This is good for AI!!!! Aaargh!!!!

There was a 2017 meme . . . hang on, let’s find it . . .

... that can easily be repurposed for current circumstances.

Here, for example, is Citigroup analyst Atif Malik on the recent campaign by incumbent AI labs to spook regulators into building them some barriers to entry:

Fundamentally, frontier AI model developers see AI safety and alignment as a compute-intensive process, implying that safety concerns may drive more, rather than less, demand for compute. This runs counter to a key investor concern that tighter safety requirements could slow the pace of frontier model development - and, by extension, compute demand.

Fundamentally, frontier AI model developers see AI safety and alignment as a compute-intensive process, implying that safety concerns may drive more, rather than less, demand for compute. This runs counter to a key investor concern that tighter safety requirements could slow the pace of frontier model development - and, by extension, compute demand.

Malik is reporting back from day one of the AI Infra Summit in Santa Clara, California. You’ll surely be shocked to learn that AI people remain bullish about AI:

Frontier AI model developers do not see compute scaling slowing down. Instead, the bottleneck is shifting from model architecture innovation alone towards securing massive amounts of compute, power, and infrastructure, while improving efficiency through platforms such as [OpenAI’s] Astra and next-generation hardware like [Nvidia’s] Vera Rubin. The implication is continued strong demand across the AI infrastructure stack: GPUs, networking, power, cooling, and data center capacity. Next-generation facilities could accommodate 30-40% more GPUs, improving infrastructure efficiency and helping support growing AI demand.

Frontier AI model developers do not see compute scaling slowing down. Instead, the bottleneck is shifting from model architecture innovation alone towards securing massive amounts of compute, power, and infrastructure, while improving efficiency through platforms such as [OpenAI’s] Astra and next-generation hardware like [Nvidia’s] Vera Rubin. The implication is continued strong demand across the AI infrastructure stack: GPUs, networking, power, cooling, and data center capacity. Next-generation facilities could accommodate 30-40% more GPUs, improving infrastructure efficiency and helping support growing AI demand.

Neat!

Over at Morgan Stanley, analyst Adam Jonas continues to type words into a computer. His big idea for clients on Tuesday was to buy SpaceX as a “doom hedge”:

SpaceX is a manufacturing and infrastructure company engaged in the conversion of energy into intelligence at scale through its vertically integrated space launch, sat comms, and computing stack. In other words, SpaceX is an AI company. As such, any material slowdown in the AI trade is - all else equal - a headwind for the stock. The majority of SpaceX’s capex and forward earnings growth is driven by the Enterprise AI business. You can’t have it both ways. However, for investors who want resilient exposure to the long-term trade while mitigating some of the nearer-term volatility, we believe SpaceX may continue to exhibit greater stability vs. other its [sic] peers in an AI portfolio. Why?

SpaceX is a manufacturing and infrastructure company engaged in the conversion of energy into intelligence at scale through its vertically integrated space launch, sat comms, and computing stack. In other words, SpaceX is an AI company. As such, any material slowdown in the AI trade is - all else equal - a headwind for the stock. The majority of SpaceX’s capex and forward earnings growth is driven by the Enterprise AI business. You can’t have it both ways. However, for investors who want resilient exposure to the long-term trade while mitigating some of the nearer-term volatility, we believe SpaceX may continue to exhibit greater stability vs. other its [sic] peers in an AI portfolio. Why?

Fair question. Jonas has five answers.

SpaceX is diversified, being mostly a telecoms network operator, but also offers “a rather long-dated option” on planetary colonisation and so forth. It has about $61bn of net cash (which Morgan Stanley helped raise) and optionality around when to go for more (which Morgan Stanley will help it raise). Moreover, the satellites and the rockets justify three-quarters of its current valuation, at $118 and $8 per share respectively, so the AI bit is nearly free.

Reason five is Elon.

During challenging times, companies with superior market intelligence and execution can stand out from the pack. From our experience in observing the management style of the SpaceX CEO over the past 2 decades, he is not one to let a good crisis go to waste. Some of Tesla’s best performances - both operationally and with respect to share price - took place in response to or during elevated market volatility and uncertainty.

During challenging times, companies with superior market intelligence and execution can stand out from the pack. From our experience in observing the management style of the SpaceX CEO over the past 2 decades, he is not one to let a good crisis go to waste. Some of Tesla’s best performances - both operationally and with respect to share price - took place in response to or during elevated market volatility and uncertainty.

Awesome!

It’s not all like this. Edison Lee, of Jefferies’ software team, reaches a less positive conclusion after doing niggly stuff like looking at data.

Lee’s most recent note highlights that, in the past two weeks, four new suppliers have appeared on Artificial Analysis’s list of top labs benchmarked by large-language model intelligence. All four have required relatively modest investment, three use open weights, and none is Chinese by origin (though one, Apodex, is backed by Chinese private capital).

LLM pricing power is being eroded, Lee says, as new entrants catch up with the frontier by using distillation (meaning they avoid having to train the model from scratch by leeching a larger LLM for probability data).

He also notes that, after Anthropic and OpenAI raised prices for their most powerful models, the average discount offered by Chinese labs versus their American counterparts widened from 60 per cent in August to 80 per cent in September:

The price hike decisions are likely driven by the need to show the capital markets their rising model intelligence would support higher pricing, so that investors would not worry about margin or ROI pressure. However, we believe the fact that the number of players (even just in the US there are 7) is rising, and the model intelligence gap is narrowing indicate this is an overcrowded market with excessive investments. Eventually we would need to see industry consolidation or collaboration to reduce industry capex so that the industry’s return could reach an acceptable level.

The price hike decisions are likely driven by the need to show the capital markets their rising model intelligence would support higher pricing, so that investors would not worry about margin or ROI pressure. However, we believe the fact that the number of players (even just in the US there are 7) is rising, and the model intelligence gap is narrowing indicate this is an overcrowded market with excessive investments. Eventually we would need to see industry consolidation or collaboration to reduce industry capex so that the industry’s return could reach an acceptable level.

The arguments against this sort of thing are already well rehearsed. Brute measures of LLM power and cost don’t capture what really matters for users, which is that their instructions are being routed through the most suitable paths for the task.

Here’s Barclays’ analyst Sahana Athreya:

As AI workflows become more complex, firms must manage the full economic cost of delivering business outcomes, rather than focusing on model pricing alone.The AI Harness is the strategic control layer. Competitive advantage will come from owning the governance, routing, data and observability framework that determines how workloads are allocated across models and infrastructure.The next AI battleground is inference economics and infrastructure flexibility. As adoption scales, value creation is likely to shift from model development and GPU scarcity toward efficient workload orchestration, power availability, resilience and avoiding vendor lock-in.

As AI workflows become more complex, firms must manage the full economic cost of delivering business outcomes, rather than focusing on model pricing alone.

The AI Harness is the strategic control layer. Competitive advantage will come from owning the governance, routing, data and observability framework that determines how workloads are allocated across models and infrastructure.

The next AI battleground is inference economics and infrastructure flexibility. As adoption scales, value creation is likely to shift from model development and GPU scarcity toward efficient workload orchestration, power availability, resilience and avoiding vendor lock-in.

Synthetic intelligence is being commodified, but trust is differentiated. The dynamic might favour western companies with familiar names, for all the same reasons that people severely overestimate the probability that they won’t know their murderer. All the recent apocalypse stuff may just be a phase in the path towards AI labs becoming state-protected cartels of IT consultancies.

On the other hand: Aaargh!!!!!!!!!!!

Further reading:
— How big is the open-model threat to AI hyperscalers?