How Many Will Actually Get Built & Is Energy AI's BIGGEST Bottleneck? | Positron AI Co-founder

Market & Valuation Facts

• Positron AI raised $875 million in Series C funding at a $5 billion valuation.
• Thomas Somers (Positron co-founder/chairman) is building a full-stack semiconductor hardware company, from chips to rack-scale systems for AI inference.
• Current AI infrastructure is heavily shifted from "compute-bound" (training) to "memory-bound" (inference) due to the auto-regressive nature of token generation.

Hardware & Technical Evolution

• Between 2014 and 2024, GPU floating-point operations (flops) improved by 120x, while memory bandwidth improved by only 17x, creating a "memory wall."
• SRAM architectural primitives have not scaled with Moore's Law for 15–40 years, unlike the compute logic surrounding them.
• Transformers introduced a paradigm where compute requirements grew quadratically with sequence length, making KV (Key-Value) caching essential.
• KV caching stores previous token computations to avoid redundant work, but managing these caches across millions of users creates massive memory pressure (e.g., 50 users on a 10T parameter model can exceed model weights in cache size).

Economics & Profitability

• Caching tokens (reusing inputs) costs roughly 1/1000th of generating new tokens, leading to extreme margins for providers; e.g., Anthropic reportedly achieves ~80% gross margin on API business.
• The "cost per token" metric is becoming less relevant than "cost per useful result" as model efficiency increases.
• The "Silicon Data token price index" dropped from ~$60 per million tokens (5 years ago) to <$1 today, but Somers argues the value of a modern token is 100x–1000x higher due to capability gains.
• Future pricing models may shift from per-token to flat-rate annual subscriptions for autonomous AI agents.

Regulatory & Geopolitical Stance

• Somers opposes the "pacing the frontier" movement, viewing it as a potential tool for centralizing technology and stifling liberal freedom ("road to serfdom").
• He characterizes the cross-spectrum "anti-data center" sentiment in the West as a Chinese psychological operation (psyop) designed to hinder Western development.
• Data centers consume significantly less water than commonly cited; an In-N-Out restaurant uses more water than the largest US data centers.
• Somers argues that export controls and pacing are ineffective because China can outpace the West by leveraging unrestricted domestic policy and open-source ecosystems.
• He fears that regulatory caps will allow a small group of entities to control technology, creating a new feudal class structure.

Model Architecture & Trends

• Frontier models (e.g., GPT-6 "Astra") continue to scale, demonstrating "magic" breakthroughs in coding, computer use (e.g., Blender, interior design), and chip design flows (RTL to GDS) previously impossible.
• While enterprises are building smaller, on-prem models, Somers predicts 80–85% of tokens will still be consumed by top frontier models in the cloud.
• Local on-device models will likely increase total token consumption by triggering more autonomous requests to cloud models for complex tasks.
• Context window effectiveness is improving; GPT-6 Astra achieves >95% accuracy on "needle in a haystack" tests, compared to ~70% for GPT-5.
• Chinese labs (e.g., DeepSeek) have led innovation in attention mechanisms (e.g., Multi-Head Latent Attention) to reduce context costs, though US labs have not widely adopted these due to potential capability trade-offs.

Energy & Infrastructure

• Energy remains the ultimate bottleneck for human progress, but current limitations are more economic (debt capacity, infrastructure build-out) than physical.
• Somers rejects space-based data centers as a near-term solution, favoring ocean-based pumped hydro (partnering with Panthelossa) and utilizing US desert lands.
• Data center buildouts are currently constrained by regulatory hurdles and misinformation regarding electricity and water usage; Somers believes these will be overcome as economic incentives (lowering grid prices) become clearer.

Future Outlook & Risks

• Somers views the "Terminator" AI risk as low probability but considers centralization of power and regulatory paralysis the primary existential threats.
• He is bullish on the continued scaling of model size (trillions to hundreds of trillions of parameters) based on observed scaling laws, despite a lack of mathematical proof.
• The next major growth phase in token volume will occur when users trust local LLMs to autonomously delegate complex tasks to cloud-based frontier models.
• Vertical integration in the data economy (e.g., frontier labs acquiring data labeling firms like Scale AI) may eventually happen, but currently, specialized data firms like Mercor and Surge are viable ($200B+ potential).