OpenAI's Jalapeño chip beats Nvidia on inference efficiency

Lovable hits $500M ARR at $13.3B valuation with 900M monthly visits, demonstrating product-market fit in AI-generated software

Lovable raised $400M Series C led by Menlo Ventures and EQT's Scaleup Europe Fund, growing from an open-source coding tool in 2023 to a full software creation platform. The company has achieved $500M+ revenue run rate with 7x year-over-year growth, 900M+ monthly visits to Lovable-created sites, and 60M projects created (~5M/month). Two-thirds of Fortune 500 companies use the platform, with users predominantly non-technical entrepreneurs running real businesses on it via native Stripe integration, built-in SEO, and security. The strong retention reflects that users don't churn when real revenue runs through their creations.

Lovable exposes app functions as MCP servers, enabling AI agents to bypass UI entirely

Apps built on Lovable can now present dual interfaces—traditional UI for humans and an MCP server exposing selected functions to AI agents—allowing agents to call app capabilities directly. This represents a fundamental shift in how SaaS will be consumed, with Lovable positioning itself as the platform for building agent-callable capabilities rather than human-facing applications.

OpenAI's Jalapeño ASIC delivers 1.5-1.9x better throughput-per-watt than Nvidia Blackwell; Microsoft's Vera Rubin enters production as custom chips challenge Nvidia's vertical integration

OpenAI's Jalapeño inference chip, announced at Hot Chips, achieves 1.5-1.9x better throughput-per-watt than Nvidia GB200/GB300 systems with lower latency, according to SemiAnalysis benchmarks. Designed in partnership with Broadcom in ~16 months using AI-assisted hardware-software codesign, Jalapeño beats every Nvidia, AMD, and Google chip tested on multiple open-source models while maintaining generalized inference capability across workloads. On perf/TCO, Jalapeño and Microsoft's Vera Rubin (now shipping) are head-to-head, though Jalapeño achieves results without speculative decoding—leaving room for further optimization. Production volumes don't arrive until late 2027. Meanwhile, Google expanded its custom-chip partnership with Marvell to cover inference accelerators, networking, storage, and memory controllers—essentially unbundling Nvidia's vertically integrated system and re-integrating around Google's TPU ecosystem. The deal can generate up to $120B in custom chip sales through fiscal 2033, with Google holding a warrant to buy up to $12.2B in Marvell stock. This challenges Nvidia's moat of coordinating the entire AI factory under one roadmap, though Google's modular approach trades integration simplicity for harder multi-vendor coordination.

Nvidia guides Q3 revenue of $108B and ~70% FY2028 growth, assuming zero China data center revenue

Nvidia forecast Q3 revenue of $108B (±2%), above analyst estimates of $104.19B, with Q2 revenue of $96.2B representing 106% year-over-year growth and Data Center revenue up 117% to $89B. CFO Colette Kress indicated customer forecasts "point to our growth doubling next year," all while assuming zero China data center compute revenue. The guidance validates unprecedented AI capex absorption—Microsoft and Meta alone reinforced expectations that Big Tech will spend more than $730B on AI infrastructure this year, up from $400B last year. However, Nvidia's dominance faces growing scrutiny as an increasing share of tech companies' spending now goes toward in-house chip efforts aimed at reducing reliance on Nvidia's costly, supply-constrained processors.

Z.ai releases GLM-5.3-Flash with 320B parameters at one-tenth the price of GLM-5.2, running entirely on Chinese Ascend chips

Z.ai released GLM-5.3-Flash, a natively multimodal model with 320B total parameters and just 18B active parameters, outperforming GLM-5.2 across benchmarks while costing one-tenth as much. Tested anonymously as "ox-alpha" on OpenRouter, it achieved a score of 57 on the Artificial Analysis Intelligence Index at $0.045 per task (discounted)—a level of intelligence previously available only at roughly 10x the cost. The model runs entirely on Chinese Ascend chips, with Z.ai achieving 3x improvement in end-to-end serving performance compared to initial baseline, reaching hardware efficiency and per-token cost comparable to mainstream Nvidia GPUs. This demonstrates that frontier-model inference can run efficiently and economically on alternative compute stacks independent of Nvidia, with architectural improvements (hybrid sparse and linear attention, Manifold-Constrained Hyper-Connections) and optimized pre-training enabling more intelligence from less compute.

China's humanoid robots shattered Olympic records while Unitree IPO'd at $50B; US lacks national robotics strategy

At the World Humanoid Robot Games in Beijing, 2,000 robots from 16+ countries competed in 55 Olympic events, with 96% Chinese-made competitors. Chinese robots broke human records across multiple events: 100m in 9.39s (vs. Usain Bolt's 9.58s), 1500m in 2:21 (vs. human record 3:26 set in 1998), standing high jump at 2.88m, and 400m in 38.15s—all running fully autonomously with zero human intervention. Year-over-year, robot count increased 300% and team count 200%, with robots demonstrating self-taught running techniques via reinforcement learning. Unitree, a leading Chinese robotics company, IPO'd at a $50B valuation with stock up 460% on first trading day, reflecting massive capital concentration in China's robotics sector. China attracted approximately 50% of global VC in humanoid robotics, with one company (Zhiyuan) shipping 40% of world humanoids last year. Meanwhile, the US has no coordinated national robotics strategy despite Boston Dynamics and Elon Musk's advocacy, and Chinese firms dominated 21 of 38 humanoid exhibitors at CES. This signals China's manufacturing dominance and state-backed robotics investment ($148B AI/robotics fund) are translating into concrete hardware superiority harder to replicate than software.

Nvidia acquires HuggingFace for $13B at 80x revenue multiple, betting on open-source model distribution as core AI stack

Nvidia is acquiring HuggingFace for $13B, roughly double its initial $7B offer from January 2026, valuing the platform at approximately 80x its $150M annual recurring revenue. HuggingFace doubled its customer base in 2026, establishing itself as an essential developer platform where creators publish, share, and download AI models from OpenAI, Anthropic, Meta, and others. The acquisition signals Nvidia's strategic bet on open-source model distribution and infrastructure as central to its AI stack dominance, following Stripe's $8B acquisition of OpenRouter and reflecting investor willingness to pay premium prices for companies at the center of the AI ecosystem even when they aren't building frontier models themselves.

Anima Anandkumar's Neural Operators enable weather forecasting on consumer GPUs, outperforming decades-old physics simulations

Anima Anandkumar's FourCastNet model demonstrates that AI can model chaotic physical systems like weather and fusion with far less data than language models require, using mathematically principled architectures (Fourier Neural Operators with spherical harmonics) rather than pure scale. This signals a distinct path for physics-based AI that prioritizes structural inductive biases over token scaling, with implications for how AI infrastructure and models should be designed for scientific computing workloads—suggesting that specialized, hardware-efficient approaches may outcompete general-purpose scaling for domain-specific problems.

FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale

NVIDIA (Bonev, Kurth, et al.) in arXiv published a paper titled "FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale". Main takeaways:

• Achieves 8-60x faster inference than leading diffusion models (GenCast, IFS-ENS) while maintaining competitive medium-range skill and superior spectral stability to 60-day lead times
• Generates 90-day global forecasts at 0.25° resolution in under 20 seconds on single H100 GPU, enabling practical large-ensemble predictions
• Scales training to 1000+ GPUs via novel hybrid model/data-parallelism inspired by classical NWP domain decomposition, addressing computational bottlenecks in probabilistic ensemble forecasting
• Addresses key ML weather modeling failure modes: retains realistic spectra, avoids small-scale noise buildup, and maintains physical consistency beyond medium-range timescales where diffusion approaches degrade

OpenAI Jalapeño: Better Than Nvidia

SemiAnalysis published a detailed technical report on OpenAI's Jalapeño inference chip, which achieved tape-out in ~16 months and now outperforms Nvidia, AMD, and Google chips on multiple benchmarks. Main takeaways:

• Jalapeño beats Vera Rubin on perf/W (tokens per MW) without speculative decoding, and exceeds GB200 across low-latency and high-throughput scenarios
• Chip achieves 700+ tokens/sec/user at concurrency 1 on DeepSeek R1 and ~1,400 tok/sec/user on GPT-OSS, using single-token prediction with no prefill-decode disaggregation
• B0 stepping (currently in fab) delivers ~25% perf-per-watt improvement over A0, with 13.4 PFLOPs MXFP4 at 700W TDP versus Rubin's 17.5 PFLOPs at 900-1,150W
• OpenAI designed for perf/W as primary constraint since datacenters are power-limited, not budget-limited; homogenous chip architecture trades theoretical efficiency for operational flexibility across changing workload mixes
• Results use 8k1k benchmarks without AgentX testing; comparison to Rubin more relevant than Blackwell given similar HBM4 adoption and development stage

GLM-5.3-Flash: Frontier Intelligence, Flash Cost

Zhipu AI in Zhipu AI (Z.AI) published a report titled "GLM-5.3-Flash: Frontier Intelligence, Flash Cost". Main takeaways:

• GLM-5.3-Flash scores 57 on Artificial Analysis Intelligence Index v4.1.1 at $0.045/task (discounted), matching Claude Opus 4.8 on coding benchmarks (29.0 vs 29.5 on Z.ai Code Bench v1.0) while reducing activated parameters from 32B to 18B
• Hybrid sparse+linear attention architecture with IndexPool compression reduces attention compute 3.0x and KV cache 4.4x versus GLM-5.3, enabling efficient 1M-token context serving on Chinese AI chips
• Infrastructure co-design with Chinese hardware achieved 3x end-to-end serving improvement, reaching per-token costs comparable to mainstream NVIDIA GPUs, demonstrating frontier-model inference efficiency on non-NVIDIA silicon at scale