Synopsys and Amazon signed a multi-year agreement valued at more than $1B under which Synopsys supplies silicon IP, EDA tools and agentic AI engineering software for Amazon’s custom chips, on a license-plus-royalty model tied to production volumes. Synopsys will also adopt Amazon EC2, Amazon Bedrock and cloud storage, and the companies will work to accelerate multiphysics solutions on Amazon’s Trainium and Graviton processors. The royalty structure ties a portion of Synopsys revenue to the number of Amazon chips produced, not to design licenses alone. (Synopsys)
Private Companies
Basetenpartnership
Baseten became one of the first open-model inference providers in OpenAI’s B2B Marketplace, announced September 29. OpenAI enterprise customers can apply existing OpenAI spending commitments to open models that Baseten serves, inside Codex or through the Responses API, and route each task to the model with the best mix of cost, quality and speed. Baseten’s inference capacity now sits behind OpenAI’s own enterprise billing relationship, which removes a separate procurement step for open-model workloads. (Baseten)
OpenAIother
OpenAI’s annualized revenue run rate is approaching $70B, up more than 70% since the start of the third quarter and from more than $40B reported in August, according to Axios. Enterprise sales more than doubled over the same period, and consumer revenue in the third quarter alone exceeded all of 2025. A run rate multiplies one month by 12, and it still gives the sector’s compute commitments a revenue base to be measured against. (Bloomberg)
Efficient Computerfundraise
Efficient Computer closed a $97M round led by TQ Ventures, with Eclipse, Union Square Ventures and Toyota Ventures participating, about seven months after its $60M Series A in February. The company makes the Electron E1, a spatial-dataflow processor that cuts data movement and architectural overhead, and says its design uses up to 100x less energy than x86 in battery-powered robots and drones. E1 is already shipping, so the capital funds production and volume. At the edge, the energy budget decides which architecture gets designed in. (SiliconANGLE)
Public Markets
Nvidia added $150B to its share repurchase authorization, leaving $235B through fiscal 2028. The Financial Times reported, as summarized by BigGo Finance, that it is also in early talks with insurers, including broker Howden Re, to cover shortfalls on GPU-backed neocloud loans, so GPU resale value now sets how far outside financing can go. (Nvidia)
Emerging
Photonics and heat: A paper from Seongwon Yoon, Pin-Jun Chen and Shimeng Yu, Thermal Tuning Overhead in Wafer-Scale Optical Interconnects for LLM MoE Training, finds that transient temperature swings in a 3D GPU and photonic stack outrun the control loops that hold microring resonators on wavelength, causing repeated tuning stalls during communication phases. Replacing thermal tuning with ferroelectric electro-optic tuning removes the stalls and yields 2.7x to 3.8x speedups across Mixtral 8x7B, Qwen-MoE 14.3B and LLaMA-MoE 6.7B in the authors’ simulations. Thermal control now belongs on the list of problems optical-interconnect suppliers must solve alongside link bandwidth. (arXiv)
Design starts: At TSMC’s OIP symposium, Li-Pen Yuan, vice president of business development, reported a 4x increase in new tape-outs, measured by the second year of a new technology across N5, N3 and N2. L.C. Lu, senior fellow, said inference token usage has grown 500x over three years and called inference the growth driver for AI. Tape-outs are the earliest measurable signal of leading-edge wafer demand, and both figures point up. (Semiconductor Engineering)