Andreessen Horowitz has launched the Machine Age Fund with a $1.1 billion commitment to back startups building the hardware that powers artificial intelligence. The move marks a shift from a long‑standing software focus to a sector where physics, not code, now limits progress.
AI racks are reaching megawatt levels
Today’s AI data‑center units consume between 100 and 250 kilowatts, a twenty‑fold jump from the 5‑10 kilowatt racks that housed earlier GPU generations. NVIDIA projects its next‑gen “Rubin” racks could hit a full megawatt within three years—a power draw comparable to a small factory packed into a few feet of metal.
Copper cabling, once sufficient, now bumps against its capacity limits, prompting a shift toward optical interconnects. Memory bandwidth must expand across the entire hierarchy, not just at the top, to keep pace with larger models.
Power delivery is also changing. Its 800‑volt DC architecture is emerging as a new standard, but it demands a wholesale overhaul of data‑center electrical infrastructure. Cooling systems built for ten‑kilowatt racks cannot contain the heat generated by a quarter‑megawatt unit, let alone a megawatt one.
Related: SpaceX chose Louisiana for orbital not tax reasons
Data‑center campuses are scaling toward gigawatt‑class facilities—sizes once reserved for utilities. Independent forecasts suggest global AI power demand could rise from 41 GW in 2026 to 66 GW a year later, while new capacity is added at roughly 15 GW per year.
Grid connections typically take five to seven years, far longer than the twelve‑to‑eighteen months needed to construct a data‑center building.
Venture capital steps into the hardware gap
Hyperscalers such as Amazon, Alphabet, Microsoft, Meta and Oracle plan to spend over $600 billion on AI infrastructure in 2026, with about three‑quarters earmarked for new builds. Yet their capital expenditures focus on proven, large‑scale solutions rather than early‑stage innovation.
That distinction matters because venture firms can write checks to founders developing next‑generation chip architectures, novel interconnects or liquid‑cooling systems that the big cloud players are not set up to fund. As the announcement notes, “Everything needs an upgrade, now.”
In the first half of 2026, physical‑AI startups raised $47.4 billion across 521 deals, an 80 percent year‑over‑year increase, according to data from Crunchbase. Hardware represented roughly one‑third of all U.S. venture investment that year, per the SVB Physical AI and Robotics Report. Within its own pipeline, hardware opportunities accounted for more than 20 percent of incoming deals.
Related: AI poses threat to democratic elections
For founders, the focus could translate into faster access to capital for projects that would otherwise sit idle waiting for grid approvals. It also means that companies tackling the cooling challenge of a megawatt rack might find a partner willing to back a prototype before any utility signs on.
From a practical standpoint, this could speed up the rollout of more efficient data‑center designs, allowing businesses that rely on AI models to scale without waiting years for power contracts. The immediate effect may be a modest uptick in the number of pilot facilities that experiment with high‑density racks, which could, in turn, provide real‑world data for larger operators.
The partners who announced the fund include co‑founder Ben Horowitz, Martin Casado, Raghu Raghuram, David Ulevitch and David George. Casado emphasized the broader impact, saying, “Every time we have one of these technical epochs, it puts pressure on the infrastructure.”
While the $1.1 billion vehicle is only about 1 percent of Andreessen Horowitz’s total assets under management, it represents a dedicated bet on a sector that many analysts view as a bottleneck for AI growth. Critics note that AI‑related capital spending exceeds $400 billion annually, yet global AI revenue hovers near $100 billion, raising questions about a possible valuation bubble.
Related: VC spends millions of tokens daily chasing unicorns
Nevertheless, the scope is broad, covering everything from novel semiconductor designs and high‑bandwidth memory to robotics, edge devices and even home‑AI appliances. Portfolio companies already receiving support include Netris, which automates GPU networking, as well as Unconventional AI, Nexthop, Volta, Atoms, Heron Power and Mind Robotics.
In contrast to hyperscaler spending, which purchases established technology at scale, venture capital aims to seed the breakthroughs that will define the next generation of AI infrastructure. As the industry pushes toward gigawatt‑scale campuses, the need for innovative solutions across chips, interconnects, power delivery and cooling is unlikely to diminish anytime soon.
Regulators and utilities are watching the trend closely. The U.S. Department of Energy has highlighted the importance of modernizing grid infrastructure to accommodate the rising demand from data centers, a concern that aligns with the fund’s emphasis on behind‑the‑meter power sources such as private solar and micro‑reactors.
Utilities are scrambling to keep pace.