GV’s Dave Munichiello On Qualcomm’s Modular Purchase, The Firm’s 10x Return And The Shift In AI Software

The artificial intelligence space saw two major developments last week that highlight how technology companies are trying to manage the soaring costs and complexity of AI computing. First, San Diego-based Qualcomm announced its acquisition of Modular, a Palo Alto, California-based software startup focused on making it easier for developers to run AI models across different types of computer chips. At the same time, reports emerged that chip startup SambaNova is finalizing an $800 million funding round led by General Atlantic , valuing the company at $10 billion. Together, the two deals underscore a growing reality in tech: As hardware remains scarce and expensive, the software layers that connect these chips are becoming just as valuable as the silicon itself. Dave Munichiello, managing partner at GV. (Courtesy photo) Watching these shifts unfold firsthand is Dave Munichiello , a managing partner at GV (Google Ventures) who led early investments and holds board seats at both Modular and SambaNova. Munichiello brings a pragmatic operational background to tech investing, having served as a captain and paratrooper in the U.S. military before transitioning to the private sector. He later worked as an early executive at Kiva Systems , helping scale the warehouse automation company through its $775 million acquisition by Amazon . With a background in mathematics and computer science from Emory University and an MBA from Harvard Business School , Munichiello has spent his venture career focused on core software infrastructure, developer tools and data systems, including early backing of companies such as Slack , GitLab and Segment . In this interview, he discusses the mechanics behind the Qualcomm-Modular deal, the practical realities of managing hardware scarcity, and what the current wave of consolidation means for the future of independent startups. This interview has been edited for clarity and brevity. Crunchbase News: The acquisition of Modular by Qualcomm highlights a massive push to decouple AI software from hardware fragmentation. Does this signal that the ultimate value in the AI stack is permanently shifting away from proprietary hardware architectures and toward developer-friendly software layers that can run across any compute environment? Munichiello: The types of hardware required for AI in the future are becoming heterogeneous. Originally, it looked like it was just GPUs from Nvidia , and then also GPUs from AMD and other players. But now, the direction hardware is going is toward “disaggregated inference,” which basically means splitting apart the different compute used for different parts of answering a question when engaging with a model. It increasingly looks like there will be three types of chips used in disaggregated inference: an AI-specific chip, a CPU and a GPU. For a player like Qualcomm, all three of those components are present, so they need a software layer that sits across them. Everywhere else, Nvidia included, they usually sell alongside CPUs and accelerators, and there hasn’t really been a software solution that works across all of those. When did you first start investing in this wave of AI infrastructure and semiconductors? Munichiello: We’ve been investing in AI since 2016, starting as early as a company called Lattice , which was Chris Ré’s first company, sold to Apple , and became part of the Siri team. After that, we invested in Determined AI , co-founded by Evan Sparks , which was later sold to HPE and became an important part of its stack. HPE actually went on to be the compute partner for OpenAI and worked very closely with CoreWeave as well. We also got excited about semiconductors early, long before this current wave, when we led the Series A for SambaNova. I first met that company when it was just three people and a slide deck. We led that round in December 2017 — after Lip-Bu Tan led the seed investment — and I’ve sat on the board since. That initial investment was $15 million at a $480 million valuation. It seems like a lot of legacy chip giants and major cloud providers are aggressively buying up infrastructure startups. What does this consolidation mean for early-stage founders? Are we entering an era where standalone startups need to plan for an early acquisition, or is there still a path to an independent IPO? Munichiello: There is definitely a path to an independent IPO. Cerebras showed that trajectory beautifully, and I’m really happy for Andrew Feldman and that team. There is absolutely a trajectory to build big, standalone businesses because the demand for compute is completely off the charts. We can’t make semiconductors fast enough, nor can TSMC . Everyone is trying to find extra capacity by making everything more efficient. Technology often emerges with a big boom in mass demand and high prices, and then we figure out how to make it cheaper. We are in that efficiency step right now. Demand for inference is everywhere, from medicine and law to coding, customer support and finance. We are trying to squeeze every last bit of value out of chips. Squeezing that value comes from using multiple types of chips: using cheaper CPUs when we can, GPUs when we need them, and the most expensive chips only for the most complicated parts of the process. We are also evaluating software across the stack to ensure every aspect of these queries is as efficient as possible. It’s not surprising that there are a lot of acquirers. The universe of buyers has expanded from just semiconductor companies buying other semiconductor companies to software companies, hyperscalers and model companies buying chip companies, too. Amazon has Trainium and Inferentia; Microsoft has Maia; Google has the TPU, and every big tech company wants to be able to say it has a chip. How does the rise of open-source models shift this dynamic? Munichiello: The universe of potential buyers expands even larger when open-source models become prolific. In the Qualcomm announcement, they talked a lot about their enthusiasm for open source — not just keeping Modular open-source, but for models to be open-sourced. When that happens, instead of enterprise companies paying hundreds of millions of dollars to model providers to do inference, the companies themselves will own their models and run them on their own hardware. So you firmly believe that IPOs are not totally off the table for early-stage tech and hardware companies? Munichiello: Not at all. Look at SpaceX , which is highly hardware-intensive. I think we will see many IPOs here in the next six months. I know of at least 15 or 20 companies that are planning to go public, so it is going to be a very busy period. In a market where valuations are multiplying rapidly based on technical metrics like chip throughput, how are you able as an investor to separate real, sustainable product-market traction from early hype? Munichiello: There are a lot of AI companies getting valuations that are disconnected from the business outcomes they are driving. True traction comes down to quarter-over-quarter execution, hitting sales demands and actually fielding physical systems for customers. A company becomes highly attractive to investors when it delivers a massive volume of technology into production environments — like data centers for major enterprise brands and devices we use every day. That, combined with incoming demand from “Neo-Clouds” (new data centers built specifically for inference), shows real traction. These players are looking for any chips they can get their hands on, and the concept of disaggregated inference — combining three different chip types to lower the total cost of ownership — is highly compelling. It also alters the competitive landscape; it shows that the market isn’t just a runaway race for one dominant player, but an opportunity for CPU providers to catch up as well. GV has a track record of backing foundational tech long before the genera