Mike Volpi on Why AI Infrastructure Will Outlast the Hype Cycle
a16z
AI supply and demand 0:00
AI capacity tokens are described as nearly infinite in demand and tightly constrained in supply. That sets up the comparison with the internet buildout of the late 1990s. The speaker says the current AI wave is more useful than the old bandwidth wave and is already showing faster revenue growth at companies such as Anthropic, OpenAI, and several open-source model providers. He also says the buildout will be fully used, perhaps even more than now, but ownership may shift.
Assets, life, and correction 4:04
He draws a line between bandwidth and compute. Fiber capacity jumped quickly through WDM, while compute moves more slowly under something like Moore’s law. He thinks AI infrastructure is more likely to remain valuable, because the chips keep earning despite earlier guesses about a short useful life. Still, he expects a market correction, since some neo-cloud balance sheets are weak and the industry may be cleaned up even as the underlying assets stay useful.
Acquisitions and integration 16:30
Most of the Cisco acquisitions were paid for with equity, and he says they were cheap relative to market cap. He argues that if the stock rose above the purchase price in a day, the deal could be repeated. Cisco bought companies that became major product lines, not just small add-ons. Networking made this easier because open protocols let a bought switch or router work with existing gear at once.
Selling and staying informed 26:00
On exits, he says the board has to judge three cases: a strong company that should not sell, a weak company that should, and the in-between case. In that middle ground, the founder’s drive matters most, along with a cold view of whether the business can still grow tenfold. He also urges founders to keep talking to buyers, rivals, and industry leaders, because the best deal timing usually comes from knowing the market well.
Supply stays scarce 33:02
Demand for AI tokens is still close to endless, while supply is tight because silicon is the main bottleneck. More fab capacity may come in two to five years, but not soon. That leaves the people with fabs in a kingmaker position. On the demand side, buyers are getting smarter about which tokens they consume. Model routers are pushing work to cheaper tokens, so price will start to matter more.
Systems and networking 35:31
AI infrastructure now works best when you treat the whole stack as one system. GPUs run steadily, links stay busy, and the old network model of uneven traffic does not fit as well. Google’s TPU setup is held up as the clearest system-level design, though Nvidia is moving that way too. The same shift is showing up in chips, where model variety and memory-heavy reasoning favor 3D RAM, stronger memory bandwidth, and tighter packaging.
Rethinking the network 49:31
The switches in today’s data centers have the right bandwidth but the wrong shape for this newer, more static workload. He thinks future designs may be simpler, cheaper, and possibly optical, electrical, or both, with less need for buffering and lookup. The real hurdle is not only the hardware but buyer inertia and getting the whole system to agree on a new architecture.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

