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

1. Key Themes

The Assets Survive the Crash; the Owners Often Don't

Volpi's core thesis on the buildout: demand and utility are real, but the capital structure is fragile. In the fiber boom, the debt-laden players went bankrupt while the fiber they bought became valuable under new owners. He expects the same rhyme with compute: a correction in ownership, not in usefulness.

"The players that were most saddled with debt went bankrupt. Now, it's interesting to note that debt was buying assets, fiber optic cables, and those assets actually turned out to be quite useful in the future, except they were operated by somebody else because the guys that went bankrupt didn't have them anymore." — Mike Volpi 00:06:03

"Last count, when I was talking to the folks at NVIDIA, there are 200 neoclouds, and those balance sheets are not good... broadly, I expect that there will be an economic correction in the market, but the underlying assets will be quite useful for a long period of time." — Mike Volpi 00:07:19

GPU Useful Life Is Longer Than Financiers Assumed, an Unintended Gift to AI Financing

The depreciation assumption for GPUs (five years) looks conservative. H100s from four years ago rent for more today than at launch. Volpi frames this as an "unintended consequence of AI being good at software": the asset life extends, which makes the asset more financeable. Guido and Raghu push back that debt investors haven't fully bought this, and that it reflects scarcity rather than efficiency.

"Maybe it has a five-year useful life, and you're renting it out at, you know, $1.80, $2 per GPU hour... And that was four years ago. So, the chip should be defunct now... Not at all defunct. Not only that, it rents at a higher rate. It's $2.50 or $3." — Mike Volpi 00:11:12

"I'm not sure I have either. I mean, right now, it seems what's mostly happening is that we have a massive scarcity of an asset, so even previous generations are still worth a lot... If we ever get to a point where we have an abundance of GPUs, that might completely change." — Guido Appenzeller 00:12:35

Volpi's key metric: "how much token revenue can you extract per watt." Hopper, Blackwell and Vera Rubin are progressively better at "getting dollars out of a watt," but while scarcity persists, older chips stay in service.

We Live in "Jensen's World": Supply Constrained by Silicon, Demand Near-Infinite

The industry structure is governed by two poles, supply and demand, and everyone sits in between. Supply is bounded by fab capacity, not just power. Fab owners are "kingmakers." The supply constraint isn't expected to ease for two to three years, perhaps five.

"We live in a world where demand for AI capacity, tokens, whatever, is near infinite, and supply is incredibly constrained. So another way to put it is we all just live in Jensen's world." — Mike Volpi 00:33:13

"The dominant limitation is silicon... unless there's more fab capacity, that doesn't change, which just means it's not likely to really change in the next two or three years, maybe five years out... those who have fab capacity are actually kind of kingmakers of the infrastructure universe." — Mike Volpi 00:34:16

Demand Will Get Smarter: Model Routing and Token Price Sensitivity

The first thing that changes the industry's shape may be demand, not supply. As buyers realize the price spread between a top frontier token and a strong open-source token is 5-7x, routers will push workloads down-market, introducing price elasticity into a market that has had none.

"The economic difference between a top-end Anthropic token and a top-end Kimi token is like 5 to 1 or 7 to 1 or whatever. It's a massive economic difference, so people are going to start smarter about consuming what... that will begin to at least put some price dependency on demand." — Mike Volpi 00:34:18

AI Is a Bigger Prize Than the Internet, and Capital Markets Are More Efficient Now

Volpi argues the economic value of "intelligence" exceeds that of "communication," and that 30 years of market evolution means far more risk capital can reach AI. Revenue ramps at Anthropic, OpenAI, and open-source inference providers (Fireworks, Together, Modal) exceed anything seen in the internet era. Billions of smartphones mean consumption is more evenly distributed.

"Back then, it was communication that was being built. Today, we're building intelligence. And the economic value of intelligence is probably higher than the economic value of communication." — Mike Volpi 00:08:04

"Seven billion people have a supercomputer in their pocket. So there's much more evenly spread consumption of the underlying benefit that's being produced." — Mike Volpi 00:09:03

Capital Structure Today Is "Cleaner," but with Vendor Financing Echoes

Unlike the '90s, when equipment vendor balance sheets bore more of the risk, today's buildout is predominantly capital-markets driven (private debt, public equity). But NVIDIA's investments look "a lot like" vendor financing. Volpi's skepticism is about whether you can reshape an industry with capital.

"I don't think you're supposed to call it vendor financing, but it looks a lot like it." — Mike Volpi 00:10:06

"No matter how much you finance them, the little ones are always weaklings... We could fund 100 airlines in America, but in the end, you're going to have three... My point is not that it's a bad idea, it's just that it's futile." — Mike Volpi 00:20:48

AI Infrastructure Is Becoming a Single System, and the Unit of Design Keeps Growing

Because AI clusters run GPUs continuously with proprietary interconnects, the whole stack must be optimized as one system. Volpi argues the Google TPU architecture is today's best, and that NVIDIA is moving in the same direction. The minimum unit has grown from a card to a rack to a pod to a whole data center row.

"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." — Mike Volpi 00:35:01

"People that can think of the whole thing as a single system and optimize the architecture of it through that have a competitive advantage." — Mike Volpi 00:36:46

Chip Diversity Will Follow Model Diversity

Model architectures are diversifying (FP4 vs. FP8, differing mixture-of-experts configurations, parameter counts), which creates room for differentiated silicon. Reasoning-intensive workloads, with large KV caches, favor memory-bandwidth-rich designs, which is why Volpi is bullish on 3D DRAM and hybrid copper bonding.

"Reasoning intensity equals big parameter counts and big KV cache. And that means that you need to talk a lot between memory and chip... I'm a big fan of 3D 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." — Mike Volpi 00:39:16

Cycle Bust Mechanics: Double Ordering and the Supply-Chain Trap

Volpi's firsthand account of Cisco's collapse is a template for how a demand bubble unwinds: customers over-order, the vendor commits to the supply chain, orders get cut, and the vendor holds the inventory risk.

"Customers over-order because they know they're going to get cut back... then, when demand dries up... They're like, 'I ordered twice as much as I wanted to anyway. So I need half of half of that.' So the orders went to like 1/4 of what they were." — Mike Volpi 00:44:05

"Stock went from $80-some a share to $9 a share, I want to say. So, it was like literally 1/10 the value over a period of 6 months." — Mike Volpi 00:45:45

2. Contrarian Perspectives

Rewriting the Data Center Network from Statistical Multiplexing Back to Circuits

The foundational dogma of networking since the packet-switching revolution is that traffic is bursty and unpredictable, so you statistically multiplex and underprovision. Volpi argues AI traffic breaks that assumption, and the dominant Broadcom Tomahawk architecture is therefore mismatched to the workload, making the switch overpriced and ripe for disruption by a newcomer.

"You don't have random arrival... Links are consistently fully utilized. GPUs want to be fully utilized. Traffic patterns are actually much more predictable than they ever were before. Particularly in training, they're incredibly predictable." — Mike Volpi 00:47:49

"The switches you buy today have the correct amount of bandwidth and the incorrect architecture, not incorrect, but a less adapted architecture... In a world where it's relatively static, you don't need that much intelligence. Bam, send it right through." — Mike Volpi 00:49:10

"It's like classic innovator's dilemma, because no matter what vendor you are today, this is not a good trend... if a newcomer says, 'Forget all that statmuxing shit, the world's different, I'm going to go with my thing,' actually, you might be able to build something interesting." — Mike Volpi 00:50:30

Industries Find Their Own Shape; Funding the Little Players Is Futile

When Guido defends NVIDIA's customer-diversification investments, Volpi argues scale businesses naturally concentrate, and capital can't override that. His airline analogy: you could finance 100 carriers, but you'll end up with three.

"An industry wants to look a certain way because when it does, it creates maximum utility, maximum economic value. Why are there three hyperscalers? Because it's a business where scale matters." — Mike Volpi 00:20:36

Acquire in Volume and Accept Failure: A Third Work, a Third Don't, a Third Fail

Most corporate M&A culture treats each deal as a bet requiring a high hit rate. Cisco's approach was a portfolio: enough shots on goal to produce "venture-like" outliers, with each deal priced cheaply relative to market cap.

"A third of them worked really well, a third of them were whatever, and a third of them failed. The benefit that we did is we were not afraid to fail, and we did it with high enough frequency that we got a few outliers." — Mike Volpi 00:14:48

"If your stock trades more than the price you paid for the acquisition in one day, then you can do the acquisition as many as you want. Nobody will care." — Mike Volpi 00:16:37

Proactively Talk to Competitors and Acquirers; Don't Pay Attention to Inbound

Conventional wisdom says never "shop" your company or talk to competitors. Volpi says the opposite for founders (coffees with competitors and big companies as "inverse intelligence-gathering"), while from the acquirer's side, inbound offers should be ignored.

"Some people say, 'Well, you're shopping yourself.' No, it's an inverse intelligence-gathering mechanism. I even encourage my CEOs to talk to their direct competitors. Go to coffee." — Mike Volpi 00:29:16

"You often have the other way where companies come to you and say, 'Hi, I'd like to get bought.' I think broadly a bad idea... we never tried to pay attention to inbound." — Mike Volpi 00:29:16

The Next Cycle Is More Intellectually Demanding and Bigger Than '99

Volpi, who retired in 2023 and then launched Hanabi Capital, says this tops 1999, in part because nothing is obvious and you must reason across the entire stack.

"I don't think I've ever been in an industry before that I had to think about how light gets concentrated on a Zeiss mirror to figure out whether my chatbot works well." — Mike Volpi 00:54:27

3. Companies Identified

NVIDIA

Dominant AI GPU and systems vendor; owner of Mellanox networking. Volpi is admiring of its M&A and strategy while skeptical of its ability to reshape the industry through financing.

"Mellanox is the world's largest networking company. Right? It's like, genius." — Mike Volpi 00:18:04 "I think the market has granted them a moment of incredible optionality, right? Basically, they get to do whatever they want, which is the optimal moment for them to take these risks." — Mike Volpi 00:19:34 Volpi also expressed puzzlement at the Hugging Face deal: "I'm not entirely sure what Hugging Face is all about, for example." 00:18:07

Mellanox

Networking company acquired by NVIDIA. Cited as a model acquisition, "genius." (Quote above, 00:18:04.)

Cisco

Volpi's former employer, the most valuable company on the planet at its peak; built through acquisitions that supplied ~50% of revenue by 2000.

"During the period that I was there... in 2000 roughly, 50% of Cisco's revenue came from acquired sources." — Mike Volpi 00:15:26 "We actually never really spent much more on a given acquisition than 4, 5% of our market cap." — Mike Volpi 00:16:28

Google (TPU architecture)

Cited as having the best AI system architecture and as the pioneer of optical circuit switching in the data center.

"The best architecture is the Google TPU architecture, because it is literally thought of as a single system." — Mike Volpi 00:35:01 "Amin has been the pioneer of specifically optical circuit switching." — Guido Appenzeller 00:51:25

TSMC

Fab owner at the center of the scarcity thesis. Raghu: "Yep, TSMC." in response to Volpi's "kingmakers" point 00:34:16.

ASML

EUV equipment supplier; mentioned as the base of the vertical stack and a potential lever for expanding capacity ("or you could make more ASML machines") 00:34:16.

Zeiss

Optics supplier in the EUV chain; cited as an example of how far down the stack investors must reason ("how light gets concentrated on a Zeiss mirror") 00:54:27.

Cerebras

Wafer-scale AI chip company in which Volpi's firm (Index Ventures) is an investor. Fast when the model fits in SRAM, but weak once you go off-chip.

"Cerebras did very well to go fast, but if the model doesn't fit in the SRAM, you know?... once you get off-board of Cerebras, it's not going to go super fast." — Mike Volpi 00:40:33 "They run SOL on the Cerebras chip, so that's a big model... my guess is OpenAI had to do some hard work to get that model to fit in that chip." — Mike Volpi 00:41:02 (Note: "SOL" is the transcript's rendering of a model name that is unclear.)

Anthropic and OpenAI

Cited for revenue ramps "far superior" to anything in the internet era 00:05:38. Also referenced as top-end token providers and as startups to meet with for strategic dialogue.

Fireworks, Together, Modal

Open-source model inference providers cited as experiencing very fast growth 00:05:38.

Kimi

Open-source model whose tokens cost 5-7x less than top-end Anthropic tokens, evidence for the model-routing thesis 00:34:18.

Broadcom

Dominant data center switching silicon vendor (Tomahawk), the incumbent whose architecture Volpi argues is mismatched to AI traffic 00:48:19.

Arista

Networking switch vendor, mentioned as a pairing with Cerebras in patchwork architectures 00:36:46.

Hanabi Capital

Volpi's new venture fund, started after retirement in 2023 00:53:47. (He is described as an Index Ventures partner in the introduction.)

Index Ventures

Volpi's venture firm per the episode intro; investor in Cerebras.

Cisco Acquisition Targets (named)

Companies Volpi referenced: StrataCom (bought, and critiqued vs. Cascade), Andy Bechtolsheim's company and Mario Mazzola's company (outliers), an optical networking company bought from Vinod that "went nowhere," and Scientific Atlanta, bought after his tenure and described as "a disaster" [00:19:34, 00:24:56].

VMware

Raghu's former company, which made 50-plus acquisitions and found integration work heavy regardless of target size 00:22:19.

Intel

Guido's prior employer; example of a chip vendor disintermediated from end customers by hyperscalers 00:19:46.

Hugging Face

Mentioned as an NVIDIA acquisition Volpi finds hard to rationalize 00:18:04.

4. People Identified

Jensen Huang (NVIDIA CEO)

Referenced as the king of the AI hill. "We all just live in Jensen's world." 00:00:00 Volpi also thanks him on behalf of portfolio companies funded by NVIDIA: "many of our companies are beneficiaries of it, so thank you, Jensen." 00:22:07

Vishal (NVIDIA M&A lead)

Volpi says NVIDIA's M&A head is "in the same spot I am": "Vishal's in the same spot I am." 00:14:48

Andy Bechtolsheim

Founder whose company Cisco acquired; cited among the outliers that delivered "venture-like returns" 00:14:48.

Mario Mazzola

Founder whose company Cisco acquired; cited as a Cisco M&A outlier 00:14:48.

Vinod (Khosla)

Referenced as the seller of an optical networking company Cisco bought that "went nowhere," an example of the acceptable failure rate 00:19:34.

Amin Vahdat (Google)

Referred to as "Amin," the pioneer of optical circuit switching at Google 00:51:25.

Raghu Raghuram and Guido Appenzeller (a16z)

Hosts. Raghu ran VMware (50-plus acquisitions); Guido is a networking veteran and former Intel executive. Volpi: "an absolute privilege to be here with two legends" 00:03:03.

David and Andy

Guido's reference to packet-switching "cheerleaders" at Cisco ("David and Andy back then") 00:47:09.

5. Operating Insights

Build a Repeatable Integration Process, and Never Half-Integrate

Companies are good at processes and bad at one-time events. Cisco created a deal team under 10 people plus a larger integration project-management team with rules and cross-functional counterparts. Before that, integration was assigned ad hoc to individuals. When Cisco later did "partial integration," it failed.

"Companies are good at processes. They're very bad at one-time events. And so, essentially, if we were going to do a large number of acquisitions, we needed to have repeatable processes that were performed over and over again." — Mike Volpi 00:22:56

"We also got into the mode of like, 'Oh, we're not fully integrating them. We're like partially integrating them.'" Guido: "That's terrible." Volpi: "Yeah, it doesn't work very well." — Mike Volpi 00:25:11

Catalog the Inventory Against a Strategy Statement

Cisco's corp dev function maintained an inventory of every possible target (growth, founders, backers, customers) and mapped it against strategic needs ("more switches than routers, therefore we need a switch"). Everyone at the company was expected to be informed.

"Every company in the market is a potential acquisition, so you catalog them, and, you know, where are they in their growth rate? What kind of founders they have? Who are the backers? What are their aspirations? Who are their customers?" — Mike Volpi 00:28:43

The Sell-vs-Hold Test for Boards: Founder Fire and "Will It 10x?"

For "tweener" companies, Volpi's two tests: does the founder have the fire to push through, and is this honestly a 10x business, not just a double?

"The most important parameter is where's the founder's head at?" — Mike Volpi 00:28:05

"Not like will it double from here, because at the end of the day, if it's double, I'll take the money today, but will it 10x from here?" — Mike Volpi 00:27:13

Have the Cell Phone Number of Every Competitor CEO

Guido's tactical one-liner, endorsed by Volpi: build the relationship before you need it.

"If you don't have the cell phone number of the CEOs of each of your competitors, that's a problem. There will be a day when you need it." — Guido Appenzeller 00:31:38

Keep Deals Small Relative to Market Cap, and Use Equity

Volpi's rule: deals rarely exceeded 4-5% of market cap, nearly all in stock, which kept risk bounded and let Cisco do many deals. The failure mode he flagged: small deals eventually become too small to move the needle, and bigger deals break the integration model.

"Small things became too small to become significant to the overall business... after my time... they tried buying bigger things and then you get into the problem... Big things are very, very hard to integrate." — Mike Volpi 00:24:38

6. Overlooked Insights

AI Being Good at Software Extends the Life of the Physical Asset

Mentioned almost as an aside: Volpi says the surprising feedback loop is that AI itself (better software and kernel optimization) extends GPU useful life, which in turn improves the financeability of the compute buildout. If true, depreciation assumptions embedded in neocloud and data center debt models are mispriced in the borrower's favor, which matters for evaluating which levered players survive the correction he predicts.

"An unintended consequence of AI being good at software is that the underlying asset life is getting longer, thereby helping it to be more financeable. And that's an interesting phenomenon that I didn't quite expect when this thing got underway." — Mike Volpi 00:11:34

The Unit of Competition Is Shifting to the Whole System, Which Locks Out Component Startups

A passing exchange reveals a structural barrier: only NVIDIA and Google can currently ship a full rack-scale system, so any networking innovation (like a circuit-switched architecture) needs a rack-system vendor to adopt it, or a "rebel alliance" of other chip vendors and buyers to standardize one. That reframes where the opening is for network and optical startups: not selling a better switch, but coordinating a buyer-backed standard.

"Only NVIDIA and Google right now have the ability to ship you a rack system. So today, they would have to adopt it... can you have a collection of everybody else that says, 'No, actually, this is a standardized network architecture that we're all going to agree upon'?" — Mike Volpi 00:52:51

"The buyers have to agree. That's the key." — Raghu Raghuram 00:53:36