2025 in AI, with Nathan Benaich
Broad 2025 AI landscape survey surfaces several themes - sovereign AI, regulation adequacy, open-weight safety - directly relevant to APS strategy and policy work.
Key points
- Annual AI roundup covers 2025 progress in reasoning models, compute infrastructure, regulation, and sovereign AI.
- Sovereign AI discussion is directly relevant to Australian AI strategy debates around compute, data, and talent.
- Podcast format with VC framing limits direct APS applicability - useful for horizon scanning, not operational guidance.
Summary
A year-in-review podcast with Air Street Capital's Nathan Benaich covers major 2025 AI developments including reasoning model advances, DeepSeek's capability gap closure, the Stargate and XAI data centre buildout, and slowing EU AI Act compliance. Key discussion threads include what genuine sovereign AI requires (energy, compute, data, talent, chip design, manufacturing), the limits of open-weight model containment, and China's evolving AI safety posture. Commercial adoption data - 44% of US businesses paying for AI, average contract values rising from $39K to $530K - provides useful benchmarking context. The item is analytical and forward-looking but delivered through a VC investment lens rather than a governance or policy frame.
Implications for Australian agencies
- Monitor Strategy teams may want to monitor the sovereign AI framing - specifically the argument that true sovereignty requires full-stack capability - as it may inform Australian AI strategy discourse.
- Consider Policy teams tracking the EU AI Act could note the claim that only three EU member states were compliant by late 2025, as this may affect how Australia references EU implementation as a model.
Implications are AI-generated. Starting points, not advice.
"2025 in AI, with Nathan Benaich" Source: The Gradient – Substack Published: 22 January 2026 URL: https://thegradientpub.substack.com/p/nathan-benaich-2025 A year-in-review podcast with Air Street Capital's Nathan Benaich covers major 2025 AI developments including reasoning model advances, DeepSeek's capability gap closure, the Stargate and XAI data centre buildout, and slowing EU AI Act compliance. Key discussion threads include what genuine sovereign AI requires (energy, compute, data, talent, chip design, manufacturing), the limits of open-weight model containment, and China's evolving AI safety posture. Commercial adoption data - 44% of US businesses paying for AI, average contract values rising from $39K to $530K - provides useful benchmarking context. The item is analytical and forward-looking but delivered through a VC investment lens rather than a governance or policy frame. Implications for Australian agencies: - [Monitor] Strategy teams may want to monitor the sovereign AI framing - specifically the argument that true sovereignty requires full-stack capability - as it may inform Australian AI strategy discourse. - [Consider] Policy teams tracking the EU AI Act could note the claim that only three EU member states were compliant by late 2025, as this may affect how Australia references EU implementation as a model. Retrieved from SIMS, 18 May 2026.