What the AI pacing debate means for the industry

Anthropic chief executive Dario Amodei has called for a more measured pace of frontier AI development, saying safeguards need time to keep up with rapidly advancing capabilities. The proposal has received support from several leading AI figures, although it remains unclear whether it will lead to voluntary safeguards or formal restrictions.

Markets initially reacted negatively, reflecting concerns that slower frontier-model development could weaken demand for AI chips and data-centre infrastructure. However, there is little evidence that developers or hyperscalers have materially changed investment plans.

A shift in the AI safety debate

Amodei's essay, titled We Must Pace the Frontier, followed a series of public comments from current and former employees of Anthropic and OpenAI expressing concerns about safety practices at frontier AI developers.

His objective is not to halt development but to reduce risks ranging from cyber misuse and economic disruption to the potential loss of control over advanced systems.

Support from OpenAI chief executive Sam Altman, Google DeepMind chief executive Demis Hassabis and Elon Musk suggests concern about AI safety is becoming more widespread within the industry. 

Investor interpretation broadly falls into two camps. The first is that frontier AI developers are genuinely concerned about capabilities emerging from the latest generation of models. Given their visibility into systems months before public release, they may be seeing risks not yet apparent to outsiders.

The second interpretation is more commercial. Development costs continue to rise, competition from open-weight and Chinese models is increasing and governments are showing greater regulatory interest. In this environment, leading developers may have incentives to shape any future regulatory framework.

The reality may contain elements of both. For now, markets appear focused on the possibility that slower development could reduce demand for AI infrastructure.

Training versus inference

Understanding the distinction between training and inference is critical to the AI pacing debate.

Training is the process by which AI models learn from vast datasets using large amounts of computing power. Inference occurs when existing models answer questions, analyse information or perform tasks.

We estimate the majority of current AI workloads are inference. Although training consumes the most advanced infrastructure, inference accounts for a much larger number of recurring tasks.

This distinction matters because calls for pacing appear focused primarily on slowing frontier-model releases rather than stopping model training altogether. Additional testing, evaluation and governance would likely extend deployment timelines but not necessarily reduce compute demand proportionately. Activities such as safety testing, auditing and interpretability research are themselves computationally intensive.

Inference demand should continue growing

Even if frontier-model development slows, inference demand is likely to continue expanding.

Existing models are already capable of software development, customer support, document analysis, research and workflow automation, while businesses remain in the early stages of adoption. AI agents could accelerate this trend because a single task may require multiple model requests.

At the same time, falling costs should make AI economical across a wider range of applications. As a result, broader adoption and heavier usage could drive strong growth in inference even if model improvements become more gradual.

What are the implications for the AI ecosystem?

Hyperscalers

A slower pace of frontier-model development would reduce the need for ever-larger training clusters. However, we believe Amazon Web Services, Microsoft Azure and Google Cloud could redirect much of that capacity towards inference, fine-tuning and broader cloud workloads.

The principal risks are likely to be lower utilisation rates, higher depreciation costs and weaker returns on capital rather than a significant decline in cloud revenues.

Semiconductor companies

The impact on semiconductor demand is likely to be more about composition than scale.

NVIDIA appears most exposed because its leadership position is strongest in premium training infrastructure, although it remains well placed in inference. Broadcom could benefit as hyperscalers invest further in custom AI chips, while TSMC appears comparatively insulated because it manufactures both merchant GPUs and custom processors.

Demand for high-bandwidth memory, advanced packaging and networking equipment could remain resilient as inference workloads become increasingly sophisticated.

The more significant risk would emerge if the safety debate ultimately reflected deteriorating economics from frontier scaling and diminishing returns from larger training runs. At present, we see little evidence supporting that scenario.

Data-centre infrastructure

Data-centre infrastructure companies face a mixed outlook.

Slower frontier development could reduce demand for the largest training campuses and associated cooling and networking systems. However, inference workloads operate continuously and should continue to support demand for power, grid connections, cooling and related infrastructure.

The result may be a shift towards geographically distributed inference facilities rather than a handful of enormous training campuses.

Neoclouds

Neocloud providers, which specialise in renting AI computing capacity, appear more vulnerable.

Many have invested heavily in fleets of AI processors supported by debt or lease obligations. If frontier training demand weakens, utilisation and rental pricing could come under pressure.

Although capacity can be redirected towards inference workloads, these generally generate lower revenue per unit of compute and face greater competition from hyperscalers. The key issue will be whether utilisation remains high enough to support existing financing assumptions.

Software companies

Software companies may be among the relative beneficiaries of slower frontier development.

A more stable model environment gives developers time to integrate AI into products and workflows without constantly adapting to new releases. Competition could shift towards customer relationships, proprietary data and industry expertise.

The biggest beneficiaries are likely to be companies already generating measurable value from existing AI capabilities rather than those dependent on future breakthroughs.

Conclusion

The AI safety debate deserves attention, but we do not currently view it as thesis-altering for investors.

A more measured pace of frontier-model development could moderate demand for premium training infrastructure and soften growth expectations in some areas of the AI ecosystem. However, there is little evidence that major AI developers or hyperscalers are materially reducing investment plans, altering product roadmaps or scaling back capital expenditure.

Most importantly, inference demand continues to grow through wider adoption, lower costs and increasing use of AI agents. As a result, AI computing demand is unlikely to disappear even if frontier development becomes more controlled.

We therefore remain constructive on the long-term adoption of AI, while recognising that outcomes may vary considerably across companies and sectors. 

The key questions are:

• Will voluntary safety measures evolve into binding restrictions?
• Will inference growth absorb existing capacity?
• Will utilisation, pricing and investment trends begin to change materially? 

For now, those risks warrant monitoring but do not appear sufficient to change our broader view of AI adoption. 

Important information

References to individual companies are provided for illustrative purposes only and should not be construed as investment recommendations.