12 Must-Know AI Issues (September 2025): GPT-5, Llama 4, Apple Intelligence, Rubin CPX, EU AI Act, Korea’s AI Basic Act

This post distills the AI latest issues (September 2025) into 12 practical themes: model releases (GPT-5, Llama 4, Gemini 2.x), platform shifts (Apple Intelligence), silicon/infrastructure (NVIDIA Rubin CPX, record data-center spend), and regulation/litigation (EU AI Act timelines and fines, Korea’s AI Basic Act, NYT v. OpenAI). Each topic includes crisp takeaways and concrete next steps for engineering, security, and compliance teams.

Table of Contents

{% toc %}

1) Foundation Models & Products

GPT-5 is live (Aug 7, 2025)

OpenAI introduced GPT-5 and “GPT-5 for developers,” positioning it as its most capable model for coding, agents, and customization. If you’re migrating, audit safety/guardrails, eval drift versus your GPT-4/4o baselines, and capacity/rate-limit changes in your stack.

Meta Llama 4 (Scout/Maverick)

Meta released Llama 4 variants with open weights, long context, and multimodality—useful for sovereign/edge scenarios. Validate license terms and performance against your workloads before replacing closed models.

Google Gemini 2.x line evolves

Google expanded Gemini 2.0 (Flash, Flash-Lite, Pro Experimental) and later 2.5 tiers. If you target cost-sensitive agentic tasks, benchmark Flash/Flash-Lite; for large-context coding and tools, trial 2.0 Pro/2.5 Pro.

Anthropic: Memory for Teams/Enterprise

Claude now supports automatic memory (opt-in) for Team/Enterprise, plus Incognito chats for privacy. Update your data-handling docs and retention tables accordingly.

2) Platforms & UX: Apple Intelligence

Apple Intelligence is rolling out across iOS/iPadOS 18 and macOS Sequoia with ChatGPT integration you can enable for Siri and Writing Tools. Verify device/language support and your org policy on external model calls via the Apple extension.

3) Silicon & Infra: Long-Context Acceleration + Record Build-out

NVIDIA Rubin CPX

NVIDIA unveiled Rubin CPX, a prefill-optimized accelerator targeting massive-context inference and disaggregated serving (prefill/decoding separation). If you host long-context models, revisit your serving topology and capacity planning.

Data centers: record spending (US)

U.S. data-center construction hit a $40B annualized record in June 2025, driven by gen-AI demand. Expect power/land constraints to shape model placement, PPA strategy, and multi-cloud regionalization.

4) Regulation & Governance

EU AI Act: timelines and fines

The EU AI Act is now on a phased application path; penalties reach up to €35M or 7% of global turnover for certain infringements. Create a living AIBoM, risk classification, logging, and human-oversight procedures now—don’t wait for the final compliance date.

Korea’s AI Basic Act

Korea promulgated its AI Basic Act on Jan 21, 2025, with enforcement from Jan 22, 2026 (some provisions differ). Non-Korean providers operating in Korea should prepare for the local representative requirement and institute compliance interfaces.

5) Litigation & Privacy: NYT v. OpenAI

On May 13, 2025, the SDNY ordered OpenAI to preserve and segregate output logs that would otherwise be deleted, with OpenAI contesting scope and implications. Practically, assume longer log retention and update your vendor DPAs, privacy notices, and incident response playbooks.

6) Enterprise Adoption & Ecosystem Moves

Signals suggest diversified model portfolios: e.g., reporting indicates Microsoft is integrating Anthropic models alongside OpenAI within productivity suites—watch multi-vendor governance and latency/cost routing.

One-Page Issue Matrix

Implementation Snippets

A. Minimal AI Bill of Materials (AIBoM) — JSON

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22

{
  "service": "claims-summarization-agent",
  "models": [
    {"name": "gpt-5", "provider": "OpenAI", "purpose": "primary"},
    {"name": "llama-4-maverick", "provider": "Meta", "purpose": "fallback"}
  ],
  "risk_classification": "limited-risk",
  "data": {
    "sources": ["sharepoint", "licensed_news_api"],
    "personal_data": "pseudonymized",
    "retention_days": 30
  },
  "governance": {
    "human_oversight": true,
    "logging": ["prompts", "outputs", "tool_calls"],
    "evals": {"red_team": "quarterly", "bias": "per_release"}
  },
  "compliance": {
    "eu_ai_act": {"provider_disclosures": true, "post_market_monitoring": "enabled"},
    "kr_ai_basic_act": {"local_rep": true, "security_plan": "documented"}
  }
}

B. Simple Python router (cost vs. accuracy)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

from dataclasses import dataclass

@dataclass
class RoutePolicy:
    max_latency_ms: int
    min_accuracy: float
    long_context: bool

def choose_model(task: str, policy: RoutePolicy) -> str:
    if policy.long_context:
        return "gpt-5"  # paired with Rubin CPX/long-context infra upstream
    if "draft" in task or policy.max_latency_ms < 400:
        return "gemini-2.0-flash-lite"
    if policy.min_accuracy >= 0.9:
        return "gpt-5"
    return "llama-4-scout"

# example
model = choose_model("draft product spec", RoutePolicy(300, 0.8, False))
print(model)

Governance Checklist (EU AI Act & Korea AI Basic Act)

• Map use-cases to risk classes; decide if any are “high-risk” or GPAI with extra duties.
• Maintain AIBoM (models, data, evals, constraints, known failure modes).
• Establish provider disclosures, incident reporting, and post-market monitoring.
• For Korea: designate local representative, align with forthcoming decrees, and prepare bilingual notices.

Conclusion

As of September 2025, AI strategy is shaped by four vectors: (1) model maturity and mix (GPT-5, Llama 4, Gemini 2.x), (2) OS-level UX shifts (Apple Intelligence + ChatGPT extension), (3) infrastructure specialization (Rubin CPX) and record data-center build-outs, and (4) regulatory pressure (EU AI Act, Korea’s AI Basic Act) plus litigation-driven privacy realities (NYT v. OpenAI). Treat these not as headlines but as requirements for architecture, governance, and budgets.

Summary

• GPT-5, Llama 4, Gemini 2.x define the 2025 model mix; benchmark before swapping defaults.
• Apple Intelligence moves AI to system UX; set a policy for ChatGPT extension use.
• Rubin CPX and $40B DC spend underscore long-context economics and power constraints.
• EU AI Act (up to 7% turnover fines) and Korea’s AI Basic Act (2026 enforcement) require early prep.
• NYT v. OpenAI preservation order changes retention assumptions; update DPAs and notices.

Recommended Hashtags

#ai #gpt5 #llama4 #gemini #appleintelligence #rubincpx #euAIact #koreaAIlaw #datacenters #aigovernance