Today's News
Techmeme
Alphabet increased capex guidance citing continued underestimation of compute demand, while Meta's CFO acknowledged persistent underestimation of AI needs despite massive capacity ramps and multi-chip strategy (Nvidia, AMD, custom Broadcom silicon). This represents an unprecedented private-sector infrastructure buildout cycle with major implications for chip demand, data center economics, and competitive positioning.
Techmeme
Microsoft reported Q3 AI revenue surpassing $37B annualized, with CFO noting AI margins are structurally stronger than cloud's early transition, and Azure cloud services accelerating to fastest growth in 15 quarters. This signals strong enterprise adoption of AI infrastructure and applications at scale.
Techmeme
Alphabet's Q1 results show Google Cloud significantly outperforming expectations, with CEO Sundar Pichai highlighting AI's role in driving all-time-high search queries and strong consumer AI subscription growth. This is significant as it demonstrates monetization velocity for AI infrastructure and models across Google's stack.
Google launches dual TPU strategy with TPU 8t for training (2.7x perf-per-dollar vs Ironwood) and TPU 8i for inference (80% improvement, optimized for MoE latency)
Liberty’s Highlights
The chips use specialized designs from Broadcom and MediaTek respectively, support 1M+ TPU clusters for training, and integrate Axion Arm CPUs, demonstrating Google's silicon diversification and competitive response to Nvidia's Blackwell dominance in both training and inference workloads.
Inference compute emerging as strategic bottleneck, with CPU demand spiking alongside GPU optimization
AINews
Intel CEO disclosed rising CPU utilization driven by agentic AI workloads, RL training, and code execution, while Jensen Huang declared "the inference inflection has arrived" with compute demand up roughly 1 million times in two years. This signals a fundamental shift from training-dominated capex to inference-dominated infrastructure, with prefill/decode disaggregation becoming standard across vendors (Nvidia acquiring Groq, Intel-Sambanova partnerships, Amazon/Cerebras deals).
OpenAI and Microsoft scrap AGI escape clause, allowing OpenAI to sell on AWS
Liberty’s Highlights
Microsoft's revenue share now extends through 2030 regardless of AGI achievement, and OpenAI's IP rights are no longer exclusive to Microsoft, freeing OpenAI to distribute models across cloud providers as Anthropic's coding monetization success demonstrated the value of platform flexibility.
GitHub Copilot and Anthropic shift to usage-based billing as agentic workloads become unsustainable under flat-rate plans
Liberty’s Highlights
GitHub moves to usage-based pricing June 1, 2026 because agentic coding sessions now demand significantly higher compute than simple chat queries, and Anthropic is similarly moving enterprise plans toward per-seat base fees plus usage billing, signaling structural pressure across the industry.
Elon Musk sues OpenAI for $134 billion over nonprofit-to-for-profit conversion, trial underway in Oakland
Limitless
The case hinges on whether OpenAI can legally convert charitable assets ($44M from Musk) into a for-profit structure now valued at ~$1 trillion, with a jury deciding if the hybrid model violates charitable trust law. This directly impacts OpenAI's planned IPO timeline and structure, plus Microsoft's commercial arrangements.
Coding agents consolidating around headless runtimes and harness engineering as primary optimization layer
AINews
OpenAI expanded Codex into a persistent-context work platform with 40% faster agentic workflows via WebSocket mode, Cursor launched an SDK for CI/CD and embedded agents, and VS Code shipped semantic indexing and skill context features. Research (Agentic Harness Engineering, HALO) demonstrates that model quality alone is insufficient—harness design now determines production performance, with gains like Terminal-Bench improving from 69.7% to 77.0% through iterative harness evolution.
Worth Reading
Google Cloud Blog
I read the full technical deep dive on Google's eighth-generation TPUs. The post details Google's two-chip strategy: TPU 8t optimized for pre-training with SparseCore accelerators, native FP4 support, Virgo Network topology enabling 4x DCN bandwidth, and scaling to 1M+ chips in a single training cluster; and TPU 8i optimized for post-training/serving with 3x more on-chip SRAM, a new Collectives Acceleration Engine reducing latency 5x, and Boardfly topology reducing network diameter from 16 to 7 hops for MoE/reasoning workloads. Includes concrete architectural diagrams, performance comparisons, and software stack details (Pallas, native PyTorch support).
Read moreArmin Ronacher's Thoughts and Writings
I read the full post on practical agent engineering lessons from building production agentic systems. The essay covers concrete technical decisions: why SDK abstractions break under real tool use, explicit cache management tradeoffs between Anthropic/OpenAI/Google, reinforcement patterns that outperform initial expectations, file-system-based shared state as critical infrastructure, and model selection guidance (Haiku/Sonnet as best tool-callers, Gemini 2.5 for document/image work). Includes specific implementation details on cache point placement, failure isolation strategies, and why output tooling is surprisingly difficult to steer—directly relevant to infrastructure and product strategy decisions.
Read moreDigitalOcean Blog
I read the full technical post on DigitalOcean's inference optimization work achieving #1 output speed on Artificial Analysis benchmarks. The report details concrete performance wins (230 tok/s output speed, sub-1s TTFT on DeepSeek V3.2 with 10K input tokens) and breaks down the full stack co-design: NVIDIA HGX B300 hardware selection, NVFP4 quantization tradeoffs, vLLM serving optimizations including kernel fusion and speculative decoding, and production validation via Workato (77% faster TTFT, 79% lower latency, 67% lower costs). This is directly relevant to inference economics and infrastructure buildout during the inference inflection.
Read moreNot Boring
I read the full essay establishing a framework for understanding the next generation of winners in physical tech. McCormick argues that just as Aggregators dominated the internet era by owning demand and modularizing supply, Vertical Integrators will dominate the physical/industrial era by doing the opposite—integrating proven-but-cutting-edge technologies into coherent systems that compete directly with incumbents. The piece uses Base Power Company as a concrete case study (manufacturing battery packs but not cells, designing for system-level optimization rather than component optimization) and positions this as fundamentally different from both legacy industrials and deep tech companies. The framework has direct relevance to understanding competitive dynamics in AI hardware, infrastructure buildout, and manufacturing strategy.
Read moreSources
I read the full interview with AWS CEO Matt Garman discussing Amazon's newly announced partnership bringing OpenAI's GPT-5.4 and 5.5 models to Bedrock for the first time. The piece covers concrete competitive dynamics—how Amazon was originally OpenAI's first cloud partner before Microsoft's investment, Garman's claim that Amazon "will be a better partner," the six-to-nine month integration timeline, and strategic implications of breaking Microsoft's exclusivity arrangement. This directly addresses hyperscaler competition for frontier model distribution and the shifting economics of AI infrastructure partnerships.
Read more
![AINews] The Inference Inflection - Latent.Space](https://www.latent.space/p/ainews-the-inference-inflection?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce489d38-e839-4154-b78d-3deb24865002_2018x1072.png&open=false){kind=link}