Technical Deep Dive
The revelation that Apple used Google's TPU v4 and v5p clusters to train its foundational models is a technical bombshell. Apple's own M-series chips, while excellent for inference on device, lack the massive parallel floating-point throughput required for training large language models (LLMs). TPUs, by contrast, are custom ASICs designed specifically for matrix operations — the core of transformer-based models. Google's TPU v5p, announced in late 2023, delivers over 400 teraflops per chip and can be clustered into pods of 8,960 chips, offering a total of 3.6 exaflops of bfloat16 performance. Apple likely used this pod architecture to train models with hundreds of billions of parameters.
From an engineering perspective, the choice of TPUs over Nvidia's H100 or B200 GPUs is telling. TPUs offer lower precision (bfloat16 vs. FP8 on H100) but superior scaling efficiency for large clusters due to Google's proprietary interconnect (ICI). Nvidia's NVLink and InfiniBand are also excellent, but Google's internal network is optimized for its own workload. Apple's decision to use TPUs suggests that raw performance per chip was less important than cluster-level throughput and cost efficiency. A recent paper from Google Research showed that TPU v5p achieves 95% scaling efficiency at 8,960 chips, compared to ~85% for Nvidia H100 clusters at similar scale.
| Training Infrastructure | Peak FLOPs (bfloat16) | Cluster Size | Scaling Efficiency | Estimated Cost/Hour |
|---|---|---|---|---|
| Google TPU v5p Pod | 3.6 exaflops | 8,960 chips | 95% | ~$1.2M |
| Nvidia H100 DGX SuperPOD | 1.8 exaflops | 4,096 GPUs | 85% | ~$1.5M |
| Apple M3 Ultra (theoretical) | 0.2 exaflops | 1,024 chips | 70% | ~$0.8M |
Data Takeaway: TPU v5p offers 2x the peak FLOPs at 20% lower cost than an equivalent H100 cluster, with superior scaling efficiency. This explains Apple's choice — but it also creates a dangerous dependency on a direct competitor.
For developers, the open-source ecosystem around TPUs is less mature than Nvidia's CUDA. However, Google's JAX framework (GitHub: google/jax, 30k+ stars) provides a Python-first approach to TPU programming, with automatic differentiation and XLA compilation. Apple could have used JAX to port its PyTorch-based training pipeline. The key GitHub repository to watch is `google/jax` for TPU-native training, and `apple/ml-aim` (Apple's own open-source model library, 8k+ stars) for inference optimizations.
Key Players & Case Studies
Apple is the central player here, but its strategy is paradoxical. On one hand, it is the world's most valuable hardware company, designing its own CPUs, GPUs, and even modems. On the other hand, it cannot afford to build its own training infrastructure at scale. Apple's total R&D spend in 2024 was $30 billion, but building a TPU-class cluster would cost $5-10 billion upfront. By renting Google Cloud TPUs, Apple avoids capital expenditure but cedes strategic control. This mirrors Apple's earlier decision to use Intel modems before acquiring Intel's modem business — a pattern of dependency followed by vertical integration.
Google is the silent winner. By providing TPU capacity to Apple, Google gains three advantages: (1) revenue from cloud services, (2) insights into Apple's model architecture through usage patterns, and (3) leverage in future negotiations. Google's own Gemini models are trained on TPUs, and now Apple's models are too. This creates a fascinating dynamic where both companies are simultaneously competitors (in mobile, search, and AI assistants) and partners (in cloud infrastructure).
Nvidia is the potential loser. If Apple — the world's most valuable company — chooses TPUs over Nvidia GPUs, it signals that Nvidia's stranglehold on AI training may be loosening. However, Nvidia's H100 and B200 remain dominant for inference, where Apple's on-device chips excel. Jensen Huang's angry remarks about token burning are a direct response to this threat. He needs every GPU to be fully utilized to justify Nvidia's $2 trillion valuation.
Anthropic has taken a different approach, focusing on values alignment through its hiring process. The company now places 'values interviews' at the core of recruitment, screening for candidates who prioritize safety over raw capability. This is a direct counterpoint to the 'burn tokens' mentality. Anthropic's co-founder Dario Amodei has stated that 'the most dangerous AI is one trained without ethical constraints.' The company's Claude 3.5 model, while competitive with GPT-4o, was trained on a fraction of the compute — 10^24 FLOPs vs. 10^25 for GPT-4 — suggesting that values-driven training can be compute-efficient.
| Company | Training Infrastructure | Model Size (params) | Training Compute (FLOPs) | Values Screening |
|---|---|---|---|---|
| Apple | Google TPU v5p | ~200B (est.) | 5 x 10^24 | No public info |
| Google | TPU v5p | Gemini 1.5: 1.5T | 2 x 10^25 | Limited |
| Anthropic | Nvidia H100 | Claude 3.5: ~200B | 1 x 10^24 | Yes, mandatory |
| OpenAI | Nvidia H100/B200 | GPT-4o: ~200B | 1 x 10^25 | Limited |
Data Takeaway: Anthropic uses 10x less compute than OpenAI for comparable model quality, suggesting that values-driven training and architectural efficiency can reduce hardware dependency. This is a critical lesson for Apple.
Industry Impact & Market Dynamics
The AI market is shifting from a single-player (ChatGPT) to a 'three-legged stool' of ChatGPT, Gemini, and Claude. As of Q2 2025, ChatGPT's consumer AI traffic share has fallen from 60% to 38%, while Gemini and Claude have grown to 28% and 22% respectively. This fragmentation means that no single model will dominate, and the infrastructure race becomes even more critical.
| AI Assistant | Consumer Traffic Share (Q2 2025) | Monthly Active Users (MAU) | Training Cost (est.) |
|---|---|---|---|
| ChatGPT | 38% | 180M | $500M/year |
| Gemini | 28% | 120M | $400M/year |
| Claude | 22% | 90M | $100M/year |
| Others | 12% | 50M | $50M/year |
Data Takeaway: The market is fragmenting, but total compute spending is rising. The 'winner' may not be the best model, but the one with the most efficient infrastructure.
Meanwhile, China's EV battery retirement wave presents a parallel infrastructure challenge. MIIT estimates that by 2027, China will retire 500,000 tons of lithium-ion batteries annually. These batteries contain critical minerals like lithium, cobalt, and nickel. If recycled properly, they could supply 30% of China's lithium demand by 2030. If not, they create an environmental crisis. Companies like CATL and BYD are investing in recycling plants, but the technology is immature. Hydrometallurgical recycling recovers 95% of materials but costs $10/kg, while pyrometallurgical recycling is cheaper ($5/kg) but recovers only 60%. The AI connection? Training large models requires massive energy storage for data centers, and retired EV batteries could be repurposed as grid-scale storage for AI clusters. This creates a circular economy: EV batteries → data center backup → recycling → new batteries.
Risks, Limitations & Open Questions
1. Hardware Lock-In: Apple's reliance on Google TPUs creates a single point of failure. If Google raises prices or restricts access, Apple's AI roadmap could stall. The open question: will Apple build its own TPU-class chip? The company has the talent and capital, but it would take 3-5 years.
2. Token Burning vs. Efficiency: Jensen Huang's demand to 'burn tokens' could lead to wasteful training — models trained on redundant data just to keep GPUs busy. This is economically and environmentally unsustainable. The AI industry needs to adopt 'compute efficiency' metrics alongside accuracy benchmarks.
3. Battery Recycling Scalability: China's battery recycling capacity is currently 200,000 tons/year, but the retirement wave will hit 500,000 tons by 2027. The gap will lead to illegal dumping or export of hazardous waste. The AI industry's energy demands will exacerbate this problem.
4. Values Alignment at Scale: Anthropic's values interviews are laudable, but can they scale to thousands of hires? And can values survive the pressure to 'burn tokens'? The tension between safety and speed is unresolved.
AINews Verdict & Predictions
Prediction 1: Apple will acquire a cloud AI company within 18 months. The dependency on Google is too risky. Apple will likely buy a smaller cloud provider with GPU/TPU capacity (e.g., CoreWeave or Lambda Labs) to bring training in-house.
Prediction 2: Nvidia will introduce 'token efficiency' metrics by 2026. Jensen Huang's 'burn tokens' rhetoric is unsustainable. Nvidia will pivot to selling 'compute per watt' and 'tokens per dollar' as key metrics, aligning with the industry's need for efficiency.
Prediction 3: China will mandate 50% recycled content in new EV batteries by 2028. The MIIT announcement is a prelude to regulation. This will create a $10 billion recycling industry and set a global precedent.
Prediction 4: Anthropic's values-first approach will become the industry standard. As AI models become commoditized, trust and safety will be the differentiator. Anthropic's market share will grow to 30% by 2027, forcing OpenAI and Google to adopt similar hiring practices.
Final Verdict: The AI industry is entering an era of 'infrastructure trilemma' — balancing compute cost, energy consumption, and environmental impact. Apple's TPU dependency, Nvidia's token-burning imperative, and China's battery crisis are all symptoms of the same problem: we are building smarter models on dumber infrastructure. The winners will be those who solve this trilemma, not those who build the biggest models.