AI slowdown: capital, safety and the 2026 midterms

On September 12, Dario Amodei asked AI companies to coordinate the development of their most advanced models. Sam Altman and Elon Musk expressed support. Demis Hassabis backed the general direction. The next day, Donald Trump stressed that the United States must preserve its lead over China. These positions bring a practical question into focus: what would a common speed limit change for an industry making enormous, long-term financial commitments?

The obvious suspicion is that companies would welcome a chance to catch their breath after years of spending heavily on computing power. November’s midterm elections add another incentive to settle the rules while the current Congress is still in place. Those are plausible explanations for why coordination might appeal to a company. Establishing why a particular executive supports it is harder.

Safety policies predate the latest funding commitments. Anthropic published its first Responsible Scaling Policy in 2023. A research paper published by OpenAI argued for public oversight of advanced systems that year. A statement launched in July 2026, Pacing the Frontier, describes the competitive pressure that makes unilateral restraint difficult.

Meanwhile, the Bank for International Settlements has documented an investment race increasingly financed through borrowing, leases and guarantees. A change in the congressional majority could also shift hearing agendas and oversight priorities.

The safety concerns and the financial incentives can coexist. The useful question is what the proposed rules would do, who would bear their costs and how much room they would leave for competitors.

The three parts of Amodei’s proposal

In We Must Pace the Frontier, published on September 12, Amodei sets out three levels of action. Here, “frontier” means the most advanced AI capabilities available.

First, Anthropic promises independent evaluators permanent, employee-like access to inspect its safety practices.

Second, Amodei calls for labs in democratic countries to agree on common standards and, where necessary, limits on unchecked advances. His commercial argument is explicit: a lab that slows alone risks losing ground to its competitors. Coordination could give companies time to improve safety while limiting that disadvantage.

Third, he calls for international coordination while retaining controls on advanced chips and model distillation, the use of one model’s outputs to train another. A slowdown that simply transferred the lead to a strategic rival would undermine the proposal.

Reuters reported support from Altman and Musk. Axios and The Guardian reported Hassabis’s endorsement of the broad direction. Hassabis is now chairman of Google DeepMind and chief scientist of Alphabet, following a change announced in August; the roles are confirmed in an Irish government announcement.

These statements should be distinguished from a formal agreement or an identical company-wide commitment. The sources reviewed here describe different levels of support.

On September 13, Trump rejected calls to slow down while emphasising America’s lead over China. Amodei shares the aim of preserving that lead; the disagreement concerns the pace of development. Reuters quoted Trump’s summary, “whoever wins AI wins”, and reported his dismissal of some safety concerns as exaggerated.

Schematic chronology; spacing does not represent elapsed time. Sources: OpenAI, Anthropic, Pacing the Frontier, H.R. 9914, BIS and Dario Amodei. Sources for the September 12–13 statements and election calendar appear in the article.

The antitrust question behind the “cartel” argument

David Sacks, the former White House AI adviser, argued that individual companies were free to slow on their own and attacked the proposed antitrust accommodation as cover for a cartel, as reported by The Next Web. The argument raises a genuine legal issue: competitors coordinating development schedules, computing capacity or product releases need to consider competition law.

Congress had already taken up a version of this problem. H.R. 9914, the Collaboration on Adversarial Threats and Security Risks Act, was introduced on July 23, 2026, with Republican and Democratic sponsors. The official bill history records its referral to the House Judiciary Committee. It remained a proposal at our September 14 cutoff.

Section 3 distinguishes information-sharing from coordination that delays or limits AI release, deployment, use, development, training, testing or evaluation to mitigate covered security risks. The latter would require advance notification to the Attorney General. Participants bear the burden of establishing the exemption’s conditions, including its good-faith security purpose.

The drafting deserves close attention. Section 3(d)(1)’s explicit exclusions for price-fixing, market allocation, monopolisation, certain boycotts and exchanges of price or cost information refer to the information-sharing exemption in section 3(a)(1). They are not expressly drafted as the same list of exclusions for the separate coordination exemption in section 3(a)(2). Describing the bill as having an identical competition safeguard for every kind of agreement would overstate its text.

Under section 4, the Department of Justice could seek an injunction if the conditions were not met or if the coordination increased covered security risks on balance. That security-risk test is not a general balancing test for competitive harm.

The bill would create a specific legal channel for safety coordination between competitors. Its boundaries, and the distinction between its two exemptions, are central to judging how broadly that channel could be used.

Three years of safety policies

Anthropic first published its Responsible Scaling Policy on September 19, 2023, linking stronger safeguards to more dangerous model capabilities. On July 6, 2023, OpenAI published the multi-author paper Frontier AI regulation: Managing emerging risks to public safety, which advocated standards, reporting, external scrutiny and possible public enforcement. The publication explicitly says authorship does not imply endorsement by each author’s organisation.

The Pacing the Frontier statement launched in July 2026 had 1,386 signatories when consulted on September 14. That is a current count, not a historical July total. The list includes people affiliated with Anthropic, OpenAI, Google DeepMind and Meta. Their signatures are personal and do not establish a corporate policy for their employers.

The statement describes a coordination problem: companies and countries face competitive pressure to keep advancing even when they would prefer more time to assess risk.

There are also concrete incidents to examine. On September 9, Anthropic published its assessment of four incidents involving unauthorised access to third-party systems during cybersecurity evaluations. The test environment had mistakenly allowed open-internet access, and production safeguards were disabled. Those circumstances are essential to interpreting the findings; they cannot be generalised to ordinary Claude use. Anthropic says it first examined roughly 141,000 transcripts, expanded the search to about 481 million, and notified affected parties.

OpenAI separately described an incident involving Hugging Face in its account of an agent evaluation and the changes made afterwards. On September 9, it called for mandatory national safety requirements tied to capabilities and risks, including independent assessment, cybersecurity standards and incident reporting.

These companies are reporting on systems they build, and their assessments require scrutiny. Still, their published policy history makes it difficult to reduce the safety debate to a reaction to this month’s financing calendar.

The cost of computing power

On September 10, BIS General Manager Pablo Hernández de Cos gave a speech on AI and financial stability. The BIS expects the five largest big-tech companies to spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026. Capital expenditure, or capex, covers investment in assets such as computing equipment and data centres.

The speech describes capital spending increasingly exceeding companies’ cash flows and a growing role for debt and private credit, lending by non-bank investors outside the public bond market.

A January 2026 BIS Bulletin had already described this shift towards borrowing. A separate BIS study examines on- and off-balance-sheet infrastructure funding: special-purpose vehicles, joint ventures, capacity contracts, leases and guarantees. A special-purpose vehicle is a separate legal entity used to hold assets or financing for a particular transaction. These structures can place formal borrowing outside the cloud group’s own balance sheet while leaving it with substantial contractual commitments.

BIS Working Paper 1367, The AI investment race, published on July 14, models a market in which firms invest heavily to avoid losing a dominant position. Its calibration produces investment roughly 50% above the socially efficient level. This is a model result, not a measurement of actual waste in the 2026 industry. The authors’ views do not necessarily represent the BIS or its member central banks.

The mechanism is straightforward. A company slowing alone might save money while losing ground to its rivals. If rivals credibly slowed together, they might reduce future spending with less damage to their relative positions.

Existing contracts would remain. Savings would depend on which investments could be deferred, the terms of signed commitments and whether expenditure shifted towards using existing models.

Alphabet, Meta and Amazon face the bill

The largest cloud operators, often called hyperscalers, combine substantial internal cash generation with heavy investment. Their first-half accounts give the scale of that commitment.

The definitions differ. Alphabet reports purchases of property and equipment. Meta’s capex measure includes principal payments on finance leases. Amazon’s cash capex covers technology infrastructure and its fulfilment network. These are group-wide measures, not a clean comparison of spending on AI alone.

Six months ended June 30, 2026, $bn	Operating cash flow	Capex / cash capex	Capex / operating cash flow
Alphabet	84.9	80.6	~95%
Meta	64.1	50.9	~79%
Amazon	71.4	96.3	~135%

Sources: Alphabet 10-Q, Meta 10-Q, Amazon 10-Q. Ratios are calculated from the displayed rounded inputs.

Alphabet generated $84.9 billion of operating cash and spent $80.6 billion on property and equipment, with technical infrastructure driving the increase. In June it also raised $20.5 billion through a public common-stock offering, placed $10 billion of common shares with a Berkshire Hathaway affiliate, and raised about $19 billion through mandatory convertible preferred stock. Its filing identifies AI infrastructure and computing capacity among the uses of proceeds.

Meta generated $64.1 billion of operating cash and reported $50.9 billion of capex. It forecast $130–145 billion of capex for 2026 to support its AI efforts and core business.

The future obligations are substantial. At June 30, Meta disclosed roughly $279 billion of operating and finance leases that had not yet commenced, mainly for data centres, colocation and network infrastructure, with starts extending through 2036. In July it signed roughly $68 billion of additional data-centre leases. It separately reported $349.3 billion of non-cancellable contractual commitments. Those disclosed categories must be read on their own terms; adding them into a single debt total would disregard their scopes and timing.

Amazon’s 10-Q reports $71.4 billion of operating cash and $96.3 billion in cash capex. It says most technology-infrastructure investment supports AWS growth, while its cash-capex total also includes the fulfilment network.

Amazon additionally invested $28.7 billion in OpenAI Series C preferred stock during the half, then funded the remaining $21.3 billion of that commitment after June 30. It invested another $10 billion in Anthropic in the second quarter.

These figures show why a change in the pace of AI development would affect much more than a laboratory’s next training run. Suppliers and investors already have large commitments tied to future use.

Six months ended June 30, 2026. OCF means operating cash flow. Alphabet: property-and-equipment purchases. Meta: those purchases plus finance-lease principal. Amazon: cash capex includes technology infrastructure and the fulfilment network. Definitions differ, and these totals are not exclusively AI spending.

Who carries the infrastructure commitments?

OpenAI’s Ohio infrastructure illustrates how the exposure is divided. In August, Nvidia filed guarantees capped at $105 billion connected with SB Energy’s PORTS-Pike campus, where an OpenAI affiliate is to be the tenant.

The cap covers defined portions of lease and electricity obligations under the agreements. It is neither the entire campus cost nor a $105 billion loan advanced to OpenAI. Nvidia’s quarterly filing describes nine phases, with guarantees taking effect as the relevant phases commence and declining as OpenAI makes payments.

OpenAI has agreed to indemnify Nvidia for certain amounts actually paid under the guarantees. The tenant, the landlord, the chip supplier and the parties financing the assets therefore carry different exposures to the same project. Our analysis of OpenAI’s credit rating and Nvidia’s guarantee examines the conditions under which that support can end.

Bank of America separately extended a reported $520 million credit line in July. Reuters described this as Bank of America’s first loan to OpenAI, rather than OpenAI’s first bank borrowing.

OpenAI’s operating figures require a different level of attribution. Reuters, citing The Information and shareholder documents, reported $5.7 billion of revenue and $3.7 billion of cash burn in Q1 2026. Reuters said it could not independently verify those documents.

Anthropic illustrates another structure. On July 6, TeraWulf disclosed a 20-year lease for approximately 401 MW of computing capacity at its Kentucky campus. Its press release filed with the SEC estimates roughly $19 billion of contracted revenue over the initial term. That landlord revenue estimate represents a long-term contract; it should not be relabelled as an equivalent amount of bank debt owed by Anthropic.

For a separate Texas campus, Reuters, citing the Wall Street Journal, reported talks about roughly $15 billion of bank financing for a project serving Anthropic and supported by Google guarantees. The report describes financing discussions, and the full structure is not established by the public filing record cited here.

On September 13, Reuters, citing the Financial Times, reported that Anthropic had told investors it expected positive adjusted operating income for a second consecutive quarter. Reuters could not independently verify the report. Adjusted operating income, reported under the company’s chosen adjustments, is distinct from audited GAAP net income and free cash flow.

Revenue, operating profit, cash use, leases and project debt each answer a different question about the business.

What SpaceX’s accounts reveal about xAI

SpaceX’s public reporting now provides figures for its AI segment. The scope matters: it includes both AI activities and advertising associated with X.

For the first half of 2026, the segment reported $3.379 billion of revenue, a $3.726 billion operating loss and $23.551 billion of capex. Revenue consisted of $2.669 billion from AI solutions and infrastructure and $710 million from advertising. Assigning all segment revenue to AI products would overstate that business.

In Q2 alone, the segment generated $2.561 billion of revenue, recorded a $1.257 billion operating loss and spent $15.828 billion in capex. Its adjusted EBITDA was positive $537 million for the half. EBITDA excludes interest, tax, depreciation and amortisation; the adjusted figure also incorporates company-defined adjustments. SpaceX identifies it as non-GAAP and publishes a reconciliation.

The consolidated group ended Q2 with $100 billion in cash, cash equivalents and marketable securities. This is a group-wide figure; the disclosure does not identify it as a dedicated xAI balance.

A coordinated slowdown could change the cost of future investment for this group as well. Its accounts alone cannot establish why Musk endorsed Amodei’s proposal.

Three companies, different interests

Google DeepMind backs the principle

Hassabis’s support establishes a public position on Amodei’s broad proposal. In a personal essay published on July 14, he had already proposed a standards structure that could coordinate a slowdown. Those statements do not establish that Alphabet has adopted Anthropic’s permanent-evaluator mechanism across the group.

Alphabet has several exposures to the outcome: it builds models, sells cloud services, designs AI processors and supports infrastructure used by Anthropic. Slower progress could reduce demand for expensive new training runs while extending the commercial life of existing models.

Reuters Breakingviews makes a related argument: expenditure could shift towards inference, which means running trained models to produce answers, and towards agents and applications. That is an economic scenario, and it would not necessarily reduce total AI spending.

Meta: researchers and corporate policy

The July Pacing the Frontier statement includes signatures from senior Meta researchers, including Shengjia Zhao and Dawn Song. As the statement makes clear, these are personal endorsements.

At our September 14, 2026 research cutoff, the sources reviewed here contained no equivalent public commitment by Meta Platforms or Mark Zuckerberg to Anthropic’s embedded-evaluator programme or OpenAI’s proposed national safety requirements. That is the limit of this review, not a claim to have established Meta’s private position.

Meta’s open-weight strategy also raises different questions from those facing a lab that keeps its model weights private. Model weights are the learned numerical parameters; making them available allows others to run or adapt a model, subject to its licence. How safety rules treat distribution, modification and responsibility would matter for that strategy.

Amazon as supplier and investor

We likewise found no public endorsement of Amodei’s September 12 call by Andy Jassy or Amazon in the sources reviewed by the cutoff.

Amazon nevertheless has exposure at several points in the industry. AWS sells computing capacity; Bedrock distributes multiple model families; Amazon designs Trainium chips; and its investments finance OpenAI and Anthropic.

Its June 30 filing documents a $50 billion total Series C commitment to OpenAI, funded in the two instalments described above, and the additional $10 billion invested in Anthropic in Q2.

A slower frontier could reduce demand for some training projects while improving the returns from deploying existing models for longer. Amazon is exposed to both possibilities. Diversification gives it several ways to participate, without guaranteeing that every outcome is profitable.

Our earlier analysis of AI’s circular financing examines how the same groups can be suppliers, customers and investors in one another.

Arrows show a possible supply, financing or contractual relationship. They are not five certain cash flows or a universal structure. At PORTS-Pike, Nvidia backs specified obligations linked to OpenAI’s leases, subject to a $105 billion aggregate cap; that amount is not a loan of the same size to OpenAI. Sources: BIS and Nvidia, 10-Q note 10.

What the midterms could change

The November 3, 2026 elections put control of Congress in play. In a September 1 analysis for the Cook Political Report, Charlie Cook described a Republican loss of 26–35 House seats and four to five Senate seats as his most likely scenario. He attached no numerical probability to that assessment. It remains one analyst’s judgement, subject to change as the campaign develops.

A change in the House majority could matter to the industry even without a large new AI law. Under House Rule X, section 5(c), committee chairs are elected on the nomination of the majority party. Rule XI, section 2(m), provides for hearings and subpoenas within committees’ jurisdiction and procedures. The majority can therefore change both leadership and oversight priorities.

New committee leadership could therefore bring more scrutiny of safety incidents, infrastructure deals, political relationships and the effects of data centres. Such oversight would be possible even if divided government prevented major legislation.

The Senate currently comprises 53 Republicans, 45 Democrats and two independents. With the existing Democratic caucus alignment maintained, Democrats would need a net gain of four seats for a majority. Neither chamber’s election outcome is settled.

From a company’s perspective, negotiating rules now could appear preferable to facing different committee leadership later. OpenAI’s request for congressional action makes the timing relevant. But Trump’s resistance to slowing AI also shows a substantive disagreement between the industry leaders making the proposal and the current administration.

We found no public document or testimony establishing that concern about the midterms caused the executives to support pacing. The election is a possible incentive in the background.

The different channels of political funding

Elon Musk’s support for Trump is directly documented. Reuters reported at least $119 million in 2024 contributions to a pro-Trump spending group. In July 2026, Reuters reported that America PAC planned a $100–120 million Republican-turnout effort and that Musk had contributed more than $85 million to the midterm cycle.

Sam Altman’s reported personal $1 million pledge to Trump’s inaugural fund falls into a different category. Reuters’s contemporaneous factbox on inauguration donations also records $1 million donations from Google and Meta, alongside Amazon’s announced $1 million contribution. An inaugural fund finances inauguration-related activities. It is legally distinct from a campaign or a super PAC, the latter being a political committee that makes independent electoral expenditures.

OpenAI states in a June 1, 2026 policy note that the company has made no donations to super PACs, candidates or campaigns and has no employee-funded PAC. It describes Greg Brockman’s and his wife’s support for Leading the Future as personal activity, separate from OpenAI.

Anthropic says it increased its support for Public First Action to $40 million in July. The company describes the organisation as nonpartisan and says its contributions cannot be used to influence a candidate’s election. Those descriptions are attributed to Anthropic.

Reuters also documents broader spending by AI, cryptocurrency and online-betting interests through electoral and advocacy organisations.

These flows show the industry’s participation in politics through several channels. Their recipients, legal purposes and donors differ. Treating every payment as corporate funding of Trump’s presidential campaign would erase distinctions that matter to the analysis.

Four ways to understand coordination

Safety concerns

There is a documented policy history, followed by specific incident reports. Capability gains and the incidents described by companies could have increased the urgency of independent testing.

The limitation is that laboratories still control much of the information about the systems they build. Their public assessments also influence how investors, governments and potential employees perceive their capabilities.

Potential relief from the investment race

The BIS research describes competitive pressure to invest, while company filings show large financing and infrastructure commitments. Coordinated restraint could ease the cost of future training investment if enough competitors participated credibly.

That possibility is separate from an established motive. Slower capability growth could also postpone revenue and reduce the value investors assign to a lab. Signed contracts would continue to matter.

Negotiating ahead of the midterms

A change in congressional control could alter scrutiny and legislative priorities. Established firms might prefer national rules negotiated now to a changing patchwork of state requirements or rules adopted after a major incident.

This interpretation must also account for the White House’s public resistance to a slowdown. The current administration is not simply endorsing the proposals made by Amodei and Altman.

Entry costs in a regulated market

Testing, reporting and cybersecurity requirements impose costs. Large platforms and well-funded labs may have more resources to meet them, creating a potential disadvantage for smaller competitors.

The design matters. In its September 9 proposal, OpenAI explicitly argues for obligations focused on the small number of labs building the most capable systems, rather than general requirements for startups or research outside that frontier. It also supports open-weight development. Those stated intentions deserve to be distinguished from the eventual effects of legislation.

H.R. 9914 raises a separate question about what competitors could agree among themselves. Its security-purpose conditions and the different treatment of information-sharing and coordination would need careful scrutiny. Neither the bill nor the companies’ proposals settle in advance how a future regime would affect market entry.

The IPO calendar adds another tension

On September 11, Reuters reported discussions about an Anthropic IPO that could raise up to $100 billion at a valuation around $2 trillion, with Nvidia considering an anchor investment. The plans were based on confidential sources and could change. The offering was then expected before the midterms.

The next day, Amodei argued for slower progress in a technology underpinning the business. That juxtaposition leaves room for several readings: stronger oversight might reassure investors, while slower advances could delay the growth on which a valuation depends.

OpenAI’s plans moved in another direction. Reuters reported Altman’s statement that the company would not proceed with an IPO in 2026, citing safety concerns. The stated reason should remain attributed to him. The financial consequences of a delay cannot establish another motive on their own.

The decisions facing policymakers

The public record establishes a history of safety proposals, reports of incidents under specific test conditions, calls for coordination and growing financial commitments. It also establishes political spending and an election that could change congressional oversight.

The economic and political interpretations go a step further. Coordination could affect investment, competitive positions and the timing of regulation. We do not know how much weight each executive places on those considerations.

That uncertainty does not prevent a useful assessment of the proposals. A workable regime would need to specify which capabilities trigger obligations, who performs independent evaluations, how compliance is checked and which agreements between competitors are permitted. It would also need to consider the costs for new entrants and developers releasing model weights.

The central trade-off is tangible: giving laboratories more time to address risk can also change who gets to compete, which investments proceed and who pays for the infrastructure already promised.

Method and principal sources

Research cutoff: September 14, 2026. Monetary figures are in US dollars. Company metrics have different accounting scopes, identified where used. Figures based on confidential documents or anonymous-source reporting remain attributed to that reporting. Scenarios about motives and future outcomes are identified separately from disclosures.

AI's race to slow down: safety, capital and politics