Investors should pivot focus from AI training to AI Inference, as specialized ASIC chips are projected to offer 10x better performance and efficiency than general-purpose GPUs. While NVIDIA (NVDA) dominates current infrastructure, its general-purpose architecture faces growing competition from specialized startups like Etched that prioritize "tokens per watt" for massive agentic clusters. To hedge against supply chain bottlenecks, look for opportunities in the TSMC (TSM) 4nm and 5nm ecosystems, which offer a less crowded alternative to the high-demand 3nm nodes used by major incumbents. High-conviction themes include the "supply chain of the token," specifically targeting companies involved in High Bandwidth Memory (HBM) and power-efficient data center infrastructure. By 2027, the shift toward millions of concurrent AI agents suggests that the most valuable companies will be those that can minimize "wall clock time" for complex knowledge work.

Etched (Private Company)

• Etched is a semiconductor startup founded by two Harvard dropouts, Gavin Uberti and Robert Wachen, focused on building specialized chips for AI Inference. • The company recently raised $120M in Series A funding (part of $800M total mentioned in the title) to compete directly with NVIDIA. • Technical Bet: Unlike GPUs (which are general-purpose), Etched is building "ASICs" (Application-Specific Integrated Circuits) specifically designed to run Transformer models (the architecture behind ChatGPT). • Key Innovations: * Low Voltage Inference: They run chips at under half the voltage of traditional GPUs, significantly reducing heat and power consumption. * Cluster-Scale Memory: A custom interconnect stack that allows thousands of chips to talk to each other with 5x lower latency than current industry standards. * Vertical Integration: They are building the entire stack—the chip, the circuit board, the cooling system, and the server rack.

Takeaways

• Inference is the Future: The founders argue that while "Training" (building models) has been the focus, "Inference" (running models for users) will be the largest market in the world. • Performance Leap: Etched claims their specialized design can offer 10x better performance than general-purpose GPUs like NVIDIA’s Blackwell for specific AI workloads. • Efficiency over Flexibility: By removing the ability to do "everything" (like graphics or legacy math), they maximize the "Model Flops Utilization" (MFU), getting more "cents on the dollar" out of the hardware.

NVIDIA (NVDA)

• The transcript positions NVIDIA as the "incumbent" that Etched is trying to disrupt. • The "NVIDIA Tax": Large companies (Google, Meta, Microsoft) are trying to build their own chips to avoid the high costs of buying NVIDIA GPUs. • Design Constraints: NVIDIA chips are described as being built on "buffer"—designed to be general-purpose for everything from data centers to IoT, which creates inefficiencies for specific AI tasks.

Takeaways

• Vulnerability in Inference: The discussion suggests that NVIDIA’s dominance in training may not automatically translate to inference, where specialized chips (ASICs) can provide better "tokens per watt" and "tokens per dollar." • Latency Bottlenecks: Current NVIDIA architectures (like Blackwell) are noted to have point-to-point latencies of ~4,000 nanoseconds, which Etched views as a major bottleneck for massive AI agent clusters.

The "Token" Economy & AI Infrastructure

• The founders believe we are moving toward a world where the most valuable companies are those that produce the most "tokens" (units of AI thought/output). • Economic Shift: Productivity will soon be measured in "agents per megawatt" rather than "GDP per capita." • Supply Shortage: There is a multi-year, possibly multi-decade, supply shortage of AI tokens.

Takeaways

• Investment Theme: Look for companies that own the "supply chain of the token"—from power and data centers to specialized silicon. • Wall Clock Time: The next frontier is shrinking "wall clock time" (e.g., turning a 6-month scientific research task into a 1-month task via faster inference). • Trillion-Dollar Data Centers: The founders predict the emergence of individual data centers costing $100 billion to $1 trillion as economies of scale continue to grow.

Semiconductor Supply Chain

• TSMC (TSM): Highlighted for its exceptional customer service and willingness to partner with small, innovative firms to improve yields. • HBM (High Bandwidth Memory): Identified as a critical and scarce component. Etched intentionally uses a different supply chain (4nm process) than NVIDIA’s latest chips to avoid zero-sum competition for the same factory lines. • Power Availability: The primary bottleneck for scaling AI is no longer just the chips, but the availability of electricity and the efficiency (tokens per watt) of the hardware.

Takeaways

• Diversification: Investors should note that not all AI chips compete for the same "leading edge" nodes; there is value in the 4nm and 5nm supply chains as alternatives to the crowded 3nm space. • Power as a Moat: Companies that can eke out more "intelligence" from a single megawatt of power will have a massive competitive advantage as data center space becomes more constrained.

AI Agents & Knowledge Work

• 2027 Prediction: The founders predict that by 2027, there will be more AI agents doing knowledge work than humans. • Agentic Teams: Future AI will not just be one chatbot, but "teams" of millions of agents working concurrently on massive projects (like building a rocket or a complex software system).

Takeaways

• Software Shift: The "COGS" (Cost of Goods Sold) for software companies is no longer zero; it is now a function of inference costs. • Productivity Explosion: The ability to run "year-long" compute tasks in days will lead to the fastest proliferation of technology in human history.