What the Internet Boom Teaches Us About the AI Buildout | Mike Volpi

NVIDIA

Across the shows: 102 speakers in other episodes, 83 bullish and 57 bearish calls, past 30 days

• Mike Volpi · Hanabi CapitalBullish

  Nvidia dominates a market where AI compute demand is near-infinite and supply is tightly constrained.

  Mike says demand for AI capacity is near-infinite and supply is incredibly constrained, so the industry lives in Jensen Huang's world. The shape of the industry will only change when demand or supply characteristics change. Since supply is limited by silicon, which is unlikely to change soon, Nvidia's central position holds. Guido calls Nvidia king of the hill.

  “We live in a world where demand for AI capacity tokens is near infinite, and supply is incredibly constrained. So another way to put it is we all just live in Jensen's world.”at 0:00

• Mike Volpi · Hanabi CapitalBullish

  Nvidia's acquisition strategy is clever; Mellanox is a standout, while Hugging Face is unclear.

  Mike calls Nvidia's Mellanox acquisition a clever deal and says it is genius. He says he is not sure what the Hugging Face acquisition is about. He thinks it is fine for Nvidia to take risks, since some deals will fail and some will work as well as Mellanox. He says the market has given Nvidia a moment of incredible optionality, so it is the right time to take these risks. He uses Cisco's experience to argue that a few big winners can justify many misses.

  “So I think they're being very clever, and you can always critique any given acquisition.”at 18:45

• Mike Volpi · Hanabi CapitalUncertain

  Nvidia's financing of smaller AI players is futile, since the industry consolidates to a few scale players.

  Mike says Nvidia's financing looks like vendor financing, and it is a way of shaping what the industry looks like. He argues the industry wants to concentrate because scale matters, as with three hyperscalers and the airline industry, so the smaller companies Nvidia funds will eventually fail. He says Nvidia's intentions are noble and the strategy could earn a good return if Nvidia exits at the right time. Raghu counters that the financing enables startups to access capacity, and Mike says many of his portfolio companies benefit, so he is not calling it a bad idea, just futile.

  “No matter how much you finance them, the little ones are always weaklings.”at 20:55

• Mike Volpi · Hanabi CapitalBullish

  Nvidia and Google are the only firms able to ship full rack systems, a capability that gives them an edge.

  Volpi says that the chip vendors are moving toward systems-level approaches, and that Nvidia wants to go in that direction too. Only Nvidia and Google can currently ship a complete rack system, with AMD possibly getting there later. Because of that, any new networking architecture would have to be adopted by them first. Other chip vendors would need to agree on a standard, and the two dominant vendors are unlikely to shift off their own approaches. This gives Nvidia a structural position in the rack-level stack.

  “only NVIDIA and Google right now have the ability to ship you a rack system.”at 52:40

Cisco Systems

Across the shows: 4 speakers in other episodes, 4 bullish and 0 bearish calls, past 30 days

• Mike Volpi · Hanabi CapitalNeutral

  Cisco's high-frequency acquisition strategy worked; acquisitions became about half of revenue by 2000.

  Mike says Cisco's team did about 70 acquisitions while he was there, and they ended up producing about half the company's revenue, which was roughly $15 billion of a $30 billion base. Cisco accepted that deals would split into a third that worked well, a third that were average and a third that failed. It made enough deals to get a few outliers, such as deals that were venture-like returns. Most deals were cheap, typically 4 to 5 percent of market cap, and were paid for in equity. Open networking protocols let Cisco integrate acquired products quickly.

  “we were not afraid to fail, and we did it with high enough frequency that we got a few outliers”at 15:31

• Mike Volpi · Hanabi CapitalBearish

  Incumbent networking vendors like Cisco can't charge premium prices if the data center workload stops changing.

  Volpi argues that switches are expensive because they handle complicated, statistically multiplexed traffic: packets arrive unpredictably, must be buffered, looked up, and sent out. In a more static AI data center, where links are fully utilized and traffic is predictable, much of that intelligence and buffering is unnecessary. As a result, the incumbent vendors, including his former employer Cisco, face a bad outcome because they can't charge a lot for products that don't change. He calls it a classic innovator's dilemma: incumbents won't pursue the change, but a newcomer could build something far cheaper.

  “Networking vendors, including my former employer, this is bad news, because when things don't change, you can't charge a lot of money for them.”at 50:01

Alphabet

Across the shows: 67 speakers in other episodes, 61 bullish and 46 bearish calls, past 30 days

• Mike Volpi · Hanabi CapitalBullish

  Google's TPU architecture is probably the best AI infrastructure design because it is built as one system.

  Mike argues that the best architecture today is Google's TPU, because it is designed and thought of as a single system. Nvidia's architecture is moving in that direction but still has swappable pieces. Other builds, like a Cerebras chip with an Arista switch, are patchworks that create inefficiencies. He says the teams that can optimize the whole infrastructure as one system have a competitive advantage. He notes that Google's optical switches make the whole data center the minimum unit.

  “In fact, today, I would argue that probably the best architecture is the Google TPU architecture because it is literally thought of as a single system.”at 36:25

Cerebras Systems

Across the shows: 2 speakers in other episodes, 1 bullish and 1 bearish calls, past 30 days

• Mike Volpi · Hanabi CapitalUncertain

  Cerebras is fast, but large models that spill off its wafer lose the memory-bandwidth advantage.

  Volpi says Cerebras has done very well to go fast, but if a model doesn't fit in its on-chip SRAM, performance suffers. Once a model goes off the chip, speed drops. Because the wafer is so large, the chip is at a significant memory-bandwidth disadvantage once it goes off-chip. Cerebras now runs a large model on its chip, but he guesses OpenAI had to work hard to make that model fit. Volpi's firm is an investor in Cerebras, so he is weighing a real limitation while holding a position.

  “You're actually at a significant disadvantage in terms of the memory bandwidth once you go off chip.”at 40:56

Broadcom

Across the shows: 10 speakers in other episodes, 10 bullish and 5 bearish calls, past 30 days

• Mike Volpi · Hanabi CapitalBearish

  Broadcom's statistically multiplexed Tomahawk switching is mismatched with predictable AI traffic; a new architecture will emerge.

  Volpi says the dominant data center networking architecture today, Broadcom's Tomahawk, is a statistically multiplexed packet-switching design. The core workload of the internet was random and bursty, but AI data centers have fully utilized links and much more predictable traffic, especially in training. He expects a new architecture that assumes predictability and full utilization, such as time-division or circuit-style switching, to emerge at some point. The current switches have the right bandwidth but a less well-adapted architecture. He sees an opportunity for a newcomer to rethink the design and make it far cheaper, though he doesn't name a winner.

  “I think that at some point, for data center networking, we're going to have a new kind of architecture that assumes a high degree of predictability and a full utilization as opposed to an unpredictable stochastic stat muxed approach.”at 48:45

AI infrastructure & data centers

• Mike Volpi · Hanabi Capital

  The AI buildout will be useful and fully utilized, even though current owners may not keep the assets.

  Mike says the fiber assets from the internet boom stayed valuable for decades even after the companies that built them went bankrupt, and he expects AI compute assets to behave the same way. He points to strong revenue ramps at Anthropic, OpenAI, and open-source inference providers such as Fireworks, Together and Modal, with growth rates far above what internet providers showed in the late 1990s. He argues that demand for AI capacity is near-infinite, so the build will be used fully. Who owns the assets is a separate question that depends on macro and geopolitical factors.

  “do I think the build out that's happening now is useful and will be fully utilized, if not more? 100%.”at 6:56

• Mike Volpi · Hanabi Capital

  Expects an economic correction; weakly financed neoclouds will fail and the industry will consolidate.

  Mike says some builders have healthy balance sheets, but he counts about 200 neoclouds, and their balance sheets are not good. He expects an economic correction that will shake out weaker players, as happened in the bandwidth bust when debt-laden companies went bankrupt. He also argues that scale businesses consolidate, so the industry will end up with only a few large players, with the smaller ones eventually failing even if they are financed.

  “some of the balance sheets are pretty healthy... there are 200 Neo Clouds, and those balance sheets are not good... I expect there will be an economic correction in the market”at 7:23

• Mike Volpi · Hanabi Capital

  AI buildout capital is cleaner than the 1990s, mostly equity and capital markets, not vendor-financed equipment.

  Mike says the internet buildout drew more capital from equipment vendors' balance sheets. Today's buildout is more capital-markets driven, with private debt and public equities. There is some Nvidia-style vendor financing, but the capital is largely equity and not tied to specific equipment purchases. He concludes that the capital structures are cleaner now, which reduces some risk compared with the 1990s.

  “I'd say the capital structures are a little cleaner now.”at 10:33

• Mike Volpi · Hanabi Capital

  GPUs stay useful far longer than their five-year depreciation life, making compute assets more financeable.

  Mike says the team once assumed Hopper H100s had a five-year useful life. Four years later they still rent for about $2.50 to $3 per GPU hour, versus roughly $1.82 earlier, so the chips are not defunct. AI demand is keeping older chips valuable, and that longer life is good for lenders whose models assumed short lives. He notes that per-watt token revenue will make older chips less efficient than Blackwell and Vera Rubin, so they will eventually stop making sense to run. Scarcity in chip manufacturing keeps them in use for now. Guido adds that an abundance of GPUs could reverse these economics.

  “we thought this asset that have a five year life, partly because of AI itself, is a useful asset for quite a bit longer.”at 11:41

• Mike Volpi · Hanabi Capital

  Copper beats optics for short in-rack links, since optics are finicky and degrade as logic chips heat up.

  Volpi believes optical interconnects between separate chips are plausible, but optics are very finicky and need a lot of experimentation. Optical components don't tolerate heat well, and as the logic chip gets hotter and leaves a comfortable temperature range, connectivity issues follow. His view is that if a connection can be done with copper, it should be. Optics make more sense for scale-up racks, where they simplify cabling, but he is talking about links inside a server and for top-of-rack connections, where he says copper is the better choice.

  “When you heat up optics, they don't like it. ... photons don't like getting hot.”at 42:44

AI chips & semiconductors

• Mike Volpi · Hanabi Capital

  Silicon fab capacity is the binding AI constraint and will stay scarce for two to three years.

  Mike says power is a limit, but the dominant constraint is silicon. Unless fab capacity grows, supply stays constrained. More output would require more wafers through existing machines or more ASML machines, which takes time. He expects that the chip-manufacturing scarcity will not change in the next two or three years, maybe five. As a result, those with fab capacity are kingmakers in the infrastructure stack, and the scarcity keeps older GPUs valuable.

  “There are some power limitations for sure, but the dominant limitation is silicon. And I would say that unless there's more fab capacity, that doesn't change... it's not likely to really change in the next two or three years, maybe five years out.”at 33:31

• Mike Volpi · Hanabi Capital

  Model architectures are diversifying, so chip architectures will diversify too, with chip and model design co-optimized.

  Volpi says model architecture is becoming reasonably diverse: some labs use FP4 or FP8 precision, some use mixture-of-experts with many or few experts, and some scale parameter counts while others don't. He argues chip design and the model optimized around it are interrelated, so more model diversity means more chip diversity. Depending on the model and its intended tasks, a different chip architecture can be selected for it.

  “What I find interesting is that model architecture is becoming reasonably diverse.”at 38:29

• Mike Volpi · Hanabi Capital

  3D DRAM and hybrid copper-bonded memory stacking are promising for memory-bandwidth-bound AI inference.

  Reasoning-intensive workloads mean large parameter counts and big KV caches, so chips must move a lot of data between memory and compute. Chips can only talk to the outside through their edges, which means fewer connections than if the top and bottom of the chip were usable. Volpi is a big fan of 3D DRAM architectures and sees hybrid copper bonding as even more interesting for higher density. He thinks this architecture has huge promise as long as KV-cache-intensive use cases keep growing, and that reasoning-driven labor, which favors strong memory bandwidth, is where most AI value is created. He also acknowledges that manufacturability is much more complex.

  “I'm a big fan of three d DRAM architectures, and some of the new stuff that's coming up with hybrid copper bonding is even more interesting because you can have even higher density.”at 39:52

AI: overall outlook

• Mike Volpi · Hanabi Capital

  More optimistic about this AI cycle than the internet buildout, but expects a correction and industry cleanup.

  Mike says this cycle is building intelligence, whose economic value is probably higher than communication, so the stakes are bigger. Markets are more efficient now, so more risk capital is willing to fund AI than was possible in the 1990s. He points to the revenue ramps of AI providers as evidence of practical applicability far above internet bandwidth at the time. He still expects a correction and consolidation of the industry structure.

  “I'm actually more optimistic about this cycle, but I also believe that there will be some correction and some cleaning up of the industry structure.”at 8:32

AI pricing & unit economics

• Mike Volpi · Hanabi Capital

  Model routing will add price sensitivity to AI token demand in the near term.

  Mike says the economic gap between a top-end Anthropic token and a top-end Kimi token is about five to seven times. He expects users to get smarter about which tokens they consume, with model routers moving work to cheaper models. This will start to make demand price-dependent. He believes this is more likely in the near term. Today demand is so high that cheaper options are not yet needed.

  “I do think that that will begin to at least put some price dependency on demand, and that's more likely to happen in the near term.”at 34:53

AI bubble & tech valuations

• Mike Volpi · Hanabi Capital

  Perceived-demand bubbles cause over-ordering, then sharp cancellations and steep corrections, as Cisco's 2001 collapse showed.

  Volpi explains that when demand is perceived as a bubble, customers over-order because they expect to be cut back. Cisco was out of capacity and took on the financial risk by placing orders across its supply chain. When demand dried up, customers didn't just stop buying, they cancelled and cut orders to a quarter of their earlier levels. Cisco then missed the December quarter by roughly 35 to 40%, having never missed a quarter while public for a decade. The stock fell from the 80s to about $9 in six months, roughly a tenth of its value. This is the pattern he is drawing on when comparing today's buildout.

  “What tends to happen when you have a bubble in demand, in perceived demand, is that customers over order because they know they're gonna get cut back.”at 43:53