### Podcast Summary
**Date:** 2026-11-11
#### Introduction of Guest and Theme
* **Raghu** and **Guido** host the podcast and welcome guest **Mike**, a former senior executive at Cisco overseeing late-90s M&A during the internet build-out, and currently an investor at **Anabi Capital**.
* The discussion centers on comparing the historical internet/bandwidth build-out of the late 90s with the current "great AI build-out," touching on M&A strategies, capital structures, infrastructure, and technology differences.
---
#### The Internet Build-Out vs. The AI Build-Out
* **Similarities:** Both eras involve building massive new technology (the internet/web then, intelligence now) expected to be universally transformative, accompanied by parallel surges in demand (connectivity/bandwidth then, AI capacity/tokens now) and large influxes of equity and debt capital.
* **Differences in Infrastructure Increments:**
* Bandwidth capacity increased dramatically via technologies like WDM (adding more colors of light to fiber optics to increase capacity 40x or 80x).
* Compute is bound by something resembling Moore’s law, making the pace of technological improvement more moderate compared to bandwidth at the time.
* **Practical Applicability and Revenue Ramps:** The practical applicability of AI today is higher than network bandwidth was in the 90s, demonstrated by superior revenue ramps at companies like Anthropic, OpenAI, and open-source model providers (Fireworks, Together, Modal).
* **Asset Utility and Ownership:** While underlying assets (fiber optic cables then, compute/GPUs now) will prove fully utilized and useful over time, it remains an open economic question whether they will be owned by the same entities currently holding them. Mike predicts an economic correction and market clean-up, noting that many of the 200 neo-cloud balance sheets are weak.
* **Scale and Economic Value:** The AI build-out is much larger (investing roughly 2% of GDP in AI infrastructure). The economic value of intelligence is higher than that of communication. Furthermore, markets are more efficient today, with widespread distribution of supercomputers (7 billion people having them in their pockets) and software having "eaten the world."
* **Financial and Capital Structure Differences:**
* In the 90s, more capital came proportionally from equipment vendors' balance sheets. Today, the AI build-out is more capital-driven by private debt, public equities, and Nvidia's vendor-like financing.
* Fiber optics in the ground lasted 20–30 years. Compute/GPUs were initially expected to have a short 5-year useful life; however, due to massive scarcity and AI's utility, older hardware (like Hopper H100s) is still heavily utilized and even renting at higher rates ($2.50 to $3.00+ per GPU hour).
* Longer useful life helps make the AI infrastructure asset class more financeable, though the key long-term metric will be token revenue extracted per watt (with Hopper being less efficient than Blackwell or Vera Rubin).
* Unlike the 90s bandwidth glut, there are currently no unused GPUs.
---
#### M&A Strategy: Cisco Then vs. Nvidia Now
* **Cisco’s M&A Playbook:**
* Mike oversaw roughly 70 acquisitions during his time at Cisco.
* The strategy relied on a high frequency of deals, not being afraid to fail, and obtaining statistically significant outcomes (roughly one-third worked really well, one-third were average, and one-third failed).
* Outliers delivered venture-like returns (e.g., buying a company for $500 million that generated $1 billion in earnings two years later).
* Acquisitions were funded entirely with equity (costing around 4–5% of Cisco's market cap at most) and eventually accounted for 50% of Cisco's revenue.
* A centralized integration group with repeatable processes was crucial, as businesses are good at processes but bad at one-time events.
* **Networking Integration Advantage:** Networking acquisitions at Cisco operated on open protocols (BGP, ISIS, Ethernet switching), allowing newly acquired switches/routers to immediately interoperate and scale via the existing sales team.
* **Nvidia's Position in 2026:**
* Nvidia enjoys incredible optionality and market valuation, allowing them to take risks on M&A (similar to their acquisition of Mellanox).
* Nvidia is actively attempting to shape the industry and avoid being disintermediated by relying on only 3 or 4 hyperscalers by funding numerous neo-clouds and startups. Mike notes that while these are noble intentions, scale businesses naturally consolidate to around 3 major players, making some industry-shaping efforts futile.
---
#### Startup M&A and Board Advice
* **Portfolio Company Exits:** When deciding whether to sell a struggling or tweening company, check if the founder still has the fire and desire to push through, and objectively evaluate whether the business can 10x from its current state.
* **Networking and Inbound Offers:** Startups should avoid inbound acquisition interest (often indicating an auction or a late realization). Instead, founders should constantly network, build relationships with big players and competitors, and conduct inverse intelligence gathering (e.g., weekly coffee chats with industry leaders at Nvidia, Amazon, OpenAI, or direct competitors).
---
#### Infrastructure Landscape and Semiconductor Technology
* **Supply vs. Demand Dynamics:**
* The industry currently lives in "Jensen's world," characterized by near-infinite demand for AI capacity/tokens and heavily constrained supply.
* Supply is governed primarily by silicon and fab capacity (constrained by ASML machines, wafer throughput, and fab limitations), which will likely persist for the next 2 to 3, or even 5 years.
* Demand is beginning to shift toward efficiency, with model routers directing traffic to more economically viable models (e.g., Kimmy vs. top-end Anthropic tokens boasting a 5-to-1 or 7-to-1 price difference).
* **System-Level Architecture:** AI infrastructure requires a single-system design approach rather than a patchwork of disparate parts. Google’s TPU architecture and Nvidia’s newer rack-level systems exemplify this "data center as a single unit" approach.
* **Silicon and Memory Evolution:**
* Model architectures are diversifying (FP4, FP8, mixture of experts, varying parameter counts).
* Reasoning intensity requires large parameter counts and big KV caches, favoring architectures with high memory bandwidth.
* Mike expresses strong support for 3D DRAM architectures and hybrid copper bonding for high density.
* While optics are useful for scale-up racks and cabling, copper remains preferred inside servers due to thermal sensitivities (photons and logic chips disliking high heat).
---
#### Lessons from Cisco’s Correction
* **The 2000 Market Correction:** Cisco experienced massive over-ordering because customers and supply chains anticipated scarcity and long lead times. When demand dropped, customers canceled orders down to a fraction of their prior volume. Cisco took the financial risk on supply chain orders, leading to a massive 35–40% revenue miss in the December quarter—the company's first miss in a decade—causing the stock to drop from over $80 to $9 a share within six months.
---
#### Networking Architecture Shifts for AI
* **The Death of Traditional Packet Switching Workloads:**
* Traditional networking (the internet) was built on circuit switching and stochastic stat-muxed (statistical multiplexing) packet switching to handle random arrivals, variable-length IP packets, and under-provisioned devices handling uneven traffic.
* AI data center workloads (especially training and batched inference) are consistently fully utilized, highly predictable, and lack random arrival patterns.
* **The Future of Data Center Switches:**
* Current data center switches (like Broadcom’s Tomahawk) use stat-muxed packet switching architectures, which are less adapted and unnecessarily complex/expensive for static AI traffic flows.
* Future data center networking will shift toward more predictable architectures (such as optical circuit switching, time-division multiplexing, or MEMS-based designs), enabling cheaper switches without large buffers and intelligence overhead.
* This trend presents an innovator's dilemma for traditional networking vendors who charge high prices for complex stat-muxed hardware.
---
#### Conclusion and Reflections
* Mike reflects that despite retiring in 2023 to buy a yacht and play golf, he launched **Anabi Capital** because he finds the current AI era to be the most intellectually fascinating, dynamic, and system-level time in infrastructure history—eclipsing even the 1999 boom—offering massive opportunities to build trillion-dollar companies.