The lone sell on Nvidia as AI euphoria peaks
One analyst sits on an island. While almost every Wall Street desk has Nvidia as a buy, Jay Goldberg keeps a sell tag on the market's favorite ticker. His case is simple: more things can go wrong from here than right.
Nvidia has climbed more than 3,000% since early 2020. It's the S&P 500's top performer, the perceived engine of AI spending, and the ticker people check before coffee. That's also why the risk is real. Blind confidence masks fragile assumptions.
Where the optimism is coming from
Demand for GPUs still looks insatiable. Out of 80 analysts, 73 are buyers, six say hold, and one says sell. The average price target sits around $220 versus a recent close near $186.26. Some, like HSBC's Frank Lee, see a much longer runway and targets as high as $320.
Underneath the headlines is a narrow buyer base. The spending engine is concentrated in six firms: Microsoft, Alphabet, Amazon, Meta, Oracle, and OpenAI. The five public names are on track to spend close to $400 billion in capex this year, up more than 67% from last year. OpenAI has talked about plans far beyond that, even into the trillion-dollar range. The question is the same across the group: what return shows up, and when?
Why one analyst says sell
Goldberg's stance comes down to concentration and timing. The stock is already priced for perfection while the buildout is still unproven. If demand flinches or capital gets tighter, the downside can move faster than the upside.
- Buyer concentration: A handful of hyperscalers drive most orders. If even one pauses, guidance shifts.
- Capex to ROI risk: Big checks today need visible, bankable revenue tomorrow. Hype doesn't pay interest.
- Power constraints: Data centers need electricity that many grids can't deliver on time.
- Leverage and counterparty risk: "Neoclouds," property deals, and financing chains create weak links. One obscure failure can cascade.
- Valuation math: When a leader is close to sold out, upside comes from price or new buyers. Both have limits.
The bull case - and why timing matters
Bulls argue we're early in a multiyear adoption cycle. More industries will use accelerators, AI workloads will expand, and software layers will grow around the hardware. If that happens, today's multiples can look reasonable in hindsight.
The timing is the catch. Shares are up more than 70% since late April. If the company is capacity-constrained, near-term beats depend on pricing, mix, or surprise supply. Any shortfall in those levers, or a pause in hyperscaler budgets, can sting.
- What would validate the bull view: New buyers beyond the "big six," recurring AI revenue with high margins, clear power build-outs, and stable unit pricing despite rising supply.
- What would challenge it: Flat or lower capex guides, slower monetization of AI features, power delays, or early signs of inventory overhang.
A familiar pattern from the dot-com era
Goldberg points to the telecom buildout during the dot-com bubble. Cisco rode a wave of infrastructure spending, then fell hard when traffic and profits lagged the investment cycle. History doesn't repeat exactly, but it rhymes. For a refresher, see an overview of the dot-com bubble.
What to watch over the next two quarters
- Hyperscaler capex guides: Direction, cadence, and any commentary on AI ROI.
- Monetization signals: Paid AI seats, model-driven ad lift, AI SKU attach rates, and usage-based revenue disclosures.
- Power and permitting: Grid interconnect timelines, regional constraints, and utility capex plans.
- Supply chain health: Credit spreads for data-center builders, cancellations, and extended payment terms.
- Inventory and pricing: Lead times, discounting, and mix between top-end and mid-range accelerators.
Practical takeaways for investors and finance teams
- Size positions to survive whiplash: Leaders can correct 30-50% and still be long-term winners.
- Separate narrative from cash flow: Track when AI features become paid products with durable margins.
- Map second-order risks: Power delays, financing stress, and regulatory bottlenecks can hit before demand does.
- Diversify exposure: Consider picks-and-shovels across power, memory, networking, thermal, and software layers.
- Set if/then rules: If capex growth slows and pricing softens, reduce; if new buyers and power capacity expand, add.
The tension in one chart
On one side: record demand, a clear product moat, and a software ecosystem growing around it. On the other: buyer concentration, physical power limits, and expectations priced like smooth sailing. Both can't be right forever. The tape will pick a side when budgets and utility timelines do.
If you're responsible for AI budgets or need practical ways to validate ROI, explore curated resources and tools for finance teams: AI tools for finance and the latest AI courses.
Bubbles aren't obvious until they are. Neither are inflection points that justify today's prices. Stay curious, keep your rules tight, and let the numbers-not the noise-do the talking.
Disclaimer: For information purposes only. Past performance is not indicative of future results.
Your membership also unlocks:
