Apple is working on an enterprise server using two or four M8 Ultra chips for the AI inference market, with discussions underway with Nvidia about networking equipment, according to The Information. Launch expected no earlier than 2029.
Data center operator Crusoe has secured a multiyear cloud capacity agreement with AI search company Perplexity AI, providing access to Nvidia GB300 processors to support the startup's model training and inference workloads.
Dutch startup Axelera AI launched its second-generation Europa chip and secured AI factory supply contracts with over 600 customers, while Euclyd raised €200M in Series A funding co-led by Samsung to develop its CRAFTWERK inference chip as an alternative to GPU-based solutions.
AI chip startup Positron AI has raised $875 million in its latest funding round, more than quadrupling its valuation to $5 billion in seven months. The company develops chips designed to run AI models more quickly.
Chinese AI company DeepSeek launched its V4.1-Flash model on September 10, 2026, featuring a 552-billion-parameter architecture that cuts memory requirements by 75% and offers inference pricing as low as $0.003 per million tokens. The model claims to match or exceed performance of competitors like GPT-6 Astra, Claude Opus 5, and GPT-5.6 Sol on various benchmarks while operating at a fraction of the cost.
Chinese AI company DeepSeek launched its V4.1-Flash model on September 10, 2026, featuring a 552-billion-parameter architecture that cuts memory requirements by 75% and offers inference pricing as low as $0.003 per million tokens. The model claims to match or exceed performance of competitors like GPT-6 Astra, Claude Opus 5, and GPT-5.6 Sol on various benchmarks while operating at a fraction of the cost.
Perplexity has launched Hybrid Compute, allowing users to combine cloud and local AI models while keeping sensitive data on their Mac, and open sourced Lily, a Rust + Metal inference engine for running Qwen3.6-35B-A3B on Apple Silicon that powers the local processing capability.
Tencent has unveiled Hy4 Preview, a 770-billion-parameter open-source Mixture-of-Experts model that uses a recursive self-improvement loop to boost inference throughput by 31.8%. The release, trained on Tencent's vast product ecosystem, positions the company back in the top tier of China's AI competition.
OpenAI published benchmark results for its custom Jalapeño inference chip built with Broadcom, claiming it delivers more AI work per watt and faster responses than Nvidia's GB300 processors with a 50% cost advantage.
OpenAI presented benchmarks at Hot Chips conference showing its 700W Jalapeño ASIC, co-developed with Broadcom, delivers up to 1.9x throughput per kilowatt and 3.6x lower latency compared to Nvidia's 1,400W flagship GPU. The chip is designed for fast AI inference at scale.
Nvidia has put its Groq 3 LPX racks into full production following its $20 billion acquisition of Groq, with the SRAM-based language processing units achieving 3,400 tokens per second and Samsung manufacturing the 4-nanometer chips. The first deployment is set for Nebius Token Factory before year-end.
d-Matrix presented Raptor at Hot Chips 2026, featuring a TSMC 4nm compute die bonded face-to-face at 36-micron pitch directly onto custom DRAM, achieving 100 terabytes per second memory bandwidth while using one-sixth the energy per bit of HBM3.
Starcloud, a startup developing satellites that can perform AI inference in orbit, has raised $250 million in an extension to its March Series A funding round, bringing its valuation to $2.3 billion. The company is building data centers in space as launch access becomes increasingly constrained.
NVIDIA has launched TensorRT Model Connect (TRTMC) as an open-source public preview tool that enables developers to convert Hugging Face or local checkpoints to TensorRT inference using just two commands, simplifying the deployment of AI models.
AI infrastructure startup Groq has raised $350 million in Series A funding at a $3.5 billion valuation, marking a strategic pivot from developing its own AI chips to becoming a neocloud provider offering GPU and AI infrastructure services. Notably, Nvidia has joined the funding round as the company shifts to using Nvidia-powered infrastructure.
French startup Kog is developing technology to extract more AI inference performance from existing GPUs, entering a competitive market where purpose-built chip makers like Cerebras recently debuted successfully with its May IPO.
OpenAI introduced Ultrafast, a new inference tier that runs GPT-5.6 Sol at up to 750 tokens per second—14 times faster than Standard mode—powered by Cerebras wafer-scale chips from their $10 billion partnership. The service creates a three-tier pricing structure targeting real-time applications like voice AI, financial research, and incident response.
IBM and startup Together AI announced a multi-year $240 million agreement to build a large-scale artificial intelligence inference cluster on IBM Cloud using Nvidia systems. The deployment will feature Nvidia's HGX B300 systems and is scheduled to launch in Q1 2027.
IBM and startup Together AI announced a multi-year $240 million agreement to build a large-scale artificial intelligence inference cluster on IBM Cloud using Nvidia systems. The deal reflects IBM's bet that enterprise AI inference—the process of running trained models at scale—represents a major growth opportunity.
AMD announced the acquisition of Toronto-based AI chip startup Taalas for an undisclosed sum on August 6, 2026. Taalas specializes in etching AI model weights directly into silicon to enhance inference performance, part of AMD's strategy to challenge Nvidia's dominance in AI hardware.
AMD announced the acquisition of Toronto-based AI chip startup Taalas for an undisclosed sum on August 6, 2026. Taalas specializes in etching AI model weights directly into silicon to enhance inference performance, part of AMD's strategy to challenge Nvidia's dominance in AI hardware.
VAST Data announced an expanded collaboration with AMD to integrate its AI Operating System with AMD's 6th Gen EPYC CPUs, Instinct GPUs, and Pensando AI NICs, targeting AI cloud providers and enterprises building high-performance AI infrastructure for inference workloads.
Cognition released SWE-1.7, its most capable AI model to date, which the company says performs within a few points of leading frontier models while offering significantly reduced costs and 1000 tokens per second speed. The company reports continued gains from scaling reinforcement learning.
Etched, a startup developing specialized AI inference chips for Transformer models, announced it has secured $1 billion in contract orders and raised $800 million in funding, reaching a $5 billion valuation after TSMC successfully manufactured its Sohu chip.
OpenAI has reportedly developed a system-level optimization method that cuts the inference costs of running its AI models by approximately 50%, a breakthrough that could significantly improve the economics of operating large language models at scale.
Etched, a company designing chips specifically for running AI models, has raised $800 million from investors including Jane Street and TSMC-linked VentureTech Alliance, while announcing it has secured over $1 billion in signed customer contracts.
OpenAI announced its first custom-built AI inference processor called Jalapeño, designed in partnership with Broadcom and optimized for running large language models. The chip is expected to be deployed at scale by late 2026 to power ChatGPT and other AI products.
Rebellions, which designs chips specifically for AI inference, raised $400 million at a $2.3 billion valuation ahead of a planned IPO later this year, positioning itself as another challenger to Nvidia's dominance.
Google released TurboQuant, a compression algorithm that reduces large language model memory usage by at least 6x while improving performance, targeting AI inference cost reduction. The announcement caused significant market impact, with memory chip maker stocks reportedly declining.
