ax curated

Zhipu open-sourced GLM-5.3 weights last night, available on HuggingFace and ModelScope. Two tech outlets covered this, both pulling from the same IT Home report — so we're looking at a single official source, not independent verification. The positioning is clear: complex coding, defensive cybersecurity, and long-horizon tasks. Zhipu's Z.ai lead Li Zixuan confirmed local deployment, fine-tuning, and commercial use are all allowed. The only catch: companies with over $10B annual revenue need a security review before offering GLM-5.3 as an external model service. That threshold is deliberately high — it basically targets Google, Microsoft, and Amazon-tier cloud providers while leaving startups untouched. On performance, Zhipu cites the Artificial Analysis Intelligence Index, where GLM-5.3 scored 60 — same tier as Claude Fable 5 and GPT-5.6 Sol, tied with Kimi K3 for top open-source model. I'd take this with a grain of salt though: AA is a composite score, and we don't have breakdowns showing how much better it actually is on coding or security specifically versus alternatives. The detail I find most credible: Zhipu ran an extra two weeks of safety evaluation before release because of the cybersecurity capabilities. That's a concrete acknowledgment of what this model can do and a real effort to prevent misuse. What's missing: parameter count, inference cost comparisons, and actual SWE-bench or coding benchmark splits. Wait for the community to run those before drawing conclusions.

I'd take this with a grain of salt since both sources are repackaging Zhipu's own announcement—no third-party benchmarks yet. But the numbers are specific: 320B total params, 18B active, AA Index score of 57, directly matching Claude Opus 4.8. Pricing at 1/40th of Opus and 90% cheaper than their own GLM-5.2. The architecture choice is the interesting part. They're using a hybrid of sparse and linear attention, which cuts KV cache and compute by 3-4x for long-context workloads. The other signal: the entire service runs on a domestic chip cluster, and they claim end-to-end performance improved 3x with per-token costs now matching NVIDIA GPUs. If reproducible, that's a real milestone for domestic silicon in production inference. What's missing: independent benchmarks beyond the AA Index, real-world user feedback, and clarity on whether that 1/40 price is a limited-time discount or the standard rate.

Two sources picked up this podcast episode, which tells me Patel's centralization thesis is hitting a nerve. He lays out specific numbers: Anthropic and OpenAI took about 30% of new compute this year, that jumps to 40-50% next year, and by 2028 they could control most of the world's usable FLOPs. The mechanism is straightforward—Anthropic is generating $50M per megawatt in inference revenue against $10-15M in costs, and all that profit gets funneled back into training. It's a flywheel that's hard for anyone else to match. Both sources framed the story identically around centralization, which makes sense since that's the headline claim from the episode. I'd take the 2028 projection with a grain of salt though. This is a podcast conversation, not a SemiAnalysis research note—Patel is sketching a trajectory, not publishing verified forecasts. The specific market-share numbers for 2028 aren't in the transcript, so the "most of the world's compute" claim is directional rather than pinned to a concrete figure. He also touched on China getting under 10% of new compute and the possibility of AI capex triggering a sovereign debt crisis, but neither outlet led with those angles. If you're using these numbers for anything serious, wait for the written SemiAnalysis piece—podcast estimates tend to be looser than their published research.