Eric Schmidt & Andrew Feldman: The Race to SuperIntelligence

Cerebras Systems Market Positioning

• Andrew Gerrandt (CEO/Founder) asserts Cerebras is the fastest AI compute provider by a significant margin ("nearly two orders of magnitude"), surpassing competitors like NVIDIA.
• The company's WSE (Wafer Scale Engine) chip is approximately 56 times larger than the largest chip previously built by the industry.
• This architecture is designed to address the unique data movement patterns of AI workloads, which involve vast amounts of simple calculations and intermediate results requiring frequent memory access.
• Cerebras has successfully moved to the "inference" phase of the AI lifecycle, where society utilizes trained models, benefiting directly from the trend of AI becoming commercially useful.

Hardware and Software Co-Design Strategy

• Cerebras adopted a "clean sheet" strategy in 2016, focusing on accelerating sparse linear algebra (matrix multiplications) rather than embedding specific architectures like 3x3 convolutions which might have become obsolete.
• The architecture predates the 2017 Transformer paper; Gerrandt confirms this strategic foresight allowed them to remain dominant despite architectural shifts.
• Co-design involves integrating system-level IO, prompt caching, and new tools alongside chip architecture, with leading AI researchers actively involved in defining what to build next.
• Cost per token served for AI answers is improving by a factor of ten annually, driven by hardware efficiency, better algorithms, and increased context window management.
• This rate of improvement compresses what historically took 70–80 years in the automotive engine sector into roughly five years in AI platform industrialization.

Industry Trends and Economic Shifts

• The traditional allocation split of 80% training and 20% inference is now considered "completely wrong," with the consensus shifting toward inference dominance as AI scales.
• Inference demand is exploding due to three factors: increasing user base, higher frequency of use per user, and the computational intensity of reinforcement learning (RL) and agentic reasoning.
• Paul Graham, Sam Altman, and Elon Musk have publicly acknowledged the economic penalty for latency; if AI systems are slow, users will abandon them for faster competitors.
• The industry is currently in the "explosion" phase of an S-curve, where revenue is high and consolidation has not yet occurred; the asymptote (where scaling stops) is unknown.
• Traditional SaaS metrics (e.g., user growth rates) are becoming inadequate as the industry invents new categories where "growth" rules no longer apply.

Open Source, Geopolitics, and Data

• A geopolitical "AI battle" is emerging between the West and China, where China is strategically deploying open-source models in Africa and Central Asia to gain market share.
• Open-source ecosystems continue to benefit startups by allowing them to compete against capital-heavy incumbents; the rise of DeepSeek has reportedly pressured OpenAI to announce open-source models.
• Open models can reduce inference costs to $1.50–$2.00 per million tokens, compared to ~$100 per million tokens for closed-source models, driving higher adoption and creativity.
• Critical data silos exist in sectors like biology (e.g., Mayo Clinic's 30 years of data) and government (e.g., UK MRI records), which are not currently accessible for broad model training unlike internet traffic exhaust.

Alignment, Safety, and Sovereignty

• Alignment (ensuring AI reflects human values) remains unsolved; current methods rely on smart humans running partial tests on raw models rather than intrinsic safety guarantees.
• A specific risk involves self-improving systems developing capabilities (e.g., cyber or biological attacks) and hiding them from detection or disclosure.
• The U.S. Trump administration has reportedly shifted focus from "safety" to "security" regarding Chinese threats, while the UK and France are establishing AI safety institutes.
• Sovereign AI capabilities are technically feasible for nations like those in Europe or the UAE, provided there is political will to prioritize the necessary infrastructure and education pipelines.
• Electricity cost and availability are identified as the primary constraints for AGI infrastructure; nuclear power is described as the closest path to powering AGI at scale.
• A single large data center requires approximately 2 gigawatts of power (comparable to a small city); the U.S. is estimated to need 92 new nuclear power plants within the next decade to meet AI energy demands, a feat not seen in the last 20 years of deployment.

Future Consumer Integration

• The transition to on-device AI (running LLMs on smartphones without connectivity) is expected to happen invisibly, similar to how smartphones replaced wallets and car keys over time.
• When AI becomes a seamless part of daily life, users will likely find it impossible to revert to pre-AI lifestyles, even if they were to regret the transition.