Given the carnage currently unfolding across the equities universe, one would be hard-pressed to get excited over a still-nebulous AI mania 2.0. Yet, this is exactly what Piper Sandler's Brent A. Bracelin is grooving to, citing NVIDIA's exploding GPU capacity and a concurrent implosion in AI model inference costs as important catalysts for unleashing the "AI wave 2."
JPM: "Our analysis suggests that about 90% of the momentum sell-off is complete when compared to previous tactical momentum unwinds (40 instances over the last 35 years). However, if this is a structural regime change, the unwind could have much further to go, as our JPM momentum…
— Wall St Engine (@wallstengine) March 10, 2025
As recently noted by Bernstein analyst Mark Li, NVIDIA shares have been "battered by a storm of growth fears, supply chain noise, and tariff and regulatory risks," including geographical restrictions on NVIDIA's GPU sales, especially in relation to China. What's more, the uncertainty introduced by Trump's tariff policy into the broader economy is now hammering confidence and investment decisions, leading to a growing chorus of recessionary calls. According to J.P. Morgan, past crowded momentum unwind episodes have entailed an average peak-to-trough drawdown of 14 percent. Meanwhile, Morgan Stanley's Mike Wilson now thinks the S&P 500 index is heading to the 5,500 level in H1 2025 before rebounding powerfully in the second half.
Against this backdrop, Piper Sandler's Bracelin has just published a much more sanguine investment note on the prospects of AI wave 2.0, citing a 10x increase in NVIDIA's GPU capacity over the past two years and the recent 35x decrease in AI model inference costs - courtesy of DeepSeek's R1 and Alibaba's Qwen AI models - as the requisite catalysts for unleashing the next stage of AI-induced bull cycle.
Barecelin uses NVIDIA's data center revenue, which increased from $3.6 billion in early 2023 to $36 billion last quarter, as a proxy for its growing GPU capacity.
On inference costs, Bracelin argues that OpenAI's GPT-4o entailed "an initial price of $15 per 1 million tokens with performance of 128K tokens/second." However, Alibaba's "Qwen QwQ-32B model on Groq this week is priced at $0.39 per 1M output tokens with performance of 400K tokens/second," equating to a ~35x reduction in inference costs.
Bracelin argues that these two ingredients "could spark much broader proliferation of agentic AI applications," and that the "application sector appears to be on the precipice of a new AI monetization wave with long-tail benefits that could accrue for the next decade."
The analyst goes on to note:
"We are encouraged by early proof points including 1) Microsoft Copilot having approached the $1B run-rate milestone, 2) Salesforce Data + Agentforce AI run-rate of $0.9B, 3) HubSpot with 35% of support tickets handled by Agentic AI, 4) ServiceNow with 1,000+ customers purchasing the GenAI Pro+ SKU, and 5) Snowflake Snowpark reaching 3% of sales on >100% y/y growth."
Of course, Bracelin is aware of the ongoing market-wide meltdown, noting that "the 1-mo IGV decline [for the software sector] of -12% (vs. -5% S&P 500) shows increasing tariff, policy, and macro concerns eroding confidence in the demand fundaments for software spend overall." Nonetheless, the Piper Sandler analyst believes that this weakness is transitory and that the upcoming monetization wave should pave the way for the gradual resumption of a secular uptrend.
It is hardly a surprise, therefore, that Bracelin has identified Salesforce (CRM), Microsoft (MSFT), and Snowflake (SNOW) as his top ideas pick for 2025.
