Nvidia's Groq 3 LPX hits 3,400 tokens/sec, but Cerebras CS-4 looms

Nvidia's Groq 3 LPX enters production with 3,400 tokens/second—but Cerebras' next-gen CS-4 may narrow the gap

Nvidia's $20 billion Groq acquisition is paying off: the Groq 3 LPX rack system achieved 3,400 tokens/second on Gemma 4 31B with a 100K-token context window, 4x faster than Cerebras' current 882 tok/s benchmark. The LPU architecture uses SRAM-heavy dataflow (150 TB/s memory bandwidth, 500 MB per chip) distributed across up to 256 accelerators per rack, optimized for the memory-bandwidth-constrained decode phase of inference. Nebius is the first customer. However, the comparison may not age well. Cerebras requires only 1–2 CS-3 accelerators for the same model (vs. Nvidia's 64 LPUs), and the company's newly announced CS-4 doubles compute, I/O bandwidth, and memory bandwidth while tripling accelerators per rack—launching later this year. Cerebras is also pairing WSE-3T accelerators with AMD Helios GPUs and AWS in heterogeneous configs similar to Nvidia's NVL72+LPX approach. For infrastructure buyers, this signals intense competitive pressure on inference economics and a shift toward hybrid GPU+accelerator architectures.

Grok 4.6 reaches frontier-model parity with GPT-5.6 Sol in blind evaluations

Lenny's independent testing places xAI's Grok 4.6 at frontier-model quality, ranking alongside OpenAI's GPT-5.6 Sol and ahead of Anthropic's Claude Sonnet 5 and Opus 5 in blind model evaluations. Claude Sonnet 5 still leads in conversational rhythm for agent interactions. This signals xAI has closed the raw capability gap with OpenAI and Anthropic.

Cost-normalized benchmarks and agent design reshape model selection and production AI strategy

Together AI's cost-per-solved-task benchmarks reveal a fundamental shift in enterprise model selection: GLM-5.3 completes 5× more work than Fable 5 on DeepSWE under a $100 budget despite similar first-try performance, while GPT-5.6 Sol Max costs $6.47/task vs. Fable 5 Max at $21.63/task on the same benchmark. This move from raw capability metrics to cost-per-solved-task is changing which models enterprises deploy in production. Simultaneously, NVIDIA research shows structural agent design correlates more strongly with real-world performance than model capability alone, while Anthropic advocates standardizing on reusable enterprise harnesses rather than bespoke orchestration. Together, these signals suggest infrastructure and workflow design are becoming more important than raw model improvements for production AI systems.

Stablecoin card spending hits $600M monthly as KAST targets global fintech scale

KAST, founded by former Circle executive Raagulan Pathy, is positioning stablecoins as infrastructure for a global fintech platform modeled on Nubank/Revolut but without geographic constraints. The company reports stablecoin-linked card spending grew from near-zero in November 2025 to $600M+ by May 2026, capturing the "digital nomad" segment while competitors like RedotPay and EtherFi target emerging markets and DeFi respectively.

Deep-moat SaaS companies monetizing AI outperform shallow-moat peers as inference costs collapse

Companies like Toast (23% revenue growth acceleration), Monday.com (~$1.4B ARR at 2–3x revenue multiple), and Salesforce show strong fundamentals despite mixed stock performance, while shallow-moat SaaS has seen 50%+ Y/Y revenue declines. The underlying thesis: if inference costs fall sharply, value accrues to the application layer where COGS collapse—positioning companies like Cursor, Harvey, and Figma as potential winners.

Chinese companies rapidly move upstream from contract manufacturing to brand ownership and vertical integration

Battery cell makers are becoming grid and datacenter operators; luxury goods manufacturers are becoming brands themselves; state-backed deeptech playbooks are maturing. This signals a structural shift in competitive dynamics that Western companies need to account for in their strategic planning.

What Nvidia's First Groq 3 LPU Benchmarks Tell Us About Its $20B Gamble

The Register analyzed Nvidia's first public benchmarks for its Groq 3 LPX inference systems following the $20B Groq acquisition. Main takeaways:

• Nvidia's LPX racks achieved 3,400 tokens/second on Gemma 4 31B with 100K token context, claimed as 4x faster than Cerebras' 882 tok/s on identical workload.
• Groq 3 LPUs use 150 TB/s on-die SRAM (vs. Nvidia GPUs' HBM), enabling extreme memory bandwidth but limiting per-chip capacity to 500 MB—requiring 64+ LPUs for a single 31B model.
• Nvidia's heterogeneous strategy splits inference: prefill on GPUs (compute-heavy), decode on LPUs (bandwidth-heavy), creating a Pareto improvement over GPU-only systems.
• Cerebras' unreleased CS-4 accelerators (WSE-3T-based, 2x compute/bandwidth, 3x density) and AMD Helios+Cerebras hybrid configs may erode Nvidia's performance claims within months.

The Electric Slide: The history, 99% decline, and future of the Electric Stack

Not Boring published a 40,000-word essay on the Electric Stack—the four core technologies (lithium-ion batteries, magnets/motors, power electronics, embedded compute) that enable electrification across products. Main takeaways:

• The Electric Stack cost curve has declined 99% since 1990 (12.6% annually), with batteries down 98.7%, motor actuation down 98.8%, power electronics down 99.5%, and compute down 99.9%.
• China now dominates three of four layers: 75% of global lithium-ion battery production, 90% of neodymium magnets, and rapidly growing share of power electronics and embedded compute, enabling dominance in EVs, drones, robots.
• America's strategic bet on AI leadership underestimates the complementarity of energy, intelligence, and action; China's bet that controlling the physical Electric Stack matters more than AI alone may be correct.
• Western companies sold critical capabilities: GM sold Magnequench (neodymium magnets) to China for $70M; A123 Systems (LFP batteries) sold to Wanxiang for $257M in 2013.
• The essay includes original Electric Slide composite cost curves and an interactive tool showing how component cost weightings vary by product (Tesla Model 3 vs. DJI drone).

The AI Engineer: A New Subdiscipline Emerges

Latent Space published an essay on the emergence of AI Engineering as a distinct professional discipline separate from ML Engineering. Main takeaways:

• AI Engineer roles are growing 10x faster than ML Engineer roles on job boards, with the ratio expected to invert within 5 years as foundation models shift work from training to application
• Highly effective AI Engineers (cited: Simon Willison, Pieter Levels, Riley Goodside) typically lack ML PhDs and PyTorch expertise, instead shipping products through API integration and prompt engineering at scale
• Foundation model availability, GPU scarcity forcing API-first architectures, and the shift from "fire-ready-aim" ML workflows to rapid prototyping with LLMs creates 1,000-10,000x cheaper product validation cycles
• JavaScript/web developer TAM expansion (parity with Python developers) via tools like LangChain.js and Vercel AI SDK doubles addressable market for AI tooling
• The AI Engineer Summit conference announcement signals professionalization of the role, with focus on AI UX, devtools, infra, agents, and open-source model deployment