Plenty of artificial intelligence (AI) and networking news sprang up out of Google Cloud Next 2023. But to what degree are the company’s announcements noteworthy and to what extent are they…not? We rounded up opinions from a handful of analysts. Here’s what they had to say.
AI advances
The big news to come out of the conference focused on artificial intelligence (AI), with Google specifically touting advances in its Vertex AI platform and Duet AI tool. You can find all the details on those announcements here.
Moor Insights and Strategy founder and CEO Patrick Moorhead told Silverlinings he’s “impressed with how Google responded to Microsoft’s first out-of-the-gate announcements back in November. While I always said this competition was more of a marathon between AWS, GCP, Azure and OCP, the rate of going GA [general availability] is a great sign for Vertex and Duet.”
Gartner VP of Cloud Services and Technologies Sid Nag added in emailed comments Duet helps developers write applications faster "by using a query mechanism based on AI...thereby reducing code development cycle times.”
Nag also highlighted the fact that by adding Llama into its model garden “Google seems to have addressed the Amazon Bedrock functionality with a Meta LLM capability with an associated query interface.”
Hardware, chips and more
In a note to investors, New Street Research’s Pierre Ferragu spotlighted Google’s GPU announcements, specifically those related to making compute based on NVIDIA’s H100 chips available next month and its early access to NVIDIA’s DGX GH200 Grace Hopper chips. He explained that many users prefer GPUs over tensor processing units (TPUs) and “Google is simply adapting to this reality.”
Still, he said Google’s massive TPU holdings are helping it build a “strong competitive advantage.” By NSR’s calculations, Google has around 1 million GPUs, comprising around 15% of its total installed base of its 6 million or so. It also has around 3 million TPUs, which offer computing roughly equivalent to 4-5 million GPUs given they offer 1.2 to 1.7 times more throughput. That gives the company a 5-6 million GPU equivalent, or “more than the rest of the GPU installed base altogether.”
The conclusion? The “GPU [is] central to Google’s infrastructure; TPU [is] leveraged for incremental scale and cost leadership.”
Catch up on all of the news from Google Cloud Next 2023 with our dedicated news hub here.
Chirag Dekate, VP and analyst at Gartner, said TPU’s are Google’s “secret sauce” as the industry and enterprises alike enter the AI era.
“If, for example, you’re trying to train a large language model, just for the sake of conversation let’s say it’s GPT…if you’re trying to use GPUs, you might need hundreds of instances of GPUs. Now, if you use TPUs, because of the performance efficiencies you might, practically speaking, need a fraction of that, maybe 60 to 70 TPUs.”
Why does that matter so much? Well, Dekate explained that if you’re the CEO of a startup that’s trying to build AI models, TPUs will allow you to have a lower cost base. This, in turn, will enable you to pass through price advantages to your customers to gain a market advantage.
“You’re now starting to see what the future ecosystem is going to look like,” Dekate said. “The future ecosystem is increasingly going to be accelerated and technology providers are going to have to specialize at the AI hardware level.”
Put another way, he believes the specifics of the computing on the back end will matter more in the AI era than they did at the dawning of the cloud age.
Google's Cross Cloud Networking
But Google Cloud Next wasn’t all AI and chips. Google also addressed the needs of enterprises operating in a multi-cloud world with its new Cross Cloud Network offering.
Nag told Silverlinings Google has adopted “a thought leadership position” with its Switzerland-like approach (the country being famously neutral, of course). This stance allows clients to connect to multiple cloud environments as multi-cloud adoption continues to rise and also lets them tap into “best-in-class capabilities from multiple clouds without having to replicate the associated data in all clouds,” he said.
Morehead seconded this opinion. “The hybrid, multi-cloud is real and what enterprises want. The challenge is that for each cloud, enterprises need a separate team for apps, networking, data, security and operations. GCP’s Cross-Cloud Network is an answer to minimizing latency between multiple clouds and given Google has the most extensive WAN, makes sense that it is so performant.”
AI Grounding
Though this announcement mostly flew under the radar, Dekate told Silverlinings the “whole notion of grounding is rather important” when it comes to AI. That’s especially the case as enterprises look to build generative AI applications that can provide contextualized responses.
“Google’s approach is differentiated in advanced capabilities like citation, where the generative AI models clearly identify the source of the data used to generate GenAI outcomes,” he said. “Examples of these include code completion and text generation where the citation of source data increases trust in generative AI systems.”