AI Safety May Require More Compute, Not Less: What It Means for Nvidia and AI Infrastructure
The debate over slowing artificial intelligence development is increasingly turning into a debate over something investors may not have expected: how much computing power will be needed to make advanced AI systems safer.
A research report from Citi, following discussions at the AI Infra Summit in Santa Clara, suggests that tighter AI safety and alignment requirements could increase infrastructure spending rather than reduce demand for chips and data centers. Safety testing, monitoring, red-teaming and other evaluation processes can themselves require substantial computing resources.
That distinction matters for investors. A slowdown in the release of frontier AI models does not necessarily mean a slowdown in spending on GPUs, networking, power and data-center capacity. In some scenarios, spending could shift from simply building larger training clusters toward infrastructure used to test, monitor and control increasingly capable AI systems.
Why AI Safety Needs Compute
AI safety covers a broad range of activities designed to identify and reduce harmful or unexpected behaviour in AI systems.
That can include testing models before deployment, running adversarial evaluations, monitoring AI agents while they work, checking whether safeguards can be bypassed and using additional models to evaluate the behaviour of another model.
These processes consume computing power.
As AI systems become more capable and operate for longer periods, the amount of information that needs to be evaluated can also increase. For example, an AI agent completing a complicated coding task may generate a long sequence of actions rather than simply returning a short answer.
Researchers are already developing monitoring systems specifically designed to examine AI-agent behaviour. Apollo Research says its monitoring work aims to translate additional compute into scalable security and has been researching systems that monitor coding agents and their trajectories.
This creates an important economic distinction: using AI to improve AI safety can itself become an AI-compute workload.
The Debate Is About Pacing, Not Necessarily Stopping AI
The recent discussion around AI safety has partly centred on whether frontier AI development should be slowed so researchers and regulators have more time to evaluate increasingly capable models.
Anthropic CEO Dario Amodei has called for a more cautious pace and stronger safeguards, while OpenAI CEO Sam Altman has also supported greater coordination and security measures, although the two executives have framed the issue somewhat differently. Other technology leaders, including Nvidia CEO Jensen Huang, have argued against slowing technological progress.
For investors, the important point is that “pacing” does not automatically mean “less AI infrastructure.”
An investment framework from Franklin Templeton makes a similar distinction: capital could shift away from some very large training runs toward safety infrastructure such as testing, red-teaming, interpretability, monitoring and cybersecurity. Those activities can still require significant computing and engineering investment.
In other words, the composition of AI spending could change even if overall infrastructure demand remains strong.
Nvidia Sees a Much Heavier Compute Burden From AI Agents
The shift from conventional chatbots to autonomous or semi-autonomous AI agents is another reason infrastructure requirements could rise.
According to Citi's summary of discussions at the AI Infra Summit, Nvidia vice president Ian Buck said agentic AI workloads can be roughly 100 times more demanding than 2023-era chatbot applications. The increased requirements are associated with factors such as longer input sequences, larger KV caches, multi-turn interactions and the use of multiple sub-agents.
The comparison should not be interpreted as meaning every AI application will consume 100 times as much compute. It describes a particular class of increasingly complex workloads.
Nevertheless, the direction is important.
A chatbot that answers one question and stops has relatively limited computational requirements. An AI agent that researches information, writes code, checks its own work, calls external tools and repeats the process can require considerably more inference and monitoring.
That creates another source of demand for AI chips.
Data Centers Are Being Designed for More GPUs
The infrastructure build-out is also changing physically.
Citi's report said next-generation data centers are being designed to accommodate 30% to 40% more GPUs, while companies are working on software and networking improvements to extract more useful work from increasingly large clusters.
That means the AI infrastructure opportunity extends beyond GPUs.
Large AI clusters require:
Advanced processors
High-speed networking
Memory and storage
Cooling systems
Electricity generation and transmission
Data-center facilities
Power-management equipment
Software for orchestration and monitoring
As a result, AI spending is becoming an increasingly broad infrastructure theme rather than a simple semiconductor story.
Meta's Expanding AI Infrastructure Shows the Scale
Meta is another example of how quickly the physical requirements of AI are expanding.
At the AI Infra Summit, Meta's infrastructure chief Santosh Janardhan discussed the company's expansion from its 1-gigawatt-plus Prometheus cluster toward a 5-gigawatt Hyperion footprint over the coming years. The company has also been working on networking architecture designed to reduce disruptions in massive GPU clusters.
A gigawatt is a measure of power capacity. A 5-gigawatt AI infrastructure footprint therefore represents an enormous electricity requirement, illustrating why energy availability is becoming an important constraint alongside chip supply.
This is one reason AI infrastructure investment is increasingly connected to utilities, power equipment, data centers and grid development.
Why AI Safety Could Increase Infrastructure Spending
There are several reasons safety investment could add to compute demand rather than simply replace existing spending.
First, a company may need to test a model before releasing it.
Second, it may need to continuously monitor the model after deployment.
Third, researchers may use AI systems to evaluate other AI systems, creating additional inference workloads.
Fourth, developers may run multiple versions of a model during safety testing rather than testing only the final version.
Finally, increasingly autonomous AI agents may require monitoring throughout their operation rather than only checking the final output.
The result is a potentially larger computing footprint surrounding every major AI model.
What It Means for Nvidia and the Semiconductor Industry
For Nvidia, the argument is particularly relevant because the company supplies much of the high-performance computing infrastructure used for AI training and inference.
If AI safety becomes a larger and more continuous workload, demand for accelerators could remain strong even if regulators or companies impose restrictions on how quickly frontier models are released.
However, this does not guarantee that Nvidia's revenue or stock price will rise indefinitely.
The semiconductor market remains exposed to competition, customer concentration, chip cycles, supply constraints, changing architectures and the possibility that customers develop more specialized hardware.
Intel is also attempting to position itself as a broader AI infrastructure provider. CEO Lip-Bu Tan said at the summit that Intel's strategy is moving beyond its traditional CPU role toward a fuller AI infrastructure offering.
That illustrates how the AI infrastructure market is becoming increasingly competitive.
The Bigger Opportunity May Be Beyond GPUs
One of the most important implications for investors is that AI safety could increase demand across the entire computing ecosystem.
If companies need more compute for monitoring and evaluation, they will also need more:
Networking: Large clusters need extremely fast connections between processors.
Power: AI data centers require enormous and reliable electricity supplies.
Cooling: High-density computing generates substantial heat.
Data centers: More workloads require additional physical capacity.
Cybersecurity: AI systems need protection against attacks and misuse.
Software: Companies need tools to coordinate, evaluate and monitor AI workloads.
This broadens the potential economic impact of AI spending.
Risks Investors Should Consider
The bullish infrastructure argument still has important uncertainties.
AI companies could develop more efficient models that require less compute for particular tasks. Specialized accelerators could also improve performance per unit of electricity.
Regulation could restrict certain AI applications, while economic conditions could force companies to delay capital-intensive projects.
There is also a valuation risk. Strong AI infrastructure demand does not automatically mean every semiconductor or data-center company will generate superior financial returns. Investors still need to consider earnings, capital expenditure, competition and the price already reflected in a company's valuation.
The AI safety debate itself also remains unsettled. Experts disagree over the probability, timing and severity of extreme AI risks, while there is broader agreement that cybersecurity, misuse and reliability deserve attention.
What Investors Should Watch Next
The most useful indicators will be actual spending rather than headlines about AI safety.
Investors should monitor:
Hyperscaler capital expenditure
GPU and accelerator demand
Data-center construction
AI inference growth
Networking equipment orders
Electricity demand from data centers
Spending on AI monitoring and cybersecurity
Semiconductor gross margins
New AI-model efficiency improvements
If safety requirements increase while AI companies continue expanding infrastructure, the result could be a market where compute demand remains strong even as the pace of model deployment becomes more controlled.
AI Safety Could Become Another Source of Compute Demand
The emerging AI-safety debate does not necessarily translate into less infrastructure spending. Evidence from industry discussions suggests that testing, monitoring, alignment and agentic-AI workloads can require substantial additional computing power.
For investors, the key takeaway is to distinguish between slower AI deployment and lower AI infrastructure demand. They are not the same thing.
If advanced AI systems require more testing and continuous monitoring as they become more capable, safety could become another significant workload for GPUs, networking systems and data centers.
The next phase of the AI boom may therefore be measured not only by how quickly new models are released, but also by how much infrastructure is required to train, operate, evaluate and secure them.
Follow the blog for more updates on artificial intelligence, semiconductor companies, technology infrastructure and global markets.
This article is for informational and educational purposes only and should not be considered investment advice.