Anthropic CEO Dario Amodei recently wrote a blog post where he argued that advancements in AI were moving too quickly for safety protocols to keep up. He argued that development of these frontier models, the most cutting edge AI models, must be “paced” to avoid the potential danger of AI doing something destructive. This blog post came on the heels of some high profile instances where AI agents “escaped” the guardrails put around them and did something unintended. In the last few months, all three of the largest frontier AI companies have had test agents hack into other companies, including Google Gemini, OpenAI’s GPT, and Anthropic’s Claude. A few researchers have even warned that without better protections, AI has some chance of killing all of humanity.
Several prominent researchers have pushed back on the extinction theory specifically, but it is clear that AI safety is going to be a key focus going forward. Within hours of Amodei’s posting, OpenAI CEO Sam Altman said he agreed that pacing frontier models was necessary. The following week, leaders from Anthropic, OpenAI and Google’s DeepMind were reportedly collaborating on new industry-wide safety recommendations.
AI safety could be one of the most important stories for financial markets in the coming months. There is a lot of uncertainty about what exactly “pacing” of these most advanced models will mean. Some even think this pacing of frontier models could burst the AI bubble. We think the reality is more nuanced. Implementation of better AI safety is more likely to shift data center demand, which could create some winners and losers. Here are our thoughts about why this story is important for investors and how we are approaching its impact in portfolios.
How much computing power is used to train AI models?
Exact measurements are unavailable, but we can estimate that training of the most advanced AI models might account for 20-25% of total AI computing power usage.
The boom in AI stocks has been a function of the huge spending on data center construction and the equipment needed inside those data centers to power AI. Computing usage for AI breaks down into two big categories:
- Training: This is the initial creation of a large language model or LLM. This requires extremely heavy computing power. The cost to train these models has grown as the models have become larger and more sophisticated.
- Inference: This is what happens when you type a prompt into an AI chatbot, or when someone uses AI to complete some task. The computing cost for inference is far lower than training, and can be metered depending on the complexity of the task. When you hear about someone paying for tokens, that is inference.
According to consulting firm McKinsey, inference and training each make up about half of AI computing power usage today. It isn’t easy to estimate exactly how much of this training power is being used by the frontier model makers who might be considering slowing training. AI research firm Epoch AI estimates that OpenAI and Anthropic are probably using something like 20% of total global AI usage. These two companies are almost certainly using more training compute than inference, so they probably account for more like 25% or 30% of training usage. If we add in other frontier model training from companies like Google DeepMind, xAI, Meta’s AI, Mistral, etc., it may be that frontier model training is accounting for 40-50% or more of total training computing power. Since training represents about half of total AI compute usage, we can guess that frontier model development is using around 20-25% of total AI data center capacity.
Note that this estimate really is just a guess based on what little public data is available. However, it is good enough to give us some context of how much data center demand could be impacted by any slowing of AI training.
How could pacing AI development impact data center demand?
While the exact impact of pacing AI development is unknown, it could shift data center demand from training to inference rather than reducing overall usage.
Certainly if demand for data center capacity suddenly dropped by 25%, that would be a huge negative for stocks. That probably would create a bubble-bursting kind of market crash. But of course, “pacing” new model training doesn’t mean eliminating training altogether. Exactly what “pacing” implies is impossible to guess, but if we were to assume that the biggest frontier AI companies slowed their new model creation by 10-20%, that would imply a 2.5-5% decline in compute demand.
A decline in training demand of that magnitude could easily be absorbed in extra demand for inference. OpenAI and Anthropic are talking about pacing new training for purposes of bolstering AI safety. That implies they will be doing more testing of models before they are released to users, which would increase inference usage.
Moreover, any net easing of compute demand could result in a lower cost for companies to access AI models. McKinsey estimates that inference demand will grow much faster than training demand in the coming years in part because the cost of compute declines. It could be that the pacing of frontier models accelerates that trend.
AI training vs. inference: which stocks could benefit?
Training an AI model requires a different kind of data center with different kinds of equipment compared to inference use. That may shift who the winners are in the AI sector.
Training an AI model is a long, involved process. Think of it like taking a final exam covering a whole semester’s worth of material. Depth and breadth are much more important than speed. Inference is more like a timed pop quiz. You need a narrow answer to a specific question, but you need that answer quickly.
This divergence in needs means that each task requires different kinds of equipment, and even different kinds of data centers. The table below compares the needs of a data center focused on training vs. one focused on inference.
| Feature | Training Data Center | Inference Data Center |
|---|---|---|
| Graphical or Central Processors (GPU and CPU) | Requires thousands of top-of-the-line GPUs linked together to act like one giant supercomputer doing heavy calculations. | Can use more standard CPUs and/or GPUs, with a greater focus on cost efficiency than computational power. |
| Memory chips | Needs enough memory to hold the vast math formulas being calculated during the learning phase. | Heavier memory use, focused on speed rather than capacity. |
| Data Center Location | Location less important, usually in rural areas due to cost of land and/or close to power supplies. | Located near major cities to improve speed to users. |
| Power Usage | Massive power needs, which need to run continuously for months at a time. | Power usage goes up during the day when people are using the app and drops at night when people sleep. |
| Cooling Needs | The chips run so hot that air fans aren't enough. Liquid coolant must be piped directly over the chips to keep them from melting, often resulting in heavy water usage. | Uses traditional air fans or lighter liquid cooling doors because the heat is spread out over more servers. |
| Data Storage | Focus on storage capacity - slower hard drives are adequate. | Focused on speed of retrieval from storage - need faster drives and/or heavier memory usage. |
Sources: McKinsey & Company, Iron Mountain, International Energy Agency
Note that today, capacity is tight at both kinds of data centers. Hence a mild slowdown in training and a commensurate increase in inference wouldn’t necessarily be a total game changer for stocks. However, this table does highlight how yesterday’s AI winners might not be tomorrow’s winners. For example, Nvidia has been focused on ultra powerful GPUs, which are ideal for training but wouldn’t be worth the expense for inference. Similarly, there have been a number of smaller companies that produce highly specialized equipment for things like cooling or power management. Some of these stocks have surged due to training demand. It is possible growth for those companies slows in the future.
For Facet, we think this is just another example of why trying to pick stock market winners is so difficult. At one time we might have thought of “AI stocks” as a monolith, but it is very possible that there’s a divergence among companies producing AI-related equipment going forward.
How will AI be regulated in the future?
Future AI regulation could be government-led or overseen by a self-regulatory industry group, with the current push for AI safety likely forming the foundation of these new rules.
Currently there is little regulation specific to AI. However, that is very likely to change. AI is already a major political issue in the upcoming mid-term elections. Almost 70% of Americans oppose data center construction in their area, according to a May 2026 Gallup poll. A separate July 2026 poll indicated that 79% of Americans expect AI to reduce the total number of jobs. If we add to this worries that AI could literally destroy humanity, it seems inevitable that politicians will want to reign in AI.
This new push for AI safety from Anthropic and others could be their attempt to get ahead of impending regulation. In his blog post, Amodei indicated support for an industry-led oversight committee which would partner with governmental agencies. There is precedence for this. The Financial Industry Regulatory Authority or FINRA is a self-regulatory entity. This means it is actually run by its “member” firms, in this case Wall Street broker-dealers. They collectively set their own rules, usually with approval from the SEC or other governmental entities. In a complex and rapidly changing industry, it may be that industry participants can react more quickly than the government can when new rules are needed. It could be that a similar system could work for AI.
Currently this approach is not favored by the White House. President Donald Trump seems focused on “winning” the AI race against China, going so far as dismissing safety concerns as a “hoax.” Instead, the Administration plans to create its own AI task force headed up by an AI Czar. The President could use anti-trust laws to prevent the major AI firms from forming their own safety committee.
We note that the anti-trust concerns have some merit. Right now, Anthropic, OpenAI and others are under extreme competitive pressure to develop ever more powerful models. Training these models is costing AI companies billions of dollars. It would probably be to their financial benefit to slow the pace of new model production in a coordinated fashion. That way they could spend less money while not losing any competitive advantage.
It isn’t easy to see exactly how all these factors will influence the eventual regulatory regime for AI. However, we expect the contours of such a regime will start to become more clear in the coming months. Investors will be scrutinizing how emerging regulation might impact investment in AI.
Does the debt used to build AI data centers create market risks?
While building data centers with debt creates risks for individual company stocks, widespread defaults are very unlikely because highly profitable big tech companies overwhelmingly guarantee this debt.
Data centers are increasingly being financed using a significant amount of debt, which does create some concerns. During the internet boom of the 1990’s, some telecom companies took on huge debt loads to build out fiber cable networks. After the dot com bust, a number of these firms went bankrupt.
Today, data center construction is overwhelmingly directly financed or guaranteed by big tech companies. The guarantors fall into three big categories:
- Cloud computing companies such as Microsoft or Amazon. These companies are hoping to profit from other companies renting this cloud computing capacity.
- AI model owners such as Alphabet or Meta. In this case, the company might mostly use this data center themselves, probably more for inference than training. They may also rent available capacity to other companies.
- Semiconductor manufacturers such as Nvidia or Broadcom. In this case, the company uses the debt guarantee to entice the data center to buy chips from the manufacturer. These are more often training-oriented data centers.
Even if a data center were going to be entirely leased by OpenAI or Anthropic, the construction costs would be underwritten by one of these more established companies. In an extreme case, where OpenAI and/or Anthropic slowed model training substantially, this could result in excess capacity at data centers, which in turn could result in weaker return on investment than the data center owner assumed.
However, even in this extreme case, it is unlikely this would result in debt defaults. All of these big tech companies are highly profitable and can absorb the cost of backing this debt without creating financial stress, even if profits from the data center are not as strong as expected.
That’s not to say there wouldn’t be financial risks in such a scenario. The equity owners of the data center, which is often a real estate company or private equity, could see their investment severely impaired if big tech has to make good on their guarantee. Moreover, the cost of paying out the guarantee would decrease profits for the guarantor company. That could certainly weigh on that company’s stock.
There are also some companies with heavier debt loads than others, such as Oracle or CoreWeave, and these companies could be more vulnerable. Part of Facet’s approach to tech investing is to underweight companies with higher debt burdens, lower profit margins, and more volatile revenue. There are a variety of scenarios where these more indebted companies could struggle, including one where there is overcapacity in data centers.
How is Facet approaching tech and AI stocks today?
Because the effects of pacing AI training are difficult to predict, Facet focuses on preparing for multiple scenarios and avoiding individual stock risks rather than trying to guess specific market outcomes.
It is possible that the pacing of AI model training markets an inflection point in demand for compute. It is also possible it merely results in a shift of the kind of compute that is in demand. That shift could be good for some companies and bad for others.
This is why we prefer to build our investment approach around a variety of possible scenarios. AI is a rapidly evolving technology. Predicting exactly how it will impact the economy as a whole, much less any specific company is almost impossible in our opinion. We believe by balancing portfolios to be ready for a variety of scenarios, we could perform reasonably well across each scenario. Gambling on one specific outcome could produce big returns, but also could take on large losses.
Here are a few key thoughts about how Facet is currenting thinking about tech and AI investing:
- Avoid individual stock risk: The history of tech stocks is littered with companies that were leaders for a period of time but then fell by the way side. AI is arguably the fastest moving technology of our lifetimes. If you are betting heavily on one stock, it may be a time to consider diversifying.
- Underweight more speculative companies: Companies with low profit margins, weak cash flow, volatile revenue, and heavy debt burdens could be most vulnerable if tech stocks hit a rough patch. Through our ETF mix, Facet is underweight these kinds of companies.
- Stay invested in tech stocks generally: The heavy spending on AI infrastructure isn’t going away, even if there’s some kind of shift. There are a lot of profits to be made by companies benefiting from this spending. Facet is avoiding the most speculative companies, but is still keeping a healthy weight to tech stocks.
