MacroVoices #446 Matt Barrie: AI’ll Be Back!

Episode Summary

Executive Summary: Matt Berry argues AI has entered a slowdown in frontier model progress because training and inference are colliding with limits in data, chips, capital, and power. He says the real money will come not from foundational models but from AI applications, system integration, and industry-specific workflows, while fraud, warfare, and energy demand become major side effects of the technology's spread.

Main Topics: Frontier model slowdown and scaling limits (Priority: 5/5): Berry says the pace of breakthrough at the foundation-model layer is slowing as models exhaust easy data, require vastly more compute, and run into physical and economic constraints. Foundational models as weak businesses (Priority: 5/5): He argues OpenAI, Anthropic, and similar firms lack durable moats and clear unit economics; their outputs are increasingly commoditized by open source and fast-moving competitors. AI value shifts to applications and system integration (Priority: 5/5): Berry repeatedly says the biggest returns will come from integrating AI into real businesses—customer support, sales, operations, and software development—rather than from owning the base models. Software development as the next major AI wave (Priority: 4/5): He identifies software engineering as the near-term 'midjourney moment,' where large codebases can be loaded into models to accelerate refactors, feature development, and product iteration. Energy and infrastructure constraints (Priority: 4/5): AI is framed as an energy trade as much as a tech trade: the grid, data centers, and nuclear/hydro power availability will determine how far the industry can scale. Fraud, identity, and the 'dead internet' risk (Priority: 4/5): Berry warns AI is supercharging scams, synthetic identity documents, voice cloning, and fake online activity, making internet authentication and trust harder. AI in warfare and robotics (Priority: 3/5): He highlights drones, robotic dogs, and autonomous systems as the most immediate military killer apps, with commercial spillovers into humanoid robots and other platforms.

Key Arguments: AI's headline breakthroughs come from scaling transformers on massive data and compute, but those gains face hard limits in available training data, chip supply, data-center capacity, and electricity. Foundational-model providers do not yet have sustainable competitive advantages; open source and smaller teams can rapidly match or exceed benchmark performance. The economic value of AI will mostly accrue to firms that embed it into workflows and industry-specific processes, especially via system integrators and freelancers. Charging consumers $20/month for model access is unlikely to support the extreme training and inference costs required for frontier AI development. Software development is the most promising near-term application because AI can ingest large codebases, analyze backlogs, and accelerate migration and refactoring tasks. AI may increase employment in many contexts by creating more work, higher-order tasks, and new support needs rather than simply eliminating jobs. Scams will become much more sophisticated through voice cloning, synthetic video, fake documents, and AI-written outreach, increasing fraud and authentication risks. Power scarcity is becoming a primary bottleneck; the future of AI infrastructure is tied to nuclear, hydro, and broader energy-system expansion.

Data Points: GPT-2 parameters: 1.5 billion - Cited as an early, relatively limited model that scraped about 8 million web pages. GPT-3 parameters: 175 billion - Used as the first major scale-up in model size. GPT-3 training data: 400 billion tokens - Berry said this represented about 3% of Wikipedia, 8% of books, 22% of the web, and 60% of Common Crawl. GPT-4 parameters: 1.8 trillion - Described as the leap that allowed top-decile or top-percentile exam performance. GPT-4 training data: 13 trillion tokens - He said GPT-4 consumed very large portions of the web plus Twitter, Reddit, YouTube, and other sources. GPT-4 training run cost: ~$80 million - Berry gave this as the last estimate he had seen for a GPT-4 training run. Gemini training cost estimate: ~$200 billion - He cited this as an estimated scale of spending for a next-generation model run. NVIDIA quarterly revenue: ~$300 billion a quarter - Berry referenced this as evidence of the extraordinary chip demand tied to AI. NVIDIA revenue concentration: 46% from four customers - He used this to argue AI capex is concentrated in a few hyperscalers. Hyperscaler AI capex: ~$200 billion per year - Berry said four big customers are driving most foundational-model spending. NVIDIA annual revenue from those customers: ~$15 billion per quarter into NVIDIA - He claimed roughly a quarter of that AI spending flows directly to NVIDIA. OpenAI valuation target: $100 billion to $125 billion pre-money - He said OpenAI needed these valuations to raise enough capital for next-generation models. Human brain synapses estimate: 100 trillion to 1,000 trillion - Used as a rough benchmark for comparing AGI scaling narratives to GPT-4's parameter count. Doublings to AGI (his rough estimate): 9 doublings - He argued that matching human-brain-scale capacity from GPT-4 would require about nine exponential jumps, along with nine doublings of energy, compute, and data. Freelancer signups: 25,000 per day - Berry cited this to illustrate how AI-driven onboarding could not be economically done by humans at scale. Freelancer web design share: 30% of work - Used to compare past internet-era demand with likely future AI development demand. Freelancer app development share: 16% of jobs - Compared as the smartphone-era analog to the coming AI-development wave. Survey result: freelancers earning more: 50% - Berry said half of surveyed freelancers reported earning more due to AI productivity gains. Survey result: earnings unchanged: 27% - He said another large group reported no material change. Data center power draw: 300 to 400 MW - He cited a friend building new data centers that require this level of power. GPT-4 relative inference cost: ~3x GPT-3 - He said GPT-4 costs roughly triple GPT-3 to run, exacerbating economics and energy use.

Pivotal Quotes: "I think the real money is going to be made just as it was in the dot-com boom in the transformation of every single industry... into AI-powered applications." — Matt Berry: Explaining why applications and system integration will outperform foundational model investing. "We're starting to run into maths, we're starting to run into economics, and we're starting to run out of bullshit in terms of being able to promote these stratospheric valuations." — Matt Berry: His blunt summary of why the AI hype cycle is cooling. "The next big thing is AI development. So, you'll have just like you've got web development, you've got app development, you'll have AI development." — Matt Berry: Describing the labor market and services opportunity for freelancers and integrators.

Implications: Investors should focus on AI adoption layers, not model glamour. Winners are likely to be integrators, vertical software firms, and infrastructure providers; losers may include overvalued foundation-model businesses and weakly defended consumer subscriptions.