‘Sell Nvidia’

Want to see something weird?

Right?! It looks like a bearish note from a bulge-bracket bank on Nvidia, the world’s most overbooked hedge fund hotel.

That kind of thing just doesn’t happen. Of the 64 research houses to publish on Nvidia shares, 58 rate the stock a “buy”. It’s been nine months since a shop other than Morningstar published a sell note.

And we should be clear that Barclays’ equities analysts are still enthusiastic on Nvidia, last week reiterating a $650-per-share target price. The screenshot above is credit research. It’s from Sandeep Gupta, a 14-year TMT veteran on the bank’s fixed-income side.

Gupta’s main thesis is straightforward. Competition for generative AI hardware is catching up, end-user deployments don’t need much power once the database is built, there’s a shakeout due soon, and Nvidia’s estimated 98 per cent market share of data-centre GPUs probably won’t last:

AI chip demand will eventually normalize once the initial training build has been completed. The inference phase of AI is going to require less computing power than the training phase.  High-powered PCs and phones could be capable enough to run inferences locally, reducing the need for growing GPU plants. [ . . . ] We think the cost of building Large Language Models (LLMs) and limited monetization opportunities could eventually force the exit of smaller (but still meaningful) customers of Nvidia from the AI market, helping to ease GPU supply constraints.

Revenue growth for big tech is expected to slow this year, while the likes of Meta and Amazon are working on their own chips. That’s bad because just five customers provided 46 per cent of Nvidia’s revenue last quarter, with Microsoft and Meta alone accounting for 28 per cent of the total.:

Nvidia “is acutely sensitive to fluctuation in the topline growth and procurement decisions of top customers given the concentration of NVDA’s sales,” says Barclays. “This is especially relevant given the aforementioned competition in the chips market and potential for NVDA to lose revenue share if its tech customers develop self-sufficient chip manufacturing capacity.”

The more novel section in Gupta’s research concerns Nvidia’s funding for AI start-ups.

Nvidia’s venture capital arm made 33 investments between January and October 2023, with 11 of those in its last fiscal third quarter, according to Crunchbase data. That’s a significant rise in activity; between 2005 and 2023, it had done at most four VC deals in any given quarter, Barclays says:

Though the total amount of funding and the specific structure of the deals (cash, stock, tech, or services) are unknown, Nvidia has become linked with a host of companies that we expect are using Nvidia technology. CoreWeave, an Nvidia-backed cloud computing startup not only uses Nvidia chips but has secured $2.3bn in financing using Nvidia H100 GPUs as collateral. This financing will be used to purchase more advanced chips, most probably from Nvidia. It is estimated that CoreWeave was Nvidia’s 7th largest customer of H100 GPUs in 2023, purchasing 40k of the processors for 4.5% of total revenues (an H100 GPU sells for $30k). Meta and Microsoft tied for number 1 at 150k each.

The core takeaway is that an increasing share of NVDA’s revenues from the past two quarters can potentially be attributed to startups NVDA has funded itself. These companies operate in robotics, machine learning, SaaS, and cloud computing, all sectors reliant on AI chips. This self-funded demand is a risk because it is dependent on NVDA’s own investment spending and potentially misinterprets independent demand for NVDA chips because NVDA is essentially funding its own customers.

The other threat flagged is regulatory scrutiny of its Chinese business. In short, authorities might be worrying more than investors about possible grey market import routes (Singapore has grown to 14.9 per cent of Nvidia’s group revenue from 9 per cent a year ago) and about Chinese buyers repurposing Nvidia gaming cards for their AI projects.

To reiterate, all of the above is credit research, which rarely gets much attention when it’s about a single investment-grade issuer. As far as we can see, no media has picked up the Barclays note since it was published late Friday.

Gupta acknowledges that Nvidia’s balance sheet is bulletproof and in the wake of its failure to buy Arm in 2022 the risk of issuance is low. His concern is only that Nvidia is too expensive. Its bonds trade tighter to the equivalent Treasury yield than AAA-rated Microsoft and Apple, whose long-term prospects are much clearer.

Meanwhile, in the shoutier world of equities, there’s wall-to-wall coverage today of a Goldman Sachs note raising its Nvidia share price target on Nvidia to $800. Goldman argues that AI capex is going to keep flowing to the industry’s “gold standard” supplier and that export-tailored hardware can circumvent sanctions on China.

At pixel time Nvidia shares are at a record high, up 3 per cent to $681, which is equivalent to about 35 times the 12-month-forward consensus EPS. Because 58 sellside “buy” ratings can’t all be wrong.

Further reading:
— There’s a hot new hedge fund hotel (FTAV)
— Find someone who loves you like hedge funds love Nvidia (FTAV)
— This is nuts. When’s the crash? (FTAV)