Gutenberg Lab — Morning Paper

AI Revenue Concentrates on 30M Coders—1.5B Knowledge Worker Market Still Untapped

Featuring

Jack Altman (Uncapped host) × David George (General Partner, Andreessen Horowitz)

Bottom line

Current AI success comes from just 30 million coders out of 1.5 billion knowledge workers, representing less than 5% B2B market penetration.

3-Line Summary

Andreessen Horowitz's David George argues we're in an "And" era where frontier models, open source, applications, autonomy, and robotics will all succeed simultaneously. OpenAI, Anthropic, and Cursor collectively generate roughly $12 billion in revenue, but this comes from just 30 million coders, not the 1.5 billion knowledge workers. With data center capacity unavailable until 2028 and structural supply-demand imbalance, George asserts "everything is going to work."

Three Key Points

1. Revenue concentration reveals untapped market: Frontier labs' ~$12 billion revenue comes from just 30 million coders out of 1.5 billion knowledge workers, with a "massive power law" where the top 1 million account for most spending—actual diffusion is under 5%. Legal AI "Harvey" is "probably twelve months behind coding" in diffusion, validating that other knowledge work categories will follow similar adoption curves. Reasoning models unlocked legal because "lawyering is reasoning."

2. Supply constraints guarantee multi-year runway: Data center capacity is unavailable until 2028 due to bottlenecks "in every step of the supply chain." If current 4 gigawatts doubled to 8, "it would all get used" at current or higher prices—this supply-demand imbalance is structural, not temporary. Customer support AI has been "better, faster, cheaper" for two years, yet penetration remains "tiny"—technology readiness precedes market penetration by years.

3. AI uniquely improves with capital injection: Unlike the SaaS era when Vision Fund capital "screwed up" companies, in AI "if you throw more money at it, it can just scale up and get better." Putting $40 billion into Salesforce means hiring sales reps (doesn't scale); putting it into model training makes models "much better." This "very unique in the moment we're in" property means higher valuations (less dilution plus more war chest) become competitive weapons.

Editorial Perspective

AI infrastructure spending has surpassed railroads as percentage of GDP, reaching roughly 3% and likely to 10x. Yet current success represents under 5% penetration of the knowledge worker market. Waymo proves autonomous vehicles are 10-14x safer than human drivers, yet blocking their diffusion is "barbaric" and "history will judge poorly." The combination of technology-adoption time lag, capital's direct product improvement effect, and overlapping waves (AI, autonomy, robotics, bio, defense modernization) signals value creation "far more than" the $25 trillion from the mobile-social-cloud-SaaS era. The question isn't whether individual layers will succeed—it's how large the simultaneous success will be.

Source: Uncapped "David George & Jack Altman on AI, Autonomy, and the Next $25 Trillion" (October 4, 2026)

https://a16z.simplecast.com/episodes/david-george-jack-altman-on-ai-autonomy-and-the-next-25-trillion-bFIBGYtV

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

• --

OpenAI bets on an agent-first internet — eliminating configuration and prioritizing safety over capability

Speakers

Lenny Rachitsky × Tibo Sottiaux (Head of ChatGPT at OpenAI)

Bottom line

The next era of AI is one where humans stop configuring systems and agents handle the majority of internet actions.

3-Line Summary

Tibo Sottiaux, OpenAI's Head of ChatGPT, announced DOTS (a persistent AI agent system) while identifying three shifts the industry is underpricing: models will get 10x better in a year, agents will take the majority of internet actions, and all modalities will integrate seamlessly—yet most builders design for today's capabilities. OpenAI, with 1.2 billion users, spends more compute on safety monitoring than primary agent work.

3 Key Points

1. Agent workflows oscillate with model breakthroughs: Sottiaux himself runs many agents in parallel when pushing the frontier, then consolidates back to a single powerful agent after breakthroughs like Ultrafast that "can keep everything in memory and learn." Current multi-agent architectures may be temporary scaffolding.

2. Loops and graphs are the wrong abstraction: The loops-and-graphs paradigm popular in current AI engineering (LangGraph, CrewAI) will become obsolete. DOTS is "an incredibly smart agent that works 24/7, understands your goals, preferences, and learns from feedback" without manual loop configuration.

3. Safety investment exceeds capability investment: OpenAI spends the majority of its API stack investment on the safety stack—more compute on secondary monitoring (preventing high-risk actions, detecting prompt injection) than primary agent work. With 1.2 billion users, "we don't gamble there."

Editorial Perspective

OpenAI's move to eliminate configuration follows the classic platform maturation pattern. The App Store started with complex developer APIs but matured into "users just open apps." DOTS attempts to shift AI usage from "pick a model, write a prompt" to "agents just work." Yet Sottiaux himself admits "you kinda need a PhD in model pickers"—the transition is incomplete. The notable choice is prioritizing safety investment over capability—treated not as competitive differentiation but as the cost of maintaining trust with 1.2 billion users.

Source: Lenny's Podcast "OpenAI's Head of ChatGPT: We're entering a new era of AI (again) | Tibo Sottiaux" (October 4, 2026)

https://www.lennysnewsletter.com/p/openais-head-of-chatgpt-were-entering

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

• --

Venice AI Processes 250B Daily Tokens as Privacy Backlash Fuels 35-40% Monthly Growth

Featuring: David (Host, Bankless) × John (Head of Strategy, Venice)

Bottom line: As AI privacy scandals hit mainstream media, Venice's structural advantage — "we can only turn over what we have" — is driving 35-40% monthly growth, double the internal forecast, while Frontier Labs face legal subpoenas and enterprise trust erosion.

3-Line Summary

• Venice AI processed 250 billion daily tokens as of September 16, 2026 (up from 100B in June), with actual growth at 35-40% MoM vs. 15-20% forecasted
• OpenAI's alleged use of mathematicians' Codex prompts (Sept 8) and Anthropic's retroactive 30-day data retention policy for enterprise customers (same day) drove VVV token from $18 to $30 peak
• VVV emissions drop to 2M/year (Oct 3, 2026), with Head of Strategy John hinting "2M is still probably higher than we need" for deflationary goal — further cuts likely mid-October

3 Key Points

1. Growth outpacing models by 2x: Internal forecast was 15-20% MoM growth; actual has been 35-40%. John says team "keeps waiting to see if we're gonna hit some sort of growth plateau" but it hasn't happened. Predicts 1 trillion tokens/day by January 2027 (conservative) or December 2026 (exponential curve)

2. Frontier Labs structurally cannot stop data collection: John argues OpenAI and Anthropic are in "prisoner's dilemma" with each other and China — slowing down appears worse than continuing. Federal judge ordered OpenAI to turn over 20M anonymized ChatGPT logs (January 2026 ruling). John: "We can only turn over what we have. There's nothing to turn over if you don't have it"

3. Programmatic burns now exceed discretionary burns for first time: Automatic token burns from subscriptions (Pro $2 / Pro Plus $5) and 5% of API purchases surpassed team-controlled burns in September. John calls this "great sign" and says long-term goal is programmatic dominance — business fundamentals now driving tokenomics

Editorial Perspective

Privacy scandals are not about technical superiority but legal asymmetry. Only those who hold no data can answer "we have nothing to turn over" in court — this structural advantage positions Venice as retail investors' entry point to AI inference demand otherwise locked in Frontier Labs' trillion-dollar IPO queues. However, the bet that "specialized model combinations beat generalized god models" remains unproven — JEV classifier examples are early signals, not validation at scale.

Source: Bankless "$VVV Community Call | September" (October 4, 2026)

http://podcast.banklesshq.com/

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

• --