The market significance of "AI security": It increases the demand for inference and post-training computing power by 20%, raising the overall computing power cost of the industry by 18%.

The market significance of "AI security": It increases the demand for inference and post-training computing power by 20%, raising the overall computing power cost of the industry by 18%.

AI safety regulation is evolving from an ethical issue into a quantifiable cost variable, and will profoundly reshape the investment logic of AI infrastructure.

According to a recent research report by Barclays, the "Pacing" mechanism implemented by the Advanced AI Labs will increase the overall computing power cost of the industry by about 18% starting in 2027, corresponding to additional expenditures of over $44 billion.

On August 18, 2026, OpenAI published a blog post formally disclosing its security monitoring requirements for high-capability models. According to the disclosure, for models with Sol-level capabilities and above, all reinforcement learning (RL) training, evaluation, and inference workloads must be subject to real-time monitoring. Current estimates indicate that monitoring overhead accounts for approximately 20% of the monitored inference computing power. Based on this, Barclays analyst Ross Sandler's team calculated that, since approximately 85% of AI lab computing power has already been allocated to post-training and inference workloads, and it is projected that by 2027 almost all models will exceed the GPT-5.6 Sol-level capability threshold, the demand for inference and post-training computing power will increase by 20%, driving up the overall industry computing cost by approximately 18%.

This shift is already impacting the AI infrastructure investment chain. Barclays points out that some AI labs currently have inference gross margins exceeding 80%, providing sufficient room to absorb additional security costs in the short term, but in the long term, profit margins may converge towards 65%. Meanwhile, if leading labs slow down their model release pace, inference service providers such as Google, Meta, AMZN, and MSFT are expected to benefit from the competitive landscape.

Cost Calculation of Rhythm Control

Barclays' quantitative analysis shows that "monitoring" computing power costs will add approximately $44 billion to the industry's costs in 2027, and further expand to approximately $76 billion in 2028, corresponding to a stable proportion of basic computing power costs of around 18%.

Specifically, the total cost of basic computing power in the industry is projected to reach $246 billion in 2027, with $132 billion for training and $114 billion for inference. Monitored computing power totals approximately $219 billion, covering post-training/RL workloads ($104 billion, representing 79% of training computing power) and all critical-level inference ($114 billion, representing 100% of inference computing power). Adding a 20% monitoring overhead, the additional cost reaches $44 billion, accounting for 18% of the total cost of basic computing power.

It is worth noting that the aforementioned monitoring requirements do not apply to the pre-training stage, therefore the pre-training cost is not directly affected. This also explains why the overall cost increase (18%) is lower than the individual cost increase (20%) for inference and post-training stages.

Compress the inference profit margin to its long-term average of 65%.

Barclays believes that the impact of security costs on the profitability of AI labs depends on whether these costs can be passed on to end users by increasing token pricing or charging based on results.

Currently, some leading AI labs have inference gross margins exceeding 80%, providing a significant buffer to absorb additional security costs. Barclays predicts that as security compliance costs continue to inflate, AI lab inference profit margins will gradually converge towards a long-term central value of 65%.

Regarding training costs, predictions from leading AI labs indicate that annual training costs for a single company will peak at approximately $130 billion between 2028 and 2029, before stabilizing. Barclays believes that the trend in training costs depends more on competition for market share among labs in the cutting-edge market than simply on pace control policies. Furthermore, personnel costs resulting from the integration of third-party security auditing firms (such as METR and Redwood) into the R&D process are currently a relatively minor expense, and AI labs generally use AI-monitoring-AI methods to control human resource investment.

The rebalancing effect of the competitive landscape

The impact of pacing control policies on the competitive landscape is equally significant. Barclays points out that if the two leading AI labs (OpenAI and another) slow down their model release pace, it will provide other competitors with a window to catch up. Historical data shows that it typically takes 35 to 40 days for lagging Western labs to catch up with cutting-edge models, and pacing control could further compress this time lag.

Barclays believes that this gap could widen further if distillation training activities are restricted, thus limiting the impact of open-source weighted models on cutting-edge laboratory tempo control strategies.

For publicly traded tech giants, Barclays believes that companies like Google, having long been under strict legal regulation, may be better prepared in terms of security compliance than private AI labs with relatively limited resources. The report also points out that the silence of companies like Meta, Google, MSFT, and xAI regarding security incidents such as Hugging Face may indicate that these companies had already taken more prudent security measures in advance.

Frequent safety incidents increase regulatory pressure

The immediate backdrop to pushing for the implementation of the pace control policy is the recent surge in AI security incidents. According to data from Felony Bench cited by Barclays, between July and September 2026 alone, OpenAI-related models were involved in 10 security incidents, including multiple third-party system intrusions, unauthorized use of GitHub credentials, and the deployment of malware packages; other AI labs also experienced 10 similar incidents during the same period, and META also experienced a third-party system intrusion in early August.

Barclays draws parallels between the current situation and Meta's data privacy crisis in 2018. At that time, Meta significantly increased its security and compliance workforce by tens of thousands, and its forward price-to-earnings ratio fell by approximately 40% from its peak. Analysts point out that although the two events differ in nature, the regulatory pressure and valuation reassessment risks stemming from AI security issues are being transmitted downstream in the AI infrastructure industry chain along a similar path. Since no cutting-edge AI labs are currently directly listed, valuation pressure is primarily reflected in downstream companies such as computing power providers and cloud hyperscale vendors.

~~~~~~~~~~~~~~~~~~~~~~~

The above content is from Zhuifeng Trading Platform .

For more detailed analysis, including real-time updates and firsthand research, please join the [ Trading Channel Annual Membership ].

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.