*Source commit: 2c47ab1*
If AI is a rocket, compute is the fuel. Without fuel, even the most advanced engine is scrap metal. In spring 2026, the battle over that fuel is entering a white-hot second half.
Anthropic's Big Bet: Locking in a Decade of TPUs
Anthropic announced agreements with Google and Broadcom to secure multi-gigawatt capacity of next-generation TPUs starting in 2027.
A gigawatt equals one billion watts; multi-gigawatt means tens of billions of watts of compute—enough to train and serve the next generation of Claude frontier models. Just as important is the timing: this is not buying spot capacity, it's pre-booking the future. Like a farmer reserving the autumn harvest in spring, Anthropic is locking in compute supply for 2027 and beyond.
Why? Look at OpenAI's situation: by 2028, frontier labs may spend over $100 billion on compute. This isn't buying computers—it's buying dedicated power plants. When compute is the bottleneck, whoever owns compute owns the future.
Also disclosed: Anthropic's annualized revenue has surpassed $30 billion. This is no longer a "burn money for the future" research game—it's a functioning commercial empire whose revenue can sustain long-term compute investment and convince investors this is no bubble.
Huawei's Counterattack: Ascend 950PR and the Chinese Chip
DeepSeek V4's plan is interesting: run natively on Huawei's Ascend 950PR, with upper layers compatible with NVIDIA programming interfaces.
It's a clever strategy. Imagine a chef used to a particular set of knives who is suddenly told they can no longer buy them. Relearning a whole new cutting technique would be too costly. DeepSeek's choice: swap the knife, keep the grip and cutting motion. The Ascend 950PR's underlying silicon is Huawei's, but the upper interface is CUDA-compatible—drastically lowering migration costs from NVIDIA.
Alibaba, ByteDance, and Tencent have placed large orders, driving chip prices up about 20%. The 950PR outperforms the H20 but falls short of the H200—not top-tier, but good enough.
More importantly, it proves China's domestic AI compute stack can close the loop. Even if some memory chips still need importing, core compute is no longer under others' control. It's a loud slap at US export controls: the more you blockade, the more we build ourselves—cheaper and easier to use.
Blackwell Arrives: Cursor's Secret Weapon
The Cursor team implemented "warp decode" for their Composer MoE model on Blackwell GPUs, boosting token generation speed by roughly 1.84x.
Think of a factory where each worker handles one step sequentially. Warp decode is like having a team (a warp) collaborate simultaneously, sharing resources and processing in parallel.
MoE (Mixture of Experts) models are already an art of efficiency: only a subset of "expert" parameters activate per token. Combined with Blackwell's new architecture, this "activate on demand" pattern is amplified further.
Tri Dao also noted that the Muon optimizer will have a fast path on consumer Blackwell cards, since the implementation can reuse the matmul main loop. Good news for anyone who wants to squeeze every drop of performance from new cards—new hardware shouldn't just run fast, it should run smart, wasting not a single watt.
The Raspberry Pi Fights Back: Small Devices Have Dignity
Community members ran LLM benchmarks on a Raspberry Pi 5 + M.2 HAT+ + 1TB SSD setup. The surprising result: Gemma4 E2B-it Q8_0 reached 41.76 tok/s on the pp512 test.
A $50-ish computer originally meant to teach kids programming can now run large models. Key points behind this:
1. PCIe Gen3 direct-attached SSD with ~800 MB/s read speed—sufficient for loading model weights 2. Reasonable quantization levels (Q4/Q6/Q8) balancing speed and quality 3. mmap reads weights directly from SSD—no need to fit all parameters in memory
As community discussion noted, mmap from SSD works, breaking the assumption that models must be fully loaded into VRAM/RAM. Perhaps in the future, models will page in and out of disk like an OS swaps memory: disk becomes cold storage, RAM becomes a hot cache, and compute focuses only on what's actually needed.
The Mac Trade-off: 31B or 26B?
A user on a 48GB MacBook Pro compared models: Gemma4 31B took 30–50 minutes per code audit, while a 26B MoE variant took on the order of 2 minutes.
Why the huge gap? The 31B is a dense model—every token processes 31 billion parameters with a huge KV cache. The 26B A4B is MoE, activating only a few experts per step, cutting compute demand by roughly an order of magnitude.
More parameters isn't better—what matters is how many are actually activated. Like a company: a 300,000-person firm with only 30,000 people actually working may be less efficient than a lean 20,000-person team.
Recommendation: combine quantization, shorter context, and sensible KV cache configuration before deciding you need 31B. For most tasks, the 26B MoE is probably good enough.
The MLX Ecosystem: Apple's Hidden Weapon
Someone ran Gemma 4 E2B at ~40 tok/s on an iPhone 17 Pro via MLX, Apple's machine learning framework optimized for Apple Silicon. The community has also ported ESM-2 (protein modeling) and Falcon Perception to MLX, running biological sequence modeling and image segmentation locally on Macs.
For scientific computing, computer vision, or privacy-sensitive offline workloads, this is a win: your data never leaves your device. Apple's ecosystem advantage shows here—unified hardware (Apple Silicon) + optimized software (MLX) + a large developer community. This isn't porting Linux tools to Mac; it's AI infrastructure built natively around the Mac.
Hugging Face's Textbook: A Guide to Parallel Computing
HF released the Ultra-Scale Playbook, systematically covering data, tensor, pipeline, expert, and context parallelism, with measured data on up to 512 GPUs. For teams upgrading from single-machine experiments to large-scale training/inference clusters, it's an extremely practical architectural reference—like someone walking through every pitfall and handing you the finished map.
Final Thoughts
The first half of the compute war was about "having or not having"—whether you could buy chips and build data centers. The second half is about "doing it well"—with the same number of cards, who can use them better, train faster, and serve inference cheaper.
Anthropic locking in long-term TPU capacity is a bet on the future. Huawei's Ascend 950PR rising proves blockades fail. Blackwell's warp decode squeezes every drop of hardware performance. The Raspberry Pi's comeback redefines the boundary of the feasible.
How does this war end? Perhaps no one can predict. But one thing is certain: compute is no longer just a technical issue—it's strategic, political, and existential. When chips become strategic resources, whoever holds the chips holds the future.
---
"Whoever masters compute masters AI's lifeline."