A16z Starts Machine Age Fund as AI Demand Hits Physical Power Limits
Ben Horowitz, Erik Torenberg, Martin Casado, Raghu Raghuram
AI scaling breaks traditional software economics by letting capital directly purchase intelligence, shifting the industry's core bottleneck from model code to physical power, cooling, and silicon.
Andreessen Horowitz has launched the Machine Age Fund to target the computing and physical stack 'south of the model'—including silicon, high-bandwidth memory, networking, power delivery, cooling, and data center facilities. (0:00)
The primary bottleneck in AI has shifted from algorithmic breakthroughs to hardware constraints, with hyperscaler CapEx projected to surge from $700 billion toward $1 trillion next year while key supply chain components remain booked solid into 2028. (5:00)
Server rack power requirements are surging from legacy baselines of 5–10 kW to 100–250 kW, rendering traditional AC power distribution and air cooling unviable and mandating high-voltage DC architectures and liquid cooling loops. (28:10)
Unlike the speculative telecom bust of the 1990s, current AI hardware is pre-sold before fabrication, and secondary markets routinely see chips trading at up to four times their original purchase price. (7:15)
Detailed notes
## Background
Published on August 28, 2026, via RSS by Erik Torenberg alongside Andreessen Horowitz partners Ben Horowitz, Martin Casado, and Raghu Raghuram, this episode discusses the launch of a16z's dedicated Machine Age Fund. The show notes outline an in-depth conversation exploring how artificial intelligence bottlenecks have shifted from algorithmic model design to the physical stack—including silicon, memory, networking, power generation, and data centers—and examine why this infrastructure cycle fundamentally differs from previous technology booms.
## TL;DR
The partners explain why a16z created the Machine Age Fund, arguing that the primary gating factor for AI progress has migrated 'south of the model' into physical systems, memory, power delivery, and data center facilities. Unlike traditional software development constrained by engineering management and Brooks's Law, frontier AI allows capital to be converted directly into raw compute and intelligence, triggering historic hyperscaler CapEx and severe multi-year hardware backlogs. They analyze the physical constraints breaking legacy computing facilities—including a shift to 100–250 kW racks, liquid cooling, and direct current (DC) architectures—and examine the emergence of the intergenerational 'systems founder.' Ultimately, they frame the physical infrastructure buildout not merely as a multi-billion-dollar market opportunity, but as a crucial geopolitical imperative to secure technological abundance and protect democratic values.
## Atmosphere & style
The discussion carries an authoritative, insider tone steeped in technical depth, combining venture-scale ambition with sober physical reality. The dynamic between Horowitz, Casado, Raghuram, and Torenberg feels like an executive war-room debrief, where grand historical comparisons to the Industrial Revolution and Henry Ford's River Rouge sit naturally alongside detailed debates over copper interconnect limits and 800V DC power delivery. A distinct sense of historical irony permeates the room as partners famous for championing 'software is eating the world' candidly admit that progress has collided head-on with the stubborn physical limits of concrete, cooling loops, and electrical grids.
## Bulletin
- `00:00` Andreessen Horowitz has launched the Machine Age Fund to target the computing and physical stack 'south of the model'—including silicon, high-bandwidth memory, networking, power delivery, cooling, and data center facilities.
- `05:00` The primary bottleneck in AI has shifted from algorithmic breakthroughs to hardware constraints, with hyperscaler CapEx projected to surge from $700 billion toward $1 trillion next year while key supply chain components remain booked solid into 2028.
- `28:10` Server rack power requirements are surging from legacy baselines of 5–10 kW to 100–250 kW, rendering traditional AC power distribution and air cooling unviable and mandating high-voltage DC architectures and liquid cooling loops.
- `07:15` Unlike the speculative telecom bust of the 1990s, current AI hardware is pre-sold before fabrication, and secondary markets routinely see chips trading at up to four times their original purchase price.
## Key takeaways
- `00:00` The operational bottleneck in artificial intelligence has moved south of the model, requiring a ground-up reconstruction of physical computing infrastructure rather than incremental server tweaks.
- `03:00` Founder behavior reflects a hardware renaissance, with elite venture deal flow in physical systems jumping from roughly 3%–5% historically to nearly 30% today.
- `03:00` Unlike SaaS startups focused primarily on customer acquisition, AI companies face virtually unlimited demand but are constrained by gross margins and operational compute efficiency.
- `05:00` Token consumption per interaction is experiencing an order-of-magnitude leap as workflows evolve from 100-token conversational prompts to reasoning loops and autonomous agents consuming tens of thousands of tokens.
- `07:15` Today's infrastructure buildout differs fundamentally from the dot-com bubble because silicon is pre-sold, data centers operate at max capacity, and enterprise penetration sits at only 5% to 10%.
- `12:35` The Mythical Man-Month engineering governor has broken down: software progress historically stalled when adding human engineers, but AI capability can be directly accelerated by pouring capital into compute and clusters.
- `15:43` Massive capital has become an asymmetric competitive weapon, enabling well-funded challengers to rapidly erase multi-year software leads by deploying multi-billion-dollar clusters.
- `20:00` AI deployment paradigms are progressing from search widgets and chatbots to autonomous digital employees that are provisioned their own machines, browsers, and identities.
- `25:00` Frontier models cost $3 billion to $5 billion to train and incur double that in inference, making it economically rational to design custom ASICs tailored to the static weights of a single model.
- `30:00` Data center racks scaling to 100–250 kW require high-voltage DC distribution, exposing a critical labor bottleneck where only 2% of US electricians are certified to work on DC at those voltages.
- `35:00` Permitting friction and grid gridlock—with 44 gigawatts of new data center demand facing only 25 gigawatts of grid additions by 2028—are pushing compute infrastructure abroad to regions like Mexico and Australia.
- `40:00` Platform incumbents like NVIDIA intentionally bypass peripheral efficiency optimizations to focus on core demand, creating massive supplier niches reminiscent of the post-Fordlandia automotive unbundling.
- `45:00` Infrastructure startups require 'systems founders' who can coordinate cross-disciplinary envelopes across silicon design, packaging logistics, manufacturing queues, and power envelopes.
- `50:00` A two-decade talent deficit in hardware and systems engineering is reversing, with young founders pairing directly with veteran physical systems operators to navigate multi-million-dollar capital cycles.
- `53:18` The long-term objective of the Machine Age Fund is to establish domestic compute and power abundance to preserve American technological preeminence and democratic values.
## Detailed walkthrough
### `00:00` Announcing the Machine Age Fund and the Infrastructure Bottleneck
Erik Torenberg opens the conversation by announcing Andreessen Horowitz's launch of the Machine Age Fund, a dedicated vehicle targeting the underlying physical and systems infrastructure powering the modern artificial intelligence boom. Introducing the fund's premise, Torenberg cites Marc Andreessen’s historical framing: “This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. The comps on this are the microprocessor, the steam engine, and electricity, or maybe the wheel.” With AI expanding exponentially, the partners argue that the central gating factor in artificial intelligence has fundamentally moved away from model design toward everything operating beneath it—including silicon, high-bandwidth memory, networking, power generation, cooling systems, and data center real estate.
Ben Horowitz emphasizes that computing shifts of this scale mandate a ground-up reconstruction of the entire physical stack. Rather than incrementally tweaking existing servers, the industry now faces resource limits that stretch from raw commodity extraction to novel materials. As Horowitz explains: “We have a whole new technology that's the most important technology ever, and you need a whole new infrastructure.” He points out that the sheer magnitude of power and compute requires radically rethinking system software, inventing new methods of power delivery, and even finding substitutes for copper.
Raghu Raghuram reinforces this view by observing how model advancement has altered the engineering landscape over the prior three years. While initial progress was bound by algorithmic breakthroughs, the models are now improving at an accelerating clip and can even assist in their own development. Consequently, the primary operational bottleneck has shifted south of the software layer. As Raghuram puts it: “Now the bottleneck is all what I call south of the model.”
```mermaid
flowchart TD
accTitle: The AI Computing Stack South of the Model
accDescr: Diagram illustrating the shift of AI bottlenecks from frontier models down to the physical infrastructure layers.
subgraph Top["Software & Application Layer"]
A["Frontier Models & AI Agents"]
end
subgraph Bottleneck["'South of the Model' Infrastructure Bottleneck"]
B["System Software & Networking"]
C["Chips & High-Bandwidth Memory (HBM)"]
D["Power Generation & Liquid Cooling"]
E["Physical Commodities & Raw Materials (e.g., Copper)"]
end
A --> B
B --> C
C --> D
D --> E
```
`03:00` The Founder Renaissance in Hardware and Physical Systems
Martin Casado points out that venture capital usually lags behind technical talent, and the formation of the Machine Age Fund directly reflects an unmistakable shift in founder behavior over the past two years. Historically, hardware pitches were an extreme outlier among top-tier entrepreneurs, representing perhaps 3% to 5% of elite deal flow. Recently, that share has exploded toward 30%, with world-class engineering teams abandoning pure software applications to tackle foundational systems and physical architecture. As Casado observes: “So like the founder community, which tends to be much smarter than the VC community, has identified this as a very active area for innovation and they're responding.”
This founder migration is driven by unprecedented market forces: demand for AI compute appears virtually boundless, subjecting every single node of the hardware supply chain to severe stress. Horowitz and Casado observe that the fundamental economic equation of AI startups differs completely from traditional SaaS companies. In standard software businesses, the central anxiety is driving customer demand and revenue growth. In the AI wave, demand is effectively guaranteed; instead, the existential challenge is gross margin and operational efficiency.
Because legacy computing architectures were never engineered to support modern tensor-heavy workloads, existing hardware imposes hard physical limits on unit economics. To deliver intelligence profitably at scale, entrepreneurs are compelled to redesign microarchitectures, memory interconnects, and cooling loops to eliminate structural overhead.
`05:00` Hyperscaler CapEx, Exploding Token Demand, and Sold-Out Supply Chains
Addressing whether the sector is caught in an unsustainable hype cycle, Raghuram highlights the aggressive capital expenditure of the major hyperscalers, who possess the clearest visibility into global enterprise, international, and frontier lab demand. Hyperscaler CapEx is projected to climb from approximately $700 billion this year to near $1 trillion next year. Across both foundation model labs and early-stage application providers, revenue growth and usage trajectories are setting historical records. Furthermore, hardware pricing defies traditional tech deflationary cycles: GPU and component prices continue to hold or rise, with only an estimated 5% to 10% of the total addressable enterprise market having deployed solutions to date.
The structural shortage is compounded by an explosion in token consumption per unit of economic work. While first-generation conversational interfaces required only around 100 tokens per interaction, complex reasoning systems and autonomous agentic workflows routinely burn thousands or tens of thousands of tokens per task. Key components across the semiconductor and systems supply chain are essentially booked solid through 2027 and into 2028, leading to extreme allocation scrambles, including multi-day competitive auctions where buyers bid up batches of just a few thousand GPUs.
- **Hyperscaler CapEx Spend:** Tracking around $700 billion collectively in the current year, with projections reaching upwards of $1 trillion next year.
- **Enterprise Penetration:** Only about 5% to 10% of the eventual addressable enterprise market is currently stood up on modern AI compute.
- **Supply Backlog:** Critical compute and infrastructure components are fully committed and sold out through 2027 and 2028.
- **Token Consumption Scaling:** Interactive chat queries consume roughly 100 tokens, whereas agentic reasoning loops consume thousands of tokens per single unit of output.
`07:15` Why This Is Not the Dot-Com Bubble: Real Demand, Pre-Sold Silicon, and Server Memory Windfalls
The group directly contrasts the current environment with the late-1990s dot-com bubble. During the 1998–1999 telecom boom, vast amounts of capital went into laying "dark fiber" in the ground speculatively, without end-user demand or applications like consumer streaming video ready to consume the bandwidth. Today's AI buildout exhibits the inverse dynamic: hardware is pre-sold before it is fabricated, data centers are immediately operating at capacity, and secondary markets see hardware trading hands at staggering premiums. As Horowitz summarizes: “This is like we're flat out and people are reselling GPUs for four times what they bought them for and this kind of thing.”
Horowitz adds that compounding the supply crunch are severe physical and geopolitical bottlenecks: the industry has run out of available electrical power, specialized cooling infrastructure, memory capacity, and data center footprint, all while facing intense regulatory and political friction against rapid physical construction.
To illustrate the sheer scale of the hardware supply imbalance, Casado shares an anecdote from an enterprise executive:
> "I was talking to a CFO of a large company, a large public company, who had historically been very resistant about going into the cloud. So they had a lot of servers. And they were doing an inventory check, and they realized that the memory in their servers had increased so much it could fund the entire migration to the cloud."
Raghuram confirms that this scarcity extends deep into high-bandwidth silicon. Citing reports from the annual Hot Chips conference at Stanford University, he notes that a market-leading memory vendor stated that satisfying their current inbound customer demand alone would require three full years of total manufacturing capacity.
`10:00` The Supply Lag and Why Previous Tech Epochs Look Small by Comparison
The discussion kicks off around the profound mismatch between supply and demand. The bottleneck is not just about forecasting future needs; even meeting today's baseline demand is fundamentally broken. When asked whether the tech industry merely underestimated the capability of these models, the group argues the gap goes much deeper: nobody could have supplied enough infrastructure because generative AI exploded onto the scene only four years ago. Chip development cycles naturally run three to four years, and securing land, approvals, and construction for modern data centers takes four to five years. Compounding that, operators must either build their own dedicated power generation, secure massive grid tie-ins, or both. Traditional infrastructure industries are accustomed to considering 20% to 30% annual growth exceptional, yet they are now expected to feed an AI software ecosystem where triple-digit growth is merely the baseline.
```mermaid
flowchart TD
accTitle: Physical Infrastructure vs AI Software Growth Mismatch
accDescr: Diagram illustrating the severe lag between physical buildout timelines and AI software demand growth.
subgraph SoftwareDemand["AI Demand Surge (Triple-Digit Baseline)"]
D["Explosion of Enterprise Adoption"]
E["Token Demand Growing ~1,000% Annually"]
end
PhysicalConstraints -.->|"Cannot keep pace with"| SoftwareDemand
```
Addressing why a dedicated Machine Age fund was not created five or seven years ago, Martin Casado reflects that while starting two years earlier might have been ideal, launching it now is right on time. In prior technology transitions, an architecture shift produced only one or two dominant infrastructure winners:
- The transition from mainframes to client-server brought up specialized players.
- The shift to the internet minted companies like Cisco and Juniper Networks.
- The rise of mega hyperscale data centers—even though largely internalized by the cloud incumbents—yielded Arista Networks.
In every previous era, the disruption was isolated to a single switch or chip company, whereas today every single layer of the stack must be reinvented. Raghu Raghuram and Ben Horowitz highlight that hardware architectures designed for prior eras are hitting hard physical limits, such as signal loss and thermal barriers in copper interconnects. With token demand expanding toward 1,000% annually and consumer adoption barely beginning, software needs have outpaced legacy physics.
As Martin Casado points out regarding the fund's timing: “We probably could have started a little bit earlier, but the amount of change is so high now that it's just an obvious thing to do.”
`12:35` The Death of the Mythical Man-Month: Money Directly Yields Intelligence
The conversation shifts to the demand side and the compounding nature of model reasoning. Moving from simple chatbots to step-by-step reasoning, autonomous agents, and multi-agent workflows has dramatically multiplied the token requirements per single user query. Erik Torenberg remarks that “nobody likes to use AI more than AI,” asking why token consumption does not level off. Martin Casado explains that current scaling is achieved primarily through inference-heavy methodologies: reinforcement learning (RL), chain-of-thought prompting, and persistent long-running agents all devour immense inference cycles.
```mermaid
flowchart LR
accTitle: Legacy Software Scaling vs Machine Age Scaling
accDescr: Comparing the engineering bottleneck of legacy software development with the capital-to-compute model of modern AI.
subgraph Legacy["Legacy Software Scaling"]
L1["Capital In"] --> L2["Hire Engineers"]
L2 --> L3["Mythical Man-Month Bottleneck: 2+ Years / High Failure Rate"]
L3 --> L4["Software Product"]
end
subgraph Modern["Machine Age AI Scaling"]
M1["Capital In"] --> M2["Massive GPU Clusters & Power"]
M2 --> M3["Inference & Training Scaling Laws (RL, CoT)"]
M3 --> M4["Raw Intelligence Generated"]
end
```
More fundamentally, the entire paradigm of computing has changed. Software development historically operated under Fred Brooks's law from *The Mythical Man-Month*—throwing more human engineers at a late software project simply makes it later. Engineering complexity functioned as a natural governor or speed limit. In AI, however, intelligence has turned into a pure resource allocation problem. As Casado highlights: “It used to be when you built something, it was an engineering problem. You'd throw a bunch of engineers at it, and that doesn't scale, and that would have a natural law of engineering physics, which is what the Mythical Man Month came out of.”
In this new paradigm, capital bypasses the human organizational friction:
- Massive capital is funneled directly into clusters, silicon, and megawatts.
- Hardware, data, and compute translate systematically into improved intelligence.
- Auto-catalytic loops emerge where developers use AI to write lower-level GPU kernels, which in turn speeds up the AI that designs the next generation of systems.
As Raghu Raghuram notes: “Any problem that you have can be solved with enough infrastructure, GPUs and power and money. And so until we run out of problems, we're not going to run out of demand.”
`15:43` Capital as an Asymmetric Weapon and the Knowledge Worker Frontier
For the past decade, Silicon Valley operated under the consensus that venture capital had become oversaturated and startups were overfunded. Ben Horowitz explains why that old playbook has been rendered obsolete by the sheer economics of AI clusters. In the traditional software world, a startup with a two-year lead could not be caught by an incumbent throwing a thousand engineers at the problem; doing so would merely implosion the organization, honoring the adage that “nine women can't have a baby in a month.”
Now, money acts as an immediate equalizer. A company can suddenly erase a competitor's multi-year lead simply by deploying billions of dollars into massive compute infrastructure. Horowitz points to how quickly models like Grok or Kimi can materialize at frontier quality when backed by immense capital: “It's taking $3 billion and like lighting up a magnificent cluster and then all of a sudden, you know, whatever, Grok can come out of nowhere and like, oh, all of a sudden it's real or Kimi or what have you. It's just like these leads, you can throw money at the problem, and you can throw money at almost any problem, and that works. And so that is just completely different than anything we've ever lived through.”
The group maps out how this compute consumption will ripple across society:
- **Casual Users to Professional Coding:** What started as consumer curiosity in ChatGPT has evolved into coding assistants heavily utilized by roughly 30 million global developers, commanding a disproportionate share of compute.
- **The Knowledge Worker Wave:** The next frontier encompasses more than one billion knowledge workers across the world who are shifting toward deep workplace automation and back-office agents. Each phase jump unlocks an order of magnitude more token consumption.
- **Computer-Use Agents & Embodied AI:** Beyond text, tools like Grok and computer-use agent frameworks sit directly inside user interfaces, driving machines like human operators. Robotics and embodied physical systems will form an even larger second wave of demand that has not yet entered the grid.
`20:00` Computer Use, the Steam Engine Analogy, and AI as the New Digital Employee
The conversation opens with an examination of how artificial intelligence is evolving beyond traditional coding into genuine computer use—handling mundane, manual digital tasks like updating credit cards across forgotten online services and canceling unwanted subscriptions. These agents represent a brand-new class of knowledge workers residing entirely within computers. Comparing this transition to historical breakthroughs, the group points out that Marc Andreessen’s comparison to the steam engine or electricity is precise: a fundamental source of power has been discovered that can be directed toward productive labor. Society is only at the beginning of a multi-decade wave of throwing compute at problems with clear reward signals. While text and software engineering were the first frontiers, computer use is just starting, and fields like materials science, biology, and creative expression represent massive upcoming opportunities.
```mermaid
flowchart TD
accTitle: Evolution of AI Deployment Models
accDescr: The progression of AI interfaces from basic search bars to chat interfaces, personal proxy agents, and finally autonomous employees.
subgraph S1["Interface Paradigms"]
A["Search Enhancement<br/>(AI as better Google)"]
B["Chat Interface<br/>(Back-and-forth conversational UI)"]
end
subgraph S2["Autonomous Entities"]
C["Personal Proxy<br/>(Shared passwords, human extension)"]
D["Digital Employee<br/>(Dedicated identity, browser, and computer)"]
end
A --> B
B --> C
C --> D
```
Martin Casado traces the progression of how the tech industry conceptualizes AI entering human workflows. Early approaches treated AI merely as a smarter search bar or an interactive chat prompt. That paradigm then shifted toward standalone assistants—such as OpenClaw—that acted as an extension of the user, sharing credentials and executing tasks on their behalf. However, Casado highlights that systems like GroqBot represent the true breakthrough by being treated outright as independent digital employees. Rather than piggybacking on human credentials, the bot is assigned its own machine and browser to leverage frontier intelligence autonomously. Casado notes that whenever an operational chore arises, his default assumption has become whether GroqBot can execute it, ranging from calendar management and booking meetings to triage across his email inbox without explicit micro-management.
Turning to the organizational and cultural impact inside Andreessen Horowitz, Ben Horowitz reflects on the steep learning curve required to integrate non-human staff alongside human counterparts:
> “We spent many, many, many years figuring out how to work with our kind of regular human employees and now we've got these other kinds of employees.”
Horowitz cautions that while these agents can be immensely productive, they also introduce significant friction: they can burn through expensive tokens without accomplishing anything, forget context, hallucinate, behave erratically, and introduce critical security vectors. Rather than aiming for full automation to eliminate human staff, the actual goal is figuring out how to augment personnel and avoid disaster:
> “We're much more going like, okay, how do we make all our humans superhuman without wrecking the place because the bots get out of control?”
- The industry is moving from conversational chat widgets to sovereign autonomous software agents that operate dedicated computers.
- Rather than sharing human keys and passwords, modern agents are provisioned their own environments, browsers, and permissions as autonomous corporate entities.
- Integrating agent workforces demands new management practices to monitor token burn, security boundaries, and erratic agent behavior.
---
`25:00` Rebuilding the Stack from First Principles: Silicon Economics and Model-Specific ASICs
The discussion turns toward the infrastructure supply side, emphasizing that legacy data centers, system architectures, and facility designs were never built to handle the unique demands of artificial intelligence. Raghu Raghuram lays out the mental model for how founders must dismantle and rebuild infrastructure from first principles. By inspecting every bottleneck in the execution loop—from how an inference engine consumes vast pools of memory to generate tokens to the arithmetic operations inside the compute engine—every component must be rethought. Engineering teams must determine how memory and compute communicate, how memory is arranged hierarchically, how much power is drawn across chips and data centers, and how to cool these systems efficiently.
Martin Casado contextualizes this shift around the staggering economic scale of frontier model development. Training a frontier model now costs between $3 billion and $5 billion in direct capital, requiring an inference payback of roughly double that—around $10 billion—to make commercial sense. Under those numbers, even a 20% efficiency optimization equates to $2 billion in real value, which completely alters the hardware equation:
> “We've actually gotten to this interesting point in the industry where it actually makes sense to build an ASIC per model just because the amount of capital investment in that model.”
Unlike traditional software, which is inherently stateful, highly dynamic, and constantly changing, neural network weights remain fixed once trained. Because billions of dollars are poured directly into a single static digital artifact, hardware architectures can become hyper-specialized around the specific math of an individual model to maximize throughput and power efficiency.
- Frontier model training runs routinely require $3 billion to $5 billion in up-front compute investment.
- Incurring roughly $10 billion in inference costs justifies building customized application-specific integrated circuits (ASICs) optimized for a single model's weights.
- Unlike traditional dynamic codebases, the static nature of trained neural network weights allows hardware architectures to be engineered around the fixed matrix multiplications of a specific model.
---
`28:10` The Physical Limits of Power and Cooling: 250 kW Racks and the Shift to DC
The technical demands of modern clusters are breaking conventional data center standards. Average server rack power densities are experiencing a monumental leap, surging from legacy ranges of 5 to 10 kilowatts up to 100 to 250 kilowatts per rack—an estimated 70x increase in compute density. This extreme concentration of power makes legacy facility engineering completely obsolete, opening massive investment opportunities across physical infrastructure.
Ben Horowitz points out that at 100 to 250 kilowatts per rack, standard alternating current (AC) power distribution fails:
> “Well, first of all, when you get to that level of power per rack, AC power doesn't work anymore.”
Data centers are forced to transition to direct current (DC) power architectures. Horowitz remarks on the historical irony of this pivot: Thomas Edison once campaigned aggressively for DC power by publicizing the lethal dangers of AC through public electrocutions of animals, yet direct current at massive data-center scale brings extreme physical danger and unique cooling requirements of its own.
Simultaneously, thermal management must adapt. Air cooling is wholly incapable of dissipating heat at these densities, making liquid cooling an absolute baseline requirement for any state-of-the-art facility. However, simply installing liquid loops is no longer sufficient; operators must now navigate environmental regulations and supply chain scrutiny, sparking demand for eco-friendly cooling fluids and advanced thermal containment designs.
- Standard server rack power requirements have risen from 5–10 kW to 100–250 kW, driving a roughly 70x surge in compute density.
- Extreme rack power levels render AC power unviable, forcing data center operators to deploy high-voltage DC power architectures.
- State-of-the-art AI facilities have fully transitioned from air cooling to liquid cooling, accompanied by rising demand for sustainable, non-toxic coolant chemistries.
`30:00` Physical and Human Bottlenecks in Data Center Engineering
The group examines how the sudden surge in AI compute density is running directly into the physical constraints of data center architecture, the power grid, and the labor pool. Data centers cannot simply consume power indiscriminately anymore; society requires them to contribute back rather than act as parasites. While bad actors make up a small portion—perhaps 10%—wasting water (though still far less than agriculture like almonds or pistachios), that era is hitting a hard stop. Next-generation server racks are so dense and heavy that the physical buildings themselves are becoming obsolete: floor loading limits must be re-engineered, walls must be made thicker to baffle deafening acoustic noise so facilities do not violate neighborhood peace laws, and soaring demand has made reinforced concrete one of the fastest-rising line items in construction costs.
```mermaid
flowchart TD
accTitle: Data Center Bottlenecks
accDescr: Physical and operational constraints colliding with modern AI data center construction.
subgraph PhysicalConstraints["Physical Facility Limits"]
A["Dense Rack Weight"] --> B["Reinforced Concrete & Floor Loads"]
C["Extreme Acoustic Noise"] --> D["Thicker Wall Insulation"]
end
subgraph ElectricalLabor["Electrical & Power Bottlenecks"]
E["800V DC Rack Power"] --> F["Dangerous High-Voltage Environment"]
F --> G["Severe Shortage: Only 2% Electricians DC Certified"]
end
subgraph FutureOps["Next-Gen Operations"]
H["Robotics & Automation"] --> I["Automated Server Assembly & Racking"]
end
```
Simultaneously, the electrical reality inside the data center has shifted dramatically. Power feeds delivering up to 800 volts direct current (DC) directly to racks have made these facilities significantly more dangerous, exposing a severe labor bottleneck: only about 2% of electricians in the United States hold certifications to work on DC power at these voltages. While companies like Meta have established training programs resembling a modern Job Corps to upskill workers for free, the irony is glaring against popular narratives about automation.
> “AI is taking all the jobs. AI is going to create a lot of new electricians.”
Cloud hyperscalers are also furiously experimenting with robotics to physically assemble and rack servers, driven by the sheer scale of the equipment. Clarifying the investment remit of a16z's Machine Age Fund, the partners emphasize that their focus remains strictly on computer science infrastructure:
- **Fund Remit:** Any computer science layer a model runs on—chips, networking fabrics, interconnects, high-throughput storage, and the foundational electrical systems driving compute.
- **Physical Interaction / Edge:** How AI interfaces with the real world through sight, sound, and speech across new physical compute form factors, whether edge devices, CDNs, or embodied robotic systems (while steering clear of heavily verticalized, regulated niche industries).
- **Facility Engineering Shift:** Prior-generation data centers are rendering themselves useless for upcoming supercomputing architectures (such as NVIDIA Feynman-class systems), where legacy building layouts cannot handle the sheer acoustic resonance, floor weight, or power requirements.
`35:00` The Gigawatt Scale, Grid Gridlock, and the "Machine Age"
The conversation turns to the sheer magnitude of power required by future infrastructure. By 2028, newly planned data centers will need roughly 44 gigawatts of added power, compared to only about 25 gigawatts of projected utility grid additions. Putting this scale into perspective, the group demystifies what a single gigawatt actually means in physical terms: it represents multiple football fields of infrastructure, capable of powering roughly 50,000 homes or fully electrifying and air-conditioning an entire mid-sized city like Flagstaff, Arizona. Despite the industry casually throwing the term around, very few true gigawatt-scale data centers actually exist today.
The reason hyperscalers and utilities cannot simply build faster comes down to hard physical lead times rather than code:
```mermaid
flowchart LR
accTitle: Power and Permitting Lead-Time Bottlenecks
accDescr: The chain of constraints delaying new gigawatt data centers from coming online.
A["Permitting & Regulatory Bidding"] --> B["Grid Connection Access"]
B --> C["Hardware Shortages: Turbines & Transformers"]
C --> D["Offshoring Compute: Mexico & Australia"]
```
Because these lead times are notoriously stubborn, compute-hungry startups are increasingly pushed abroad to countries like Mexico or Australia, creating massive economic and employment windfalls elsewhere while local US jurisdictions block or delay construction. The solution is establishing a symbiotic standard where data centers provide their own baseline generation, feeding electricity back into municipal grids during daytime peak demand and drawing cheap baseline power back at night when municipal consumption plummets.
Finally, the partners unpack the philosophy behind naming the initiative the "Machine Age Fund" rather than an AI fund. Coining the phenomenon "machine intelligence" instead of artificial intelligence is an overdue correction: modern models are massive caches and synthesis engines of recorded human thought rather than independent, clean-slate human cognition. Furthermore, there is profound irony that the venture firm famous for declaring that "software is eating the world" now sees capital running face-first into the absolute limits of physical machines.
> “I mean, there's kind of this deep irony. And this is from, you know, the software's eating the world people that, you know, you've really come to a place where you pour money into something and then you're limited by the actual machines below it.”
`40:00` Incumbent Blind Spots, Silver Bricks, and Fordlandia
Given the colossal capital expenditures already pouring into AI infrastructure, Erik Torenberg raises whether the window has closed for startups to capture meaningful market share. With multi-trillion-dollar incumbents like NVIDIA dominating the space, one might wonder why existing giants would not simply absorb every adjacent layer. Raghu Raghuram answers that sustaining historical compounding rates—whether measured in tokens per second per dollar, tokens per watt, or tokens per rack—demands step-function breakthroughs rather than incremental tweaks. Achieving a 10x improvement across any of these physical metrics requires first-principles thinking, which traditionally originates from hungry startup founders rather than established giants. Martin Casado frames this through fundamental market dynamics: when incumbents achieve multi-trillion-dollar market caps, capturing even a 5% sliver represents an enormous standalone business. While NVIDIA theoretically *could* tackle those peripheral bottlenecks, they have little incentive to divert focus from the core 90% that powers their massive growth engine. This parallels earlier platform shifts, when observers repeatedly asked why Microsoft or Amazon would not simply build every emergent software category themselves.
```mermaid
flowchart TD
accTitle: Expansion and Fragmentation of Technology Markets
accDescr: Diagram illustrating how rapid platform expansion creates niches that lead to vertical fragmentation.
subgraph Core["Incumbent Focus"]
A["Core Platform Scale<br/>Multi-Trillion Dollar Incumbents"] -->|"Focuses on Top 90%"| B["'Gold Bricks'<br/>Core Compute Demand"]
end
subgraph Periphery["Startup Opportunity"]
A -.->|"Ignores Margin Niches"| C["'Silver Bricks'<br/>Niche & Edge Solutions"]
C --> D["New First-Principles Architectures"]
D --> E["Independent Multi-Billion<br/>Dollar Companies"]
end
B -.->|"Market Expands"| F["Market Fragmentation"]
E --> F
```
Ben Horowitz illustrates this incumbent mindset by recounting an anecdote from their partner, Alex Rampell:
> “He had this startup called TrialPay. He was trying to sell it to or sell that it's his services to Meta and then Facebook. And Dan Rose, who was the head of CorpDev at the time, said, ‘Alex, that's great. It sounds like you can collect a lot of silver bricks. But I'm like, I have so many gold bricks, I can't even pick them all up. So the last thing I'm doing is looking at a silver brick.’ And NVIDIA's in that position.”
This dynamic also mirrors historical industrial evolutions. Ben Horowitz recounts Henry Ford's 1913 River Rouge complex, an ultra-verticalized manufacturing marvel where iron ore, coal, and raw materials entered at one end and finished automobiles rolled out the other. In pursuit of total vertical integration, Ford famously established Fordlandia—an entire prefabricated American town planted deep in the Amazon rainforest to control his own rubber tree supply. Complete with suburban bandstands and ice cream parlors, the social engineering project collapsed once managers attempted to impose rigid industrial punch-card work schedules on local laborers, who promptly rejected the regime. In time, the automobile industry unbundled into deep, specialized supplier tiers. “As markets expand, they fragment,” Ben Horowitz notes, creating room for challenger companies across the stack.
Martin Casado points out that while standard software companies enjoyed software gross margins that flowed almost automatically as a natural byproduct of distribution, the AI era is fundamentally different: hardware and infrastructure optimizations directly govern company margins and commercial viability.
- **Tokens per watt and per dollar:** Continued AI expansion requires continuous 10x improvements in compute efficiency that incumbent architectures cannot satisfy alone.
- **Incumbent prioritization:** Frontier incumbents deliberately leave multi-billion-dollar opportunities behind because chasing marginal efficiency distracts from harvesting core enterprise demand.
- **Fordlandia historical parallel:** Henry Ford’s attempt to own everything from Amazon rubber trees to finished cars proved unsustainable; expanding industries inevitably decentralize into distinct supplier ecosystems.
---
`45:00` The Era of the "Systems Founder" and Heavy Iron
When breaking down the sub-sectors where new category-defining infrastructure companies will emerge, Raghu Raghuram outlines a stack far broader than isolated compute silicon. Constructing a modern AI accelerator requires rethinking the entire architecture: specialized memory hierarchies, advanced networking fabrics, power delivery systems, cooling loops, and the extensive software layers required to automate, orchestrate, and manage distributed fleets. However, financing these infrastructure ventures diverges sharply from traditional software venture capital. Early financing rounds routinely require hundreds of millions of dollars before a working product ever ships. While frontier model labs also consume vast pre-product capital, chip and hardware architectures present higher structural execution risk and demand founders with deep, specialized backgrounds. As Ben Horowitz observes of the talent pool, “The guys who know how to make memory, they're not young.”
Raghu Raghuram emphasizes that these demands have given rise to a distinct persona: the systems founder. Success in this domain cannot come solely from an academic researcher or a pure software developer. Instead, leaders must coordinate complex multi-variable equations spanning silicon design, system-level packaging, foundry queue logistics, supply chain availability, and downstream thermal and electrical envelopes:
> “They got to be able to architect and design the chip or the system, whatever it is. Then you got to think about how is this thing going to actually get manufactured, right? Who's going to be supplying this? And a whole bunch of these downstream things, which normally, if you're building software, you don't know how to think about all of these things.”
Raghu Raghuram points to Jensen Huang as the archetypal "Michael Jordan" of this philosophy, someone who designs chips around an exhaustive vision of the broader data center ecosystem.
Martin Casado highlights two shifting environmental tailwinds:
1. **Unprecedented lab demand:** Frontier model labs are so compute-starved that they actively sign binding commercial agreements with early-stage startups long before physical silicon tape-outs or working hardware exist—a complete departure from five years prior, when enterprise hyperscalers dismissed nascent hardware.
2. **Abundant capital for systems:** Institutional follow-on capital has mobilized behind heavy infrastructure, dramatically lowering late-stage financing risk.
Addressing Patrick Collison’s observation that 20-something founders seem far less common today than during the personal computer and consumer internet waves of Bill Gates or Mark Zuckerberg, Ben Horowitz points out that building hardware with physical supply chains requires hard-earned operating reps. Even legendary founders like Elon Musk or Travis Kalanick cut their teeth on pure software businesses before graduating to operationally unforgiving, capital-intensive engineering domains. Building a successful enterprise is brutally difficult even when the founder fully grasps the technology; attempting to master both organizational leadership and frontier physical engineering simultaneously without prior experience is an almost insurmountable hurdle.
`50:00` The Founder Age Curve and the Talent Rebirth in Systems and Hardware
The discussion centers on the demanding learning curve required to build infrastructure and hardware companies compared to consumer software. While a brilliant young founder can launch a purely software-driven AI product right out of school, orchestrating massive, capital-heavy physical systems typically demands decades of operational scar tissue—the kind possessed by leaders like Elon Musk or Travis Kalanick. Over the past twenty years, both the technology industry and academia largely turned their attention away from low-level systems, hardware, and semiconductors. Because it was not viewed as a primary venture growth sector, a generational paucity of talent emerged; students did not graduate with chip-building internships, and few large tech incumbents offered programs that trained junior engineers in deep hardware fundamentals.
That talent landscape is now pivoting rapidly as AI infrastructure demands force a generational revival. The speakers observe that the frontier companies currently reshaping the industrial complex will leave behind a legacy of trained talent far broader than their own corporate footprints, seeding an entirely new cohort of founders capable of designing hardware and low-level computer science architectures. Even when founders launch infrastructure startups in their twenties, the winning formula involves immediately surrounding themselves with seasoned veterans who understand the nuances of tape-outs, supply chains, and large-scale manufacturing. As one speaker emphasizes about the leadership dynamic in these hybrid environments: “It doesn't necessarily have to be the founder with experience, but that founder better be able to tap into that experience in a real way.”
- **The Twenty-Year Defocusing:** Academia and mainstream venture capital spent two decades heavily incentivizing application-level software, creating a talent vacuum in basic chip design and physical infrastructure.
- **The Intergenerational Model:** Successful young technical founders in the hardware space do not attempt to learn legacy physical engineering from scratch; they build multi-generational offices where twenty-something visionaries pair directly with veteran industry operators.
- **The Talent Flywheel:** The new wave of AI infrastructure unicorns is actively training a rising generation of junior engineers who will exit to launch the next era of hardware startups.
---
`52:15` Why No New GPs Were Needed for the Machine Age Fund
Erik Torenberg points out that despite the Machine Age Fund being a major new dedicated vehicle for Andreessen Horowitz, the firm did not hire external General Partners to spearhead it. Instead, the existing partnership chose to lead the effort directly. The partners explain that their own foundational backgrounds—spanning distributed systems, enterprise virtualization, networking, and core infrastructure—have always lived at the heart of their careers. While the market often associated the firm primarily with software, a16z had quietly written early checks into ambitious physical and hardware bets over the years, including SpaceX, Anduril, Astranis, and Waymo.
For the partners, returning to deep systems and hardware investing was not an attempt to invent a new competency, but rather the reactivation of long-held institutional memory and personal passion that had waited for market conditions to catch up. One partner reflects on how their natural inclination toward deep tech historically required deliberate restraint during the software boom: “Well, it's kind of funny. I think we almost had to be warned against it almost just because our backgrounds are from hardware. And I think the reason that we needed a reminder is because all of this has been so much in our careers, existence, and heart, we were kind of drawn to that.”
---
`53:18` The Five-to-Ten-Year Vision: Infrastructure Abundance and American Competitiveness
Asked what success looks like in five to ten years if the Machine Age Fund achieves its goals, the group frames the outcome around radical infrastructure abundance and national competitiveness. The objective is an ecosystem defined by hyper-efficient, eco-friendly data centers alongside a surplus of silicon, memory, and clean power that removes compute bottlenecks entirely.
```mermaid
flowchart LR
accTitle: Capital and Compute Abundance Flywheel
accDescr: Capital infusion into infrastructure unlocks compute abundance, sustaining American technological leadership and democratic values.
subgraph Inputs ["Foundational Buildout"]
A["Machine Age Capital"]
B["Power & Green Data Centers"]
C["Next-Gen Silicon & Memory"]
end
subgraph Output ["Ecosystem Impact"]
D["Compute Abundance"]
E["Global AI Leadership"]
F["Preservation of Free-World Values"]
end
A --> B
A --> C
B --> D
C --> D
D --> E
E --> F
```
The stakes extend far beyond venture returns to the preservation of global technological leadership. The partners emphasize that the United States remains an unparalleled engine where individuals can arrive with nothing and build world-changing enterprises. If the country cedes its lead in the underlying compute stack to rival nations with authoritarian governance models, the global norms surrounding technology will inevitably shift. As one partner concludes: “I think that doesn't continue to go if we lose our lead in technology. I think we'll be in another era and it'll be another country. And maybe they have a different set of values around that.”
- **The Vision of Abundance:** Five to ten years out, success looks like a massive domestic surplus of chips, memory bandwidth, energy, and ultra-efficient data centers operating sustainably.
- **Geopolitical Imperative:** Technological preeminence directly underpins American democratic values; losing the infrastructure race risks allowing foreign powers with divergent values to dictate the rules of the next industrial era.
- **Wrap-up:** Host Erik Torenberg concludes the discussion, thanking Ben Horowitz, Raghu Raghuram, and Martin Casado for detailing the fund's philosophy, before closing out with standard compliance disclosures regarding a16z's portfolio holdings.
## Notable quotes
> `12:35` Any problem that you have can be solved with enough infrastructure, GPUs and power and money. And so until we run out of problems, we're not going to run out of demand. — Raghu Raghuram
> `00:00` We have a whole new technology that's the most important technology ever, and you need a whole new infrastructure. — Ben Horowitz
> `25:00` We've actually gotten to this interesting point in the industry where it actually makes sense to build an ASIC per model just because the amount of capital investment in that model. — Martin Casado
## Mentioned
Machine Age Fund, Andreessen Horowitz, Ben Horowitz, Erik Torenberg, Martin Casado, Raghu Raghuram, Marc Andreessen, Alex Rampell, Dan Rose, Jensen Huang, Henry Ford, Elon Musk, Travis Kalanick, Patrick Collison, Thomas Edison, Fred Brooks, The Mythical Man-Month, Cisco, Juniper Networks, Arista Networks, NVIDIA, Meta, Microsoft, Amazon, SpaceX, Anduril, Astranis, Waymo, Grok, Kimi, GroqBot, OpenClaw, ChatGPT, Fordlandia, Hot Chips
## Video description
Ben Horowitz, Martin Casado, Raghu Raghuram, and Erik Torenberg discuss the launch of a16z's new Machine Age Fund and the infrastructure buildout behind AI, from chips, memory, and networking to power, cooling, and data centers.
Why a dedicated fund now? The group argues that the bottleneck in AI is increasingly shifting from the models themselves to everything beneath them. Hyperscaler CapEx is surging, critical components are booked years in advance, and each new generation of reasoning and agents requires dramatically more compute. They unpack why this cycle looks different from previous infrastructure booms and how AI is turning problems once constrained by engineering into problems that can increasingly be attacked with capital and compute.
They also explore where the next generation of infrastructure companies could emerge, why founders are returning to hard technical problems across hardware and systems, and what it will take to rebuild the computing stack for the Machine Age.
Andreessen Horowitz partners argue that artificial intelligence progress is no longer bottlenecked by algorithmic model design, but by the physical computing and power systems underneath it. Through the firm's Machine Age Fund, they contend that sustaining the artificial intelligence expansion requires rebuilding the physical hardware, electrical distribution, and data center stack from first principles.
Physical Limits and Data Center Constraints (0:00)
Ben Horowitz, Raghu Raghuram, and Martin Casado state that AI development has shifted 'south of the model,' where constraints now center on silicon, high-bandwidth memory, networking fabrics, and raw commodities such as copper. Data centers face immediate physical hurdles: average server rack power requirements have surged from legacy levels of 5–10 kilowatts up to 100–250 kilowatts—an estimated 70-fold leap in compute density. Horowitz notes that this concentration renders alternating current (AC) power unviable, compelling data centers to shift to high-voltage direct current (DC) architectures and mandatory liquid cooling.
Capital Scaling and Hardware Scarcity (5:00)
Unlike traditional software startups focused on customer acquisition, AI startups face virtually guaranteed demand, making operational gross margins and unit economics their existential hurdle. Raghuram notes that hyperscaler capital expenditures are projected to climb from approximately $700 billion to nearly $1 trillion annually, even though only 5% to 10% of the enterprise market has deployed AI compute. Critical infrastructure components are booked through 2027 and 2028, with one leading memory vendor stating that current customer demand alone would consume three full years of total manufacturing capacity.
The partners contrast this environment with the 1990s dot-com bubble, observing that modern compute is fully pre-sold and operating at capacity rather than sitting idle as speculative dark fiber. Furthermore, Casado points out that Fred Brooks's traditional software law—where adding personnel delays late projects—no longer holds: capital poured directly into hardware clusters and megawatts systematically yields expanded raw intelligence, especially as reasoning models and autonomous workflows multiply token consumption per query.
Incumbent Blind Spots and Systems Founders (40:00)
Addressing whether dominant companies like NVIDIA will capture every layer, the partners argue that multi-trillion-dollar incumbents prioritize their core platforms—picking up 'gold bricks'—while leaving multi-billion-dollar efficiency optimizations ('silver bricks') to startups. Horowitz draws a parallel to Henry Ford’s failed, vertically integrated Amazon rubber venture, Fordlandia, illustrating that rapidly expanding industrial markets inevitably unbundle into specialized supplier ecosystems.
Capturing these opportunities demands what Raghuram terms 'systems founders,' who can coordinate silicon design, packaging logistics, foundry queues, and downstream thermal envelopes. Because two decades of venture and academic focus on application software hollowed out junior hardware talent, successful infrastructure startups rely on an intergenerational model, pairing younger visionaries directly with experienced operators who possess hard-won manufacturing and supply chain experience.
Autonomous Agents and National Competitiveness (20:00)
Casado outlines how AI interfaces are advancing from conversational chat prompts to autonomous 'digital employees' that are provisioned their own dedicated machines, browsers, and permissions to execute tasks independently. While these agents multiply token demand and workplace leverage, Horowitz warns that integrating non-human workers introduces operational friction, including token waste, erratic outputs, context loss, and security risks that require human augmentation rather than unchecked automation.
Looking out five to ten years, the partners define success as achieving broad domestic compute abundance across sustainable data centers, clean power, and advanced silicon. They frame this infrastructure buildout as an urgent national priority, warning that if the United States cedes its compute and infrastructure lead to authoritarian rivals, global norms and values governing future technological eras will be determined elsewhere.
While the partners present capital investment in physical infrastructure as the direct path to compute abundance, the thesis ultimately depends on surmounting stubborn real-world bottlenecks—including multi-year utility interconnect queues, high-voltage labor shortages, and severe supply backlogs—before physical constraints curtail the pace of intelligence scaling.
Detailed notes — zh
## Background
本期内容由 Ben Horowitz、Martin Casado、Raghu Raghuram 与 Erik Torenberg 于 2026 年 8 月 28 日在播客及网络平台发布。官方简介指出，几位嘉宾共同探讨了 a16z 旗下全新设立的“机器时代基金”（Machine Age Fund），并深度拆解了从芯片、内存、网络到电力、冷却和数据中心等支撑 AI 浪潮的整套底层基础设施建设。
## TL;DR
对话深入探讨了 AI 产业瓶颈正全面从上层模型转向“模型以南”的物理世界基础设施，包括万亿美元 CapEx 带来的芯片短缺、电网扩容断层与散热极限。与以往互联网周期的投机性泡沫不同，现代 AI 展现出近乎无限的算力需求，并正从受限于组织协同的纯软件工程演变为由海量资本与算力直接驱动的“资源转化”范式。随着从简单对话演进至长程推理与自主智能体，现有计算架构、机架供电与数据中心规范正面临彻底重构。面对巨大的物理挑战与地缘政治竞争，系统型创始人与硬科技老兵正在重回舞台中央，通过在芯片架构、高带宽内存及绿色能源等领域的底层突破，力求实现计算要素的全面丰饶以捍卫技术主导地位。
## Atmosphere & style
整场对话兼具硅谷顶级投资人的宏大视野与底层工程老兵的务实硬核，语调在历史纵深感与紧迫感之间无缝切换。几位合伙人不仅旁征博引汽车工业史、电网技术演进与早年互联网泡沫，更频繁穿插如旧内存条在二级市场暴涨抵消云迁移成本、亚马逊丛林橡胶镇等极富戏剧性且幽默的现实轶事。现场氛围坦诚而极富感染力，听众能深刻感受到嘉宾们对于“软件吞噬世界”最终撞上“物理机器天花板”的深层反思，以及将算力基建上升至国家科技领导力维度的宏大雄心。
## Bulletin
- `00:00` a16z 正式推出“机器时代基金”（Machine Age Fund），专注于计算芯片、网络互连、存储架构及电力配套等“模型以南”的计算机科学物理基础设施。
- `05:00` 超大规模云服务商（Hyperscalers）当前年度资本支出达 7,000 亿美元且明年或破万亿美元，关键 AI 算力与基础设施产能已被锁定至 2028 年。
- `00:00` 顶级创业团队中主攻复杂硬件与重系统架构的项目占比已从过去的 3%~5% 激增至约 30%，AI 研发范式正从传统“工程约束”转向“资本与算力堆叠”。
- `28:10` AI 数据中心单机架功率已飙升至 100~250 千瓦并转向 800V 直流电，导致全美仅 2% 的电工掌握相关资质，物理基建、冷却与劳动力断层正成为核心发展瓶颈。
## Key takeaways
- `00:00` AI 瓶颈已从模型本身全面转移至“模型以南”（South of the model）的物理硬件与基础设施层，涵盖芯片架构、互联介质、散热材料、供电系统乃至原材料采掘。
- `04:12` 与以往科技周期忧虑需求与获客不同，当前市场对 AI 算力的需求接近无限，企业真正的焦虑转向了被硬件物理极限锁死的利润率（Margin）与效率（Efficiency）。
- `07:18` 当下的算力短缺并非互联网泡沫时期的投机性“暗光纤”，所有产出的 GPU 均被客户全额预购，行业甚至出现旧服务器拆卸内存套现足以覆盖上云成本的极端现象。
- `10:00` 底层基础设施传统年增速仅 20%~30%，而 AI 需求基准增速呈三位数、Token 需求年增率接近 1000%，传统长达数年的半导体与电厂建设周期拉大了产业断层。
- `13:30` 软件研发不再受限于布鲁克斯法则（《人月神话》）的人力协调内耗，AI 扩展律转向高密度推理与强化学习，资本、计算资源与电力可直接转化为智能产出。
- `20:00` 企业与 AI 的协同形态正在由搜索框、个人助手（如 OpenClaw）演进为具备独立身份与办公环境的“数字员工”（如 GrokBot），组织核心在于如何在机器不失控的前提下让人类化身“超人”。
- `25:00` 前沿模型单次训练成本已达 30 亿至 50 亿美元，单体数字人工物的极高资产价值正倒逼硬件走向专用 ASIC，以微小推理效率提升换取数十亿美元的成本节省。
- `28:10` 数据中心机架功率跃升至 100~250 千瓦，迫使供电全盘向危险的高压直流电（DC）切换，同时催生液冷标配、高承重结构以及钢筋混凝土防噪加固等严苛物理改建需求。
- `34:25` 到 2028 年全美新建数据中心电力需求增量约 44 吉瓦，而电网预期增量仅 25 吉瓦，严重的审批和电力缺口正迫使新一代 AI 初创公司将算力部署流向海外。
- `37:11` “人工智能”（AI）本质应称为“机器智能”（Machine Intelligence），它并非人类心智机理的复刻，而是人类全部智慧经验的“超级缓存”，技术的质跃最终仍受物理机器制约。
- `40:00` 遵循“市场扩张必然导致碎片化分工”的历史规律，NVIDIA 等巨头无法吞下全部场景，即使占据其未顾及的 5% 边缘垂直市场，也足以孕育百亿美元级的独角兽。
- `43:30` 软件超高毛利率神话正在终结，大模型端到端的高能耗让硬件层与互联架构的微小优化直接成为企业盈利生死存亡的胜负手。
- `45:00` “机器时代”对创业者提出全新要求，单纯写算法或搞学术的年轻极客难以应对复杂的全球实体供应链与流片周期，行业更需要深谙工程落地的“系统型创始人”或新老搭配的复合团队。
- `52:15` 实现底层计算芯片、高带宽存储与绿色清洁能源的全面充沛，是守住前沿技术领导力、捍卫自由开放创新土壤的战略根基。
## Detailed walkthrough
### `00:00` Machine Age Fund 的诞生与 AI 底层物理基础设施的重构
播客由 Andreessen Horowitz（a16z）联合创始人 Ben Horowitz、通用合伙人 Martin Casado 以及前 VMware CEO、现 a16z 投资合伙人 Raghu Raghuram 共同开篇。主持人 Erik Torenberg 宣布 a16z 正式推出全新的“机器时代基金”（Machine Age Fund），专门用于投资驱动下一代 AI 浪潮的底层基础设施。在引入该基金时，Erik 引用了 Marc Andreessen 的一段极具历史分量的话：“这是我人生中所见过的最大技术革命，这显然比互联网还要庞大；能够与其相提并论的只有微处理器、蒸汽机、电力，或者轮子的发明。”（“This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. It's The comps on this are the microprocessor, the steam engine, and electricity, or maybe the wheel.”）
Ben Horowitz 指出，每当人类迎来一种全新的、颠覆性地使用软硬件基础设施的范式时，都需要一套全新的底层架构，而这一次的基础设施重构具有前所未有的超高冲击力。当前的困局不仅仅是缺乏新芯片或新的系统软件，而是涵盖了整个物理链条——从能源供给方式的革新，到寻找能够替代铜的传导材料，无所不包。Raghu Raghuram 进一步打趣并强调，过去计算机从业者谈论基础设施时，指的无非是服务器、存储和网络，但眼下的 AI 基础设施瓶颈“一路延伸到了最底层的铜矿”。过去三年里，模型能力以惊人的速度迭代，如今模型本身不再是瓶颈，真正的瓶颈已全面转移到了他所称的“模型以南”（south of the model）——即承载模型运转的物理世界底层。
Martin Casado 从投资风向和创业者动向切入，指出风险投资机构本质上是在跟随创业者的步伐。过去几年，攻坚复杂硬件与系统架构顶层难题的顶级团队比例出现了极其剧烈的跳跃。Martin 粗略估算，过去顶尖创业者投来的项目中硬件类大概只占 3% 到 5%（现场甚至认为 5% 都算高估了），而如今这个比例已经骤增到了 30% 左右。谈及这一趋势，Martin 评价道：“因此，比风投群体聪明得多的创业者群体，已经把这里看作最具创新活力的阵地，并纷纷做出了响应。”（“So like the founder community, which tends to be much smarter than the VC community, has identified this as a very active area for innovation and they're responding.”）
- **顶级创业项目转型比例**：顶尖创始人提交的硬件/重系统项目占比从过去的约 3%~5% 暴涨至目前的近 30%。
- **瓶颈转移论**：“模型以南（South of the model）”，涵盖芯片架构、互连介质、散热材料、供电系统乃至原材料采掘。
---
### `04:12` 无限需求撞上物理极限：算力、能耗与商业模式的重塑
讨论深入探讨了驱动这一风向转变的宏观现实。Ben Horowitz 剖析道，当下市场对 AI 智能的需求在理论上是“无限”的，从而导致整条供应链的每一个环节都处于极度紧绷和重压之下。这种无限需求给商业逻辑带来了一个极反直觉的变化：在传统的科技创投周期中，投资人最焦虑的是企业能否实现增长、能否说服客户购买产品（PMF 与商业化扩展）；然而在 AI 周期中，没有人担心需求和增长，真正让人寝食难安的是“利润率（Margin）”与“效率（Efficiency）”。
AI 时代的单位经济效益，在很大程度上被硬件的物理极限死死锁死。由于现有的数据中心、服务器架构、网络排布和供电方式根本不是为了承载现代 AI 这种极高吞吐、密集矩阵运算的工作负载而设计的，导致现存系统在计算和推理任务中损耗了巨大的能耗与资金成本。这就演化成了一个全球性的共识：如果要实现可盈利的规模化增长并释放 AI 企业的商业价值，就必须彻底从物理层面对系统组件进行革命性的改写。
```mermaid
flowchart TD
accTitle: AI 商业逻辑与基础设施物理约束链路图
accDescr: 展示 AI 无限需求如何通过硬件物理限制传导至企业利润率并倒逼底层重构
subgraph DemandSide ["需求端"]
D1["无限 AI 需求与爆发式增长"]
D2["企业无须担忧获客增长"]
end
subgraph Bottleneck ["物理层瓶颈"]
B1["传统数据中心与硬件并非为 AI 设计"]
B2["功耗、发热、内存墙与互连物理瓶颈"]
end
subgraph Outcome ["商业与架构影响"]
O1["企业利润率承压（效率低下）"]
O2["倒逼芯片/组件/供电全面底层重构"]
end
D1 --> D2
D2 --> B1
B1 --> B2
B2 --> O1
O1 --> O2
```
---
### `05:00` 真实需求还是泡沫？超大规模云厂商万亿 CapEx 与排到 2028 年的产能
面对“当下的供给短缺究竟是真实的结构性失衡，还是又一次短暂的技术狂热周期”这一尖锐问题，Raghu Raghuram 列举了一系列强有力的硬性指标反驳泡沫论。首先是拥有全行业最高视野的超大规模云服务商（Hyperscalers）。这些巨头汇聚了来自前沿模型实验室（Frontier Labs）、AI 数据公司、跨国大企业以及全球各区域的真实算力诉求。数据显示，各大 Hyperscalers 今年的资本支出（CapEx）总额已达约 7,000 亿美元，而明年据推测可能直接突破 1 万亿美元大关。若不是看到了前所未有的全方位刚性需求，这些巨头绝不可能以这种规模激增资本支出。
不仅买方在疯狂下注，供给侧的紧张程度也令人咋舌。核心硬件组件的供货期普遍已经排到了 2027 年甚至 2028 年，市场上甚至出现了数千张 GPU 需要通过多日竞价拍卖才能抢到的荒诞景象。与此同时，AI 所处理的工作单元（unit of work）价值正在提升，但其背后的 token 消耗量呈指数级暴增——从原本简单聊天窗口（Chat）的几百个 token，直接跨越到复杂自主智能体（Agent）动辄数万乃至数十万个 token 的循环消耗。不仅单次任务的吞吐消耗扩大了数个数量级，受益人群也从程序员迅速扩散至全体知识工作者。
- **超大规模厂商资本支出（Hyperscaler CapEx）**：当前年度总计约 7,000 亿美元，预测明年合计可能触碰 1 万亿美元。
- **供应链排期**：关键 AI 算力与基础设施核心组件的产能基本已被提前预订至 2028 年。
- **计算消耗跃升**：交互形态由简单的百 token 级别聊天机器人演变为消耗成千上万 token 的多步骤推理 Agent。
---
### `07:18` 算力荒与互联网泡沫的本质差异：被透支的物理资源与荒诞的旧内存套现故事
当主持人追问历史上是否曾出现过整个行业被预售一空数年之久的类似情景时，嘉宾们将其与 1998-1999 年互联网泡沫时期的光纤大跃进进行了深层对比。Ben Horowitz 回忆道，当年电信繁荣期虽然也存在带宽不足的表象，但彼时在地下大规模铺设的光缆绝大多数是投机性的，最终沦为了大量闲置的“暗光纤（Dark Fiber）”。当时互联网根本没有足够的用户基数，高带宽消耗的杀手级应用（如在线流媒体视频）受限于终端条件压根无法落地，那是“需求端缺乏真实消费者”的单向狂欢。
与之截然相反，如今制造出来的每一块 GPU 在出厂前就已经被客户真金白银买走，市场上甚至有人以原价四倍的溢价转售现货 GPU。除了算力，整个产业更是彻底耗尽了电力与散热指标，同时还在面临极其严峻的地缘政治与合规阻力。Ben 感慨道：“在我整个职业生涯中，我们从未遇到过任何类似的事情，这确实是史无前例的。”（“So it's really unprecedented in my career that we've had anything like this.”）
为了展现当前硬件短缺有多么离奇，Martin Casado 分享了一个真实的荒谬故事：他曾与一家大型上市公司的 CFO 交谈，该公司在历史上极其保守、极度抗拒将业务迁移到公有云，因此其自建数据中心里堆满了大量旧服务器。近期该机构在做固定资产与库存盘点时震惊地发现，由于全行业硬件严重短缺，他们机房旧服务器里拆下来的老内存条在二级市场上的升值总额，“居然足以资助他们完成整套向云端迁移的全部开销”。Raghu Raghuram 对此补充道，行业正全面告急，不论是电力、冷却、GPU 还是内存均已见底。就在此时于斯坦福大学举办的 Hot Chips 行业顶级半导体峰会上，全球领先的内存厂商更是直言不讳：“目前积压在他们手头的市场需求，需要足足耗费他们三年的全部产能才能彻底供应完成。”（“The leading memory runner said that, the demand they have, Today, it will take them three years of capacity.”）
### `10:00` 算力供需失衡与底层硬件的物理极限
讨论一开始聚焦于当前 AI 算力供求之间的严重错配。与会者指出，当前面临的芯片与基础设施短缺甚至不是为了满足“未来需求”，仅仅是应对当下的用量就已经捉襟见肘。之所以陷入这种境地，并不是大家低估了模型的效果，而是因为整场 AI 变革来得太快，目前仅仅进入了第四个年头。半导体芯片的研发制造周期通常长达 3 到 4 年，而建设数据中心从破土动工到完工交付也需要 4 到 5 年。更严峻的是电力限制：数据中心要么必须自建电厂，要么必须锁定稳定的电力来源，更多时候甚至需要两者兼备。一个以往年增速 20% 到 30% 就算表现极佳的传统基础设施行业，突然被直接挂接到年增速达三位数的 AI 软件产业上，两者之间的断层自然越拉越大。
针对为何 a16z 到现在才推出专门的 Machine Age 基金，大家回顾了计算机架构演进的历史。在主机向客户机/服务器（Client-Server）转型的时代，以及互联网爆发催生 Cisco 和 Juniper 的时期，每个技术代际都会崛起少数独立的基础设施巨头；即便在超大规模数据中心阶段，云厂商主导垂直整合，市场也跑出了 Arista 这样的网络新贵。然而，以往的变革通常较为局部，可能仅涉及单一一类芯片或交换机。如今则完全不同，算力需求正以极其陡峭的“垂直曲线”爆发，Token 消耗量预计每年将保持接近 1000% 的增速，任何现有供应链都无法跟上这种膨胀。更根本的问题在于，现有硬件系统的架构全部是为另一个计算时代设计的，从铜导线互连等底层物理特性来看，传统技术路径均已逼近物理极限，必须依靠全新的底层架构突破来实现升级。
- **芯片与建设周期：** 芯片设计迭代需 3~4 年，大型数据中心建设需 4~5 年。
- **产业增速错配：** 底层基础设施传统年增 20%~30%，而 AI 软件需求基准增速为三位数，Token 需求年增率预计接近 1000%。
- **硬件物理瓶颈：** 从供电、散热到以太网铜缆互连，现有计算堆栈正全面撞上物理设计的上限。
### `13:30` 从“工程约束”到“资源堆叠”：Token 爆炸的内在逻辑
```mermaid
flowchart TD
accTitle: AI Scaling Paradigm Shift
accDescr: 比较传统软件工程模式与现代AI扩展模式的差异
subgraph Traditional["传统软件工程模式"]
A["资本投入"] --> B["招聘大量软件工程师"]
B --> C["遭遇《人月神话》瓶颈<br/>(协调内耗、进度迟缓)"]
C --> D["产品产出周期漫长且易失败"]
end
subgraph AIScaling["现代 AI 算力堆叠模式"]
E["资本投入"] --> F["采购 GPU 与锁定电力"]
F --> G["大规模强化学习与推理 (Token 消耗)"]
G --> H["直接生成智能与下游应用"]
end
```
谈及需求端为何随着聊天机器人（Chatbots）、推理（Reasoning）、单智能体（Agents）到多智能体系统（Multi-agents）的演进，单项任务消耗的 Token 呈现数量级倍增且毫无放缓迹象时，讨论揭示了一个深刻的本质转变：当前 AI 扩展律（Scaling Laws）主要由海量推理驱动。强化学习（RL）、思维链（Chain of Thought）以及长时间运行的自主智能体，本质上全都在进行高密度的持续推理；更甚者，AI 还展现出“自催化效应”（Auto-catalytic effects），例如利用 AI 编写底层的 GPU 算子来加速新的 AI 研发，即“没有任何人比 AI 更喜欢消耗 AI”。
这标志着计算范式的根本更迭。过去构建软件是个纯粹的“工程问题”，受制于布鲁克斯法则（《人月神话》），企业无法靠堆砌工程师来跨越研发周期，工程组织本身是天然的节流阀。然而在 AI 时代，系统更像是一个纯粹的资源转化器：只要持续投入资金、芯片和电力，就能直接产出智能。Martin Casado 深刻总结道：“过去做东西是个工程问题，你把一堆工程师扔进去，但那无法规模化扩张……而在这里，感觉它真切地是一个资源限制问题。” Raghu Raghuram 紧接着强调：“只要你有足够的基础设施、GPU、电力和资金，你遇到的任何问题都可以被解决。因此，在我们耗尽问题之前，我们绝不会耗尽需求。”
### `16:30` 资本直接转化为壁垒与广阔的知识工作者空间
针对过去十年创投圈广泛存在的“初创公司融资过剩、估值过高”的质疑，Ben Horowitz 指出规则已经被彻底改写。在过去的软件世界里，面对拥有两年先发优势的竞争对手，盲目扩招 1000 名工程师只会把公司拖垮——“九个女人没办法一个月生出一个孩子”。然而在今天的算力堆叠时代，这套逻辑失效了。企业不需要招聘 10 万名工程师，只需筹集 30 亿美元，点亮一个超大规模计算集群，像 Grok 甚至 Kimi 这样的模型就能平地而起，直接抹平甚至反超以往的壁垒。行业不得不经历一场巨大的心理调试：资本、计算资源与最终成果之间不再隔着不可控的工程延宕，砸钱变得切实有效。
从终端需求来看，算力不仅被 ChatGPT 的数亿活跃用户或全球约 3000 万开发者所驱动，更在向庞大的下游快速渗透：
1. **开发者工具：** 编程已成为 AI 深度渗透的先锋阵地，极高密度的代码生成迅速拉高了全天候算力负荷。
2. **知识工作者：** 全球有超过 10 亿名知识工作者，办公自动化、报表生成及内容分析正处于爆发前夕。
3. **后台智能体与自主系统：** 智能体深入企业后台接管复杂业务流程，使得每一步决策都需要倍增的推理支持。
4. **计算机控制与具身智能：** 类似于 GrokBot 或支持计算机使用（Computer Use）的智能体，能够像人类一样直接坐在电脑前操控桌面界面，正在释放又一波不可估量的交互计算需求。而在未来，具身智能（Embodied AI）和机器人进入物理世界，更将彻底打开算力消耗的无底洞。
Ben Horowitz 用最直白的话点明了这种由资本硬推所带来的行业剧变：“你可以朝问题砸钱，而且你可以朝几乎任何问题砸钱，并且这确实管用。这与我们经历过的任何事情都完全不同。”
### `20:00` 计算机使用（Computer Use）的觉醒与长达数十年的算力浪潮
讨论从 AI 的具体使用场景切入：比如让 AI 代替自己更新一系列拖延已久的服务绑定的信用卡、批量取消流媒体订阅等。这已经脱离了单纯的代码编写，而是货真价实的“计算机使用”（Computer Use）。随着这些场景的爆发，大量知识工作者仿佛被直接植入到了计算机系统内部工作。Marc Andreessen 曾将这波技术变革与蒸汽机或电力的发明相提并论，与会者对此表示高度认同。人类获得了一种全新的生产力要素，可以直接将其转化为实际产出，但目前我们仅仅处在语言、代码生成以及初步计算机使用的起步阶段，未来面向科学探索、新材料研发、生物医药以及创意生成等领域的探索将持续三十至四十年。
传统软件工程的关键瓶颈已经被突破，尽管未来瓶颈会转移到其他复杂领域，但把算力投入到各类现实问题中的长期趋势才刚刚开始。正如一位发言人所断言：“So let's expect this compute need to persist for decades.”（因此，我们要做好这种算力需求将持续数十年的准备。）
```mermaid
flowchart LR
accTitle: AI 交互演进路径
accDescr: 从搜索引擎、浏览器扩展到具备独立系统身份的虚拟员工演进过程。
subgraph Phase1 ["早期交互阶段"]
A["搜索栏 / 传统 Chat 模式"]
end
subgraph Phase2 ["扩展阶段"]
B["类 OpenClaw / 个人外挂"] -->|"共享密钥与密码"| C["人类身份延伸"]
end
subgraph Phase3 ["组织化阶段"]
D["类 GrokBot / 独立虚拟员工"] -->|"独立浏览器与系统"| E["执行高阶自主任务"]
end
Phase1 --> Phase2
Phase2 --> Phase3
```
随后，Martin Casado 深入拆解了 AI 走进工作与生活的三次认知迭代：
- **第一阶段（搜索与对话框）**：行业早期把 AI 当作功能塞进现有产品，本质上只是一个搜索栏，或者做成你问我答的 Chat 模式。
- **第二阶段（个人的数字外挂）**：以 OpenClaw 为代表，人们尝试让 AI 成为独立存在的实体，但它更像人类的外挂副手，共享你的密码、账号与密钥，替你执行你的个人日常事务。
- **第三阶段（具备独立身份的“员工”）**：以 GrokBot 为代表，AI 不再共享用户的个人密钥，而是拥有自己独立的电脑、独立的浏览器，作为一个真正的“数字员工”存在。因为底层跑着世界顶尖的大模型，凡是普通人类员工能做的事情，它几乎都能胜任。
Martin Casado 补充指出，当需要办某件事时，下意识的第一反应变成了“GrokBot 能不能帮我做”，答案往往是肯定的。比如不仅是管理日程、预订会议等基础任务，还可以让它通读整个邮箱做邮件分类（Triage），它甚至懂得在没有明确指令的情况下，自主判断并在执行分拣前先与人类确认关键事项。
---
### `23:14` 数字员工带来的组织文化变革与协同挑战
当新型“数字员工”大规模进入组织，企业的管理机制与文化面临全新考验。Ben Horowitz 从组织构建的角度指出，企业花费了数十年时间才摸索出如何与真实的人类员工协同工作，现在突然迎来了一大批全新的非人类员工。这种全新物种存在不可忽视的学习曲线和负面风险：它们可能会疯狂消耗 Token 带来巨额开销却一事无成；它们会遗忘关键信息；它们会产生幻觉胡编乱造；它们会有不良行为表现，甚至造成严重的数据安全漏洞。
但与此同时，它们的生产力上限又极高。因此，核心课题不是盲目搞全盘替代，而是如何让它们与人类团队顺畅协作。Ben Horowitz 极其直白地阐述了自己的组织哲学：“We're much more going like, okay, how do we make all our humans superhuman without wrecking the place because the bots get out of control?”（我们现在的思考方式更多是：好，我们如何在不让机器人生乱失控的前提下，让我们所有的人类员工变成超人？）
整个机构曾尝试过多种将智能体（Agents）接入业务系统的方式，最终证实最有效、最持久的路径恰恰是 Martin Casado 的洞察：把智能体真正视作有独立工作环境的“人”来对待并直接指派任务。
---
### `25:00` 算力底层重构：打破非 AI 时代的基础设施枷锁
话题随后转向供给侧瓶颈。当前的数据中心、芯片微架构、系统级软件以及配套机房基础设施，最初无一是为了 AI 工作负载量身定制的。要为 AI 彻底重构计算架构，必须从第一性原理拆解推理引擎的核心需求：推理过程消耗海量内存，并在计算过程中不断生成新的 Token。
目前全行业优秀的创始人正在沿着硬件与系统栈层层拆解瓶颈：
- **计算与矩阵乘法**：究竟需要什么样的算力单元，如何针对矩阵乘法进行极致的硬件优化。
- **分层内存架构**：推理逐步生成 Token 需要不断调度内存，内存与计算核心如何高效互连，多级内存应该如何分层排列。
- **能耗与互连传输**：单颗芯片内部、跨芯片互连乃至跨数据中心互连需要多大功耗，每传输一次数据要付出多少电力代价。
与此同时，模型的经济账正在倒逼硬件架构走向专属定制（ASIC）。训练一个前沿模型（Frontier Model）如今需要消耗 30 亿到 50 亿美元，为了让投资具备经济合理性，推理端创造的价值通常需要达到该成本的数倍（例如 100 亿美元级别）。如果开发专用 ASIC 能带来 20% 的推理效率提升，对应省下的成本就是 20 亿美元，而这笔资金完全足以从零流片定制一款专属 ASIC。
传统软件具有大量动态状态，而大模型的权重是完全固定的（Fixed Weights）。Martin Casado 强调了这一历史性拐点：“Like, I don't think in the history of the industry, we've ever created a digital artifact with something like $5 billion that went directly into that artifact.”（我认为在整个工业史上，我们从未创造过任何一件直接砸进约 50 亿美元才打造出来的数字人工物。）这种数额空前的单体资本投入，正在对硬件架构提出前所未有的定制化苛求。
---
### `28:10` 供电与冷却极限：从交流电到直流电、从风冷到液冷
计算密度的狂飙突进正在剧烈重构数据中心物理层面的工程逻辑：
- **机架功率密度暴增**：单机架功率需求从过去的 5 到 10 千瓦（kW）一路飙升至 100 到 250 千瓦。
- **计算密度提升**：计算密度呈现约 70 倍的爆炸式增长。
- **散热方式跃迁**：传统风冷彻底失效，液冷已经成为所有前沿数据中心的标配门槛。
随着机架功率迈入 100 至 250 千瓦区间，传统的交流电（AC）供电体系已经无法支持，必须全盘切换为直流电（DC）。然而直流供电自身不仅需要专属冷却，危险系数也极其惊人。Ben Horowitz 调侃起科技史上的反讽一幕：当年托马斯·爱迪生为了推销自己的直流电、抹黑交流电，公开通过电击处决大象等动物来展示交流电的危险，结果历史兜兜转转，如今超高功率场景下最危险、也最强大的反而成了直流电。
在冷却技术方面，液冷已成为新一代数据中心的既定事实，但仅靠液冷本身还远远不够。在当下的能源监管与环境政策约束下，行业对冷却液的环保特性、毒性控制及新型热管理系统提出了更高的物理与化学工程要求，由此催生了整个数据中心物理设施层的全新投资窗口。
### `30:00` AI 数据中心的物理极限：承重、噪音、高压电与劳动力断层
构建下一代 AI 数据中心所面临的挑战，早已超越了单纯的“算力扩容”，而是演变成一场与物理规律和现实基础设施的硬碰硬。当前的许多数据中心在资源消耗上表现恶劣，虽然浪费水资源的情况可能没有网上宣传的开心果或杏仁种植业那么夸张，但部分数据中心充当着电网的“寄生虫”，只索取而不反馈。随着机架功率密度不断飙升，传统数据中心的建筑构造已经彻底过时——地面结构必须重新设计以承受沉重的超密机架；高功率设备运转产生的剧烈噪音迫使建筑必须采用更厚的墙体隔音，否则根本无法通过社区和州政府的环保审查。由于这种对硬件物理防护的极端需求，市场上甚至出现了一个耐人寻味的现象：除了内存芯片在疯涨，钢筋混凝土的价格也成为涨幅最快的领域之一。
```mermaid
flowchart TD
accTitle: AI 数据中心物理基础设施瓶颈
accDescr: 展示 AI 超密机架对物理建筑、电网和技术劳动力带来的连锁限制与升级压力。
subgraph Hardware["硬件与电力负荷"]
DenseRacks["超高密度机架 (如 Feynman 架构)"]
HighVolt["800V 直流高压供电"]
end
subgraph Physical["建筑与环境要求"]
Floors["高承重地面设计"]
ThickWalls["超厚防噪墙体 (混凝土需求暴增)"]
end
subgraph Labor["劳动力与安全瓶颈"]
Dangerous["极端运维危险性"]
Shortage["高压直流电工极度匮乏 (全美仅 2% 资质认证)"]
end
DenseRacks --> Floors
DenseRacks --> ThickWalls
DenseRacks --> HighVolt
HighVolt --> Dangerous
HighVolt --> Shortage
```
更致命的瓶颈发生在电力工程与技能储备上。当机架电压推升至 800V 时，现场危险程度急剧上升，而全美具备高压直流电（DC）作业资质的电工仅有 2%，形成了严重的技能断层。尽管 Meta 等巨头已经开设培训项目，扮演起为行业输送高压电工的“免费技校”，但缺口依然巨大。与此同时，云厂商为了规避高风险人工并提升效率，正狂热地测试在数据中心内使用机器人来组装和安装服务器。讨论者们打趣道，人人都担心“AI 会夺走所有工作”，但眼下它却实打实地拉动了对电工等蓝领技能的庞大需求。
- 随着英伟达下一代 Feynman 等先进架构到来，现有数据中心绝大部分都将因无法承载其物理与电力标准而直接被淘汰。
- 800V 高压直流电进入机架，全美仅有 2% 的电气工程师或电工掌握相应操作认证。
- 钢筋混凝土由于隔音和防震承重需求激增，价格出现飞速上涨。
- 大型云数据中心正全面加速引入机器人自动化作业，处理服务器搬运与高危机架装配工作。
### `32:51` Machine Age 基金定位：聚焦计算机科学底层架构与实体智能
针对 a16z 正在募集的 Machine Age 基金，合伙人们明确界定了其投资边界与核心使命。该基金纯粹聚焦于“计算机科学基础设施”（Computer Science Infrastructure），即模型运行所依赖的一切技术底座——涵盖芯片、网络、互连技术、存储，乃至紧贴计算架构的电力系统。与关注国防、航天、垂直制造等受强监管领域的 American Dynamism（美国活力基金）不同，Machine Age 基金专注于纯粹的计算机科学平台创新。
AI 的革命性突破在于它让计算机第一次真正具备了与物理世界交互的能力：它不仅能“看”、能“听”、能“说”，还能在现实空间中做决策。这种进化催生了前所未有的新型平台，而大众常用的“边缘设备”（Edge Device）一词并不能准确概括其实质。在他们看来，未来的平台既可能表现为手机、笔记本或 CDN 节点，也完全可以演变成能在物理世界穿梭自如的具身智能设备（Embodied Device）。只要是能够将 AI 算力与推理能力向外推进的底层计算机科学平台，都是该基金的核心涉猎范围。
> “By the way, to be clear on the actual fund that we're raising, our focus is on computer science infrastructure. So anything a model runs on, that's computer science, right? So think, you know, chips, network, interconnect, storage, all the way down probably to the electricity.” —— Martin Casado
### `34:25` 1 吉瓦的震撼：电网扩容滞后与被迫“出海”的数据中心
行业常常张口闭口“100 吉瓦”，但鲜有人真正意识到 1 吉瓦（Gigawatt）意味着什么。Martin Casado 拿自身成长经历进行了极其直观的度量：他长大的亚利桑那州弗拉格斯塔夫市（Flagstaff）拥有 4 万到 6 万居民，全城全年的总耗电量都达不到 1 吉瓦。换言之，仅仅 1 吉瓦的电力就足以维持 5 万户家庭或整座城镇从照明到全天候空调的全部运转。然而，到 2028 年，全美新建数据中心预计将需要新增约 44 吉瓦电力，而同期国家电网的预期增量仅有区区 25 吉瓦，缺口深不见底。
公用事业公司和超大规模云厂商（Hyperscalers）根本无法在短期内填补这一断层。建造一座超级数据中心远非写软件那样靠工程师“周末加班”就能解决，它受困于漫长的实体工期、严苛的许可证审批、受限的天然气与电网接入配额，以及全球范围内变压器和涡轮机的严重短缺。眼下算力需求正以每年 10 倍的速度爆炸式增长，而电力供应链完全无法跟上节奏。其直接后果是，美国的审批和电力限制正在倒逼新一代 AI 初创公司前往墨西哥、澳大利亚等海外国家部署 GPU，相当于把关键算力、高端就业和长期经济红利拱手让人。
> “I grew up in Flagstaff, Arizona, which is a town of 40,000 to 60,000 people depending on the universe. We have less than a gigawatt of power consumption. I mean, this is true. So you can basically light up and air condition your entire town for a gigawatt.” —— Martin Casado
> “This is not a software problem. It's not just like a bunch of engineers, can't you, like, work weekends and that type of stuff. Like, this is, not that that works anyway, but... There are real bottlenecks in this, and these lead times are not that easy to compress.” —— Ben Horowitz
### `37:11` 共生电网模式与“机器时代”的真正内涵
面对地方社区对数据中心进驻的抗拒，Ben Horowitz 指出，正确的破局之道并非全面禁止，而是树立严苛但可持续的标准：数据中心必须为社区创造净效益——不仅做到零噪音、无水资源破坏和创造就业，还要反哺电网。现实中已有成功范例：部分数据中心自建发电系统，白天在城市用电高峰期向电网反向输电，夜间城市负荷骤降时再借用基础电网过剩的基荷电力（Base Load），因为发电厂总是保持峰值产能运转。这种日夜互补机制促使当地居民的用电费率不升反降，形成了真正互利共生的合作模式。
谈及基金为何命名为“机器时代”（Machine Age），团队展开了深刻的反思。他们一致赞同 Ben Horowitz 的观点——过去数十年来大家常挂在嘴边的“人工智能”（AI）其实是一个用词不当的错误概念，它本质上应当被称为“机器智能”（Machine Intelligence）。当前的系统并非拥有与人类相同的心智机理，也无法在毫无先验知识的情况下凭空重构出一门语言；它实质上是人类既有全部思想与智慧的“超级缓存”（Cache），它是在消化吸收人类经验后进行高效率的工具化应用。
此外，作为曾经提出“软件正在吞噬世界”的投资机构，面对当下的产业格局体会到了一种深层的反讽：当资本无限涌入 AI 时，整个产业的真正天花板不再是算法构想，而是被脚下最实打实的物理机器、硅片、钢铁与电缆所死死卡住。“机器时代”这个名字，正是向这种物理硬件在智能浪潮中所扮演的决定性角色致敬——下一阶段的技术质跃，注定将由物理机器的质量与突破来书写。
> “Artificial intelligence was the wrong word. We shouldn't have called it. It's machine intelligence... it is a cache of how humans thought, is a collection of humans thoughts.” —— Martin Casado
> “I mean, there's kind of this deep irony. And this is from, you know, the software's eating the world people that, you know, you've really come to a place where you pour money into something and then you're limited by the actual machines below it. And so I think it is kind of a nod to like the hardware component is so significant in this wave.” —— Martin Casado
### `40:00` 市场扩张与碎片化：为什么 NVIDIA 捡不完所有“金砖”
面对当下 AI 基础设施领域天量资本支出（CapEx）和极高的资本密集度，Erik Torenberg 抛出了一个核心疑问：新入局的初创企业是否已经错过了以实质性体量切入市场的窗口期？像 NVIDIA 这样的行业巨头是否会直接吞食大部分市场蛋糕？对此，Raghu Raghuram 指出，无论是每秒每美元生成的 Token 数（Tokens per second per dollar），还是每瓦特生成的 Token 数（Tokens per watt），要想实现跨数量级的十倍（10x）飞跃，传统改良路线注定受限，必须依赖打破常规的第一性原理创新。
Martin Casado 从自由市场的底层规律进一步剖析：即便当下的芯片巨头拥有数万亿美元的市值，哪怕初创企业只分食其中 5% 的边缘市场，也足以诞生规模庞大的非上市独角兽。巨头不会做、也不该做所有事情。Ben Horowitz 随即分享了 a16z 合伙人 Alex Rampell 过去创办 TrialPay 时的生动趣事：当时 Alex 试图向 Meta（当时的 Facebook）推销其支付服务，时任企业发展（CorpDev）负责人的 Dan Rose 这样婉拒道：“Alex，这听起来很棒，你确实能捡到很多银砖；但我眼前有太多的金砖，多到我都搬不过来，所以我最不可能做的事就是低头去捡银砖。”对于 NVIDIA 而言，也是同样的处境，他们必须全力聚焦在驱动核心增长的那 90% 的主要业务上。
```mermaid
flowchart TD
accTitle: 基础设施市场演进模型
accDescr: 展示市场规模扩张引发的垂直整合瓦解与生态碎片化分化过程
subgraph S1["市场扩张阶段"]
A["行业巨头聚焦核心业务<br/>（拾取主流金砖）"]
B["海量差异化用例涌现<br/>（边缘高价值银砖）"]
end
subgraph S2["历史规律演进"]
C["完全垂直整合模式<br/>（如福特早期橡胶与汽车工厂）"] --> D["多层级供应链解耦<br/>（汽车零部件专业化分工）"]
end
subgraph S3["新一代创业机遇"]
E["第一性原理技术创新<br/>（能效比/成本十倍级飞跃）"]
F["系统级垂直优化方案<br/>（芯片/内存/网络/电力协同）"]
end
A --> C
B --> E
D --> F
```
Ben Horowitz 进而用汽车工业史上的经典案例佐证了“市场扩张必然导致碎片化分工”的历史必然。在 1913 年福特汽车的胭脂河（Rouge River）工厂时期，亨利·福特曾狂热地推行完全垂直整合，从水、煤炭运入工厂，直接开出整车；他甚至在亚马逊丛林中买下了整座橡胶园，建立了一座拥有全套美式冰淇淋店与露天音乐台的微型城镇“福特兰迪亚”（Fordlandia），试图完全自控轮胎原料，直到严苛的工作纪律激怒了当地劳工导致计划流产。Martin Casado 总结了这背后的核心规律：“As markets expand, they fragment.” 随着 AI 的应用场景呈指数级裂变，任何一家巨头都绝无可能独自服务所有差异化的高价值场景。
- **边缘市场的巨大价值**：数万亿美元市值巨头所忽视的 5% 市场缝隙，年化体量足以支撑多家百亿美元级的新兴领军企业。
- **历史周期的重演**：如同 PC 时代的微软、云计算时代的亚马逊 AWS，行业巨头一旦体量见顶，其边缘创新空间便会迅速释放给挑战者。
- **垂直整合的极限**：单纯依靠单体公司通吃整个堆栈不仅难以持续，也无法满足各垂直场景对成本与能效的极致优化需求。
---
### `43:30` 硬件层重构与软件利润神话的终结
谈及 AI 基础设施与过往纯软件周期的本质差异，Martin Casado 揭示了一个常被行业忽视的认知误区：过去几十年里，软件公司引以为傲的超高毛利率（Gross Margins），其实在很大程度上是软件交付方式自然带来的红利，而非刻意攻克技术底层产生的结果。无论是一次性交付的套装软件，还是托管在云端的 SaaS，只要商业模式跑通，由于边际分发成本趋近于零，高利润率几乎是唾手可得的。
但在 AI 时代，这一神话正在被打破。大模型的训练和推理持续消耗着海量的物理资源，硬件架构的优化程度不再仅仅是一个底层技术细节，而是直接决定了业务盈利水平与企业生死存亡的关键变量。Martin Casado 深刻指出：“We may actually be entering an era where the optimization in the hardware is absolutely meaningful to the upside of the business in a way that we haven't seen in the past.” 硬件层、互连层以及底层架构的效率改进，成为了决定 AI 商业可行性的胜负手，这为新一代专注底层硬核创新的创业团队创造了前所未有的窗口。
---
### `45:00` 万亿级基建赛道与系统型创始人的崛起
对于 a16z 设立的 Machine Age Fund 所专注的细分赛道，Raghu Raghuram 详细拆解了底层基础设施的生态矩阵。在当下，仅仅设计一款算力芯片（Compute Chips）已经远远不够，团队必须具备交付完整端到端系统的能力：
1. **计算与芯片（Chips & Silicon）**：针对特定工作负载打造创新的加速单元。
2. **内存体系（Memory Innovation）**：攻克高带宽内存与容量瓶颈。
3. **网络架构（Networking）**：支持海量集群超大规模互联的通信拓扑。
4. **电力与基础设施运维工具（Power & Automation Tools）**：包括数据中心供电、散热以及管理大型算力集群的系统级自动化软件。
与过去传统的企业级软件投资相比，这些公司的运作模式截然不同。最显而易见的特征便是极高的启动门槛——首轮融资动辄数亿美元，公司在拿出成型产品之前就必须吞吐海量资本。更重要的是，这催生出一种全新的创业者画像，即“系统型创始人”（Systems Founders）。
Raghu Raghuram 强调，做纯科研的学者或写算法的工程师无法胜任这类公司的掌舵人。创始人不仅需要主导芯片与架构设计，还必须在画出第一张设计图纸前，就通盘考虑供应链供货、先进制程代工以及下游交付生态。在这一点上，英伟达的黄仁勋就是极致标杆，Raghu Raghuram 评价道：“And of course, Jensen is the Michael Jordan of this, right? They think the entire ecosystem right from the get-go before they start designing the chain.”
此外，外部市场环境也发生了两项决定性的质变：
- **前沿实验室极度饥渴**：主流 AI 大模型实验室对新算力的渴望达到了白热化程度，他们甚至愿意在初创公司硬件尚未正式流片出厂前，就签署大额采购或测试协议，这为投资人提供了极强的早期验证信号。
- **资本供给高度充裕**：市场对硬核基建重构达成广泛共识，后续轮次的大规模资本储备极为充沛，彻底打破了过去几年硬件创业“融资无门”的阴霾。
---
### `47:00` 为什么 20 岁的年轻创始人正在减少？
针对 Stripe 创始人 Patrick Collison 此前关于“如今像当年扎克伯格、比尔·盖茨那样二十岁出头建立标志性企业的年轻创业者似乎正在减少”的观察，Erik Torenberg 向合伙人们询问这一趋势背后的原因。
Ben Horowitz 明确指出，这主要是由于创业主战场的技术复杂度和系统复杂性发生了根本迁移。当创业项目涉及极其盘根错节的全球供应链、实体制造流程以及高度尖端的底层硬件与物理定律时，深厚的行业经验是不可替代的护城河。即便是最具天赋的天才人物，如埃隆·马斯克（Elon Musk）或特拉维斯·卡兰尼克（Travis Kalanick），他们在二十岁出头的创业早期做的同样是纯软件业务（如 Zip2、X.com/PayPal），直到在商业管理和技术研发上积累了足够多的经验与挫折洗礼后，才逐步毕业并跃迁到造车、造火箭等极其复杂的硬核物理领域。
Ben Horowitz 深入剖析了创业者认知负荷的极限：“Look, when you're learning how to build a company, it's hard enough if you completely understand the product. If you don't completely understand the product and have to learn it while you build the company,” 创业本身需要同时学习如何组建团队、融资、销售和运营，这已经极其艰难；如果在搭建公司架构的同时，创始人连最底层的技术和复杂制造系统都尚未理解透彻，那么在多重维度的极端压力下，创业失败几乎是注定的。因此，在需要深厚经验累积的机器时代基础设施浪潮中，年长且具备完整工业界履历的资深专家正在重返舞台中央。
### `50:00` 硬件创业的人才断代与代际融合
对于涉足底层硬件与物理基础设施的全新创业者而言，其面临的学习曲线极其陡峭。对话者们对比了两种典型的创业者肖像：一边是年轻且才华横溢的年轻创始人（例如像 Michael 这样的软件奇才），他们构建的是纯粹的软件或 AI 应用；另一边则是拥有几十年深厚行业经验、能调动复杂工程与供应链的领袖（如埃隆·马斯克或特拉维斯·卡兰尼克）。年轻创业者若要在当下直接掌舵复杂的重型基础设施公司会异常艰难，或许需要再历练十年才具备足够的掌控力。更深层的原因在于，过去二十年间整个行业和学术界几乎都忽视了硬件与底层系统领域，高校不再将此作为风口，学生实习时也根本没有机会去亲手设计和制造芯片。
然而，这种人才断代现象正在发生深刻逆转。如今涌现出的一批硬科技与重工业先锋企业，正在为整个科技界培养全新的工程人才梯队，这些人在更年轻时就进入系统级战场，其对产业生态的辐射效应甚至将超越公司本身。对话者指出，当下最理想的硬件创业模式并非必须由老将亲自担任创始人，而是富有野心的年轻一代与资深行业老兵的深度绑定。a16z 投资的一家代表性企业便是由两位二十多岁的创始人创立，但只要走进其办公室，就会看到坐满了行业经验极其丰富的资深工程师。“It doesn't necessarily have to be the founder with experience, but that founder better be able to tap into that experience in a real way.” 创始人自身未必需要数十年的硬件资历，但必须具备在现实中撬动、组织并驾驭这批核心资深专家的领导力与协作能力。
- 行业过去二十年的偏向：学术界与硅谷资本长期聚焦于纯软件与应用层，硬件与底层系统一度沦为非增长领域，导致全行业精通芯片架构与大型系统的中青年人才严重匮乏。
- 人才溢出效应：正如 SpaceX 和特斯拉为航天与制造业重塑了供应链和人才池一样，新一代计算与硬件先锋企业正在批量培养深谙硬件系统工程的年轻骨干。
- 团队构建法则：成功的硬科技初创团队往往呈现“双层结构”——敢想敢闯、具备敏锐技术嗅觉的年轻创始人，搭配拥有数十年工程落地经验的资深老将。
### `52:15` 机器时代基金的底色与美国科技霸权愿景
谈及此次设立专注于底层基建的“机器时代基金”（Machine Age Fund），主持人提出一个引人注目的事实：尽管这是一支规模极其庞大的新基金，a16z 却没有为此引进任何新的通用合伙人（GP）。本·霍罗威茨与马丁·卡萨多解释道，这是因为合伙人团队本身就成长于硬核技术与底层架构领域，重系统、重工程本就是他们整个职业生涯的底色与心血所在。在过去相当长的时间里，市场普遍追逐轻资产软件，他们甚至需要刻意克制自己冲向硬件的本能；而在早期投资历程中，他们其实早已低调布局了 SpaceX、Anduril、Astranis 及 Waymo 等硬科技企业。因此，掌舵这支新基金并非临时扩充外部能力，而是让团队重返其最具基因优势的原生战场。
```mermaid
flowchart TD
accTitle: 机器时代基础设施支撑科技繁荣的飞轮模型
accDescr: 展示从底层基础设施突破到国家科技领导力维系的全链条因果关系。
subgraph CoreInfra ["底层要素全面丰饶"]
Chips["低成本/高性能计算芯片"]
Mem["超大带宽存储架构"]
Power["清洁电力与高效能数据中心"]
end
subgraph TechAdvantage ["技术与创新繁荣"]
Models["万亿级推理与自主智能体"]
Pioneers["全球顶尖人才汇聚"]
end
subgraph GlobalStakes ["地缘与社会愿景"]
Leadership["维系美国科技领导地位"]
Values["捍卫开放机遇与自由价值"]
end
CoreInfra --> TechAdvantage
TechAdvantage --> Leadership
Leadership --> Values
```
当被问及若该基金在未来五到十年实现目标、世界将呈现何种图景时，本·霍罗威茨给出了极具地缘战略视角的回答。他直言，最根本的期许是美国在底层基础设施竞赛中赢得胜利，并在全社会实现芯片、存储、电力以及超环保高效数据中心的极度充沛。“America is a special place and we're important not only to everybody here, but anybody in the world who wants to kind of make a contribution and do something bigger than themselves.” 霍罗威茨坚信，美国之所以特殊，在于它为全世界怀揣梦想、一无所有的人提供了做出深远贡献的舞台；然而，一旦在底层计算基础设施的竞争中失去领先地位，全球科技时代的话语权将不可逆转地滑向另一个国家，而那个国家很可能推行着一套截然不同的价值观。
- a16z 团队的原生基因：管理合伙人均出身于底层网络、虚拟化与系统架构，运作该基础设施基金是其技术基因的直接兑现，此前已早期投资 SpaceX、Anduril、Astranis 等物理与硬件前沿企业。
- 未来十年的基础设施愿景：彻底打破算力与能源桎梏，达成芯片制造、高带宽存储、绿色能源与超高能效算力中心的“全面丰饶”（Abundance）。
- 科技霸权与地缘价值：算力基建是捍卫自由世界科技领导力与开放创新土壤的核心堡垒，失去技术主导权将直接威胁到基于自由机会的现代社会底座。
## Notable quotes
> `00:00` This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. It's The comps on this are the microprocessor, the steam engine, and electricity, or maybe the wheel. — Marc Andreessen
> `25:00` Like, I don't think in the history of the industry, we've ever created a digital artifact with something like $5 billion that went directly into that artifact. — Martin Casado
> `37:11` Artificial intelligence was the wrong word. We shouldn't have called it. It's machine intelligence... it is a cache of how humans thought, is a collection of humans thoughts. — Martin Casado
> `43:30` We may actually be entering an era where the optimization in the hardware is absolutely meaningful to the upside of the business in a way that we haven't seen in the past. — Martin Casado
> `40:00` As markets expand, they fragment. — Martin Casado
## Mentioned
Ben Horowitz, Erik Torenberg, Martin Casado, Raghu Raghuram, Marc Andreessen, Alex Rampell, Dan Rose, Patrick Collison, Elon Musk, Travis Kalanick, Jensen Huang, Henry Ford, Thomas Edison, a16z, Machine Age Fund, American Dynamism Fund, VMware, Meta, Facebook, NVIDIA, Cisco, Juniper Networks, Arista Networks, Microsoft, Amazon AWS, TrialPay, SpaceX, Tesla, Anduril, Astranis, Waymo, Stripe, ChatGPT, Grok, Kimi, GrokBot, OpenClaw, Hot Chips
## Video description
Ben Horowitz, Martin Casado, Raghu Raghuram, and Erik Torenberg discuss the launch of a16z's new Machine Age Fund and the infrastructure buildout behind AI, from chips, memory, and networking to power, cooling, and data centers.
Why a dedicated fund now? The group argues that the bottleneck in AI is increasingly shifting from the models themselves to everything beneath them. Hyperscaler CapEx is surging, critical components are booked years in advance, and each new generation of reasoning and agents requires dramatically more compute. They unpack why this cycle looks different from previous infrastructure booms and how AI is turning problems once constrained by engineering into problems that can increasingly be attacked with capital and compute.
They also explore where the next generation of infrastructure companies could emerge, why founders are returning to hard technical problems across hardware and systems, and what it will take to rebuild the computing stack for the Machine Age.
[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/a2703003-7e2d-4dd3-b92b-99a1aebc67ed/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=a2703003-7e2d-4dd3-b92b-99a1aebc67ed&feed=JGE3yC0V)
Source quotes
[0:00] Marc Andreessen: This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. It's The comps on this are the microprocessor, the steam engine, and electricity, or maybe the wheel.
[25:00] Martin Casado: Like, I don't think in the history of the industry, we've ever created a digital artifact with something like $5 billion that went directly into that artifact.
[37:11] Martin Casado: Artificial intelligence was the wrong word. We shouldn't have called it. It's machine intelligence... it is a cache of how humans thought, is a collection of humans thoughts.
[43:30] Martin Casado: We may actually be entering an era where the optimization in the hardware is absolutely meaningful to the upside of the business in a way that we haven't seen in the past.
[40:00] Martin Casado: As markets expand, they fragment.
Source quotes
[12:35] Raghu Raghuram: Any problem that you have can be solved with enough infrastructure, GPUs and power and money. And so until we run out of problems, we're not going to run out of demand.
[0:00] Ben Horowitz: We have a whole new technology that's the most important technology ever, and you need a whole new infrastructure.
[25:00] Martin Casado: We've actually gotten to this interesting point in the industry where it actually makes sense to build an ASIC per model just because the amount of capital investment in that model.
Original transcript
[0:00] We have a whole new technology that's the most important technology ever, and you need a whole new infrastructure. Normally when we talk about the infrastructure, we're talking about the servers and the storage and the network. Here it goes all the way down to the mines, copper mines. That's how widespread this thing is going to be. It used to be when you built something, it was an engineering problem. And here it feels like it really is a resource limitation. So whether it's tokens or not, we're pouring a ton of money into systems and then those systems are producing a result. to actually match the resources we're pouring into them. The leading memory runner said the demand they have today will take them three years of capacity to supply. If this fund does what we think it will do, how do we see the world in five to ten years? America wins in the infrastructure dig, and that would be awesome. Today, A16Z is announcing the Machine H Fund, a new fund dedicated to the infrastructure powering the next era of AI. I'm joined by Ben Horowitz, Ragul Raguram, to explain why we're launching it now. and why we believe the next major bottleneck in AI isn't necessarily the model. It's everything underneath it. Chips, memory, networking, power, cooling, and data centers are all being pushed beyond what they were originally designed to handle. At the same time, AI is changing an old rule of technology. Throwing money at an engineering problem didn't necessarily make it move faster. Increasingly, capital can be converted directly into compute, intelligence. We unpack what that shift means, where new infrastructure companies can break through, and why a new generation of founders is returning to some of the hardest problems in computing. Ben, Martin, Raghu, welcome. Thank you. All right. Thank you. I want to start with a Mark quote to introduce this new fund. This is the biggest technological revolution of my lifetime. This is clearly bigger than the internet. It's The comps on this are the microprocessor, the steam engine, and electricity, or maybe the wheel. Guys, the Machine Age Fund... Please introduce it. Ben, start us off. Well, basically, What's happened is we have a whole new technology that's the most important technology ever. And What happens every time, There's a dramatic... new way of using all of the things that we love, infrastructure, you need a whole new infrastructure. And never has it been more high impact as it is on this one. So not only do we need new chips, new system software. We need new ways of doing power. We need to replace copper. I mean, like, it's absolutely everything. So it's a very exciting time. So particularly for the kind of hardware aspects of this new era, we needed a new approach. Yeah, I would agree. Normally, at least in the computing, when we talk about the infrastructure, we're talking about... the servers and the storage of the network. Here it goes all the way down to the copper mine. That's how widespread this thing is going to be. That's number one. Number two, I think what we have seen over the last three years is a steady increase of the capabilities of the models. where the model is no longer the bottom line. And in fact, using AI, these models are getting better faster and faster and faster. Now the bottleneck is all what I call south of the model. And so that's why we need to work on that. The only thing I'd add very quickly is we tend to follow founders. And we've been watching over the last couple of years as the number of very strong teams going after complex hardware problems has increased. I don't know the actual numbers, but I was trying to estimate it over the weekend. So I think we'd get maybe 5% of the deals from top founders would come in, would be hardware before. 30% right now. So like the founder community, which tends to be much smarter than the VC community, has identified this as a very active area for innovation and they're responding. I think 5% is probably generous. Yeah, it was very low. It was very low. Yeah, 3%, yeah. And explain some of the macro conditions that have led to this change in terms of the surplus of fibers pursuing these ideas. Like what are they seeing? Well, I mean, the obvious is the demand for AI is basically infinite. And as a result of that, every part of the supply chain is under duress. I mean, everything, used to make things like memory. It's also very interesting, there's something unique about AI, which because the demand is infinite and growth is infinite, what you tend to worry about is the margin of companies, which is how efficient it is. Like normally you worry about growth. Can I just, can I just get people to buy this stuff? You don't have to worry about that here. The question, can you do this in a way that's profitable? And a lot of the efficiencies are actually strictly a physical limitation of hardware. And so even the business model of the AI wave is systems because they weren't built for AI, they weren't built for those workloads. And I think there's just this global observation that we actually need to change the
[5:00] or components to get that efficiency to help drive the growth and to drive the value of the businesses. Yeah. And how did we know that demand is actually outpacing supply here rather than this being another hype cycle? Well, I mean... Yeah, there are a number of security. Firstly, It is that some of the smartest... Judges of demand. are cutting huge purchasers. I mean, if you look at the hyperscaders, right? Their CapEx spend has been... exploring next year, supposedly it could be a trillion dollars. collectively across the bay. Hyperscalers this year, it's about $700 billion, right? And if you think about the hyperscalers position in the industry, Thank you. they see demand from everywhere, right? They see, obviously, the Frontier Labs wanting their compute. They see the AI data companies. They see the enterprise. They see the U.S. geography, the international geography. So if anybody has visibility, it is that. And they've been jacking up their CapEx, because it's never been seen before, right? So that's a clear, clear sign. And secondly, if you look at the companies that we see on a day-to-day basis, they are all ripping. All the application companies, the growth is insane. The Frontier Labs, the growth is insane. It's been documented. So I would say on the demand side, the signals have never been clear that this is not a hybrid. And prices are going up. up on chips. If GDP prices went down, they always go down. As long as it goes down, if you look at the code, It's like this and then we're stuck with this. And we know only like 5-10% of the address of the market have stopped today. I mean, if you look at the supply across the board, it's basically all booked out to 2028. I mean, it's so bad. We've actually seen multi-day auctions for a few thousand GPUs. The other side of that, of course, is demand. And as Raghu said, we've seen the fastest growing companies you've seen in the history of the industry. And also the unit of work that AI can do. the value of that unit of work keeps increasing. But underneath the covers, the number of tokens that are consumed is going by orders of magnitude, right? If it's 100 tokens for chat or agent, it's thousands of tokens, right? So you've got expansion on both sides of demand. One is the unit of work is becoming more and more consumptive. Yeah. of tokens, and then secondly, a number of people there for that are going to be a benefit. It's not just the developers. It's going to be all knowledge workers and all of beyond that. So that's what we see. You said the key components in supply are sold out to 2027, maybe 2028. What does it mean for an entire industry to be sold out that far like I don't know if this has ever happened before. Do you guys recall? I mean, remember in the internet days when we were doing massive build out, the majority that was actually being put in the ground was speculative and was dark. Remember the dark fiber in here? Basically every GPU that's being created is already pre-sold. Yeah, we weren't quite there. I mean, there was a lack of bandwidth. like in the 98, 99 timeframe. But There wasn't that much real demand for it because there just weren't that many people on the Internet. So it was a two-sided thing. And the companies were all rushing there and needed more bandwidth theoretically. But there weren't. the users on the other side to consume it necessarily. And then to really consume a lot of bandwidth, you have to do high bandwidth things like video, which weren't really viable for a number of reasons that had nothing to do with how much bandwidth was in the data center. So It smelled... Similar, but it wasn't this. This is like we're flat out and people are reselling GPUs for... four times what they bought them for and this kind of thing. Like it's just not, and then we're also out of power and cooling. And then, um, On top of that, it's really hard to build because there's these incredible political headwinds going into it. So it's really unprecedented in my career that we've had anything like this. I want to give you a quick anecdote. So I was talking to a CFO of a large company, a large public company, who had historically been very resistant about going into the cloud. So they had a lot of servers. And they were doing an inventory check, and they realized that the memory in their servers had increased so much it could fund the entire migration to the cloud. So I just feel like we're in a very unusual situation. That's right. We're out of many things. Power, cooling, memory, GPUs, like you name it, we're out of it. conference for the industries, the one called Hot Chips, is going on in Stanford. And the leading memory runner said that, the demand they have, Today, it will take them three years of capacity.
[10:00] and supplied. It is just today. It's not even future demand. So in terms of being about everything simultaneously, Is it because people just underestimated how good models would be, how useful they would be? They just couldn't have received the demand? Well, I don't even think it's that. I mean, this stuff came out of nowhere, right? We're only four years into this. So even if we had a perfect Oracle once it started working, I don't think we could have built the capacity. We could have built the capacity. There's no way. And we're talking about like chip cycles, which tend to be three to four years. We're talking about breaking ground and building data centers, which is four to five years. power source. So like you either have to build your own power or like, usually both, you've got to build your own power and have a power source, which is, not easy. You know, I'm a, industry that's used if it's growing at 20 30 percent it's a great growth rate Right? And it's being connected to an AI software industry that's like Triple digits is the base. So you can see the disconnect, right? So it's just widening. And so why didn't this fund exist five years ago or seven years ago? Or why was it not a great category to invest in the same way, Brad? Well, I would say we're probably, I'd like to think we're just in time, but we probably would have been Well, seemed to have at least a couple of years ago. I mean, I will say you can actually point on basically every epoch to an independent company that came up, right? Clearly, they moved from a mainframe to the client server. We saw a bunch of companies come up. They moved to the internet. That's why we've got Cisco and Juniper. Even in the mega data centers, which, by the way, was largely driven by the incumbent cloud providers verticalizing. You saw the arising of Arista. because the change has been relatively minor, like one chip company, one switch company, where here, everything is. And so I agree with Ben. We probably could have started a little bit earlier, but the amount of change is so high now that it's just an obvious thing to do. And the other thing is the demand... for intelligence and is so vertical. And, with. Really no end in sight. I mean, because every company that's adopted it, is growing very fast in its usage and then most companies haven't adopted it to a high degree And then consumers are just getting started. And so it's going to probably, the demand for tokens is probably going to grow close to 1,000% a year, which you cannot grow supply that fast. Like we're not, like the amount of just... work we're going to have to do across the board. to get to the point where we can grow like infrared, at that kind of rate is pretty vast. So I think there's a lot of investing opportunity on the way. And by the way, the other thing is like all the architectures, of the hardware systems. were built for a whole different era of computing. And so more than... just we need more capacity we need capacity to build There's lots of opportunities to build different kinds of of infrastructure. Yeah, they're all reaching Therefore, the physics limits for what they were designed. Bye. like what Ben was talking about, copper and so on and so forth. And you could go across every one of these categories, and you could find, okay, this is the limit of this type of technology. So now, You got to... some technical breakthroughs to get to the next one. I want to dive deeper on the demand side for a second. As we've moved from chatbots to reasoning to agents to multi-agents, each step has multiplied the number of tokens a single task takes up by increasing orders of time. Nobody likes to use AI more than AI. Why does that keep happening instead of leveling off? And do you just see that happening indefinitely, just continuing to? Well, so there's a couple of answers. The first one is, for sure right now, if you look at the way we're achieving scaling, the way we're doing it is through a lot of inference. So through a lot of token, right? If you think about what RL is, it's a lot of inference. If you think about chain of thought, it's a lot of inference. If you think about long-running agents, of course, it's a lot of inference. approaches that we've been using to scaling. I think if you want to step back and say, what is the macro trend here? It used to be when you built something, it was an engineering problem. You'd throw a bunch of engineers at it, and that doesn't scale, and that would have a natural law of engineering physics, which is what the Mythical Man Month came
[15:00] And here it feels like it really is a resource limitation. So whether it's tokens or not, we're pouring a ton of money into systems, and then those systems are producing a result. And right now we're bottlenecked on those systems' ability to actually match the resources we're pouring into them. And so I think tokens right now is probably where we're on the scaling curve. But we don't have a natural regulator like engineering like we did before. So I think we should expect this to continue, and we have to build a supply to support it. Yeah, like the similar way to think about it is – Any... problem that you have can be solved. with enough... infrastructure, GPUs and power and money. And so until we run out of problems, we're not going to run out of demand. And that's the challenge. AI's answer to getting better and better is to use more AI. Influence is one. basic building block path that it keeps using over and over and over again. And so that's why these tokens multiply. Yeah, even the auto catalytic effects, so even the idea of using AI to create more AI, like creating GPU kernel, of course, is just using more AI as part of the process. So again, One way that we think about it is in the past, money would come in, You have an engineering problem. We know that it takes two years, normally fails, it's a national governor, and then you get the product on the other end. There's nothing between the money going in and then the hardware creating intelligence. And so now we're just limited by our ability to create supply. It's a very, very different dynamic. Give me more GPUs and I'll solve this problem. Yeah, that's right. As long as you have the money, the GPUs, and the data, for the foreseeable future, you'll be able to scale these things. It's fascinating because over the last decade, it feels like there are so many people and the pervasive sentiment was there's too much money going to startups. We're overfunding these startups. There's too much money in venture capital. Say more, Ben, about what that means because there used to be this sort of skepticism that the more money you put into the industry that there would be bigger outcomes. And now we're saying at the offsite that there's to some degree the market is as big as we collectively contribute to it. Yeah, so this is – look, the one thing we all knew – in startup world is that If I have a two-year lead on you and you try and cast me by hiring 1,000 engineers, you're going to wreck your company. That never works. It's a mythical man month. Nine women can't have a baby in a month. That's it. That never works. Okay, now that works. But it's not hiring 100,000 engineers. It's taking $3 billion to... and like lighting up a magnificent cluster and then all of a sudden you know whatever grok can come out of nowhere and like oh all of a sudden it's real or or kimmy or or what have you It's just like these leads, you can throw money at the problem, and you can throw money at almost any problem, and that works. And so that is just completely different. than anything we've ever lived through. So we're, by the way, we're all psychologically adjusting to this. the ChatGPT app has a billion weekly actives. There's about 30 million developers. who are using relatively a big portion of compute demands. How do we think about commute compute demand needs now and in the future in light of what people are actually doing with air. That's the progression, right? So, charge agreement was a casual act and calling for professionals, right? Now, Using coding, Yamaha built amazing tools for knowledge workers. So that's the next frontier. And now there are over a billion knowledge workers in the Y, right? And with that, it's a long ways to go for that demand. And by the way, the work that they do, all this work, and automation and so on, and then they get to the back office, which is all the agents. So progressively, each of these things unlocks everything. I would say an order of magnitude more than that. I mean, that's just the start of this. Well, and now you have GrockBot. which is, is kind of, uh, You know, what happened with coding is... kind of happening with all use of computer via GrokBot and So we're in a whole nother wave of demand and most certainly there's going to be more to come. So it does seem quite unlimited at the moment. And we haven't even gotten into embodied AI or robots, which are going to be another source of demand. Martin is an expert, but my understanding is that Glock uses computer use, which is just like you don't have been sitting inside the computer talking to Ed. I literally used it over the weekend
[20:00] to update my credit card with a bunch of services that I've been lazy to do and cancel a bunch of subscriptions. I mean, this is not coding or whatever. This is true computer use. All of a sudden, you create a lot of knowledge work case, except they're all sitting inside of the computer doing work. I do think that Mark Andreessen is right. The right analog here is like the steam engine or electricity. In the following way, we've introduced this new thing that you can turn to work, and there are some very obvious applications now. But there's probably 30, 40 years of throwing computer problems. I didn't think with a clear reward signal. And we're just starting. Like, we've got language and code. That's it. And just starting computer use. But, like, what else are we looking at? We're looking at in terms of science, materials, biology. I mean, of course, creativity is a massive use. And so, listen, we're on the very, very early part of a very long journey. And we've removed this key bottleneck, which is, you know, traditional software engineering. Now, of course, bottlenecks will move, and there'll be kind of more complexity elsewhere. But I think we're at it very early in a very long run of throwing computer problems. So let's expect this compute need to persist for decades. Because we mentioned it. Marty talk about GragBot because we were talking at the offsite about what struck you about it Obviously we involved in every possible way You could be involved But what did you find so interesting about it So I think we've, as an industry, gone through kind of multiple realizations for how AI enters our lives, right? And very early on, we're like, okay, well, you add AI to a product, and it's like a search bar. And then you do chat with it, and it chats back. because that's kind of the traditional way to do it. And then OpenClock kind of showed up, and that was earlier in the year. And with OpenClock, I said, okay, well, maybe it just being like Google but better. Maybe that's not the full embodiment of it. How about we'll have it be a standalone thing, but it'll be an extension of you, and it'll share your keys, and it'll know your passwords, and it'll just kind of do stuff that you would do, right? So it's kind of an extension of you, but it's more like a human extension of you. And then what I think GrokBot got really right is, no, how about it just actually an employee? So now you have this thing that's an entity, and it doesn't have special access to your keys or whatever. It has its own computer, and it has its own browser, and because these are the smartest models in the world, it can do whatever an employee can do. And it's kind of interesting because now, actually, if I want something done the first thing I think is well can GrockGock do it for me And often the answer is yes even if it something you wouldn expect So the obvious ones are like whatever it like manage my calendar it like book a meeting But there also non ones as well. So for example, I'll have it read through my email and do triage. And I don't tell it how to do that, but it will know to check with me before actually doing the triage. So like these things are sophisticated enough that you can give it relatively high level tasks and it'll do kind of, you know, like sophisticated things as a result. Ben, I know you're thinking a lot about this and how this works in the organization. You think a lot about culture, of course. What are your thoughts here? Well, I mean, I think if you just look at us, you know, it's like having a new kind of employee and there's going to be a lot of them. And we have to, just like, you know, we spent many, many, many years figuring out how to work with our kind of regular human employees and now we've got these other kinds of employees. And, you know, there is a learning curve with them. So they can burn a lot of tokens and spend a lot of money and get nothing productive done They can forget stuff They can make stuff up They can have good behavior They can have bad behavior. They can create security problems. So there's all those aspects to it, but they can also be super-duper productive. So I think figuring out how to integrate them in, have them work nicely with the people that they're working with, the actual humans, is all something that we're learning how to do. I mean, I don't want to sit up here and say, I've cracked the code, we've got this marvelous loop, and the whole firm is just completely automated now, and I'm going to slowly get rid of all the humans because I can. Like, that's not at all where we are. We're much more going like, okay, how do we make all our humans superhuman without wrecking the place because the bots get out of control? Yeah, and it's interesting because we've done, We've tried a couple of different ways. It's how best to get agents into the system, if you will. And eventually, it was Martins inside, just treat them as people and get it done. And that's what we're doing. That's turned out to be the most durable way of getting this thing going inside of an organization.
[25:00] Thank you. I want to go back to the supply side and go deeper into the bottlenecks. We were talking about how, you know, in terms of the data centers, the chip architecture, system software, facilities themselves, that none of them were designed with AI in mind. What would it look like for them to be designed with AI? Like, what is sort of the mental model for thinking about what that could mean? Yeah, so I mean, if you... start with the statement that he just said, "Hey, original model of infrastructure on any of these models has to change. You can go up So... category by calling, you can see varied breaks, right? And then you start unlocking the bottlenecks in each one of these things. So eventually, you have to get to a system where, if you look at what an inference engine does, right, it takes up a lot of memory. It generates new tokens along with the computer. And so you can just think about how do I optimize all of this? What does the memory need to be? What does a computer need to be? How do they need to talk to each other? How much power does each of them need? And if they need all of these power, how do you code each of these, right? And then how do you put the collections of these things together? That is the exercise that's underway in the industry right now with a lot of the founders. So they're breaking down the problem into its fundamental components and saying, what is the exact nature of the compute that's getting down okay, how can we measure multiplications? How do I optimize my computer and what kind of a scenario. And they all need memory progressively to generate these tokens, what is the best way of hierarchically arranging this memory, right? And then how does the power consume? I mean, then you've got to connect it together. What are the ways of connecting it? on the same chip but across chips and across data centers. How much power does each of these data transmission take. So you have to progressively break it all down, and rebuild it from these fundamental building blocks. And that's what we see underway, and that's where we see that generally. I'm going to get an interesting mental model to think about on how the landscape's changed. So today, to build a frontier model costs, let's say, $3 to $5 billion, right? And that's to train it. And so the inference has to pay back at least that, of course, right, in order for any of this stuff to be viable, let's say, two times that. to make... $10 billion. So if you can save 20% of efficiency on that, That's $2 billion. And you can easily build an ASIC for $2 billion, right? So we've actually gotten to this interesting point in the industry where it actually makes sense to build an ASIC per model just because the amount of capital investment in that model. And then unlike traditional software, traditional software has a lot of state and a lot of, you know, it's very dynamic. These models are fixed. The model weights are fixed. And so we don't know if model of how you would evolve the architecture to be far more bespoke for these massive capital investments we're doing. Like, I don't think in the history of the industry, we've ever created a digital artifact with something like $5 billion that went directly into that artifact. And so, you know, like this, I think is going to put the greatest demands on hardware that we've ever seen. Well, To that end, rack power requirements are moving from roughly 5 to 10 kilowatts to 100 to 250 kilowatts. Compute density is climbing something like 70x. Cooling is moving from air to liquid as a requirement. What are the investment opportunities as a result of this? Well, first of all, when you get to that level of power per rack, AC power doesn't work anymore. So, like, that's a pretty wild thing. So now you're into DC power, which, by the way, also requires its own cooling and is, like... by the way, super fucking dangerous, which is kind of ironic because Edison promoted DC power by... Claiming how dangerous AC power was and demonstrating it by like electrocuting animals and things. The horse, yeah. The horse, yeah. Yeah. So... But he was right, but around his own kind of power, which is extremely powerful. It's the good news. So, you know, just starting with power. Yeah, that's going to be like... very, very different. I think with cooling, And this gets into – So yes, we're going air cooling to liquid cooling. I think we're already at liquid cooling for any state-of-the-art data center. That's already kind of a done thing. But it gets into... okay, you know, given the political environment and so forth, you, like – liquid cooling isn't enough. It's got to be eco-friendly liquid cooling and, you know, kind of TCP
[30:00] Power's not enough. It's got to be... power that contributes to the power of society not takes away from it and so you have data centers who have been behaving badly, Small percentage, actually, probably 10%. wasting a lot of water. Not as much as pistachios or almonds and so forth as people demonstrate on the internet, but they could be a lot more efficient with that. And then there are ones that kind of are parasites of power and don't contribute power back. I think all that's going to end. It's going to have to end just because we've kind of gone through a one-way door on that. So that requires a level of engineering that, you know many haven't invested in yet so that's coming uh and then you know like if racks are that dense, there are other things that, like the way the floors are designed, have to support that kind of weight. You know, that kind of thing is actually for real. And then, you know, I think that, There's... You just need a lot of everything. And so there's going to be – also the things are really loud. So you have to build the data center with thicker walls or you're going to disturb the peace in the neighborhood, which is not going to be acceptable. I don't think any state is going to allow that. And so a lot of the ways people have architected and designed the buildings themselves – are already completely obsolete. Like once we get to Feynman, a much smaller percentage of the data centers that we have today work. In fact, I mean, everybody talks about memory prices, but one of the fastest areas where prices are increasing is reinforced concrete. Yeah. The other thing that happens when... these data centers are sending 800 wolves to the rack. is It's become so dangerous, number one. But secondly, we don't have enough electrical contractors that have the expertise. to deal with 800 rules inside the data center. Because this is... High voltage. Only 2%. of electrical engineer or electricians in the u.s have been certified on dc power so like that gives you an idea now met has got a whole program to train people up and so forth which is great it's like a new job corps where they train people for free to do this job but You know, it's funny. AI is taking all the jobs. AI is going to create a lot of new electricians. Yeah, I think we're just doing something in space, too. The guys that own the big cloud data centers, big data centers, they all are... furiously experimenting with robots, right? To do the work of assembling or putting servers into the data center, etc. And so you will see, That increasing as a result of the evolution in AI. By the way, to be clear on the actual fund that we're raising, our focus is on computer science infrastructure. So anything a model runs on, that's computer science, right? So think, you know, chips, network, interconnect, storage, all the way down probably to the electricity. Yeah. And say more about the robotics arm in terms of what we'll be doing versus maybe American Dynamism or how to think about that. that AI will run on. Like one of the great breakthroughs that AI does is it allows computers to interact with the physical world, right? It can see, it can hear, it can talk, right? And this means new platforms, right? And the simplest way, people will say edge device, but that doesn't really mean anything, right? I mean, it could be a mobile device, it could be a CDN, it could be a laptop, but it also could be an embodied device that goes around. don't do heavy regulated industries or more verticalized industries, but any sort of computer science platform that's going to push AI further out, we're quite interested in. Yeah. Going back to the data centers, by 2028, new data centers are going to need something like 44 gigawatts of additional power against maybe 25 gigawatts of expected grid additions. Hold on, hold on. We use that word gigawatt. No, it's like, oh, we'll have about 100 gigawatts. Martin, what's a gigawatt? I mean, how big is it? It's multiple football fields. I mean, it's massive. It's 50,000 people. What do you mean? What is it? The equivalent is like 50,000 houses? 50,000 homes. I grew up in Flagstaff, Arizona, which is a town of 40,000 to 60,000 people depending on the universe.
[35:00] We have less than a gigawatt of power consumption. I mean, this is true. So you can basically light up and air condition your entire town for a gigawatt. Yeah, I mean, this is just throwing them around. No, but by the way, everybody talks about the gigawatt. There's very few gigawatt data centers that are actually up. I mean, we've got a long way to go. To put that in, why can't utilities and hyperscalers just build faster? There's so many things. Well, there's First of all, right now, you need humans to build them. So there's just like the regular construction. But much more than that, you need... Permits. Um, You need access to power that you can plug in. So you're either, you're doing a combination of, you've got to get access to power, which is a massive kind of regulatory system, bidding struggle. There's very limited kind of amounts and things you can tap into in terms of natural gas, power grids, what have you. But then you also have to build your own power. And guess what? We've got shortages of transformers and turbines and everything that goes into that. So, It's just... you know, like, you've got to get all that stuff. It's not, this is not a software problem. It's not just like a bunch of engineers, can't you, like, work weekends and that type of stuff. Like, this is, not that that works anyway, but... There are real bottlenecks in this, and these lead times are not that easy to compress. And look, we have the best minds in the world trying to figure out how to compress them, but... It's not easy. It's not easy. And the demand is not slowing down. So we're already behind. The demand is growing. 10x a year right now and The supply just can't grow that fast. By the way, it is so bad that right now, if we have new companies going for GPUs, it's often in Mexico or Australia or in other country just because it is so difficult in the United States. Yeah, we're creating huge job, both job and long-term economic opportunity in other countries by banning data centers here. I think, look, the right answer would be to set a standard where a data center... contributes back to the community. Like that power gets better, there's no noise, there's no water issue, and it's adding jobs. Like that ought to be the standard. And then everybody ought to be just held to that standard. By the way, like there are... There are data centers that do that now. Like that's not a... you know, like a futuristic dream or something. Energy rates have gone down every year they're there, and the reason is they provide their own power. They give power to... to the state during the day. And then at night, they borrow power from the state when the state doesn't need it because the way power plants work is you're always generating peak power. capacity. And since the data center has steady capacity during day and night, and the city goes way up in the day and way down at night, that's a symbiotic relationship. Zooming out of the Why do we think the, you know, we were batting around the name for a little bit. Why do we think machine age is a compelling term for what we're doing here? Well, listen, let me take it. So the first one is, I think Ben's absolutely right. Artificial intelligence was the wrong word. We shouldn't have called it. It's machine intelligence. Say more about that. Why is that? Because it's not how humans... think necessarily, right? I mean, it is a cache of how humans thought, is a collection of humans thoughts. But like, to date, we don't know how to take an AI with no knowledge and put it out in the world and have it reconstruct language, right? Like, that's not what we've done, right? We've built something that can learn off of everything we've already learned and then use that in a productive way. And listen, AI is a general term that goes back 70 years in computer science And of course, it's got a lot of baggage, either from science fiction or from, you know, Nick Bostrom, who wrote about it or whatever. And so the first one is just an acknowledgement like this really is machine intelligence. And then you want to emphasize the machine part of it. I mean, there's kind of this deep irony. And this is from, you know, the software's eating the world people that, you know, you've really come to a place where you pour money into something and then you're limited by the actual machines below it. And so I think it is kind of a nod to like the hardware component is so significant in this wave. And we want to acknowledge that. Yeah, I think that's what's going to create the next breakthroughs. demographics is the quality of the machines are not neat. So that's basically losing to the name. It's also a cool name. It sounds good. futuristic. Yeah. The,
[40:00] Um. Given how much has been spent on AI infrastructure to date and how capex intensive these businesses can be, are we past the point where new companies can break in at, at, at sort of material levels? Uh, you know, Why not incumbents like NVIDIA, Correve, etc. Just take the lion's share of these markets. They're all doing. Well, there's no question about it, right? But to our discussion earlier, when you're Get. You need fundamentally new innovations to keep the growth continuing or the pace of improvement continuing, whether it's tokens per second, per dollar, or tokens per watt. Filters for Rock, Right. or power, you take any metric. If you want to have a 10x on those metrics, You got to have new innovations. And new innovation traditionally comes from Boozian founders thinking about solving the problem from first principles in a different way. Right. And that's what's needed here for the next jump. and an innovation. I mean, this is the law of markets, right? I mean, let's assume that the existing Silicon incumbents are multi-trillion dollars in market cap, which is absolutely the case. Even... 5% of that is a massive private company. Massive private company, right? We're talking, you know, annual. And you could say, well, but NVIDIA could do that. They could, but why would they? If they're focused on things that are in the 90%, which is also driving the same amount of growth. And you always ask these questions. We ask these questions during the cloud days, right? Like, well, wouldn't Amazon did this? You would ask these questions during the Microsoft days. Why wouldn't Microsoft do this? There's a very natural law of markets is that once you get to a certain scale, there's tremendous opportunity for innovation. at the margins. Yeah, there's a funny quote from our partner, Alex Rampell. He had this startup called TrialPay. He was trying to sell it to or sell that it's his services to Meta and then Facebook. And Dan Rose, who was the head of CorpDev at the time, said, Alex, that's great. It sounds like you can collect a lot of silver bricks. But I'm like, I have so many gold bricks, I can't even pick them all up. So the last thing I'm doing is looking at a silver brick. And I videos in that position. 100%. Yeah. We were talking about as it relates to model providers that if you're you know, in the sweet spot of what OpenAI Arthropic can do, you know, one of their main sort of interest areas that might be a tough place to be. But anything outside of those maybe, you know, three to five areas might, you know, As markets expand, they fragment. Right. And it happens all the time. And remember, in the early days of Ford, there was the 1913, there was the Rouge River plant. Literally, this was like made cars. It went like water, coal, and rubber trees, and out came cars. By the way, he bought, a whole rubber tree plantation in the Amazon jungle. And there's a great book called Fordlandia, because he wanted to own the complete vertical thing. where he created this city, called Fordlandia. in the Amazon jungle, which had like, was all Americanized, bandstands and ice cream and all this kind of stuff. And it actually worked for a while until he made people like show up to things on time. And then they were like, screw this, get the fuck out of here. So now if you look at the car industry, of course, there's multiple levels of supplier and there's a bunch of companies. And this always happens. So, you know, as markets expand, they fragment. And then once that growth It can be like new challengers rise up. And that is the everlasting cycle of private markets. Yeah, because the use cases are multiplying. And there's no way, like if you're the biggest company, you can get to the biggest use cases. But there's so many use cases. And all, as Martin was saying, very valuable use cases. It's just very hard to get to in a great way. Yeah, even in France, you want simple architecture, right? You know, it's so complex now. It's inevitable that you can optimize things in a different way. By the way, here's a very interesting thing. People often don't understand that margins kind of fell out of the standard way of doing the technology with software. Right. Like it wasn't really a technology problem. Like once you got the business working, you tended to have pretty good margins because that's just kind of how software, certainly when you shift it, but even as a service. And that's not necessarily the case with AIs. We may actually be entering an era where the optimization in the hardware is absolutely meaningful to the upside of the business in a way that we haven't seen in the past. So there's a lot of opportunity here. Let's get deeper in talking about the types of companies we'll be investing in. Um. Maybe we could start by either illustrating the subsectors or if we can talk about a few or a couple investments that we've made. And there's some that haven't been announced yet. But Ravi, do you want to take us down? Yeah.
[45:00] The sub-sectors, as we've been talking about, is every one of these categories, right? The obvious ones are computer chips. But these days, it's not enough to build a chip. You need to build a full system, right? And then, therefore, what goes into the system? There is potentially memory innovation. There is potentially networking innovation. There is... potentially power tools, and so on and so forth. So each one of these categories are categories where you can see public company style companies emerging. And those are all the things that we are looking into. And then once you put it all together, there's a lot of software around it to automate all of these things, to manage these fleets, and so on and so forth. So that is another important area. So these things keep building on each other, but every one of these categories is important. Talk about what's different about these kinds of companies from the usual company. I mean, one thing you can tell from the companies we announced is their first rounds have been massive, you know, hundreds of millions. Is it a different kind of founder or what else is different as we think about just the practice of, you know, building and investing in these kinds of businesses relative to our traditional software? Well, I think the big thing is you hit on one of the big things, which is a lot of money goes in. Yeah. before they get to a product. And that's just kind of the nature of it. Now, that's true on big models, too. But I would say... That's a little more of a known path, whereas this has got a little more risk and a little more money than some of the other things that we've done. And, you know, look, a lot of the chip founders – are here from the past. Yeah. You know, like the guys who know how to make memory. They're not young. So that part is different too, but it's kind of exciting. Yeah, the other thing about these founders – We have algatmé. Systems Funders. So what I mean by that is you can't just be – the researcher or a great continuous artist, right? You got to be able to architect and design the chip or the system, whatever it is. Then you got to think about how is this thing going to actually get manufactured, right? Who's going to be supplying this? And a whole bunch of these downstream things, which... Normally, if you're building software, you don't know how to think about all of these things. So to really the best founders, um, And of course, Jensen is the Michael Jordan of this, right? They think the entire ecosystem right from the get-go. before they start designing the chain, right? because of the nature of the bottlenecks and all these things that have come together. So that's a big characteristic that is different. There's two environmental factors that are important to you. The first one is the labs are so desperate that they will engage with startups. And so we actually have quite a bit of signal early on because the labs are inking deals with companies before they actually have hardware available. And that's a big, big shift than five years ago, right? Like you just didn't go and sell your kind of janky hardware thing to Google or whatever. The second one is the capital availability has loosened up a lot. I think there's general consensus that it is the time to reshape this stuff. And so follow on rounds, there's a lot of capital available, which, of course, you want to be investing in two areas where there's capital available. And so the atmospherics are also just different. Patrick Coulson remarked a few years ago, he said, "Hey, it feels like there's less younger founders today." In the way Zuck in college, building the next Facebook or Gates, in the same way with Microsoft. And of course, the Michael Trulles of the world, there's still some young founders building iconic companies. But it does seem, to your point, that there's more older founders building these companies or less 20 year olds. I'm curious if it resonates and why. Well, I think it's Ragu's point that if... Thank you. if you're building something that has a very complicated supply chain, has to manufacture things, and is technically complicated, that some experience helps and helps. You know, if you look at, Elon or Travis Kalanick, they're Companies when they were young were software companies. It wasn't until they got like a lot, even those guys, the best guys, needed some experience in building a company's building technology and so forth to kind of graduate to the much more kind of complicated or I would say elaborate domains. much more, there are many more moving parts in these things. And so, look, when you're learning how to build a company, it's hard enough if you completely understand the product. If you don't completely understand the product and have to learn it while you build the company, um,
[50:00] that's just such a steep learning curve for a brand new entrepreneur. So I think that what we're seeing is you see Michael on the one hand, who is a very young guy, brilliant, But what he built was kind of a pure... software AI thing. And then on the other end, you have like an Elon or a Travis who can, who's got enough experience. I think Michael could probably do that, you know, 10 years from now, but today that would have been hard. It's important to remember, like it's been defocused by the entire industry and academia for the last 20 years, right? It just, it just hasn't been the same opportunity. Like it's been there, but like, it's never been a growth area. The growth areas And so I also think we just have a positive people coming out of the universities or having experience at large companies that have done this. And there's just not that many. Like you don't go intern and like build a chip. So but a lot of that's changing now. Like, listen, we're going to create a whole generation of of founders that come from these new companies that will know how to do this. And, you know, they'll be hired in much more junior. that are changing the entire industrial complex, maybe even greater legacy than the companies themselves. And I think we're going to see the same thing for computer science and hardware. As a matter of fact, one of our investments was started by a, two founders in the 20th, but If you go walk into their offices, you see the experienced people as well. So it's ideal combination here. Yeah, yeah. It doesn't necessarily have to be the founder with experience, but that founder... Better be able to tap into that experience in a real way. Get the people with it, yeah. Yeah, yeah. Well, and then be able to work with them and... They have to be good and all these kinds of things. It's complicated. Speaking of experience, This is a big new fund we're... launching, and there's no new GP's. we're sort of collecting. It's because you guys have a lot of experience and the rest of the group in this field that has been kind of latent and dormant. Well, it's kind of funny. I think we almost had to be warned against it almost just because our backgrounds are from us hard. And I think the reason that we needed a reminder is because all of this has been so much in our careers, existence, and heart, we were kind of drawn to that. And so, listen, we've been clearly invested in Harvard other years, right? We're in SpaceX, we're in Adrenaline. These are very early checks. We're in astronauts, we're in Waymo. You know, so even early on, we did a number of those investments. But like, you know, this is because it's so much in our DNA. And so I don't think this is necessarily to increase the team of competencies just for us. If this fund does what we think it will do. How do we see the world? changing her look in five to ten years. Well, you know, hopefully... America wins in the infrastructure day. And we have lots of like super eco-friendly, efficient data centers out there, and lots and lots of an abundance of chips, an abundance of memory, an abundance of power. And, you know, that would be Awesome. It goes back to what we really think. America is a special place and we're important not only to everybody here, but anybody in the world who wants to kind of make a contribution and do something bigger than themselves. It's kind of the best place to come with nothing and do something profound. So we'd like to keep that going. I think that doesn't continue to go if we lose our lead in technology. I think we'll be in another era and it'll be another country. And maybe they have a different set of values around that. Thanks for this round. Great. For seeing me, it was fun. Our team, Ben, Bargo, thank you. Ben too. Thank you. and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our Substack at a16z.substack.com. Thanks again for listening, and I'll see you in the next episode. As a reminder, the content here is for informational purposes only, should not be taken as legal business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a16z.com forward slash disclosures. . Bye. . Thank you. Bye.
