Pacing AI development doesn’t guarantee slower adoption

Key Takeaways

• The AI debate went public. Anthropic's Dario Amodei called for slowing AI development so safety can catch up, backed by OpenAI's Sam Altman and SpaceX's Elon Musk. The extent of any slowdown remains uncertain, but the public discussion could shape the pace and guardrails of future model development.
• Demand should shift, not shrink, and may cut both ways. Slower model development raises uncertainty, but we expect demand to move from building AI to using it. Reallocated spending could even lift lab revenue and cash flow. The main risk: a prolonged slowdown re-prices the priciest AI stocks and weighs on economic growth.
• Tech trades back to its ChatGPT-launch valuation. The sector has de-rated from 32x forward earnings to 21x, back to late-2022 levels, even as earnings have climbed. A significant amount of uncertainty is already priced into shares.
• Stocks keep moving out of lockstep. S&P 500 names are trading on their own merits more than at any point in 30 years, in part due to AI. Expect the market to keep seesawing between perceived winners and losers.
• A bumpy ride, but AI adoption is still early. While 88% of organizations now use AI in some form, only about 7% have fully scaled it, per McKinsey, a sign the opportunity is still unfolding, not behind us.
• Stay balanced, keep an open mind. There are a lot of moving parts, and this once-in-a-generation technology is evolving at a blistering pace. We're holding our long-term positive view while staying willing to adjust as the facts change.

What happened

Over the weekend, Anthropic CEO Dario Amodei published an essay, We Must Pace the Frontier, arguing that AI development should slow so safety, alignment, and oversight can catch up, driven by model safety and cybersecurity concerns. Peers led by Altman and Musk quickly voiced support. In practice, this could mean the leading labs deliberately build new models more slowly. Sam Grelck, CFA, CMT® Equity Strategy Analyst Keith Lerner, CFA, CMT® Chief Investment Officer Chief Market Strategist One distinction drives the entire discussion: training is the expensive work of building an AI model, while inference is the everyday use of one, from chatbot prompts to AI-powered tools. The question is whether slowing the first shrinks overall demand for computing power or simply shifts it to the second.

Our Take

We have remained positive on the companies that power AI, including chips, data centers, and infrastructure, as our research points to accelerating demand. Slower model development injects uncertainty, but we are skeptical it reduces total demand over time. The debate centers on two questions:

• Will AI use grow fast enough to absorb the capacity freed up from building it? Given how fast companies are adopting AI and already-tight capacity, we expect it will.
• Will slower model-building lead labs like OpenAI and Anthropic to buy less capacity over time? While this is unresolved, we expect total demand to grow meaningfully.

The bear case

If the leading labs genuinely pull back, future demand for computing power could be lower than expected, and Amodei openly floats capping how much labs use for training.

Since training has historically consumed a large share of demand, a lasting reduction would raise real questions about future capacity purchases. Even if inference absorbs all available supply today, long-term demand might fall short if training slows for good.

It's too early to know, but markets don't wait for certainty, and the risk is that investors re-price the most expensive AI stocks on the uncertainty alone.

We expect demand will continue to grow

Even with a slower building pace, we expect AI adoption to keep accelerating and for capacity to remain constrained for the foreseeable future. Because labs have devoted so much capacity to training, inference has been underserved; we expect a shift in where demand goes rather than a drop in total demand.

A real-world example: in August, OpenAI paused some model-building over cybersecurity concerns, and rather than sit idle, that capacity was redirected to safety and security work. This suggests freed up capacity can be put to productive use even before allocating to growing outside demand.

The shift should also help the labs' bottom lines. Model training is a pure cost; customer inference generates revenue, so a higher mix of usage generates higher revenue and cash flow from existing infrastructure, a helpful offset to the financing and credit worries weighing on the group.

Importantly, businesses and consumers adopt AI for productivity and reliability, not because a shiny new model just launched. As most use cases run on older or open-source models, not cutting-edge (and far pricier) versions, slower releases shouldn't dent adoption much.

Our work also suggests personal AI assistants will be a major new driver of consumer demand. Combined with steadily rising business adoption, our view that total demand keeps climbing stays intact.

What it means across the market

The shift from building to using will create winners and losers:

• Likely to benefit: Software companies (a slower pace gives them time to catch up), particularly data infrastructure, cybersecurity, and "observability" firms that help monitor and manage AI systems, as we wrote last week. Memory chip suppliers could also gain, since inferencing workloads are memory-intensive.
• Likely under pressure: The most expensive, high-valuation AI stocks, especially if rates stay elevated. We see "neoclouds," the smaller AI-focused data-center providers, as most exposed, given their reliance on the leading labs and training workloads.

What we're watching

While a slower pace may pose a near-term headwind, our long-term optimism is intact.

The following will shape our view:

• How much will training truly slow? While AI leaders are calling for slower model development, the extent of any reduction in training activity remains unclear.
• How long would a slowdown last? A few months means something very different from a few years.
• Is the industry aligned? Leaders at Anthropic, OpenAI, xAI, and Google have backed different steps, and it's far from clear they'll agree on a shared path.
• Will AI become a political flashpoint? Expect AI, safety, jobs, energy demand, and competition with China, to move up the policy agenda heading into the 2026 midterms and the next presidential cycle. Regulation is the wildcard: it could shape the pace of development as much as the labs themselves.

Market positives in the sea of gloom

Stepping back, even as uncertainty has risen, both sentiment and technology valuations have undergone a significant reset.

• The tech sector's forward price-to-earnings (P/E) has fallen from 32x last October to 21x today, a roughly 35% contraction that brings valuations back to levels seen when ChatGPT was introduced in late 2022.
• As a result, investors are paying no more for technology stocks than they were before the AI era began. In our view, much of today's uncertainty is already reflected in valuations.
• Meanwhile, tech earnings momentum remains strongest among all of the S&P 500 sectors. Over the past three months, forward earnings estimates have increased 19%, compared with just 10% for the next-strongest sector.

A model development slowdown could cut both ways. A reallocation from training to inference workloads will likely drive higher near-term revenue and cash flow for the model builders. We see this as a helpful offset to the financing and credit worries that have weighed on AI-related stocks.

Stocks are also moving more on their own merits than in decades, less in lockstep than at any point in 30 years, as the market sorts AI winners from losers. We expect that to continue.

And even with new guardrails, our view is that AI use is still in its infancy. While 88% of organizations now use AI in some form, only about 7% have fully scaled it, per McKinsey.

Whole sectors, including healthcare, financial services, legal, manufacturing, and beyond, are just getting started, and even the models already deployed inside these companies are nowhere near their full potential.

Bottom Line

A slowdown in AI model training would represent a shift in demand, not the end of it. The near-term ride may be bumpier and the most expensive names more vulnerable, but the valuation reset and still-early state of AI adoption keep us constructive on the long-term story. We'll stay balanced and open-minded as this once-in-a-generation shift plays out at a blistering pace.