OpenAI released its GPT-5.6 lineup — Sol, Terra and Luna — on Friday, but only to roughly twenty partners whose access had been cleared in advance with the Trump administration. The company called Sol its strongest model yet, with major gains in coding, biology and cybersecurity, and said it intends to widen access "in the coming weeks." For now, though, the rollout looks less like a product launch and more like an export-licensing process.
The request came from the White House two weeks after the Commerce Department ordered Anthropic to pull its Fable 5 and Mythos 5 models offline entirely, over fears that their cybersecurity capabilities could be misused if accessed by foreign nationals. Officials reportedly view GPT-5.6 Sol as carrying similar risk. OpenAI says the model is "better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks," and that it stays below the company's own "Cyber Critical" threshold.
In its own blog post, OpenAI pushed back gently on the precedent being set: "We don't believe this kind of government access process should become the long-term default." No formal review framework exists yet — the administration's executive order from earlier this month only asks labs to voluntarily submit frontier models for assessment 30 days ahead of release. Commentators have started describing the result as a strategic-materials regime applied to software: the most capable models are gated less by subscription tier now than by government comfort.
Google told Meta around March that it could not meet the full Gemini compute quota Meta wanted to purchase, according to the Financial Times. Other Google Cloud customers were affected too, but Meta — by far Google's heaviest Gemini user — took the brunt of it, and has since told staff to use AI tokens more efficiently.
The episode lands alongside Google Cloud's own numbers: revenue hit roughly $20 billion last quarter, yet CEO Sundar Pichai has said compute constraints capped even faster growth and contributed to a backlog that has nearly doubled. The takeaway for builders is blunt — even the companies selling the picks and shovels are short of capacity right now.
ChangXin Memory Technologies has signed a long-term agreement to supply Tencent with server DRAM worth more than 20 billion yuan, running anywhere from three to five years depending on the source. The deal arrives just as CXMT prepares to raise 29.5 billion yuan on Shanghai's STAR Market, and gives the Hefei-based chipmaker one of its strongest commercial endorsements yet.
DRAM has been in a severe global squeeze — contract prices jumped roughly 95% quarter-on-quarter in early 2026, and UBS expects the upcycle to run until at least late 2027. CXMT's revenue rose 700% year-on-year on the back of it. The Tencent contract is also a quiet signal of China's push toward a domestic memory supply chain that can hold up under continued US-China tech restrictions.
NVIDIA's RTX Spark fuses a 20-core Arm CPU with a Blackwell GPU and up to 128GB of unified memory on a single chip, connected over NVLink-C2C. NVIDIA says it can run large local agents with context windows stretching to a million tokens, render huge 3D scenes, and edit 12K video — all without a cloud round trip. "For forty years, you launched apps. Click. Type," said NVIDIA CEO Jensen Huang. "With RTX Spark and Microsoft Windows, you ask — and the PC does the work."
Microsoft is building native Windows agent support around the chip, including new security primitives and an "OpenShell" framework so local agents only touch the data and tools a user has explicitly approved. Dell, HP, Lenovo, Asus, MSI and Microsoft's own Surface line will ship RTX Spark laptops and small desktops starting fall 2026. Analysts cited by Reuters caution that premium pricing and the ongoing memory shortage could limit it to a niche category rather than a mass upgrade cycle.
Microsoft Scout is the company's first entry in a new category it calls Autopilots: agents with their own Entra identity that act continuously in the background rather than waiting for a prompt. It is grounded by Work IQ, a workplace intelligence layer that reads signals from email, chats, files, meetings and calendars to figure out what a person actually cares about. "Work IQ is the intelligence layer that understands your data, your tools, and your organization," said Charles Lamanna, Microsoft's EVP of Copilot, Agents and Platform.
Scout is built on the open-source OpenClaw framework, with Microsoft contributing governance and policy-conformance code back upstream. It is currently in Frontier preview, with general availability targeted for October 2026, and Work IQ's own APIs went generally available on June 16 — meaning any third-party agent built on Copilot Studio or Foundry can now tap the same organizational context layer Scout uses natively.
Eighty years after Penn researchers built ENIAC, the world's first general-purpose electronic computer, a different Penn team led by physicist Bo Zhen has demonstrated all-light switching using "exciton-polaritons" — hybrid quasiparticles formed when photons couple strongly with electrons inside an atomically thin semiconductor. The switching used about 4 quadrillionths of a joule of energy, far less than it takes to briefly power a small LED.
Most photonic AI chips already use light for raw speed, but still have to convert optical signals back into electrons for any decision-making step — a costly round trip. If the Penn approach scales, it could let chips process light directly from cameras, cut the power bill of large AI systems, and even support basic quantum computing functions on the same hardware.