Bilingual

GLM 5.2 and the coming AI margin collapse
15 days ago
- DeepSeek's R1 model caused a market reaction due to low training costs, but AI economics are more complex.
- Training costs are fixed, while inference costs scale with demand and have high gross margins for frontier labs.
- GLM 5.2 is a strong open-weights competitor to models like Opus and GPT, with near-equivalent quality.
- Drawbacks of GLM 5.2 include slower speed for interactive use, lack of vision support, and poor web search capabilities.
- Migration to open-weights models like GLM 5.2 is easy due to OpenAI/Anthropic-compatible endpoints, with low switching costs.
- Cost savings are significant: GLM 5.2 is priced at ~$4.40/MTok, less than 20% of Opus, and may become even cheaper with optimizations.
- Enterprise concerns about data privacy with providers like Z.ai can be mitigated by using other providers or on-premises hosting.
- Inference margin collapse could reshape the AI industry, benefiting those who leverage lower costs.
The End of Compute Scarcity? Not So Fast
14 days ago
- AI compute scarcity may not be ending; SpaceX (xAI) and Meta renting out capacity likely reflects idiosyncratic issues rather than systemic excess, with high prices and short-term deals indicating continued scarcity.
- The rise of open-source models and end of token maxing may bifurcate AI usage: low-cost models could handle most tokens, while frontier models from labs like Anthropic and OpenAI capture premium revenue from mission-critical work.
- AI spending is highly concentrated; data shows top firms spend significantly more per employee, suggesting frontier tokens will dominate revenue despite a smaller share of total tokens.
- Overall median EV/NTM revenue multiple for SaaS companies is 3.5x, with high-growth firms at 19.7x, mid-growth at 5.3x, and low-growth at 2.8x, reflecting valuation based on growth expectations.
- The author anticipates AI's growth trajectory will outweigh short-term optimizations like token maxing, but notes a risk from the 'own your weights' trend for large AI labs.
The US-China AI arms race has taken an unexpected turn
7 days ago
- DeepSeek's open-source R1 model released in January 2025, which rivals top US AIs and is free, caused a $1 trillion drop in US tech stocks and prompted US lawmakers to propose banning it on government devices.
- Z.ai's GLM-5.2 release in contrast did not cause similar panic, indicating a shift in the US-China AI arms race where China now catches up, similar to other technologies like smartphones and electric cars.
- Chinese AI models are typically open-source and free for local use, while US models are often cloud-based, fee-charging, and closed, leading to differences in accessibility and innovation strategies.
- GLM-5.2 ranks as the most intelligent open-source AI by benchmarks, outperforming GPT-5.5 in software engineering tasks but slightly lagging in intelligence tests, highlighting industry-wide standard testing issues.
- DeepSeek's latest model leads in usage on OpenRouter, with seven of the top 10 LLMs built by Chinese companies, showing growing adoption despite potential security and compliance concerns for large enterprises.
- Experts warn that US restrictions on AI access may backfire by pushing others to develop independent ecosystems, as seen with Huawei, while China's open-source approach pressures Western closed models.
- Business inertia and risk aversion in big companies favor established US tech firms like Microsoft due to compliance and reliability, even if Chinese models are free and capable on laptops.
- Europe risks becoming an 'AI colony' dependent on foreign systems without developing its own AI industry under European legislation, facing a national security threat more serious than the nuclear arms race.
Hugging Face hacked: Blue Team turned to Chinese LLM after US models blocked
a day ago
- Hugging Face's production infrastructure was breached by an autonomous AI agent system, exploiting code-execution paths to gain access and harvest credentials.
- The security team used China's open-source GLM 5.2 model for forensic analysis after commercial frontier models' safety guardrails blocked incident response efforts.
- Hugging Face recommends defenders have a vetted model ready on their own infrastructure to avoid guardrail lockout and protect data from leaving the environment.
- The breach involved lateral movement into internal clusters using short-lived sandboxes, but no tampering with models, datasets, or supply chain was found.
- The incident report coincided with the release of Chinese AI models like Kimi K3, which are cheaper and perform competitively against U.S. models.

Hasty Briefsbeta

#open-source models

GLM 5.2 and the coming AI margin collapse

The End of Compute Scarcity? Not So Fast

The US-China AI arms race has taken an unexpected turn

Hugging Face hacked: Blue Team turned to Chinese LLM after US models blocked