Chinese cloud hyperscalers slash infrastructure spending

China’s cloud hyperscalers like Alibaba and the country’s internet companies more broadly “really, really slowed down infrastructure investment this year” chipmaker NVIDIA’s CEO Jensen Huang said as the GPU specialist missed its Q3 revenue forecast by $1 billion. He primarily blamed the miss on a slump in revenues from its gaming segment but also pointed to ongoing supply chain issues that continue to make it hard to get chip components.

(China’s zero-Covid policy has hit major manufacturing hubs and continues to disrupt global supply chains. The country's GDP fell by 2.6% in the three months to the end of June from the previous quarter. NVIDIA's earnings and warning on Chinese investment came as Huawei's founder Ren Zhengfei warned that a looming global recession threatened the company' survival, joining other founders in emphasising profit over revenues.)

Despite the Chinese retrenchment, data centre and cloud GPU demand was a glimmer of light for NVIDIA amid macroeconomic headwinds and as talk of a potentially deep recession continues across the technology industry.  Whilst its gaming segment revenues of $2.04 billion were down 33% year-on-year, NVIDIA data centre sales were $3.81 billion – up 61% year-on-year; NVIDIA executives played down that achievement however.

See also: Meta’s SuperCluster scaling to a monstrous 16,000 GPUs

CFO Colette Kress told analysts on an August 24 NVIDIA earnings call: “Although a record, this was somewhat short of our expectations as we were impacted by supply chain disruptions. Revenue from hyperscale customers nearly doubled year-on-year. Sequentially, sales to North America hyperscale and cloud computing customers increased but were more than offset by lower sales to China hyperscale customers affected by domestic economic conditions. Vertical industries grew… Key workloads driving growth include natural language processing, recommender systems, autonomous vehicle fleet, data processing and training and cloud graphics.”

As Bank of America Securities’ Vivek Arya noted: “There is a concern that data center CapEx could be the next shoe to drop in this rolling correction in semiconductors. [A concern exacerbated by sharply rising energy prices]”, with Arya asking NVIDIA execs: “I’m curious… what is the risk that there could be a correction in the data center given some of the macro caution expressed by some of the hyperscaler and enterprise customers?”

To CEO Jensen Huang the shifting shape of cloud workloads meant demand would continue.

He responded: “We hear fairly broadly that GPU supply is in shortage in the cloud… that demand for GPU rentals far exceeds current supply. [Cloud users can “lease” GPU-powered instances for their workloads as well as CPU-based ones, the former often for AI inference-type workloads that require GPU-capable processing].

As he put it: “The number of use cases for GPUs in the cloud has grown quite a bit. If you look at one particular segment in just managing -- collecting data and managing the data of the AV fleet and using that data to train AI models, using that data to reconstruct HD maps, the usage of GPUs in the cloud for just that one application has grown a lot. And furthermore, deep learning-based recommender systems has demonstrated such significant effectiveness; conversational chat, Q&A summarization, text generation, image generation.”

Follow The Stack on LinkedIn

A typical example, Huang said, was at Pinterest, whose 400 million users have “pinned” a massive 300 billion images: “Pinterest transitioned to 100x larger recommender models by moving its inference from CPUs to NVIDIA GPUs. Its ability to deploy a higher-quality model at high throughput and low latency resulted in a 16% increase in engagement, a critical metric for the company” the NVIDIA founder said on its Q3 earnings call.

The company continued to be “challenged this quarter with a fair amount of supply chain challenges because as you know, we don’t just sell the GPU chip, but these systems are really complex with a large number of chips in the system components” he added. “ all of the components have to come together for us to be able to deliver the final component… these data centers sit idle until the last piece comes together. And the last piece includes very complicated switches and very complicated NICs and networkings and cables. Building these high-performance computing data centers at very large scale for the world’s cloud is not particularly easy," he said.

"So the supply chain challenges have been somewhat disruptive. But the demand is there.”

That 33% fall in gaming revenues meanwhile? Largely an inventory issue...

"The end market gaming demand is solid" said Huang, but has falled a little from highs early this year.

The company's optimism got a little ahead of itself: "Bcause we were building for such a vibrant market, we found ourselves with excess inventory. And so, our strategy is to sell in well below the current sell-through levels in the marketplace to give the channel an opportunity to correct. We’ll do that for a couple of quarters or so. We believe that by the end of the year, we’ll be in a good shape going into next year" the CEO said.