Gutenberg Lab — Morning Paper

a16z Launches Dedicated AI Infrastructure Fund as Demand Outpaces Supply by 1000x

Speakers

Ben Horowitz × Raghu Raghuram × Martin Casado (a16z General Partners)

Bottom line

AI represents the biggest technological revolution since electricity, but every layer of the computing stack—from copper mines to data centers—was designed for a pre-AI era and is now the bottleneck.

3-Line Summary

• a16z launches the "Machine Age Fund," a dedicated infrastructure fund targeting chips, memory, power, cooling, and networking across the full stack
• Token demand is growing at ~1000% annually while GPU, memory, and power supply are sold out through 2028; the leading memory vendor reports "today's demand alone would take three years of capacity to fulfill"
• Capital now converts directly to intelligence—a $3B cluster investment can overcome a two-year technical lead, breaking the traditional Mythical Man-Month constraint in software engineering

3 Key Points

1. Per-model ASICs have become economically rational

Frontier model training costs $3-5B, requiring $10B in inference returns. A 20% efficiency gain saves $2B, "and you can easily build an ASIC for $2 billion," Casado calculates. Because model weights are fixed (unlike dynamic traditional software), bespoke silicon makes sense. "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."

2. The shift to DC power creates a human capital bottleneck

Rack power is moving from 5-10 kW to 100-250 kW, making AC power unviable. DC power is more efficient but dangerous, and only 2% of US electricians are certified to work with it. Horowitz quips: "AI is taking all the jobs. AI is gonna create a lot of new electricians."

3. 20-30% of top founders have shifted to hardware

Casado estimates hardware deals from top founders jumped from ~5% historically to 20-30% today. "The founder community, which tends to be much smarter than the VC community, has identified this as a very area for innovation." Labs are now signing deals with startups before hardware ships, de-risking the category in a way that didn't exist five years ago.

Editorial Perspective

This is the first platform transition in history where physical constraints become the source of competitive advantage. During the birth and rise of the internet (the dot-com era) and the App Store economy, developer ingenuity ultimately determined winners. This time is different—GPUs sold out through 2028 and memory fabs with three years of backlog create a ceiling that even the best model can't overcome. When Horowitz declared the Mythical Man-Month dead (nine women still can't have a baby in a month, but $3B in cluster spending can buy back two years of technical lag), he meant capital can now purchase time. But the paradox is that the physical infrastructure where that capital flows has become the new scarce resource.

Source: a16z Podcast "The Infrastructure Behind the Machine Age" (August 28, 2026)

https://a16z.simplecast.com/episodes/the-infrastructure-behind-the-machine-age-M3g5NamY

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

Meta ships first complete AI agent, but privacy distrust blocks adoption

Guests

Nilay Patel (Editor-in-Chief, The Verge) × Jake Kastrenakes (The Verge) × David Pierce (The Verge)

Bottom line

Meta beat OpenAI, Google, and Apple to market with a working consumer AI agent, but its reputation for data exploitation creates an insurmountable trust barrier that "phones don't have a light" deflections cannot overcome.

3-Line Summary

• Meta launched Muse, the first functional consumer AI agent, beating competitors to market—but half the Verge editorial team refused to try it, citing distrust of Meta with personal data
• Mark Zuckerberg's defense that Ray-Ban smart glasses have a camera indicator light "and phones don't" ignores that the light is barely visible and social norms on surveillance have moved far ahead of outdated privacy law
• AI agents require 24/7 monitoring to be useful, which is the privacy problem in miniature—Meta is gambling that product utility will overcome user distrust rather than addressing the trust deficit

3 Key Points

1. Muse is technically successful but faces a credibility wall

Muse topped the App Store charts and early downloads outpaced ChatGPT. Nilay credits Meta with building "the thing this industry has been promising to make for quite a long time," but David refused to use it: "I do not feel comfortable with Meta having my information." Half the editorial team declined to try it despite being assigned coverage. Meta hired Signal founder Moxie Marlinspike to build secure virtual machines and repeatedly invokes his name as credibility, but the company's history of data exploitation—using information "to make you even more sad," as David put it—hangs over the product.

2. "Phones don't have a light" is a thought-terminating cliche

In multiple interviews, Zuckerberg defended Ray-Ban smart glasses by repeating "we built the light into it. Phones don't have a light." Nilay called this "a phrase meant to end debate rather than address concerns." Jake noted the light "is not enough"—it's "a little pinpoint" that's "very easy to miss" unless it's dark, contrasting it with Japan's mandatory loud camera shutter sound. The legal framework ("you don't have an expectation of privacy in public") was "written ten million years ago before cameras existed," and social norms have moved ahead—people are rejecting Flock cameras, Ring doorbells, and Meta glasses regardless of legality.

3. Google Books' lock-in is more intense than Apple's ecosystem

Google announced five Google Books laptops launching October 4, 2026, starting at $900 (Acer model). Deep Android phone integration (access phone file system, run phone apps on laptop) is the core feature, but David argued "the lock-in here is actually more intense than the iPhone to the Mac. If you have an iPhone, Google Books are basically useless." Without Android phone integration, you're left with "a Chromebook with a backstop of Android apps" at premium pricing. The target is Windows laptop buyers, but Google may not be ready—no specialized software, developing app ecosystem, and the Play Store now tells users when an app is "better as a website."

Editorial Perspective

Meta got the technical answer right but refuses to address the trust question. For AI agents to be truly useful, they need to see everything in your life—but people choose who they trust with that access. The strategy of bashing through distrust with utility may work in the early days of a platform shift, but when social norms catch up, it may be too late to rebuild credibility.

Source: The Vergecast "Phones don't have lights" (September 25, 2026)

https://www.theverge.com/the-vergecast

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

a16z Launches CoSign to Make Silicon Valley's Reputation Graph Visible and Actionable

Featuring

David Booth (a16z, CoSign product team) × Eric Stromberg (a16z, CoSign product lead) × Olivia Moore (a16z investor, infrastructure team) × Josh Elman (a16z partner, former LinkedIn)

Bottom line

In the AI era, the scarce resource is human endorsement itself, and CoSign is an attempt to transform Silicon Valley's informal endorsement networks into a durable, searchable format.

3-Line Summary

a16z announces CoSign, a new platform to address a fundamental information asymmetry: the most valuable career data—who has conviction in whom—currently lives in private conversations and ephemeral social posts inaccessible to most people. By making endorsements durable, granular, and discoverable, CoSign aims to democratize access to opportunity and enable reputation capital for early talent discovery. The counterintuitive insight: as AI makes everything easier to generate, human signal becomes the bottleneck.

Three Key Points

1. Twitter followers from elite VCs trump Stanford degrees: Eric Stromberg asks, "What's more impressive to you? Someone who went to Stanford or someone who is followed by Patrick Collison, Elon Musk, Marc Andreessen?" Peer-to-peer credentialing from proven operators beats institutional gatekeeping. This is the core thesis: reputation exhaust from social platforms is more valuable than traditional credentials.

2. Information asymmetry is the real problem, not talent scarcity: Olivia Moore states that VCs "work really hard to find that signal" about people through back-channel references, but "someone who is maybe evaluating between five startup job offers" has no access to this information, "especially if you're like an entry level engineer and yet that could be the difference between joining like a complete rocket ship and like a company that isn't doing as well." The problem isn't finding good people—it's that good people can't access the information to make good decisions. Olivia and her twin sister Justine ran a Sunday newsletter aggregating startup job postings for six and a half to seven years, receiving 2,000 responses, giving her firsthand experience with the scale of this asymmetry.

3. AI makes human signal scarce, not abundant: Olivia Moore predicts that with AI agents and browsers, job applications will go from "2,000 now" to "10,000...a month from now" to "100,000." In this environment, "the scarce thing is...the human endorsement or the human conviction or the human knowledge." This is a contrarian view: most assume AI makes everything abundant, but CoSign's thesis is that AI makes human judgment the bottleneck.

Editorial Perspective

The history of LinkedIn's connection graph degrading from "I vouch for this person" to "we met once at a party" illustrates the trade-off between network effect durability and signal quality. CoSign's choice of "positive only" as a product strategy is a clear differentiation—competing on celebration rather than critique. However, how to prevent endorsement inflation remains unresolved, and the migration path from the initial 10,000-person vetted community to a broadly useful platform is not specified.

Source: a16z Podcast "The Reputation Graph of Silicon Valley | Introducing Cosign" (September 25, 2026)

https://a16z.simplecast.com/episodes/the-reputation-graph-of-silicon-valley-introducing-cosign-qLcU00IT

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --