BofA and Citi Both Back NVIDIA: Per-Rack Price Hikes Cover HBM Costs, 18x PE Hits Seven-Year Low
Alina Collins
Bank of America and Citi issued concurrent bullish reports on Nvidia, reaching the same conclusion: rack-level price increases of $2–3 million dwarf HBM cost additions, and the stock's 18x forward P/E — a seven-year low — already over-prices a bearish scenario.
Memory costs eating margins? BofA did the math
From Blackwell to Rubin, per-rack HBM — high-bandwidth memory, the specialized memory that feeds data to AI chips at speed — costs only $200,000–$300,000 more.
Nvidia's per-rack price, however, jumped from roughly $3–4 million to $6–7 million, a $2–3 million increase.
This means → the price hike is nearly 10× the cost increase. Memory inflation barely dents margins; BofA expects gross margin to hold in the mid-70s percent.
How cheap is 18x PE, really?
Nvidia's forward P/E sits at just 18×, a seven-year low — 30–35% below AI peers Amazon, Meta, and Google.
BofA flags an implied 30–35% earnings discount baked into consensus for fiscal 2027–2028, calling it "indefensible."
In plain terms = the market is pricing Nvidia as a company about to decelerate. Both banks say that assumption is wrong. BofA reiterates Buy with a $350 target.
Citi checked the supply chain — has the roadmap changed?
After speaking with Nvidia's investor-relations team, Citi confirmed the Vera Rubin Ultra roadmap is fully intact. The NVLink domain — a scale-up co-packaged optical switch — shown at Computex is unchanged.
Co-packaged optics (CPO) — embedding optical communication directly inside the chip package — for scale-out networking is already in production via Spectrum-X, with high customer adoption.
Starting with the Feynman architecture in 2028, customers can choose CPO-enabled NVLink or copper. This means → Nvidia is expanding customer options, not being forced to pivot.
Custom chips stealing share? What does history say?
Google's TPU launched in 2015, Amazon's Trainium in 2020, Meta's custom chip in 2023 — yet Nvidia's GPU accelerator revenue has grown roughly 700× since 2015.
Nvidia's hyperscaler sales grew 115% year-over-year, about twice the rate of cloud capex growth.
This reflects a pattern: custom chips have not replaced Nvidia — they have grown alongside it as the total AI-compute market expands. BofA projects Nvidia will hold 65–70%+ of AI-compute capex long term.
Who else is buying beyond the hyperscalers?
Citi notes accelerating demand from AI labs, emerging cloud providers, sovereign nations, and enterprise on-premises deployments — beyond the hyperscalers driving the bulk of current rollouts.
Management believes that as physical AI — robotics, autonomous driving, and other real-world AI applications — takes off, the non-hyperscaler share of the market must eventually grow larger.
This means → Nvidia's customer base is diversifying from "a handful of mega-cloud buyers" to a global, multi-segment mix. The demand foundation is broader than the market assumes.
Earnings beat expectations — so why did the stock drop?
Nvidia's latest earnings disclosed $95.2 billion in purchase commitments to lock in capacity and approved an $80 billion share buyback — both above expectations.
Despite the beat, shares pulled back in after-hours trading. Several trading desks called the dip a positioning window, advising clients to buy.
In plain terms = the numbers are fine; what sold off was sentiment. The core message from both banks: panicking at 18× forward PE makes less sense than reading the balance sheet. The next checkpoint is TSMC's upcoming earnings.
Content is for reference only, not financial advice.