Goldman Sachs 1-Delta: AI Model Price War Suppresses Momentum Trades
Alina Collins
OpenAI and Anthropic are selling tokens far below cost to grab users. Goldman's trading desk warns the AI value chain's pricing logic is unraveling — momentum strategies are under pressure, hardware leverage is too high, and the market must rethink which link to bet on.
How aggressive is the price war?
Top AI firms are handing startups cloud-computing and token credits that in some cases exceed $3 million — roughly the median U.S. seed round.
SemiAnalysis data puts it starkly: Anthropic's $200/month Claude Max plan lets users consume up to $8,000 worth of API tokens; OpenAI's same-priced ChatGPT Pro 20x plan allows up to $14,000 worth.
This means → both companies are paying users to show up, and both have IPO plans — the loss-making land grab cannot last indefinitely.
Why can cheaper models shake the entire valuation story?
Goldman 1-Delta desk head Rich Privorotsky warns that cheaper models — mainly from China — keep improving, casting serious doubt on the massive capex forecasts of frontier-model builders.
In plain terms = if a cheaper model does the job, the market will ask: can the hundreds of billions spent on compute ever earn a return?
This reflects a deeper signal — the "tokenmaxxing" narrative (more token consumption = more revenue) is losing its foundation.
Which link in the chain deserves the bet?
Privorotsky's core call: scarce compute is a good business, but once scarcity fades, platform ownership is the superior model.
He favors hyperscale cloud providers (AWS, Azure, GCP) over emerging cloud players, because the hyperscalers own infrastructure, software, distribution, and customer relationships all at once.
In plain terms = "They own the toll road, not just a car" — the shift to open-weight models and routing architectures will funnel even more traffic to the big platforms.
What is the warning sign in hardware trades?
Privorotsky warns that hardware-linked trades have built up excessive leverage. His example: Samsung beat profit estimates by a wide margin, yet the stock failed to rally.
This means → passive flows and rebalancing demand have largely been absorbed; positioning remains crowded, and further upside is limited.
DeepSeek is developing a low-cost inference chip aimed directly at Nvidia — confirming his view that "engineering problems will eventually be solved" and hardware's scarcity premium is shrinking.
Is the "industrial AI" theme real?
Privorotsky is skeptical of the many mining and industrial stocks trading on an AI label: AI capex has already crowded out the broader investment cycle.
In plain terms = if you are long many industrial names today, you are implicitly long AI infrastructure spending — and if that narrative cools, those positions will come under pressure together.
This is the core tension in his warning: if the "de-tokenmaxxing" trend persists, the AI-spending narrative that propped up U.S. (and Korean) equities for months will be hard to sustain.
Content is for reference only, not financial advice.