A top VC just put $1.1 billion behind a bet that AI hardware stays scarce. That's a signal to order your switchgear before design is even final

Andreessen Horowitz closed a $1.1 billion fund on August 28 aimed at one problem: the physical hardware AI runs on — chips, memory, networking, cooling, power gear, data centers, robotics — can't be built fast enough to meet demand. That's a venture capital firm putting real money behind the same shortage your electrical sub has been describing verbally for the past year, and it's worth treating as a second data point, not just industry noise.

What did a16z actually announce?

The Machine Age Fund, led by general partners Martin Casado and Raghu Raghuram, is a16z's first fund built specifically around physical infrastructure rather than software. Its stated case, per Bloomberg and TechCrunch's reporting on the launch: compute density per rack has jumped roughly 28-fold going from an Nvidia H100 rack to the newer Rubin generation, and the hardware supply chain — historically able to grow capacity 20-30% a year — is being asked to keep up with AI compute demand growing in the triple digits. a16z's bet is that this mismatch doesn't close soon, and that whoever builds the chips, memory, cooling, and power systems to close it captures years of pricing power. Hardware now makes up more than 20% of the firm's deal flow, up from a sliver a few years ago.

Why should a GC building data centers care what a VC fund believes?

Because the thesis isn't speculative — it matches what's already showing up in procurement. Transformer and switchgear manufacturers aren't quoting a soft shortage; they're quoting years.

Transformer manufacturers entered 2026 with order books already filled into 2028 and 2029. On the IT side — GPUs and high-bandwidth memory, which the tenant or hyperscaler procures, not the GC — Samsung and SK Hynix have separately warned memory shortages tied to AI demand could stretch past 2027. Two different supply chains, two different sets of executives, and now a venture capital firm putting $1.1 billion behind the same conclusion: this isn't a 2026 blip that eases next year.

What does this change about how you sequence a data center job?

The old sequence — finish design, then issue equipment purchase orders — doesn't survive contact with these lead times. If a generator step-up transformer takes 150 weeks and your job has a 24-month schedule, that order has to go out before the electrical design is even fully locked, on a preliminary load estimate the engineer will refine later. That's a real risk (ordering the wrong size unit is expensive to unwind), but it's smaller than the risk of a data hall sitting finished and empty for a year waiting on switchgear.

• Get long-lead equipment identified and priced during early design, not at 100% CDs. Ask your electrical engineer for a preliminary one-line and equipment list as soon as the load estimate is stable enough to act on.
• Expect deposit terms to tighten, not loosen. Comparable high-demand equipment categories are already seeing suppliers ask for non-refundable deposits 9-12 months before ship date — build that cash commitment into the owner's draw schedule conversation early.
• Don't assume a 2027 or 2028 job start means this resolves itself. Manufacturer order books already extend into 2029; a fund built on the premise that the shortage runs multiple years just got $1.1 billion more committed to that premise being right.
• Treat "AI-related" demand as broader than data centers. The same transformers and switchgear serve grid interconnection, manufacturing expansions, and EV charging buildouts — all of which are competing for the same constrained factory capacity, so waiting doesn't put you at the back of a data-center-only line.

None of this means the shortage is permanent — new fab and factory capacity is coming, just not before 2028 at the earliest on the memory side. For now, the practical read for anyone sequencing a data center job is the same one Cerebras's rack-density numbers pointed to a week ago: the equipment list, not the drawings, is now the long pole in the schedule.

For the design-side version of this same capacity crunch, see our look at what 140kW-per-rack systems mean for MEP scope.

Forward this to whoever owns your long-lead procurement log.

Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com.