Clear Ventures managed to stay out of the AI chip startup investment fray of the last several years.
What it saw then was a lack of focus on software and usability, an observation that persists as it looks to the future AI infrastructure. The hype around AI chips over the last several years clouds the vision for practical enterprise AI — and it’s far less about hardware performance than anyone might have thought.
Rajeev Madhavan, founder and general partner at Clear, sees some of the AI hardware startups finding success, but these stories will be few and far between. The emphasis on feeds and speeds is only a small slice of the value. Real adoption beyond national labs or with cloud builders means pulling along a whole suite of software and services — something few hardware companies know how to do, he says.
One might consider him an expert on taking a software-first approach to change how large companies tack on new innovations without completely ditching traditional platforms. Before moving into capital, Madhavan founded and chaired several companies that built toolchains to automate elements of the chip design process (Magma, which was acquired by Synopsys; LogicVision, acquired by Mentor Graphics; and Ambit Design Systems, which Cadence picked up).
That track record has a clear pattern: identify a market that requires automation (in his case EDA) and can hook into existing platforms and traditional processes. Innovate on what exists based on need — on what actual enterprises use — and move on to the next problem.
That pattern can also be more broadly applied to AI. The goal isn’t to make it a standalone workload or target for investment. It’s about finding the startups that can bring it in as part of a larger, practical workflow, often based on traditional systems, and make it “acceleration as a service” rather than a niche hardware/software gambit.
“Anyone doing hardware acceleration today has to think about how to take software, use a hardware accelerator, and provide all of that as a service. That means a lot of those that started as AI chip companies are going to have to make a painful transition to becoming part of a broader solution,” Madhavan says.
The Clear Ventures view is that by 2024, the AI chip ecosystem will be a “deserted island” with a lot of money raised but no comprehensive product that can be readily consumed and integrated beyond the hallowed halls of large research institutions and hyperscalers. Madhavan says if we look at large corporations using, for instance, SAP HANA, this is how opportunity can be better gauged.
In other words, finding ways to help large companies integrate AI in to existing systems, making it palatable by making it a service that is connected with broader enterprise infrastructure, is the goal — and the area where their investments will go. Madhavan says his own experiences from startup to IPO showed him that 90 percent of a company’s funding supports services like ERP, for instance. The goal is to build automation into existing platforms and seamlessly accelerate the various pieces and specific applications (databases, networks, specific workloads).
Newcomer to Clear Ventures, Vijay Reddy (ex-Intel Capital) says that from a datacenter perspective, what people care about has little to do with hardware performance or energy efficiency. “It’s about TCO — time to value and the cost to get there at a software level. Hardware is just one small aspect of that.” What the market needs, and what Clear is looking for, is something that is “software first and sold as acceleration as a service,” Reddy adds.
“If you build a software-first company, there is an opportunity even for datacenter-focused companies. It’s difficult to build that for any hardware company — adding a culture of software. Even Intel hasn’t been able to do that in its own long history. But we think there is an opportunity for those who can get this right for specific areas. Databases, data acceleration in general, AI, the network, these are all opportunity areas,” Madhavan explains.
“We are focused on AI-plus-X” says Reddy. “We’re interested in where AI is being applied to a vertical or industry; that can be networking, IoT, large telcos, the places where the underlying systems don’t have the intelligence to do things right.” He adds that a trend toward disaggregation complicates enterprise IT but creates opportunities for startups that can bridge the gap and make deployment of AI and some of its offshoots in graph neural networks or time series data analysis with AI techniques easier.
“A lot of activities are manual in these environments,” Reddy adds. “We’re trying to invest in companies that can automate and to look to ways existing platforms can be made better, so for instance, taking an older visualization platform like Tableau but do it better with AutoML and newer technologies.”
While the Clear Ventures portfolio is not stacked with the sexiest AI startups with new devices and software stacks, it is one of the more practical. It also speaks to the “acceleration as a service” concept Madhavan espouses. From network acceleration via a networking-centric OS from Arrcus that can support that disaggregated, distributed vision to domain-specific software platforms that accelerate legacy and emerging workloads, Clear’s AI emphasis has far less to do with hardware than orchestration, acceleration, and automation.
It could be that we are nearing the end of seeing big hardware investments in new AI chips for the datacenter, at least for a while. If Clear has the right idea, the shift to making AI practical could constitute the next wave of AI startups. Those stories aren’t as glossy without hardware metrics and wild performance predictions but they do show an AI segment that is cooling off and (finally) getting real.