One of the major questions around the crypto boom, as it drove GPU prices into the stratosphere, was exactly how much additional business Nvidia and AMD were hauling in and how much of it accrued to each company. According to Nvidia, it made $602M from crypto sales from April 2017 to July 2018.
But RBC analyst Mitch Steves doesn’t think this is accurate. “We think NVDA generated $1.95 billion in total revenue related to crypto/blockchain,” RBC analyst Mitch Steves said in a note out Wednesday. “This compares to company’s statement that it generated around $602 million over the same time period.”
Steves totaled the actual hash rates of the various cryptocurrencies over time, then calculated Nvidia and AMD’s estimated actual exposure to the cryptocurrency market. AMD appears to have correctly calculated its own exposure to the market. Nvidia does not.
Because the overwhelming majority of mining GPUs are simply repurposed consumer cards, Nvidia doesn’t necessarily have a great window into exactly which consumer GPUs are being sold for which purposes. And this, in turn, could explain both how Nvidia ended up with a glut of Pascal cards in the first place, and why Nvidia raised prices with its Turing GPU family.
To understand why, it’s important to remember that everyone in the GPU business knew cryptocurrency was going to be a flash in the pan. The boom cycle of 2017-2018 followed the same basic trajectory as the 2013-2014 surge. The principal difference, in this case, is that Nvidia was actually a player. Jen-Hsun and the other Nvidia executives aren’t idiots; they’d watched the 2013-2014 cycle play out, and they knew ASIC mining or BTC’s intrinsic price instability would eventually cause GPU-related sales to fall. Both Nvidia and AMD made comments to this effect as soon as analysts even raised the question.
If we grant Steves’ hypothesis for a moment, Nvidia would have seen a surge in gaming-related revenue for a GPU generation that wasn’t even particularly new. A sustained influx of new sales in Chinese markets might have been read as being related to the overall growth of PC gaming in China as a whole. PC gaming is far more popular in China than consoles, and while mobile gaming is the largest segment ($23B), PC games were expected to generate $14.4B in sales in 2018. That’s nothing to sneeze at.
Companies like AMD, Intel, and Nvidia model demand carefully. Stronger-than-expected demand across an entire product line can indicate that the GPUs in question are priced systemically too low. The implication of this scenario is that consumers were willing to pay more for GPUs than Nvidia actually charged them. This is known as a producer surplus.
Under this assumption, Nvidia was taking the economically prudent course of action. I don’t mean to imply that there were no technological reasons for the company to raise prices — its GPU dies were larger with Turing and GDDR6 is still more expensive than GDDR5. But Turing GPUs are more expensive, relative to Pascal, than what would expect if Nvidia were simply passing along costs. Nvidia may have believed that it had excellent reason to raise GPU prices whether AMD had competitive cards in-market or not.
Nvidia has already acknowledged that uptake for its new RTX GPUfamily has not been as strong as the company anticipated. In its earnings warning last week, the company informed investors that it expected revenue to be ~500M lower than previously communicated. The company attributed its earnings miss to “deteriorating macroeconomic conditions, particularly in China,” and stated that “sales of certain high-end GPUs using Nvidia’s new Turing architecture were lower than expected.”
The other piece of evidence supporting Steves’ theory is the Pascal glut. There is, to be sure, an argument that says Nvidia raised Turing prices simply because AMD had nothing to compete against the RTX family with. There’s probably at least some truth to this, in that Nvidia may have felt safer raising prices without a high-end competitor to undercut it. But there’s no situation in which it made sense for Nvidia to overproduce Pascal cards to such a degree that it wound up choosing not to build any for an entire quarter. That’s exactly what hammered the company’s Q4 prospects the first time, before it warned on revenue last week.
This theory that Turing’s price structure is explained in part by Nvidia’s potential failure to read the crypto market correctly is my own speculation. Steves only addresses the idea that the company may have misunderstood or failed to acknowledge its own cryptocurrency exposure. It’s possible he’s mistaken, and if so, obviously my own theory wouldn’t work, either. But given the evidence, it seems plausible that Nvidia misread what was driving a surge in its own GPU revenue, and laid what it believed were economically reasonable product plans based on that mistaken hypothesis. As a certain Vulcan once said, “It would explain a great many things.”