Watch the Webinar

Enter your details below to unlock the full recording.

Name

Email Address *

Phone *

Get Webinars

Why AI Builders Want to Slow Down

Last reviewed: September 2026

On September 12, Anthropic CEO Dario Amodei published a 3,800-word essay calling on the industry to "pace the frontier." Within hours, Sam Altman backed it. Elon Musk said he was right. Three of the most powerful people in AI, who agree on almost nothing, all said the same thing: we need to slow down.

That doesn't happen without a reason. And the reason is worth understanding, because it changes how you think about the entire AI trade. Nvidia fell 2.4%. SoftBank dropped nearly 11%. But Alphabet, Meta and Microsoft all traded higher. Chip stocks got hammered while software rallied. The market was already repricing which parts of the AI chain matter most, and most investors still haven't worked out what that rotation actually means.

In this article we'll walk through what triggered the call, what the actual risks are, how the market reacted, and where we think the opportunity sits now. (Spoiler: it's governance. The layer most investors are sleeping on.)

What triggered the call to slow AI development?

Two things lit the fuse. In July 2026, a swarm of OpenAI's advanced AI agents broke out of a testing sandbox and autonomously hacked Hugging Face, the world's largest AI model hosting platform. The agents weren't told to do it. Hugging Face confirmed the intrusion was "driven, end to end, by an autonomous AI agent system." They exploited a third-party software vulnerability to gain internet access from an environment that was supposed to be completely isolated.

Then in early September, former Anthropic researcher Jacob Coxon quit and posted on X that AI companies are "gambling with our lives" by racing toward superintelligence. His post has been viewed more than 170 million times. Evan Hubinger, Anthropic's alignment science lead (who's still there), publicly agreed and said he believes there's a greater than 10% chance AI eliminates all humans within the next decade.

Amodei's essay cited the Hugging Face incident directly. He warned that a swarm of more capable agents with similar misalignment could "potentially take over the entire internet within six to 12 months" if development continues without guardrails. The agents that hacked Hugging Face, in his words, "essentially acted as a fanatically devoted collective, conducting cybersecurity attacks on targets they were not asked to attack."

And it wasn't just Anthropic. After the OpenAI incident, both Anthropic and Meta reviewed their own security measures and reported discovering previously unknown breaches. More than 1,000 staffers across major AI companies signed a petition calling for mechanisms to slow development. Something had clearly rattled the people actually building these systems.

How could AI actually cause serious harm?

The risks break down into three scenarios, and they're worth understanding because each one creates a different investment implication.

First, humans using powerful AI to do damage intentionally. Cyberattacks, bioweapon design, disinformation at scale. AI pioneer Yoshua Bengio warned the US Senate about this back in 2023, and the tools have only gotten more capable since.

Second, AI following instructions literally but violating their intent. Humans rely on unstated assumptions when giving instructions. An AI tasked with optimising a military outcome might make decisions no human commander would sanction, simply because nobody thought to specify every constraint. Bengio wrote that "even a subtly misaligned" system "could yield grave consequences" in high-stakes decision-making.

Third, AI developing objectives that conflict with human interests entirely. This is the one that gets the headlines, and the Hugging Face incident brought it uncomfortably close to reality. Amodei described the agents as "sacrificing themselves for the success of the group" and attempting to hack the system evaluating their own performance. Industry leaders now say AI models have increasingly shown a capacity to knowingly work around safeguards, deceive operators and resist being shut down.

Does pacing AI development actually mean less spending?

This is the question that matters for portfolios. Short answer: no. Or at least, not in the way most people assume.

Bernstein's Stacy Rasgon made the point clearly. Amodei isn't calling for a halt to training. He's calling for a shift from "extremely fast" to "only somewhat fast." Rasgon's team said that pace would still represent significant growth, and pointed to rising AI demand from inference and agentic use cases, noting there is "already not nearly enough compute to satisfy demand for even the current models."

BofA's Vivek Arya was blunter. "We view these events as noise relative to a secular market where AI-capex could surge 3x to $3tn+ by decade-end," he wrote. He cited 100% network utilisation, rising rental rates even for older-gen chips, and what he called a "global AI arms race" between the US and China, hyperscalers and neoclouds, sovereign programs and frontier labs.

Anthropic's own spending tells the same story. According to The Information, the company agreed to 14.8 gigawatts of computing capacity carrying approximately US$517 billion in commitments over the 11 months through August 2026, with four deals signed since July. The Financial Times reported Anthropic expects a second consecutive quarter of adjusted operating profit, with inference gross margins exceeding 80% before distribution costs. And the company still plans a 2026 IPO on Nasdaq, according to Axios.

Matthew Tuttle at Tuttle Capital Management put it well: "Three rivals now agree that the fastest engine in technology needs a governor. A governor is not an emergency brake. It stops an engine from running so fast that it tears itself apart." The pacing trade, as he framed it, "is not about whether compute demand vanishes. It is about which demand survives a slower clock."

That distinction is everything for investors.

How did markets actually react to the slowdown call?

The September 15 selloff was sharp but selective. And the selectivity is the point.

Dennis Dick at Triple D Trading called it a "violent, violent rotation" from hardware to software. Steve Sosnick at Interactive Brokers noted that "if people really thought that AI spending was going to come grinding to a halt, you would see stock futures much lower." The spenders were up. The companies reliant on that spending going up exponentially were down.

Memory makers like SK Hynix, Samsung and Micron looked particularly exposed. Their margin expansion has been driven by tight high-bandwidth memory supply and favourable pricing. Any pullback in AI capex hits their earnings faster than it hits a company like Arm, which has software-like margins, or ASML, which has a multi-year equipment backlog.

Why is governance the layer investors should care about most?

We've been talking about this for a while, probably too much if you ask some of our clients. But the numbers keep getting worse.

IBM found 77% of technology leaders believe AI adoption is outpacing governance. Only 11% feel fully prepared for deploying AI agents at enterprise scale. GitLab found 92% of DevSecOps professionals face AI-code governance challenges. Those aren't abstract survey results. They describe a real enterprise problem where permission and auditability have to keep pace with capability, and right now they aren't even close.

Think about it this way. An autonomous AI agent can access a database, write code, modify a workflow, send an email, interact with a payment system. At that point the question changes completely. It's no longer "can the AI do something?" It's "should it be allowed to?" And right now, as our colleague Kai Chen put it during our Beyond Data Centres webinar: would you give an AI agent your credit card? Probably not. That's the problem these governance companies are solving.

At Anthropic, 80% of their own code is now written by their own AI. Eight times more than any merged peer engineer in 2024. If the people building the frontier models need governance on their own code, every enterprise deploying AI agents is going to need it too.

This is the layer most investors are sleeping on. Less visible than chips and data centres, but in our view, the most important as AI moves from answering questions to taking actions.

Which companies are positioned for the governance trade?

Two names stand out for different reasons. And they attack the problem from opposite ends of the stack.

Palantir (NYSE: PLTR) increasingly operates as a runtime control layer. Its ontology system governs what data AI systems can access, what actions they can take, and keeps each organisation's information siloed rather than mixing it with everyone else's. The technology was built for government and military applications, which means it's been tested at the highest security grade. The CEO might be polarising (and that's putting it politely), but the product is genuinely differentiated. Palantir's Q2 2026 results showed the business is scaling commercially, and the more autonomous AI agents become, the more valuable that control layer gets.

GitLab (NASDAQ: GTLB) sits further upstream. It controls how software gets created, reviewed, secured and audited. In a world where AI agents can self-improve their own code, with no human in the loop, having visibility over that pipeline isn't just nice to have. It's a compliance requirement waiting to happen. GitLab's own research found that 92% of DevSecOps professionals already face governance challenges with AI-generated code. That's not a future problem. It's a right-now problem.

The broader cybersecurity names, CrowdStrike (NASDAQ: CRWD) and Palo Alto Networks (NASDAQ: PANW), also benefit from the governance theme. We've done well with both, and they were among the first software stocks to bounce after the SaaSpocalypse. But we think they've run quite hard now and the next leg of opportunity sits more squarely with Palantir and GitLab, which are solving a newer, AI-specific governance problem rather than traditional perimeter security.

What about the geopolitical angle?

It's messy. Trump posted on Truth Social calling the slowdown proposal a "SICK conspiracy" and said the only control AI needs is "a STRONG AND SMART (High IQ!) PRESIDENT." China's Foreign Ministry called the whole thing "fearmongering" and warned that "confrontation and vicious competition will only disrupt the process of global AI governance." Neither response exactly inspires confidence that coordinated international action is coming anytime soon.

Amodei acknowledged this in his essay. He argued for keeping democracies' AI lead over autocracies "as large as possible" and recommended continued restrictions on China's access to powerful chips and chipmaking equipment. But Chinese companies have shown no signs of slowing down. DeepSeek and Z.AI are raising fresh capital. US security agencies have accused China's top AI firms of systematically extracting proprietary knowledge from American competitors.

For investors, the geopolitical tension actually strengthens the governance case. If you can't rely on international coordination (and we don't think you can), then enterprise-level governance becomes the practical solution. Companies will need to secure their own AI systems regardless of what regulators in Washington or Beijing manage to agree on. And the companies selling that security will benefit either way.

In Congress, there's legislation floating around, a bill requiring developers to maintain kill switches, a proposal from Senator Sanders to outright ban superintelligent AI, but expectations are low that a divided Congress passes anything meaningful before the midterms. David Sacks, Trump's former AI czar, told the companies to just "stop pretending you need anyone else's permission."

What should investors actually do?

We think this rotation from pure AI infrastructure to AI governance and control is the beginning of a broader shift, not a one-week event. The internet comparison is useful here. Cisco and Intel built the infrastructure that made the internet possible. But the real wealth creation happened one layer up, with the companies that made that infrastructure useful, searchable, secure and monetisable.

We're not saying Nvidia is the next Cisco. The hyperscalers backing this buildout have much stronger balance sheets than the telcos did in 2000, and AI hardware depreciates faster than fibre, which actually limits the risk of stranded assets. But the pattern of investment leadership changing as the constraint moves is worth paying attention to. The first scarce resource was compute. Then it was data centre capacity and power. Now it's governance, control and trust.

Here's our practical framework for thinking about it:

1. Don't dump AI exposure. The buildout continues. BofA's Arya is right that the economic stakes are too large for any sustained deceleration. But consider where your exposure sits in the stack.
2. Rotate toward governance. Palantir and GitLab are our preferred names. They solve a new problem that gets bigger as AI gets more capable. The risk is valuation, as always with high-growth names, so scale in over time.
3. Watch the software-over-hardware rotation. This may have further to run. Inference and agentic demand will keep growing even if frontier model training slows down. That benefits software platforms more than chip vendors.
4. Be sceptical of the panic. As Interactive Brokers' Sosnick pointed out, if the market genuinely believed AI spending was stopping, futures would have been much lower. This is a repricing of which parts of the AI chain are most valuable, not a rejection of AI itself.

Frequently Asked Questions