Blackwell, China, and the Future of AI Compute: Why Distributed Strategies Matter
The recent Podchemy conversation with Gavin Baker, highlighted by Patrick O’Shaughnessy’s post, has sparked intense debate about the trajectory of AI compute. Baker’s focus on Nvidia’s Blackwell GPU as a gamechanger for U.S. companies highlights the brute-force scaling model dominating current discourse. But when we zoom out, the picture is more complex — especially when considering China’s ambitions, alternative compute paradigms, and the brittle risks of hyperscaler-only strategies.
🔑 What Baker Emphasized
- Nvidia Blackwell: A leap in GPU architecture, cementing U.S. leadership in AI compute. Baker frames it as central to the scaling laws driving AI progress.
- Performance Gains vs Efficiency: He highlights Blackwell’s performance improvements over Hopper, but the discussion is framed in terms of raw throughput rather than power efficiency. The efficiency dimension — watts per token, sustainability of scaling — is left underexplored.
- SME and HBM Chokepoints: He stresses that semiconductor manufacturing equipment (SME) and high-bandwidth memory (HBM) are critical bottlenecks. Export controls here are decisive in limiting China’s ability to catch up.
- China’s Position: Domestic GPU efforts are advancing but remain behind Nvidia, AMD, and Google TPUs. Without SME and HBM, China faces structural barriers.
- Hyperscaler Economics: Baker warns that SaaS firms risk repeating the mistakes of bricks-and-mortar retailers. Hyperscaler economics are brittle, and challengers can undercut them by deploying AI differently.
- Edge AI as Bear Case: Baker identifies the rise of on-device models (e.g., pruned-down Gemini 5 or Grok 4 running on phones) as the most plausible bear case for explosive demand in centralized compute. Apple’s strategy positions the iPhone as a privacy-safe AI distributor, calling on cloud models only when necessary. If “good enough” models (~115 IQ equivalent) run locally at 30–60 tokens/sec, demand for hyperscaler-scale compute could flatten.
- Scaling Laws vs Usefulness: Baker contrasts the bullish case (scaling laws continuing, enabling breakthroughs like extremely long context windows) with the bear case (edge AI dampening demand). He suggests progress is harder to perceive for non-experts, shifting emphasis from “more intelligence” to “more usefulness.”
🧩 What Baker Did Not Cover
- Alternative Compute Paradigms: He did not discuss thermodynamic, neuromorphic, or photonic approaches — those remain speculative but potentially disruptive.
- Distributed AI Analogy: While Baker covered edge AI, he didn’t frame it as “rooftop solar.” That analogy extends his bear-case argument by highlighting resiliency and sovereignty.
📊 Comparative Table: GPU Market Positions
Category |
Nvidia Blackwell (US) |
China Domestic GPUs |
Alternative Paradigms (Extropic, Neuromorphic, Photonic, Quantum) |
Performance |
Leading-edge, optimized for AI training with HBM |
2–3 generations behind, limited by SME/HBM access |
Extropic efficient for probabilistic AI, Neuromorphic excels at edge, Photonic high throughput, Quantum task-specific |
Efficiency |
Higher throughput vs Hopper, but energy-intensive |
Less efficient, catching up slowly |
Extropic radically efficient, Neuromorphic ~25× GPU efficiency, Photonic low heat, Quantum not yet practical |
Supply Chain |
Dominated by US firms, reliant on SME/HBM |
Vulnerable to export controls, domestic ecosystem still maturing |
Emerging startups, research labs; supply chains not yet mature |
Strategic Risks |
Concentration in hyperscalers, brittle if disrupted |
Geopolitical chokepoints, sanctions |
Early-stage, uncertain scalability, but potential leapfrogging |
Best Use Cases |
Frontier AI model training, hyperscaler clusters |
Domestic AI, sovereign compute |
Extropic: generative AI; Neuromorphic: robotics/edge; Photonic: LLM training; Quantum: optimization |
🧩 PESTLE Risks of Mega AI Data Centers
Relying solely on hyperscaler or even space-based mega centers is brittle across every dimension:
- Political: Geopolitical chokepoints, sanctions, orbital vulnerabilities.
- Economic: Capital intensity, margin erosion, rising energy costs.
- Social: Public backlash over land, water, and inequality.
- Technological: Single points of failure, latency, unresolved space challenges.
- Legal: Data sovereignty, antitrust, liability in orbit.
- Environmental: Gigawatt-scale carbon footprints, water stress, space debris.
A dual-track strategy — mega centers for frontier model training, distributed edge/fog AI for inference and resilience — is far more robust.
📌 Author’s Commentary
Efficiency-First Paradigms: Startups like Extropic.ai and initiatives such as ZSCC.ai are pioneering radically efficient compute models. These could disrupt the brute-force GPU scaling narrative by aligning hardware with probabilistic AI workloads.
Distributed Resiliency: Baker’s bear case (on-device models) aligns with the rooftop solar analogy — local compute reduces hyperscaler dependence, increases sovereignty, and reframes resiliency as both a technical and economic inevitability.
🚀 Conclusion
Baker’s analysis underscores Nvidia’s dominance, the chokepoints that keep China at bay, the brittle economics of hyperscalers, and the bear case for edge AI. But the conversation leaves out critical dimensions: alternative paradigms and distributed resiliency. The hype around Blackwell is justified, yet incomplete. The future of AI compute will not be decided by brute-force scaling alone — it will hinge on different physics, smarter economics, and distributed resilience.
