AI News Roundup: OpenAI Ships GPT-5.5, Google Unveils TPU 8 Chips, SpaceX Plans In-House GPUs

OpenAI Releases GPT-5.5, Its Fastest Model Turnaround Yet

OpenAI released GPT-5.5 on Wednesday, calling it the company’s “smartest and most intuitive” model to date. The launch comes just six weeks after GPT-5.4 debuted — an extraordinarily fast cadence that signals how aggressively frontier labs are iterating to win enterprise customers.

Internally codenamed “Spud,” GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise ChatGPT users immediately. API access is coming “very soon,” though OpenAI says it requires additional safeguards. Pricing lands at $5 per million input tokens and $30 per million output tokens — double GPT-5.4’s rate, but offset by tighter token usage and stronger reasoning per call.

OpenAI says the model excels at agentic coding, data analysis, software operation, online research, and document creation. The company evaluated it with nearly 200 early-access partners and calls this release its strongest set of safety guardrails to date.

Google Splits Its 8th-Gen TPU Into Two Specialized Chips

At Google Cloud Next 2026 in Las Vegas, Google unveiled a first-of-its-kind approach to its custom silicon: the TPU 8t for training and the TPU 8i for inference. It’s the first time Google has split its flagship accelerator into purpose-built variants rather than shipping a single general-purpose chip.

The TPU 8t scales to 9,600 chips in a single cluster with 2 petabytes of shared high-bandwidth memory, delivering nearly 3x the compute of the previous Ironwood generation. Google says it can cut frontier model development from months to weeks. The TPU 8i, meanwhile, tackles the “memory wall” with 288GB of HBM and 384MB of on-chip SRAM — three times its predecessor — allowing entire model working sets to reside on-chip for ultra-low latency in multi-agent workloads.

Google Cloud Next: $750M Agent Fund, Gemini Enterprise Platform

Beyond chips, Google Cloud Next was packed with agentic AI announcements. Google launched the Gemini Enterprise Agent Platform with tools for building, governing, and optimizing AI agents, including a dedicated agent inbox, long-running agents, and an Agent Designer. Sundar Pichai declared the industry has entered the “agentic era.”

Google also committed $750 million to its 120,000-member partner ecosystem to accelerate agentic AI development, and unveiled “Agentic Defense” — combining Google Threat Intelligence and Security Operations with Wiz’s cloud security platform for AI-powered threat detection and response.

SpaceX Reveals In-House GPU Manufacturing Plans Ahead of $1.75T IPO

SpaceX disclosed plans to manufacture its own GPUs in its draft S-1 filing with the SEC, listing GPU production among “substantial capital expenditures” as the company prepares for what could be Wall Street’s largest-ever IPO at a $1.75 trillion valuation.

The effort centers on Terafab, an advanced AI chip manufacturing complex planned for Austin, Texas, developed jointly by SpaceX, xAI, Tesla, and Intel. The facility aims to vertically integrate chip design, fabrication, packaging, and testing, targeting over 1 terawatt of annual compute output for both terrestrial and space applications. It remains unclear exactly when production will begin or how manufacturing responsibilities will be divided between the partners.

Mira Murati’s Thinking Machines Lab Lands Multibillion-Dollar Google Deal

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, signed a multibillion-dollar infrastructure deal with Google Cloud for access to Nvidia’s latest GB300 chips. The deal — valued in the single-digit billions — covers model training and deployment infrastructure, though it’s non-exclusive.

Murati left OpenAI in late 2024, launched Thinking Machines in February 2025, and raised a $2 billion seed round at a $12 billion valuation. The company’s first product, Tinker, automates the creation of custom frontier AI models. The Google deal makes Thinking Machines one of the largest cloud infrastructure customers outside the hyperscalers themselves.

Geoffrey Hinton Warns AI Is “A Fast Car With No Steering Wheel”

Nobel laureate Geoffrey Hinton delivered a pointed speech at the Digital World Conference in Geneva this week, calling for urgent AI regulation. “Unregulated AI is like a very fast car with no steering wheel,” Hinton told the audience at the event co-organized by the UN Research Institute for Social Development.

Hinton said it remains unclear whether humanity can coexist with superintelligent AI and warned that job losses from AI are accelerating — citing call centers as a sector where AI already matches human performance. He called regulation the “steering wheel” that the technology urgently needs, pushing back against industry voices who frame oversight as merely “a brake.”

ICLR 2026 Opens in Rio de Janeiro

The International Conference on Learning Representations (ICLR) 2026 kicked off today in Rio de Janeiro, with 3,462 papers accepted out of 11,617 submissions (a 29.8% acceptance rate). Two outstanding papers were recognized: “Transformers are Inherently Succinct” offers a new theoretical lens on why Transformers outperform RNNs, while WebDevJudge stress-tests LLM-as-a-judge approaches in open-ended web development.