Why defensibility is the hook for VC data center investments

Why defensibility is the hook for VC data center investments Leah Hodgson Fri, August 14, 2026 at 7:37 PM GMT+2 5 min read AI's infrastructure boom has already been running at a staggering pace, and this week it received a further $500 billion vote of confidence from Wall Street's biggest names. Nvidia signed memorandums of understanding with six of the world's largest asset managers and investment banks, including Blackstone , KKR and Apollo Global Management , to raise the capital to fund what it is calling "AI factories. " While not guaranteed, the deal signals that demand for AI infrastructure is showing no signs of slowing.

VCs are backing that same buildout by funding the semiconductors that power it. Global funding for AI and machine learning semiconductors reached $14. 1 billion in the first half of the year, according to PitchBook's latest AI Report , on track to surpass last year's annual total by almost 50%.

Both Q1 and Q2 this year are the two highest quarters on record for investment. Deal count is also projected to land above 2025's figure. "AI is not just a tool, it redefines technology," Sriram Viswanathan, founding managing partner of deep tech VC firm Celesta Capital , said.

"If you accept the notion that everything is going to change, the question is, what is the tip of the spear? That is semiconductors, because that powers everything. " That conviction is rooted in the belief that AI's growth trajectory faces an increasing number of bottlenecks that can't be solved by software improvements alone.

Hardware, on the other hand, offers something increasingly rare elsewhere in AI: real defensibility. Bottlenecks fuel bets The scale of the semiconductor opportunity is hard for any investor to ignore. According to a base-case estimate from McKinsey in March, the industry could reach $1.

6 trillion in revenue by the end of the decade. It is also a market that is becoming increasingly valuable as AI faces more constraints, according to Viswanathan. The demand for computers significantly outweighs the supply.

AI data center demand is set to exceed capacity by 2027, and one-fifth of planned projects could be delayed due to a lack of grid connections. He argues that AI faces limitations at every layer of the advanced computing stack, from the processor and the power that feeds it to memory and the network that connects the chips. These constraints are where VCs are finding opportunities for returns, leading to significant deals.

In March, Ayar Labs closed a $500 million Series E to speed up data center transmission by using light to connect GPUs and chips. Last week, Olix , which uses lasers to move data between large clusters of chips faster, raised a $312 million Series B . XCENA , which develops computational memory architectures that reduce data bottlenecks, secured $135 million for its Series B in May.

Story Continues While demand is increasing, the barrier to entry has also lowered. Producing chips has become significantly cheaper, allowing a new wave of startups to manufacture samples for clients, said Federico Fini, a deep tech investor at 360 Capital . Multi-wafer projects, where multiple companies can share a silicon wafer and therefore split the cost of printing, have also become more common.

Chasing Nvidia According to PitchBook's AI Report, the largest share of funding within the semiconductor segment is going to chips, with $12. 8 billion trailing 12 months as of June 30. Chips are also the fastest-growing category within the segment, with an 84.

4% year-over-year increase in deal value. According to Fini, Nvidia's dominance in AI chips is fueling investment. "Companies like AMD are aggressively looking for a way to beat Nvidia," he said.

"Startups that have a product that can do the same thing but with lower energy consumption, faster processing time or computing power are seeing a lot of engagement. " Beating Nvidia at chips may be a pipe dream, but there is one edge startups can exploit: specialization. Startups such as Etched , which hit a $10.

3 billion valuation last month following its $300 million Series C, are moving away from general-purpose chips by building systems for specific stages of the computing process, like prefill and decode. Etched offers a prefill chip for understanding prompts and a decode chip for writing out the answer. Olix's first chip is also focusing just on the decode step.

Canadian startup Taalas , which AMD agreed to acquire this month, develops chips with hardwired AI models to significantly accelerate inference. The shift toward specialization is being reinforced from the demand side. Inference cloud providers, which can match AI workloads with the chips best suited to their performance and cost requirements, are playing an increasingly important role in shaping semiconductor demand, according to Runa Capital investor Denny Gabriel.

Their growing presence could open further opportunities for startups building chips suited to specific tasks, rather than one general-purpose design. Beyond LLMs Not every bet will pay off, and, according to Gabriel, funding will likely concentrate around a few winners in the long run. But VCs argue that the underlying case for semiconductors holds regardless of which startups come out on top, particularly compared to some software bets.

"The application layer is becoming a more difficult place to invest," Amelia Armour, partner at deep tech VC Amadeus Capital Partners, said. "The fear there is that one of the AI labs releases a model that just completely wipes you out. There's no IP protection in those top-layer companies, whereas there is in hardware.

Investors move in and out of semiconductor investing, but the significant value of IP is changing investor appetite. " Hardware durability extends beyond today's models and applications. As much as semiconductors power LLMs, they will also underpin the next wave of AI, which investors believe will center around the physical world .

More capital is flowing into areas such as world models and robotics. None of it will be possible without the chips underneath, which makes the technology hard to displace. "Semiconductors are not like software; once a product is adopted, it's very sticky," Armour said.