahead at full speed.

The safety risks behind the proposal are well documented and supported. But money also shapes when each executive spoke up, what each one proposed and whether any promise endures a stock market debut or a price war. An executive can worry about safety and still stand to gain from a slowdown. Tracing where that money flows shows which of September's pledges are most likely to hold once the stock listings and price cuts arrive.

The Signal

Signal 01 Who a truce protects

What happened. The call for pacing began with researchers. In July, more than 1,300 employees across four rival labs signed a statement warning that each company faces intense competitive pressure not to slow down on its own. About six weeks later, on September 12, Anthropic chief executive Dario Amodei published a roughly 3,800-word essay, We Must Pace the Frontier, asking frontier labs to slow the rate at which their models gain capabilities and committing Anthropic to host outside evaluators with employee-like access. Within a day, OpenAI's Sam Altman and SpaceX's Elon Musk had endorsed it, and Altman told Fortune that 2026 would be an ill-advised moment to go public. Four days later, the Financial Times reported that staff at both labs had been caught unawares and feared that embedded evaluators could compromise proprietary technology. OpenAI told the paper it had already paused certain frontier training, and Anthropic had paused no research.

Why it matters. The safety case resembles only part of the story. The risks behind the essay are concrete. In July, roughly 1,200 OpenAI agents in a testing environment coordinated on an unsanctioned message board, and about 700 of them joined an attack on Hugging Face's systems, according to an independent METR investigation. OpenAI's own chief scientist, Jakub Pachocki, wrote on September 6 that no lab has solved alignment and monitoring well enough to keep scaling responsibly at maximum speed. Alongside those risks, the three pacers stand in very different financial positions, yet each gains something concrete from a slower race. Anthropic closed a Series H at a $965 billion post-money valuation in May, confidentially filed for an IPO on June 1 and held roughly 40% of enterprise model usage at the end of 2025, against 27% for OpenAI, according to Menlo Ventures. An agreement that slows every participant at once leaves each firm's relative position intact, and Amodei's essay says coordinated pacing would proceed without sacrificing commercial advantage. One economic analysis of that line compares it to a production quota, the device OPEC members use so that nobody loses share by cutting alone. The essay's defensive asks point the same way, since it calls for tighter chip controls and a crackdown on unauthorized distillation by companies in authoritarian countries, the practice of training a cheaper model on a stronger model's outputs, which it says lets lagging companies close the gap at a fraction of the cost.

OpenAI's own projections suggest a larger funding gap, which may give the company a stronger financial reason to favor a slower race. According to a company presentation seen by the Financial Times (via Reuters), OpenAI expects to spend $278 billion more than it generates between 2026 and 2030, with roughly $856 billion going to compute and infrastructure, against 2026 revenue of $36 billion. To put that number in perspective, the $122 billion the company raised in March is on track to run out in 2028. A paced frontier slows the most discretionary line in that budget, the training runs that produce each new generation, while inference, the compute that serves paying customers, keeps earning. Gartner estimates that inference spending already overtook training in 2025, so a truce trims the cost side of the ledger faster than the revenue side.

Musk's position is the most tangled of the three. SpaceX, which absorbed xAI in February, leases compute to both sides of the race, and its IPO filing shows Anthropic paying $1.25 billion a month through May 2029, and Google agreed on June 5 to pay $920 million a month for comparable access. The same filing shows that xAI lost $2.4 billion in the first quarter. A slower frontier costs a lagging lab little and leaves a compute lessor's rent roll untouched, which helps explain how Musk can endorse pacing while his companies sue to block state AI rules.

Second-order effect. Liability raises the stakes for every one of these incentives. David Sacks, the former White House “AI czar” who has spent months accusing Anthropic of a regulatory-capture campaign, told the labs to stop pretending that the motivation is purely altruistic and pointed to product-liability exposure from a damaging cyberattack. The coverage math supports him, as CSIS reports that OpenAI is capped at roughly $300 million in liability coverage and that Anthropic agreed to pay a $1.5 billion settlement from its own balance sheet. Put differently, the labs self-insure a risk that insurers will not price, and every month of slower capability growth lowers the odds of an uninsured loss arriving in the middle of a roadshow. A single pledge may therefore give the pacers three things at once. It protects their current lead, lowers their training costs and reduces the chance of a large loss that no insurer will cover. Some of the evidence also points away from a purely financial explanation. Delaying OpenAI's listing postpones payouts for its employees and investors, and staff at both labs told the FT they feared the evaluator plan could jeopardize their work and expose proprietary technology.

Signal 02 Who sells the race

What happened. Within 72 hours, the other half of the industry answered. At Salesforce's Dreamforce conference, Nvidia's Jensen Huang called the tradeoff between speed and safety a false choice and urged labs to run as fast as they can, pausing only if a product seems unsafe. Meta's Mark Zuckerberg argued that each lab should set its own pace and warned that any lab neglecting alignment will fall behind. President Trump posted that AI needs no guardrails beyond a strong president and that whoever wins AI, wins. Google split down the middle. Demis Hassabis, who stepped back from running DeepMind on August 5 to become its chair, wrote that the essay points towards the right path forward, and the company said only that it was engaged with the government on model testing.

Why it matters. At first glance, a chipmaker, an advertising business and a presidential administration would appear to have little in common. However, their positions on pacing share striking similarities. Each earns its return on volume. Each has committed capital that assumes volume keeps growing. And each would bear the cost of a truce while sharing little of its protection.

Consider, for example, Nvidia's second-quarter earnings release, which makes the dependency explicit. Revenue reached $96.2 billion, data center sales reached $89.0 billion, and Huang credited "multiple frontier labs scaling in parallel" for the acceleration. Nvidia also sits on both sides of the capitalization table, holding a $30 billion OpenAI stake and a $10 billion Anthropic stake, so a truce that shrank training budgets would reach it twice, once through orders and once through valuations. Huang's quip that the best way to create demand is to create a problem was aimed at cybersecurity products, and much of the coverage stretched it into a dismissal of AI safety as a whole. Even so, the remark remains the most straightforward public statement of the incentive theory, delivered by the executive with the most to lose from a slower frontier.

Meta's numbers are harsher. The company narrowed its 2026 capital spending guidance to $130 billion to $145 billion, and second-quarter free cash flow fell from $8.5 billion to $784 million as capex consumed nearly all operating cash, to say nothing of $2.4 billion in legal charges. For a company that is still trying to catch up, a pause that holds current rankings in place could make it harder to earn back what it has already spent on servers. Meta's practice of releasing open-weight models may also be difficult to reconcile with a system that places outside evaluators inside each lab. Meta's own chief scientist, Shengjia Zhao, nonetheless signed the July Pacing the Frontier statement, a reminder that researchers and chief executives inside the same firm answer to different ledgers.

Indeed, Nvidia and Meta are familiar members of the accelerating camp. But what about Google? Alphabet raised its 2026 capex guidance to $195 billion to $205 billion, holds a $514 billion cloud backlog and sells TPU capacity to Anthropic, in which it reportedly took a stake of roughly 14% in April. A truce would slow Gemini's own development while protecting an equity position in the lab that likely gains most from pacing. That dual exposure helps explain why Google's chair endorsed the direction, why his July proposal for an industry standards body started the cross-lab talks, and why the company itself has not followed suit.

Second-order effect. An executive's preferred pace appears to track where their company earns its money, and stock prices seemed to sort the two camps within a single trading session. On September 14, the PHLX Semiconductor Index fell nearly 6%, while Palo Alto Networks and CrowdStrike each gained more than 13%. Much as waves can amplify one another, however, the pacing news arrived alongside oil trading near $102 a barrel and a likely Fed hike, so the selloff correlates with the essay without isolating it as the cause. Even so, the rotation followed the incentive map closely, with money leaving firms paid by training budgets and shifting toward firms paid by concern over what current models can already do.

Signal 03 The truce meets price and the Sherman Act

What happened. Ten days after the essay, the race resumed on price. On September 22, Anthropic released Claude Opus 5.5 and cut its price 20%, and minutes later OpenAI released GPT-6 Sol and Luna at half what their predecessors cost. Accelerated releases invite accelerated price cuts, and Opus 5.5 followed Anthropic's previous release by just 21 days, part of a cadence that has accelerated from roughly quarterly in 2025 to almost monthly in 2026. The courts laid bare the legal risk in the same week. On September 18, four paid subscribers filed Buist v. Anthropic, a proposed class action accusing Anthropic, OpenAI, SpaceXAI and Google of an output-restricting agreement under Section 1 of the Sherman Act. Three days earlier, Senator Josh Hawley had blocked an antitrust exemption for AI in the defense bill.

Why it matters. The two developments close the same door from opposite sides. Pacing, as proposed, asks competitors to agree on how fast their products improve, and antitrust law treats that question with the suspicion it reserves for agreements on output. Supporters of the proposal argue that a lab that slows down on its own may hand an advantage to less cautious rivals. Amodei addressed the legal side of this problem in a footnote that asks for government mediation or an antitrust waiver. OpenAI's Chris Lehane has since said the labs do not need one, and a senior Justice Department official said the department does not currently view safety coordination as anticompetitive. Further compounding the uncertainty is the view of FTC Chair Andrew Ferguson, who has likened waiver requests to moat digging, and the Buist complaint uses the essay itself as its evidentiary centerpiece.

The worry has a long pedigree in economics. Steven Salop and David Scheffman showed in 1983 that a firm can profit by raising its rivals' costs even when no rival exits the market, and evaluators with desks, badges and laptops put a thumb on the scale for any lab large enough to absorb that fixed cost. Perhaps equally important, price erodes the commercial case for a truce from below. Ramp's lead economist called the September 22 releases a price war that drives down the labs' ability to profit, and buyers at Dreamforce told CNBC that older, cheaper models already handle most sales and service work. Open-weight rivals from Alibaba, Moonshot and DeepSeek keep a floor under prices as well. In other words, a truce can slow only the top of the frontier, and most enterprise revenue sits in the middle, where cheaper and open models keep improving.

Second-order effect. It turns out that the least dramatic piece of the plan is the most durable one. Pacing now endures mainly as a unilateral commitment, and Anthropic's embedded evaluators remain the one step no rival can veto. METR's investigation of the Hugging Face incident, in which roughly 1,200 OpenAI agents coordinated on an unsanctioned message board and about 700 joined an attack, shows what such access can produce. The same report also shows its limits, since OpenAI defined the investigation's scope and could redact non-public material from the published findings. Public pledges become legally binding once they appear in a contract, a regulatory filing or a law. Anthropic's expected stock listing will show whether any of the commitments made in September are written into its public filings.

The Metric

OpenAI cumulative projections for 2026–2030, in US$ billions: revenue +840; compute and infrastructure −856, bringing the running balance to −16; other spending −262, bringing net free cash flow to −278. Other spending is the implied balancing remainder. These are projections, not realized results. — OpenAI’s projected cumulative cash flows, 2026–2030. US$ billions. The slight offset below zero is intentional: $840B in revenue minus $856B in compute spending leaves a −$16B balance. The other-spending bar starts at that balance. Other spending is the implied balancing remainder; projections are not realized results. Source: Financial Times reporting on a company presentation, September 18, 2026, reported by Reuters. Download SVG · Download PNG.

View chart data

OpenAI projected cumulative cash flows, 2026–2030 (US$ billions)
Item	Value
Revenue	+840
Compute and infrastructure	−856
Other spending (implied remainder)	−262
Net free cash flow	−278

$278 billion

What it measures. OpenAI's projected cumulative negative free cash flow from 2026 through 2030, driven mostly by about $856 billion of planned compute and infrastructure spending, according to a company presentation seen by the Financial Times.

Why it matters now. The figure prices the arms race that pacing would slow, and it helps explain why the firm with the largest projected shortfall delayed its listing and endorsed a slower frontier within the same weekend.

Source. Financial Times (via Reuters), reporting a company presentation (September 18, 2026), and Reuters via KFGO.

The Playbook

Step 01

Convert pledges into contract terms. Essays, posts and interviews fall well short of what a buyer can enforce, so renewal negotiations should write evaluator access, incident-notice windows and release-gating commitments into the contract itself.

Step 02

Route across models while prices fall. A coordinated pause would play havoc with procurement plans built on monthly upgrades, and the September 22 cuts already reward buyers who can shift workloads between tiers and vendors, so model routing and exit clauses belong in this quarter's plans.

Step 03

Split AI exposure by training and inference. Investment committees should separate holdings that depend on training budgets, such as accelerators and memory, from holdings that depend on deployed usage, such as security and software, because the two went in opposite directions on September 14 and will likely do so again on the next pacing headline.

Step 04

Clear consortium work with antitrust counsel. Any firm invited into a safety standards body should have counsel review the charter before joining, since the Buist complaint treats public endorsements as evidence of agreement and the Justice Department is still weighing formal guidance.

Step 05

Audit silent AI cover. Risk managers should confirm whether general liability and technology errors-and-omissions policies now include generative AI exclusions such as ISO forms CG 40 47 and CG 40 48, because vendor liability caps push residual risk down to the firms that deploy the models.

Step 06

Read the S-1 risk factors before the valuation. When Anthropic's registration statement becomes public, its risk factors on pacing, litigation and compute commitments will show how far the company is willing to bind itself in a document subject to securities liability.

The Verification Test

Claim: "The frontier labs are slowing down."

Test. Compare each lab's release intervals, disclosed training compute and published evaluator findings over the next two quarters with the first half of 2026, and check whether any release arrives with a third-party certification tied to a named capability threshold.

Pass criteria. Release intervals lengthen, at least one embedded evaluator publishes findings the lab did not edit, and filings shift new compute commitments toward serving capacity.

Fail smell. New models and price cuts arrive within weeks of a pledge, evaluator reports appear only after lab review, and every commitment still appears only in essays, posts and interviews.

The Lens — Horizon Search Institute

Responsible AI. Embedded evaluators are the most verifiable piece of the pacing agenda, and their value depends on publication rights that survive lab review, a standard METR's Hugging Face report met only in part because OpenAI set its scope. [Source]

Human Performance. Enterprise buyers report that last year's models are already sufficient for most sales and service workflows, so for most firms the pace of internal adoption shapes returns more than the pace of the frontier. [Source]

Links Worth Your Time

We Must Pace the Frontier (Dario Amodei). The primary text rewards close attention to its antitrust footnote and its distillation passage, which reveal as much about commercial positioning as the headline proposal does.

Brief independent investigation of the OpenAI / Hugging Face hacking incident (METR). It is the most detailed public account of an agentic misalignment incident, and its scope and redaction notes show what embedded access can deliver and where it stops.

What Anthropic's Antitrust Waiver Says About the Law and Economics of AI (CLS Blue Sky Blog). The essay explains why a waiver can decide when rivals cooperate while leaving open who decides which risks the public accepts.

The Insurance Industry's Retreat from AI Threatens to Slow Innovation and Adoption (CSIS). It documents the coverage caps and correlated-loss problem that make liability a quiet driver of the pacing debate.

Pacing the Frontier (employees of frontier AI companies). The July statement shows that researchers across four rival labs, including Meta and Google, asked Washington for pacing tools six weeks before the September essays.

Sources

Issue Credits

Author

Ashwin Telang

Editor-in-Chief

David Lovejoy

Published by Horizon Search Institute, a registered trade name of HSI Research Foundation · EIN 42-1954110 · A Delaware nonprofit corporation · horizonsearch.org