SoftBank Doubles Down on AI Compute

Gavin Baker, founder of Atreides Management and long-time tech investor, tells Invest Like the Best that the single best indicator of an AI bubble is TSMC’s capacity decisions. He highlights Anthropic’s extraordinary ARR growth, argues AI monetization is shifting toward pay-per-use enterprise models, and warns that power and wafer supply are AI’s two ceilings—pointing to on-ground power expansion, orbital compute, new fabs like Terafab, and TSMC’s cautious expansion as critical factors. Baker praises Amazon’s Trainium and sees robotics improving retail margins, but notes systemic risks: algorithmic breakthroughs that remove scaling needs, widening tail risks, inequality in AI access, and geopolitics. His view mixes strong AI optimism with pragmatic infrastructure-focused caution.

Intel is reportedly planning a special “Nova Lake” processor variant aimed at edge AI, featuring an 8E+12Xe configuration that drops performance CPU cores in favor of eight efficiency cores and a large GPU matching the family’s top models. The leak, posted by @金猪升级包 and reported by IT之家, says the design prioritizes GPU compute for on-device inference and SLM local AI workloads common in edge deployments. Intel has previously released edge-focused SKUs that don’t map directly to client product lines (for example, Bartlett Lake 12P), suggesting this Nova Lake variant follows that pattern. If true, it signals Intel tuning architectures specifically for edge AI performance over traditional CPU-centric desktop tasks.

NVIDIA remains the dominant hardware choice for local LLM inference in 2026, but the landscape is shifting as alternatives close the gap. Users report NVIDIA GPUs still lead on mature software support, optimized CUDA toolchains, and widespread compatibility with frameworks and model runtimes—key for low-latency, high-throughput inference. However, competitors such as AMD, Intel, and specialized AI accelerators have improved driver stability, open runtimes (e.g., ROCm, oneAPI), and price-performance for certain model sizes, making them viable for cost-sensitive or on-prem deployments. The debate matters because choice affects costs, power, and integration complexity for developers and enterprises running local models, and software ecosystem maturity remains the decisive factor.

Intel’s AI Superbuilding team on May 21 unveiled SuperClaw, a hybrid AI agent solution for AI PCs and edge devices that emphasizes privacy and lower operating costs. SuperClaw uses a local-first hybrid architecture that reportedly cuts cloud token consumption by up to 70% and can identify sensitive data with 99% accuracy. The solution targets on-device inference and sensitive-data detection to reduce cloud usage and privacy risk for edge deployments. A beta download is scheduled for the second half of June 2026. This matters for device makers, enterprises, and developers seeking cost-effective, privacy-preserving AI at the edge and reflects Intel’s push into hybrid AI tooling for consumer and industrial hardware.

AMD CEO Lisa Su said mainland China accounts for about 20% of AMD’s revenue and remains a very important market. Speaking around AMD’s AI DevDay in Shanghai, Su predicted a sharp rebound in the CPU market driven by AI infrastructure and inference needs, forecasting annual CPU growth above 35% over the next five years. She highlighted AMD’s broad business in China across PCs, gaming and some data center segments, and noted AMD’s Greater China R&D footprint of over 4,000 engineers and AI centers in Beijing, Shanghai, Shenzhen and Taipei. The remarks underscore China’s strategic role for AMD amid global AI-driven demand surges.

Ben Thompson’s Stratechery roundup spotlights 'The Inference Shift', arguing AI will move from human-centric 'answer inference' to 'agentic inference'—autonomous AI acting without human input—which will reshape compute architecture and favor new players and geographies. The newsletter also covers Anthropic’s compute deal with xAI/SpaceX, exploring strategic implications for Elon Musk, SpaceX’s potential market for space-based data centers, and the broader competitive dynamics among AI labs (including OpenAI). Other items include industry takes on deployment-focused AI companies, Apple-Intel economics, and commentary on Musk’s lawsuit with OpenAI. The pieces matter because they address where compute demand, architecture decisions, and corporate strategies are headed in AI’s next phase.

Ben Thompson argues the biggest tech shift isn’t training vs. inference but two kinds of inference: fast, human-in-the-loop “answer inference” and slower, autonomous “agentic inference.” Agentic inference — where humans aren’t involved — will reshape compute architectures, favoring different trade-offs and expanding markets (good news for China and space, less so for incumbents like Nvidia). The newsletter also covers Anthropic’s compute deal with xAI/SpaceX, Elon Musk’s lawsuit with OpenAI, U.S.-China tech diplomacy, and industry moves such as OpenAI forming a deployment company and Apple’s pragmatic ties with Intel. These pieces highlight how compute sourcing, deployment strategy, and geopolitics are remaking AI infrastructure and commercial dynamics.

AMD CEO Lisa Su projects the CPU market will grow over 35% annually through 2031, up from 3% to 4% historically, driven by AI inference and agentic AI demand (Cheng Ting-Fang/Nikkei Asia)

AMD CEO Lisa Su told a Taipei event that surging AI infrastructure demand has pushed CPU supply into a “tight” state and that the CPU market could grow more than 35% annually over the next five years. Su said CPUs lagged recent growth as the industry focused on GPUs, but as AI moves toward inference and intelligent agents, CPU demand has spiked far beyond last year’s forecasts. She also highlighted related data-center bottlenecks—memory and power among them—but expressed confidence these constraints will be resolved quickly. The remarks underscore shifting hardware needs as AI workloads broaden beyond GPU-centric training to wider deployment and inference.

Cheng Ting-Fang / Nikkei Asia : AMD CEO Lisa Su projects the CPU market will grow over 35% annually through 2031, up from 3% to 4% historically, driven by AI inference and agentic AI demand — TAIPEI — AMD CEO Lisa Su is predicting the market for central processing units (CPUs) will grow massively over the next five years …

CJ Haddad / CNBC : Meta, Broadcom, Applied Materials, GlobalFoundries, and Synopsys launch a $125M “Semiconductor Hub” at UCLA to advance AI chip research and more — Broadcom, Meta, Applied Materials, GlobalFoundries and Synopsys are joining forces to launch a $125 million “Semiconductor Hub” at the UCLA Samueli School of Engineering.

AMD CEO Lisa Su said CPU demand is growing faster than forecasts and expects the market to expand strongly over the next few years, with AMD projecting annual CPU market growth above 35% for the next five years. Su made the comments at an event in Taipei, framing the surge as aligned with AMD’s strengths. The statement underscores rising compute demand driven by AI, data centers, and client devices, and signals continued investment and competition in processors from major vendors. For the tech industry, sustained high CPU growth implies stronger chip sales, supply-chain and design activity, and intensified rivalry among semiconductor firms.

4月中这篇聊CPU的推文应该是近期财富密码最多最密集的一篇文章了。从4月中开始聊到到现在： intel、澜起科技都已翻倍甚至涨的更多； arm接近翻倍； amd、深南电路、海光也都有几十个点的涨幅。 昨晚Arm创新高，4月中聊到Agentic AI时代Arm服务器cpu是会更受益，首先其技术架构和功能更符合agentic ai对cpu的需求，然后是包括英伟达在内的几大云厂商的cpu都用的是Arm架构。之前Arm的走势还可以但没这么强势，直到这两天几个催化事件：

AMD计划在台湾的AI领域投资超过100亿美元

据《The Information》报道，Anthropic正就使用微软的人工智能芯片进行谈判

据《The Information》报道，Anthropic正就使用微软的AI芯片进行谈判

据《The Information》报道，Anthropic正就使用微软的AI芯片进行谈判