MacroVoices #473 Matt Barrie: AI of The Storm

Episode Summary

Executive Summary: This episode centers on the rapid commoditization of AI, with Matt Berry arguing that foundation models, cloud APIs, and even hardware moats are eroding as open source, China, and custom chips reshape the market. He emphasizes a shift from chat to reasoning to agents, plus rising privacy concerns that could push AI to the edge. The post-game shifts to market technicals, warning that the recent equity bounce may be a failed rally and reviewing hedging, dollar, oil, gold, uranium, and copper setups.

Main Topics: AI commoditization and the collapse of moats (Priority: 5/5): Berry argues that foundation models are no longer durable monopolies: open-source efforts, DeepSeek, Grok, and other entrants have dramatically lowered the barriers to competing with OpenAI-like systems. Progression from chat to reasoning to agents (Priority: 5/5): The discussion frames AI evolution as moving from simple chat interfaces to reasoning systems that break tasks into steps, and then to agents that can perform workflows traditionally done by humans. AI business-model stress and pricing pressure (Priority: 5/5): OpenAI and peers are struggling to monetize expensive models via API calls and subscriptions, with usage caps and timeouts signaling that current pricing may not cover costs. Hardware disruption and edge AI (Priority: 5/5): Chinese chip development, U.S. sanctions backfiring, and vertical integration by cloud giants and customers are challenging NVIDIA’s dominance, while Berry expects more AI to run locally for privacy and latency reasons. Privacy, authenticity, and social disruption (Priority: 4/5): Berry warns that cloud-based AI will make corporations and consumers uneasy about data exposure, driving demand for local/private AI and human verification amid rising deepfake, scam, and synthetic-content risks. Market technicals and hedging strategy (Priority: 4/5): Patrick and Eric interpret the S&P 500 bounce as potentially a dead-cat rally near key moving averages, and they discuss collars and option hedges as volatility declines. Commodity and macro trade setups (Priority: 3/5): The post-game reviews dollar weakness, oil’s headline-driven bounce, gold’s strong trend, uranium’s vulnerable consolidation, and copper’s COMEX/LME divergence likely tied to tariff anticipation.

Key Arguments: Foundation AI is becoming easier to build and harder to monopolize, because smaller teams and open-source projects can now challenge leading Silicon Valley models. DeepSeek’s emergence shows that smart engineering and lower GPU usage can rival frontier models at a fraction of the cost, reducing the advantage of massive training budgets. AI’s next phase is agents that execute real workflows, not just answer questions, which will transform support, accounting, research, and sales. Current AI subscription pricing is likely uneconomic; caps, outages, and quota limits suggest providers are still searching for a sustainable business model. The most durable AI opportunities may be in vertical applications rather than in model vendors, because every industry will need AI-enabled workflow software. AI will increasingly move to the edge because users and companies want privacy, low latency, and local control over sensitive data. NVIDIA remains dominant, but customer-built chips, Chinese hardware, and vertical integration are eroding its moat, making a Cisco-like outcome possible. Consumers and firms may eventually reject cloud-based AI if systems can access too much personal or proprietary information, creating demand for private or on-device AI. Synthetic content, phishing, voice scams, and fake online identity will accelerate as AI lowers the cost of deception and manipulation. In markets, the recent equity rebound may simply be a reflexive rally within a larger correction, so hedging via collars becomes more attractive as volatility falls.

Data Points: Episode number: 473 - Macro Voices episode identifier Production date: March 27, 2025 - Podcast release date SP 500 weekly move: up 65 basis points - Macro scoreboard summary U.S. Dollar Index: 104.66, up 115 basis points - Week-over-week market update WTI crude oil: $69.65, up 410 basis points - Weekly close Arbob gasoline: 224, up 323 basis points - Weekly close Gold: 3022, down 62 basis points - April gold contract close Copper: 524, up 275 basis points - Comex copper close; all-time high noted U.S. 10-year Treasury yield: 4.35% (435 basis points), up 10 basis points - Weekly close Core PCE release: Friday, March 28, 2025 - Key data watch item mentioned Next week macro data: ISM Manufacturing, ISM Services, and U.S. jobs numbers - Upcoming calendar DeepSeek engineering team size: 160 engineers - Berry describes DeepSeek V3/R1 development team DeepSeek GPU usage: about 2,000 GPUs - Rumored training resource level versus larger competitors Typical GPU requirement: about 20,000 GPUs - Berry’s comparison for a standard comparable model run DeepSeek training budget: around $5–6 million - Estimated cost for model training Typical frontier-model training budget: around $100 million - Berry compares with Silicon Valley models OpenAI API pricing: about $30 per million calls for GPT-4 - Illustrative API pricing mentioned ChatGPT Plus: $20/month - Consumer subscription tier ChatGPT Pro: $200/month - Higher-tier consumer plan AI agent marketplace tiers: $2,000/month and $20,000/month - OpenAI’s new agent-oriented pricing mentioned OpenAI reported burn: $7–8 billion last year - Berry’s estimate of operating losses/costs OpenAI revenue estimate 2023: about $1 billion - Estimated annual revenue figure cited OpenAI revenue forecast 2024: about $2.2 billion - Forecast/reported estimate cited Huawei Ascend 910 C performance: about 60% of NVIDIA H100 - Hardware comparison described Huawei chip production scale: hundreds of thousands per year - Berry says China is producing chips en masse NVIDIA customer concentration: four customers generate about 46% of NVIDIA revenue - Mentioned as a key competitive risk S&P futures correction: about 10.5% - Patrick’s recent market drawdown reference Subsequent rally: about 6% - Rebound off the lows described VIX move: from mid/high 20s near 28 to 17–18 - Options volatility and hedging cost discussion Longer-dated VIX: around 20% - Six-month volatility pricing discussion Collar example width: 5% down / 5% up - Illustrative S&P collar structure Gold measured move target: as high as 3,200 - Technical upside target discussed Spot/contract roll effect in gold: about $30 - Contango captured in the contract roll

Pivotal Quotes: "Creating a foundational AI model seems to be akin to opening yet another Thai restaurant, albeit potentially a very good Thai restaurant, in Thailand while at the same time handing out the recipe book and the business plans." — Matt Berry: Berry’s metaphor for how foundation-model competition has become crowded and commoditized "The AI has infinite patience, infinite time and access to all the world information in its knowledge base, you actually see the quality and the empathy of the customer support queries or the sales engagements are an order of magnitude better than what a human would ever do." — Matt Berry: On why AI agents can outperform humans in support and sales workflows "I think every major industry is going to be transformed by this." — Matt Berry: His investment thesis that opportunities lie in vertical applications rather than just model makers

Implications: AI is moving from novelty to utility, but profits may accrue less to model vendors and more to vertical software, edge devices, and private AI infrastructure. Investors should watch commoditization, privacy-driven decentralization, and market hedging as volatility and macro risks remain elevated.