Mike Volpi: AI infrastructure can endure a market correction

Mike Volpi expects today’s AI infrastructure to remain useful even if investors or operators change. The former Cisco executive compares the GPU and data-center buildout with the fiber-optic boom that preceded the internet crash.

Raghu Raghuram and Guido Appenzeller ask Volpi what carries over from the late 1990s, when he helped lead Cisco’s acquisitions and strategy. Then, as now, people expected a foundational technology to enable products that did not yet exist, while equity and debt rushed in to finance capacity.

The internet was expected to deliver text, images, audio, video, commerce and social networking. Demand for connectivity and bandwidth was rising. But fiber capacity could also jump dramatically when wavelength-division technology put many colors of light through the same cable.

Volpi sees a difference in computing: he expects chip capacity to improve more gradually than fiber capacity did. He also argues that AI has more immediate practical uses than network bandwidth had at the comparable stage, pointing to the revenue growth of model and infrastructure providers.

When the internet boom broke, highly indebted operators went bankrupt, but the fiber they had financed remained useful under new owners. Volpi expects AI infrastructure to be heavily used as well, while saying its eventual ownership depends on economic and geopolitical conditions outside builders’ control.

He predicts a market correction, particularly among smaller cloud providers with weaker balance sheets. Yet he is more optimistic about the technology cycle itself. In his analogy the internet buildout created communication capacity, while today’s spending creates intelligence, whose economic value may be greater.

The hosts ask how financing differs. Volpi recalls a larger proportion of the earlier buildout coming from equipment-vendor balance sheets, while today’s spending draws more heavily on capital markets. He considers present financing structures somewhat cleaner, though corporate and vendor-like financing still exist.

A fiber cable can remain useful for decades; the life of a GPU is less obvious. Volpi recalls expectations that NVIDIA’s H100 would have about a five-year useful life and rent for roughly $1.80 to $2 per GPU-hour. He says it has remained useful longer and cites higher current rental prices.

Volpi suggests that AI-assisted software improvements may extend hardware life, which could improve the economics of financing those assets. Appenzeller cautions that scarcity may be keeping older GPUs valuable; if supply becomes abundant, the economics could change.

They propose revenue per watt as a useful measure. Newer chips can generate more useful output from the same electricity, so an older GPU might be unattractive once enough efficient capacity exists. For now, the speakers say, strong demand and manufacturing constraints keep older equipment in use.

Asked what NVIDIA might learn from Cisco’s acquisitions, Volpi describes a portfolio approach. In his rough recollection, about a third of Cisco’s deals worked very well, a third were middling and a third failed. High frequency and tolerance for failure produced a few exceptional outcomes.

Some acquisitions became major product lines. Volpi recalls that by about 2000 roughly half of Cisco revenue came from acquired sources, while individual purchases were generally small relative to its market value and were financed mostly with shares.

Open networking protocols helped Cisco integrate products: an acquired switch could communicate with existing routers and be sold through the same sales organization. Volpi says that made it possible to move quickly, though he cautions that not every acquisition succeeded.

He praises NVIDIA’s Mellanox acquisition and says a company in a strong position can afford to make some uncertain bets. The hosts also debate NVIDIA’s investments in the wider ecosystem. Volpi doubts capital alone can prevent a scale-driven industry from consolidating; the others note that financing can also give startups access to capacity.

Raghuram asks how Cisco integrated so many businesses. Volpi describes a small acquisition team alongside a larger project-management group that coordinated with business units. The key was a repeatable process, rather than assigning an unfamiliar integration to an executive as a one-off task.

As Cisco grew, smaller targets could become too small to affect the overall business. Volpi says larger, established companies were harder to integrate because they brought their own cultures and practices; in his view, abandoning or only partly applying the integration process created problems.

For a founder considering a sale, Volpi distinguishes strong companies that may be better kept independent, struggling ones for which a sale offers an exit, and ambiguous cases. In the middle group, he starts with the founder’s desire to keep going and an honest assessment of the business’s potential.

He asks whether the business could grow many times over, rather than merely double, and whether its leaders have the energy to reach the next stage. Those judgments can be emotional for investors who have worked with a company for years.

Cisco tried to stay informed about possible targets before a transaction: their products, growth, founders, backers and customers. It compared that inventory with a strategy, such as the need to add switching or optical networking, rather than relying only on offers from companies seeking buyers.

Volpi encourages startup leaders to maintain relationships with larger companies and competitors even when they are not selling. Conversations can provide market information and make a future decision easier. Raghuram agrees that a network of relationships matters before it is urgently needed.

Turning to infrastructure, Volpi describes a stack from manufacturing equipment to AI applications, and a value chain constrained by supply and shaped by demand. He says demand for AI capacity currently exceeds supply and gives NVIDIA unusual leverage.

In his assessment, semiconductor manufacturing is a central near-term bottleneck, alongside power. He does not expect fabrication capacity to change quickly. On the demand side, he expects customers to become more selective about model quality and cost, using routing to choose among differently priced models.

Volpi argues that today’s AI infrastructure works best when treated as a system: processors, interconnect and networking are closely coupled. He considers Google’s TPU architecture an example of whole-system design and says assembling disparate components can introduce inefficiencies.

The hosts observe that the minimum practical unit of infrastructure is moving beyond an individual accelerator toward racks or larger groups. As networking becomes integral to the system, component choices affect the performance of the whole installation.

Volpi expects more competition in AI silicon as model architectures and workloads diversify. A chip optimized for one mix of model design, numerical precision or workload may not be the best choice for another; hardware and software design influence each other.

He is particularly interested in three-dimensional memory architectures and dense bonding because reasoning-heavy workloads need substantial movement between memory and processors. The hosts also point out the manufacturing and thermal difficulty of packing logic and memory close together.

Optical connections are another possibility, but Volpi says temperature sensitivity and integration make optics challenging inside a server. For longer rack-level connections he accepts that optical links may be necessary; where copper suffices inside a system, he would prefer it.

Raghuram asks what caused Cisco’s fall after its internet-era peak. Volpi recalls customers ordering more than they needed because equipment was scarce and they expected allocations. Cisco in turn committed to its suppliers and took on financial risk to secure parts.

When demand weakened, customers canceled or reduced orders they had inflated. Cisco was left with supply commitments just as demand fell. Volpi remembers a quarter that missed expectations badly and a steep fall in the stock over the following months.

For data-center networks, he contrasts the uneven, unpredictable traffic that shaped conventional packet switching with AI training traffic, which can be more sustained and predictable. Inference can differ, though batching may make some traffic more regular.

He expects new network architectures to exploit that predictability instead of assuming all traffic is random. Circuits, time-division techniques or combinations could reduce some buffering and switching complexity. He presents this as an opportunity for lower-cost designs, not a completed industry transition.

That shift would challenge incumbent networking vendors because their existing products are built around a different problem. Optical, electrical or hybrid approaches are possible, but a switch alone is not enough: the surrounding hardware and users must adopt the architecture.

The hosts discuss Google’s optical circuit switching and debate switching speed and use cases. They also ask whether smaller vendors could agree on a common architecture, since today’s leading system providers have their own approaches. Buyers would have to support any such standard.

Volpi says he has started a new venture fund because the current cycle is unusually interesting. Understanding what happens next requires systems-level reasoning across manufacturing, power, chips and software; a bottleneck in one layer can change opportunities in another.

He says he did not expect another period in his career with this combination of technological change and uncertainty. The hosts agree that the discussion could continue much longer and close the conversation.