How to Build the Perfect AI Workstation

The AI market is in its infancy. While Microsoft was a first mover with AI, it screwed up the product launch because one of the two tools it was supposed to launch wasn’t ready. Additionally, the products didn’t address Microsoft’s primary market, which is developers. It was as if everyone in Microsoft missed the old Steve Ballmer talk about developers, developers, developers, developers…

So, most development is cloud-based, and while companies are moving to on-premises, much of the work is still being done on shared server resources instead of on dedicated workstation hardware, which is very different than most advanced software development today.

There is no perfect AI workstation because the companies driving AI at the moment haven’t defined what that workstation should look like.

Problem and Opportunity

When you’re in an early market like AI is, there is often a problem with regard to what a tool for that market should look like. Too often, the market does its typical trial-and-error approach to defining that product until the market matures around an eventually successful offering.

The problem with that approach is that it’s largely luck, and most people will get the outcome wrong. The opportunity is for one powerful vendor to create a product that is initially defined by a company like NVIDIA and then aggressively market the result as the ideal product for the new segment. This is what Apple did with the iPod and iPhone and tried to do with the iPad but under-executed.

It is easier to lead if you are already in front yelling “follow me” than if you are in the back trying to figure out where to go. So, the problem is that no one seems to have defined the perfect product yet, but the opportunity is there for some company to create it. The three players that could do this are arguably NVIDIA, AMD, and Intel. I put Intel last because it’s also the most distracted at the moment, making it very hard to focus on opening a new market, though it has the reach and funding to lead if it could focus.

The Risk to Existing AI Vendors

Workstations are used predominantly to build new hardware and software offerings. It’s a relatively small but very lucrative market that is influential well beyond its size. The engineers who use workstations are, by their very nature, creators of the future, and the technology they use to create should, and generally does, have an inside track to becoming the technology used at volume once what they create makes it to market.

In many cases, this isn’t a huge advantage. For instance, just because a technology is used for CAD/CAM (Computer Aided Design/Computer Aided Manufacturing) doesn’t mean the same technology will be used in what is created and manufactured with CAD/CAM. However, with products that are computer-oriented and tightly tied to hardware, such as AI, that connection should be far stronger. Even if it isn’t, the perception would be that it is. This is fueling NVIDIA’s valuation at the moment; AI models are almost exclusively using their technology, so those creating those models are believed to favor NVIDIA as a result for anything intending to run those AI models.

Microsoft did create kind of a mess with its non-NVIDIA approach to AI, which has muddied the market somewhat and provided an uplift to Qualcomm, but Qualcomm has not positioned itself as an AI creator so much as a platform that can run what is created. This provides a huge opportunity for someone else to step in, and NVIDIA is planning to do exactly that with ARM.

But ARM isn’t yet at critical mass on the desktop, and AMD is far better positioned to address the entire platform given it’s a fast NVIDIA follower and it has an x86 solution. If AMD creates a Threadripper-like AI platform specifically for developers, it could move on this opportunity, given NVIDIA’s workstation solution isn’t cooked yet and ARM isn’t yet validated in the workstation space.

Wrapping Up: Building the Perfect AI Workstation

Workstations are defined by developers, but these developers have not decided what a perfect workstation should be, opening an opportunity for an OEM to create a unique product and market it as the perfect workstation. To do that, they’d need heavy backing from NVIDIA or a version of AMD willing to market the result. AMD could move into this market, but it would need to execute at a marketing level that is beyond what it has now. NVIDIA wouldn’t need as much marketing, but the ARM part of its solution is a much more difficult lift. Ideally, a configuration backed by both companies would have the greatest potential, but accomplishing that may be a bridge too far for any OEM.

So, to build the perfect workstation, which could be a foundational element for future AI growth and use, I’d go to both AMD and NVIDIA and assess the degree of difficulty and how much help I’d get from either vendor. I would then assess how much of a marketing budget I’d need to flip the market. Finally, I’d have to decide if I wanted to lead it or let someone else break this market open and slipstream that effort with an offering that matched as well or better than what they successfully released or addressed the shortcomings if the launch failed.

Whoever owns the AI workstation market will likely benefit significantly from the following AI implementation wave. This isn’t without risk, but the rewards could be legendary.