Dylan Patel breaks down OpenAI's compute ceiling, China's robotics dominance, and AMD's uphill battle

Summary

OpenAI's Compute Ceiling and the Stargate Dependency

OpenAI's next training cycle is constrained by a hard physical limit. The Orion / GPT-4.5 run was so large it consumed the entirety of available cluster capacity, meaning no larger pre-training run is possible until Stargate comes online in Texas later this year. In the interim, OpenAI is extracting value through architectural efficiency rather than raw scale — GPT-4.1 was framed as an efficiency gain over GPT-4, and a further step-up is expected before Stargate is operational.

The action has shifted decisively to reinforcement learning. Models like o3 and the forthcoming o4, along with deep research and multi-agent systems, are where OpenAI is concentrating compute spend. RL scaling is decoupled from model size — the pre-trained weights are fixed, but the compute applied on top continues to grow.

The geopolitics around compute access are unambiguous. Amazon will not sell compute to OpenAI. Google will not. The OpenAI-Microsoft relationship has fractured to the point where Stargate moved from a planned Microsoft facility in Wisconsin to an independent build in Texas. The Windsurf acquisition is cited as a live example of partners cutting access mid-deal.

DeepSeek's Transferable Innovations

OpenAI had already implemented FP8 training by at least 2023, predating DeepSeek's public release. Where DeepSeek genuinely advanced the field was in mixture-of-experts sparsity. GPT-4 ran a 1:8 expert activation ratio; GPT-4 later moved to approximately 1:32; DeepSeek pushed to 1:64. The open publication of its reinforcement learning verifiable reward paradigm, efficient inference systems, and novel attention mechanisms are the more durable contributions that Western labs can absorb.

DeepSeek's consumer chatbot experience is, however, deliberately slow — approximately 20 tokens per second versus 100–200 tokens per second on OpenAI, Anthropic, and Google products. The throttle is a conscious batching decision to maximise user concurrency on constrained compute. As a result, Chinese users frequently route around domestic products, accessing Anthropic and other Western models through OpenRouter and similar services.

Anthropic's Strategic Gaps

Anthropic is described as too focused on AGI-path research to prioritise consumer surface area. The company does not compete meaningfully in voice or image generation. Its current litigation with Reddit over training data scraping is framed as a rational financial decision — OpenAI and Google both pay approximately $60–100 million per year for Reddit data access, representing roughly 20% of Reddit's total revenue. By litigating rather than licensing, Anthropic delays that cost by years and may ultimately settle for less. The calculus: $70 million annually buys roughly 70 senior researchers or thousands of GPUs.

XAI and the Elon Attention Deficit

A Tesla manufacturing employee went nine months without a meeting with Musk during his Washington period — an anecdotal but concrete signal of operational neglect across Tesla, Neuralink, and to a lesser extent XAI. The reversion of Musk's attention back to his companies is net positive for execution velocity, though XAI's government-facing revenue line may soften as political proximity to the administration fades.

XAI's near-term revenue base is primarily X platform subscriptions. The path to enterprise-scale revenue through developer tooling is underdeveloped — Grok is not materially integrated into Cursor or Windsurf, the dominant coding tools. Cursor recently crossed $500 million ARR; Windsurf was acquired at a significant multiple; even GitHub Copilot reportedly runs below $500 million in annual run rate despite Microsoft's distribution. Code and software agents are where Anthropic and OpenAI are concentrating commercial bets over the next six months.

The XAI-X merger is structurally defensive. X carried substantial debt with insufficient earnings to service interest payments. The merger provides XAI a proprietary, real-time data source that no competitor can replicate at scale — positioned as materially faster and fresher than any alternative training corpus.

China's Robotics Lock-In

China's dominance in robotics manufacturing is not a near-term risk — it is an established structural reality. Five years ago, China produced roughly as many robots as Germany, South Korea, Japan, and the US combined. Today, China produces more than all of them combined, at dramatically lower unit cost.

Unitree is producing comparable hardware to Western competitors at approximately $10,000 per unit versus $75,000–$100,000 for Western equivalents. Unitree is actively shipping wheeled robot dogs and previewed a wheeled humanoid at a Singapore conference. The cost differential compounds with supply chain depth — the US lacks domestic manufacturing capacity for the motors and actuators that underpin robotics, and building that capacity domestically costs an estimated 10x the Chinese equivalent.

The policy contrast is stark. China's industrial subsidies operate at the company and output level — tax holidays on fab profits, land subsidies, per-unit manufacturing incentives. The US CHIPS Act is characterised as structurally less efficient. A specific example: China's national railway company, redirecting operating profits into semiconductor fabs under tax-free incentive structures, is now the third-largest power chip manufacturer in China and a top-10 global player, with a projected path to top-five globally within a few years.

American robotics startups are largely described as purchasing Chinese components and repackaging them domestically rather than building genuine supply chain depth.

Sovereign AI's Consumer Problem

National AI efforts outside the US and China face a structural distribution failure. Mistral's Le Chat is not meaningfully adopted in France despite government backing. Countries that invest in sovereign model infrastructure but stop short of banning ChatGPT or Gemini find their populations defaulting to American products. Google models dominate in India; Perplexity has outsized penetration in Indonesia. Building the model is not sufficient — without controlling the application layer or restricting foreign access, sovereign AI investment does not translate into sovereign AI usage.

Agent Timelines

The transition to autonomous agents is not a discrete event but a continuous curve. Human-model interaction windows have expanded from seconds to minutes — Claude's deep research runs exceed 30 minutes — and will continue extending. The commercial unlock is not a categorical shift to full autonomy but a gradual reduction in required human checkpoints per task. Multi-agent systems operating with minimal human oversight remain the stated destination; the current trajectory is incremental compression of that interaction requirement.