Mike Volpi on Why AI Infrastructure Will Outlast the Hype Cycle
Summary
- AI capacity should remain useful even if today’s owners fail financially. Guido says AI has greater practical applicability than network bandwidth did; Volpi points to provider revenue growth that is much stronger than in the earlier cycle. He expects an economic adjustment among roughly 200 neoclouds with weak balance sheets, while the underlying assets remain useful. Who owns them depends partly on financing, macroeconomics and geopolitics.
- GPU longevity is supported by scarcity, but may not last once supply catches up. The panel discusses H100s still working four years into an assumed five-year life and renting for more than the earlier $1.82 per GPU-hour. Guido says dollars of token revenue per watt will determine which generations remain worth running: with enough Vera Rubin capacity, H100s would be a less efficient choice.
- Silicon supply and system design shape infrastructure competition. Raghu identifies silicon as the dominant supply constraint and says fab capacity is unlikely to change meaningfully within two or three years, perhaps even five. Guido expects customers to become more selective about token consumption, where high-end human-centered tokens can cost 5:1 or 7:1 as much as Kimi tokens. The discussion also points to integrated systems as an advantage.
- Cisco’s acquisition record favors repeatable execution and tolerance for failure. Volpi’s team completed 70 acquisitions, with acquired businesses contributing roughly half of revenue by around 2000; about one-third of deals failed. His exit test for founders is continued commitment and credible tenfold upside, rather than merely doubling.
- NVIDIA’s effort to broaden its customer base may support innovation, but scale can still drive consolidation. Volpi calls attempts to preserve a permanently broader customer base “pointless” where scale economics favor concentration. Guido’s counterargument is that hyperscalers pulled away from Intel’s end users and learned to build superior chips, giving NVIDIA a reason to cultivate alternative customers.
- More predictable AI traffic could weaken the case for complex packet switching. Guido argues that continuously utilized GPU fabrics need less statistical multiplexing and could use cheaper, more deterministic architectures. Adoption requires agreement across the system and its buyers; NVIDIA and Google currently ship rack systems, while AMD may reach that point.
- Cisco’s collapse illustrates how shortage-driven orders can exaggerate demand. Volpi recalls customers ordering ahead because they expected rationing, while Cisco committed across its supply chain. When actual requirements fell, orders dropped to about one-quarter; results were roughly 35–40% off, and the stock fell from $80 to $9 in six months.
Deep dive
1. Useful capacity does not guarantee solvent owners
Volpi’s internet comparison starts with a supply difference: WDM increased fiber capacity 40–80-fold, whereas computing improves more gradually. Guido says AI has greater practical applicability than network bandwidth did; Volpi points to Anthropic, OpenAI, Fireworks, Together and Modal as evidence of revenue growth well above the earlier cycle, while noting margins may differ. The hosts put infrastructure spending at about 2% of GDP; Guido says intelligence probably has greater economic value than communication.
Volpi distinguishes useful assets from solvent owners. Fiber remained valuable after indebted builders failed; he expects AI capacity to be useful too, while ownership depends on financing, macroeconomics and geopolitics. He cites a recent conversation at NVIDIA about roughly 200 neoclouds, whose balance sheets were not good. Raghu says more capital came from equipment vendors’ balance sheets in the 1990s; Volpi says today’s capital structure is cleaner because most funding is equity and not tied to specific equipment.
The panel contrasts fiber’s potential 20–30-year life with the shorter expected life of computing infrastructure. Guido recalls H100s being discussed as having a five-year lifespan and rented for $1.82 per GPU-hour. Raghu notes that four years later the chip still works; Volpi says it rents for about $2.50 or $3. Raghu calls the longer asset life an unintended consequence of AI’s software progress that can help asset-financing businesses. Volpi says investors have not widely accepted the idea, though neoclouds appear to believe it.
Guido cautions that current rents may reflect scarcity and could change once GPUs are abundant. He says the key metric is token revenue per watt: Hopper is less efficient than Blackwell, which is less efficient than Vera Rubin. With sufficient Vera Rubin supply, there would be “no reason to run H100.” Raghu adds that high demand and semiconductor manufacturing constraints keep less efficient assets in use for now; Volpi agrees that continued scarcity extends asset life.
2. Cisco’s rise and reset frame NVIDIA’s strategic choices
Volpi’s acquisition record: 70 deals, acquired businesses contributing roughly half of Cisco’s revenue by around 2000, and individual purchases rarely exceeding 4–5% of market capitalization. About one-third worked very well, one-third were acceptable and one-third failed. Outliers justified repetition: a $500 million purchase produced $1 billion of profit two years later. Almost all consideration was stock. Open networking protocols helped acquired products work with Cisco’s existing products and allowed the sales team to sell them quickly.
Raghuram, recalling VMware’s more than 50 acquisitions, presses on integration. Cisco used a deal team of fewer than 10 people and a larger central integration group with repeatable rules and guidance. “Companies excel at processes but struggle with one-off events.” Larger, established cultures proved harder to absorb after Volpi left the M&A role; he calls Scientific Atlanta “a huge mistake” and says partial integration also worked poorly.
Volpi calls Mellanox a smart NVIDIA acquisition, but doubts that financing clouds can defeat scale economics: fund 100 airlines and ultimately only three survive. Guido counters that hyperscalers separated Intel from end users and learned to build superior chips; NVIDIA understandably wants more customers. Volpi accepts that these investments can enable innovation and benefit startups, but calls forcing a permanently broader customer base “pointless.”
Volpi’s account of Cisco’s collapse centers on orders and supply-chain commitments. Customers over-ordered because they expected rationing, and Cisco committed money across its supply chain. When demand dried up, orders fell to about one-quarter of their previous level. Volpi mistook December weakness for seasonality, until January failed to recover: “Oh no, I messed up.” He recalls results roughly 35–40% off expectations and shares falling from $80 to $9 in six months.
3. Founder conviction and tenfold upside determine exit decisions
Volpi’s sale framework starts at the extremes: discourage a strong company from selling; help a financially troubled one execute an exit. The difficult middle turns first on whether founders still want to fight through the next stage. If leadership is tired and ambivalent, “let’s sell this and move on.”
The second test is whether the business can grow tenfold, rather than merely double. After four, five or six years of recruiting friends and committing yourself, objectivity becomes emotionally difficult. Nevertheless, the board must assess the remaining opportunity honestly. Where the founder’s mind sits remains “the most important factor.”
Cisco cataloged potential targets before they sought a sale and paid little attention to unsolicited approaches. Volpi encourages founders to meet large companies and direct competitors as a “reverse information-gathering mechanism.” Those relationships improve industry understanding and make eventual transactions less mysterious; founders should already know the relevant people when circumstances change. He suggests coffee with people at NVIDIA, Amazon and OpenAI once or twice a week.
4. Silicon scarcity rewards companies that optimize whole systems
Raghu says demand for AI capacity and tokens is virtually limitless while supply is extremely limited. He identifies silicon as the dominant constraint, despite power limitations, and says meaningful change requires more fab throughput or greater production of ASML machines. He considers a change unlikely within two or three years, or even five: “We’re all living in Jensen’s world.” Guido adds that companies with fab capacity are kingmakers of the infrastructure world.
Guido expects demand to adjust sooner than supply. Customers may become more selective than simply consuming frontier tokens, using model routers to account for price differences. He cites a 5:1 or 7:1 economic difference between high-end human-centered tokens and high-end Kimi tokens.
The conversation describes a shift toward integrated systems. Raghu says that in networking, workloads were bursty and easier to multiplex, while today many GPUs run continuously and components rely on proprietary standards such as NVLink. He says Google’s TPU architecture may be the best example of treating the system as a whole, and that NVIDIA is moving in that direction. The group discusses the smallest optimization unit expanding from cards to racks and, in Google’s optical-switching architecture, potentially the data center.
Volpi’s investment lens runs from EUV equipment to applications: light focused through a Zeiss mirror can affect whether a chatbot works. Having planned to retire in 2023, he instead finds the interdependencies intellectually compelling. Unlocking bottlenecks could create companies worth $1 trillion or $500 billion, but understanding them requires thinking “at the system level.”
5. Reasoning makes memory bandwidth a central chip advantage
Volpi says diversity in model architectures—including FP4 versus FP8, different expert counts and parameter sizes—should support different chip designs for different workloads. Guido favors 3D DRAM and hybrid copper bonding because reasoning and larger KV caches require intensive memory communication. Connections through the top and bottom of chips can offer higher density than connections around the edges. As long as KV-cache-heavy workloads persist, Guido expects high-bandwidth architectures to benefit.
The discussion uses Cerebras to illustrate the tradeoffs: models that fit within SRAM can run quickly, while going off-wafer creates a substantial memory-bandwidth disadvantage. Raghu says the large model described as “Sol [?]” is running on Cerebras; Volpi guesses that OpenAI must have done considerable work to fit it.
The speakers discuss thermal complexity in packaging logic and memory together, as well as the sensitivity of optical interconnects to heat. Guido says, “if it can be done with copper, then it should be done with copper,” while the conversation also notes rack-scale benefits for bandwidth and cable routing.
6. Predictable AI traffic challenges packet-switching economics
Guido explains that internet packet switching exploits bursty, unsynchronized traffic through statistical multiplexing. AI workloads have changed: links are fully utilized, and training traffic is highly predictable. Inference is less predictable, but batching can make it more so. Guido says the fundamental principles of workloads have changed.
Guido identifies Broadcom’s Tomahawk as a statistically multiplexed packet-switching architecture. More deterministic approaches could use optical circuit switching, time-division multiplexing or combinations. Packet switches justify their complexity through memory writes, lookups and forwarding decisions; in a relatively static world, a network may need less intelligence. “You just send it as is.” Guido sees an opportunity for a much cheaper network and calls it “a classic example of the innovator’s dilemma” for existing vendors, including Cisco.
The discussion points to Google’s optical circuit switching as an example, while Raghu asks why reconfiguration takes about a minute. Guido says he does not fully understand the delay and suggests current use cases may be fairly static; Raghu disputes that characterization. Guido says, “give me a few milliseconds, not a second,” and Raghu replies that the technology may exist but lack use cases.
Adoption requires the rest of the system, rather than an isolated switch. Raghu says NVIDIA and Google currently ship rack systems, while AMD may reach that point. The speakers discuss whether additional chip vendors could agree on a standardized network architecture—a “Rebel Alliance”—and Raghu says buyer agreement is essential.