Source commit: 2c47ab1
If AI is a rocket, compute is its fuel. Without fuel, even the most advanced engine is scrap metal. In spring 2026, the battle over that fuel is entering its heated second half.
Anthropic's Big Bet: Locking in a Decade of TPUs
Anthropic has signed agreements with Google and Broadcom to secure multi-gigawatt capacity of next-generation TPUs starting from 2027. One gigawatt equals one billion watts; multi-gigawatt means tens of billions of watts of compute — enough to train and serve next-generation Claude frontier models.
The key is the time dimension: starting from 2027. This is not buying off the shelf — it is pre-booking the future. Like a farmer reserving the autumn harvest in spring, Anthropic is locking in compute supply for 2027 and beyond.
Why? Consider OpenAI's situation: by 2028, frontier labs may spend over $100 billion on compute. This is not buying computers — it is buying power plants, dedicated ones. When compute becomes the bottleneck, whoever owns compute owns the future.
Anthropic also disclosed that its annualized revenue has exceeded $30 billion. This is no longer a "burn cash for the future" research game but a real, value-creating business — revenue large enough to sustain long-term compute investment and convince investors it is not a bubble.
Huawei's Counterattack: Ascend 950PR and the Chinese Chip
DeepSeek V4's plan is interesting: run natively on Huawei's Ascend 950PR, with an upper layer compatible with NVIDIA programming interfaces.
It is a clever strategy. Imagine a chef used to a particular knife who is suddenly told that knife can no longer be purchased. Relearning a whole new cutting technique is too costly. DeepSeek's choice: swap the knife but keep the grip and technique unchanged. The Ascend 950PR is Huawei silicon underneath, but the upper interface is CUDA-compatible — greatly lowering migration costs from NVIDIA.
Alibaba, ByteDance, and Tencent have placed large orders, driving chip prices up about 20%. The 950PR outperforms the H20 and 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 imports, core compute is no longer hostage to foreign suppliers. It is a slap in the face of US export controls: the more you block, the more we build our own — cheaper and better.
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 ~1.84x.
What is warp decode? In a factory, each worker handles one step in sequence. Warp decode is like having a team of workers (a warp) collaborate simultaneously, sharing resources and processing in parallel.
MoE (Mixture of Experts) models are themselves an art of efficiency: only a subset of "expert" parameters activates per token rather than all. Combined with Blackwell's new architecture, this on-demand activation is amplified further.
Tri Dao also mentioned that the Muon optimizer will have a fast path on consumer Blackwell, since the implementation can reuse the matmul main loop. Good news for anyone who wants to squeeze full utilization from new cards — new hardware should not just run fast, it should run smart, wasting not a single watt.
The Raspberry Pi Strikes Back: Dignity for Small Devices
A community member benchmarked LLMs 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.
Key factors:
1. A PCIe Gen3-attached SSD delivers ~800 MB/s read — sufficient for loading model weights 2. Reasonable quantization levels (Q4/Q6/Q8) balancing speed and quality 3. mmap reads weights directly from the SSD — no need to fit all parameters in memory
As community discussion noted, mmap-based weight loading from SSD works. This breaks 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 operating system swaps pages: disk becomes "cold storage," memory becomes a "hot cache," and compute focuses only on what is currently needed.
The Mac Trade-off: 31B or 26B?
A user on a 48GB MacBook Pro compared: Gemma4 31B takes 30-50 minutes for a code audit, while a 26B MoE model takes on the order of 2 minutes.
Why the huge gap? The 31B is a dense model processing 31 billion parameters per token with a huge KV cache. The 26B A4B MoE activates only a few experts per step, cutting compute demand by an order of magnitude.
The lesson: more parameters are not always better — what matters is how many are actually activated. Like a company: 300,000 employees where only 30,000 work may be less efficient than a lean 20,000-person team.
Consider quantization, shorter contexts, and sensible KV cache configuration before deciding you need the 31B. For most tasks, a 26B MoE is likely good enough.
The MLX Ecosystem: Apple's Hidden Weapon
Gemma 4 E2B has been run 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 protein-modeling model ESM-2 and Falcon Perception to MLX, enabling local biological sequence modeling and image segmentation on Macs. This is a win for scientific computing and computer vision users who want offline, privacy-preserving AI — data never leaves the device.
Apple's ecosystem advantage shows here: unified hardware (Apple Silicon) + optimized software (MLX) + a large developer community. This is not porting Linux tools to Mac — it is AI infrastructure built natively around the Mac.
Hugging Face's Textbook: A Guide to Parallel Computing
Hugging Face released the Ultra-Scale Playbook, systematically covering data, tensor, pipeline, expert, and context parallelism, with measured data on up to 512 GPUs.
For teams moving from single-machine experiments to large-scale training/inference clusters, this is an extremely practical architectural reference. Parallel computing is nothing new, but systematically summarizing best practices and quantifying the trade-offs of different strategies remains very valuable — like having someone step on every landmine first, then hand you the map.
Closing Thoughts
The first half of the compute war was about "whether you have it" — can you buy chips, can you build data centers. The second half is about "how well you use it" — with the same number of cards, who trains faster and serves inference cheaper.
Anthropic locking in long-term TPU capacity is a bet on the future. Huawei's Ascend 950PR rising proves sanctions ineffective. Blackwell's warp decode squeezes every drop of hardware performance. The Raspberry Pi redefines the boundaries of what is feasible.
How does this war end? No one can predict. But one thing is certain: compute is no longer just a technical issue — it is strategic, political, and existential.
When chips become strategic resources, those who own chips own the future.
---
"Whoever controls compute controls AI's lifeline."