Apple’s AI ambitions in Europe have hit a regulatory brick wall. The European Commission’s rejection of Apple’s proposed AI framework—citing “insufficient transparency” and “platform lock-in risks”—forces the Cupertino giant to either overhaul its approach or cede ground to Google’s Gemini in the continent’s $1.2 trillion digital economy. This isn’t just about Siri’s future; it’s a battle over who controls the next generation of on-device AI, and the stakes couldn’t be higher.
The Commission’s decision, confirmed by De Standaard and De Morgen, arrives as Apple’s internal AI roadmap—once touted as a “privacy-first” alternative to cloud-based models—faces three existential threats: Google’s Gemini’s dominance in enterprise adoption, the EU’s push for open-source compatibility, and a fragmented Siri rollout that leaves iPhone users with a “dumbed-down” version of Apple’s AI. The timing couldn’t be worse: Apple’s SiriKit 5.0 API, released this week in beta, reveals a model architecture that’s 30% slower in latency tests than Google’s Vertex AI on identical ARM-based hardware.
Why Apple’s EU Rejection Is a Chip War Casualty
The Commission’s frustration stems from two technical realities: Apple’s closed NPU architecture and its refusal to adopt ONNX runtime for third-party models. Unlike Google’s Tensor Processing Units (TPUs), Apple’s Neural Engine in the M-series chips lacks standardized quantization support for models trained outside Apple’s ecosystem. This forces developers to recompile models in Apple’s Core ML format—a process that adds 12–18 hours of preprocessing time per model, according to benchmarks from Ars Technica.
Key technical gap: Apple’s NPU excels at int8 inference (8-bit quantization) but struggles with FP16 mixed-precision training, a requirement for state-of-the-art LLMs like Llama 3. This limits Apple’s ability to deploy >7B-parameter models on-device without significant accuracy loss.
“Apple’s NPU is a marvel for Apple’s own models, but it’s a walled garden for everyone else. The EU isn’t just regulating AI—it’s forcing Apple to choose between control and compatibility.”
The Siri Schism: Why Mac Users Get the Future, iPhone Users Get Left Behind
Apple’s decision to limit Siri’s AI upgrades to macOS is a strategic gamble with regulatory and market risks. The company’s internal documents, leaked to HLN, reveal that Siri’s on-device LLM—codenamed “Project Marzipan”—will ship with only 3.5 billion parameters on iPhones, compared to the 13 billion-parameter model on Macs. This isn’t just a performance difference; it’s a capability gap:
- Mac Siri: Supports end-to-end context windows of 4,096 tokens (enabling multi-turn conversations).
- iPhone Siri: Limited to 512 tokens, forcing Apple to rely on cloud fallback for complex queries—a direct contradiction to its “privacy-first” messaging.
- API Latency: Mac Siri achieves 85ms response time for local inference; iPhone Siri hits 220ms due to thermal throttling on the A17 Pro’s NPU.
The EU’s rejection hinges on this fragmentation. De Morgen reports that Commission officials explicitly cited Apple’s “two-tiered AI access” as a violation of the AI Act’s fairness principle. “If Apple wants to compete in Europe, it can’t treat its customers like second-class citizens,” said a source familiar with the discussions.
Gemini’s Silent Victory: How Google’s Open Approach Wins Developer Hearts
While Apple’s NPU remains unmatched for single-model efficiency, Google’s Gemini has quietly won the developer ecosystem war. Here’s why:
| Metric | Apple (Siri) | Google (Gemini) | EU Preference |
|---|---|---|---|
| Model Compatibility | Core ML-only (closed) | ONNX, TensorFlow, PyTorch (open) | ✅ Open standards |
| Enterprise API Access | Limited to macOS/Enterprise Plan ($20/user/month) | Free tier + $15/user/month for Pro | ✅ Cost transparency |
| Latency (ARM Neoverse N2) | 220ms (iPhone), 85ms (Mac) | 140ms (consistent across devices) | ⚠️ Apple leads on-device, but fragmentation hurts |
| Training Data Ethics | Apple’s dataset curated in-house (limited diversity) | Public + proprietary (broader coverage) | ✅ Bias mitigation |
Why this matters: The EU’s AI Act mandates interoperability for high-risk AI systems. Apple’s refusal to support ONNX or PyTorch directly conflicts with this requirement. Google’s Gemini, by contrast, runs on 90% of EU cloud infrastructure (AWS, GCP, OVH) and supports standardized fine-tuning—a critical factor for EU regulators.
“The EU isn’t just about regulation; it’s about choice. Apple’s approach is a non-starter for developers who need to deploy models across platforms. Gemini’s open API is the future—and Apple’s rejection confirms it.”
The Chip War Escalation: Apple’s NPU vs. Google’s TPU in the EU Market
Apple’s NPU advantage—4.6 TOPS/W efficiency on the M3 Ultra—is being undermined by two factors:
- Regulatory Pressure: The EU’s AI Act requires vendors to disclose energy consumption metrics for AI models. Apple’s NPU’s power efficiency becomes a liability when pitted against Google’s TPU v5e, which achieves 9.3 TOPS/W in mixed-precision workloads.
- Developer Migration: Core ML’s 3% annual adoption growth (per Stack Overflow’s 2025 Developer Survey) is being outpaced by ONNX’s 22% growth in enterprise deployments. “Developers aren’t waiting for Apple,” said Dr. Vasileva. “They’re building for the open ecosystem.”
The EU’s stance aligns with a broader trend: open architectures win in regulated markets. In China, Huawei’s Ascend 910 NPU dominates because it supports MindSpore, an open framework. Apple’s closed approach risks the same fate in Europe.
What Happens Next: Apple’s Three Possible Moves
Apple has three options, each with technical and market trade-offs:
- Option 1: Compromise on the NPU
- Add ONNX runtime support to Core ML.
- Open-source the Neural Engine compiler.
- Risk: Dilutes Apple’s IP moat; could reduce NPU efficiency by 15–20%.
- Option 2: Double Down on Cloud Hybrid
- Shift iPhone Siri to a cloud-first model with local fallback.
- Leverage Apple Silicon Cloud for inference.
- Risk: Undermines “privacy-first” narrative; increases latency and data exposure.
- Option 3: Abandon Europe
- Limit AI features in EU markets to SiriKit’s basic intents.
- Redirect enterprise AI spend to Apple Silicon Cloud.
- Risk: Loses $50B+ in EU revenue; accelerates Gemini adoption.
The most likely outcome? A hybrid approach: Apple will open its NPU to ONNX for enterprise models while keeping its proprietary LLMs (like Siri) locked in. This satisfies regulators without fully ceding control.
The 30-Second Verdict: Who Wins, Who Loses?
Winners:
- Google: Gemini’s open API and Vertex AI gain traction in EU enterprises.
- Open-Source Community: Projects like Hugging Face see increased adoption for ONNX-based models.
- EU Developers: No longer forced to choose between Apple’s ecosystem and open standards.
Losers:
- Apple’s iPhone AI Ambitions: Siri’s fragmentation deepens, ceding ground to Google Assistant.
- Closed Ecosystem Developers: Apps relying on SiriKit face slower innovation cycles.
- Apple’s Brand Perception: “Privacy-first” narrative weakened by cloud-dependent AI on iPhones.
The Big Picture: This isn’t just about AI—it’s about who controls the next decade of computing. The EU’s rejection of Apple’s framework marks the first major regulatory blow against a closed hardware-software stack in the AI era. If Apple loses this battle, the dominoes could topple for Apple Silicon’s dominance in enterprise and consumer markets alike.
Actionable Takeaway: For developers, now is the time to future-proof for ONNX and test Gemini’s API. For enterprises, the EU’s stance signals that interoperability is no longer optional. Apple’s next move will determine whether it remains a leader—or becomes a relic of the walled-garden era.
