NVIDIA Delays Kyber to 2028 🔥, Anthropic Finds J Space in LLMs 🤖, 68% RAG Context Pruned 💻

Billing lag costs more than you think. See exactly how much with the Tabs Billing Lag Calculator — benchmarked against top SaaS companies.

See exactly how much

One trip, and your team stops guessing on AI adoption, platform, security, and tooling. 

WeAreDevelopers World Congress — San José, CA, September 23–25, 2026, in the heart of Silicon Valley. 500+ speakers and 10,000+ developers working through what actually ships in production: coding agents, evals, MCP, observability, agent security.

Kelsey Hightower. Thomas Dohmke (fmr. CEO, GitHub). Christine Yen (CEO, Honeycomb). Mathias Biilmann (CEO, Netlify). Olivier Pomel (CEO, Datadog). The people actually building the tools you use every day — all on one stage.

Bring a shortlist of decisions; leave with answers and contacts you can still reach after the event. Prices rise as we get closer. 

Use code GITPUSH26 for 10% off.

Get Your Pass

Holiday weekend piled up the news. NVIDIA hit manufacturing walls, Anthropic found something weird inside Claude's brain, and open source models keep eating into frontier margins.

NVIDIA's Next Gen Kyber Rack System Delayed to 2028 (5 minute read)

NVIDIA pushed its Kyber rack scale architecture back over a year due to circuit board manufacturing challenges in Taiwan. The system packs 144 Rubin Ultra chips into a single rack that functions as one supercomputer. NVIDIA's annual release cadence is finally colliding with the physical limits of manufacturing.

Scortier's Take: This is the first real crack in NVIDIA's "we ship on schedule, every year" narrative. If you're planning infrastructure purchases around Rubin Ultra availability, add a buffer. And watch Broadcom's custom ASIC business pick up the slack.

Anthropic Discovers "J Space": A Global Workspace Inside Language Models (8 minute read)

Anthropic found an internal neural pattern in Claude that emerged during training without anyone designing it. They're calling it "J Space," and it acts like a shared workspace where the model coordinates reasoning across different tasks. The research lets you distinguish deliberate thinking from automatic pattern matching inside the model, which has real implications for safety monitoring.

xAI Rebrands to SpaceXAI (3 minute read)

Musk's AI division dropped the xAI name and merged under SpaceXAI. The rebrand ties AI development directly to SpaceX's infrastructure. Their internal Grok Build coding harness runs daily improvement cycles on Grok 4.5, using production environments across Musk's companies as live evaluation infrastructure.

Broadcom and Apple Extend Chip Partnership Through 2031 (4 minute read)

Broadcom will supply custom ASIC silicon to Apple through 2031, covering multiple generations of products. Apple plans to deploy advanced AI servers as early as 2027. This deal locks in Broadcom's position as the go to partner for custom AI chips across both consumer and cloud hardware.

How Kapa.ai Taught a Small LLM to Throw Away 68% of RAG Context (7 minute read)

Kapa.ai deployed a small, cheap LLM to filter unnecessary context chunks before they reach the main model in their RAG pipeline. The pruner discards 68% of chunks while maintaining 96% recall. If you're building RAG systems and your context windows are bloated with irrelevant docs, this is a direct cost and quality win you can copy.

GLM 5.2 and the Coming AI Margin Collapse (10 minute read)

GLM 5.2's cost efficiency is so aggressive that it's forcing a rethink of AI economics. Smaller, cheaper models are becoming good enough for routine tasks, which means the premium you pay for frontier models only makes sense for the hardest problems. This is the beginning of AI becoming a commodity utility, not a premium product.

Bringing PyTorch Monarch to AMD GPUs (6 minute read)

PyTorch Monarch now supports elastic, fault tolerant distributed training on AMD GPUs. If a node dies mid training, the job recovers automatically instead of crashing. For teams running large scale training and tired of NVIDIA lock in pricing, AMD just became a real option for billion parameter models.

React Compiler Memoization: What Actually Broke (8 minute read)

The React compiler automates useMemo and useCallback, which sounds great until you learn what actually breaks. DevTools memoization indicators don't guarantee optimization worked, and some patterns silently fail. If you're planning to adopt the React compiler, read this first to know what to watch for.

Getting Started with CSS Anchor Positioning (10 minute read)

Josh Comeau's walkthrough of the Anchor Positioning API, which lets you position tooltips, dropdowns, and popovers relative to any element with almost no JavaScript. It handles viewport awareness and fallback positioning natively. This replaces a surprising amount of Popper.js and Floating UI code.

Otari: OpenAI Compatible Gateway for 40+ LLM Providers (3 minute read)

Mozilla's open source LLM gateway gives you a single endpoint for 40+ AI providers with built in API key management, budget enforcement, and usage tracking. If you're juggling multiple AI providers and want one unified interface with cost controls, this does it out of the box.

Price per 1M Tokens Is Meaningless (5 minute read)

Tokenization varies wildly across models, so comparing cost per million tokens is comparing apples to oranges. A task that costs 1000 tokens on GPT 5.6 might cost 1400 on Claude. The only honest comparison is cost per task completion, and almost nobody is doing that math.

Big Tech Quietly Flips on the AI Jobs Wipeout Narrative (6 minute read)

Tech CEOs who predicted mass AI driven layoffs are now walking it back, shifting to "AI makes everyone more productive" messaging. Whether this is genuine recalibration or just PR cleanup is anyone's guess, but the data backs it: companies with heavy AI adoption show 10% headcount growth over two years.

State of CLI Coding Agents, Mid 2026 (7 minute read)

Claude Code, Codex CLI, and Omp deliver comparable quality across repo analysis, multi file editing, and failure recovery. They differ in how they handle ambiguous tasks, repo requirements, and tool exposure. If you're picking a CLI coding agent, the differences are smaller than you'd expect.

Tencent's Hy3: 295B Parameter MoE Model, Free on OpenRouter

21B active parameters, outperforms comparable open source models, free through July 21.

TeraWulf Signs $19B Infrastructure Lease with Anthropic

Anthropic's largest infrastructure deal, with capacity coming online late 2027.

Microsoft Cuts 3,000+ Xbox Jobs

Roughly 20% of the Xbox division eliminated in the latest round of cuts.

Alibaba's AI Models Popular but Unprofitable

Open source Qwen models gain massive adoption, but Alibaba can't figure out the business model.

California Deploys Claude Across All State Agencies

Largest US state government AI deployment in history, with Anthropic offering a 50% discount.

Data Infrastructure Spending to Hit $100B+ by 2030

AI has shifted the bottleneck from compute to data, making high quality datasets the new strategic moat.

Apple's iPhone Ultra: Foldable at $2,400

Double the price of the Pro Max, expected to sell out immediately when it ships in October.

PACE: Evaluate AI Agents 99% Cheaper

Framework predicts agentic benchmark scores using tiny evaluation subsets with under 4% error.

🔥 Job Openings

Follow me on Youtube ; LinkedIn ; X ; Instagram to stay updated.

See you in the next one!
Scortier, Signing Off!