
AI News Top 10
18 November 2026
Nvidia reports record AI chip revenue, ends H200 sales to China and loses its chief executive, while OpenAI’s new flagship draws harsh reviews from power users.
Nvidia doubles revenue as board refuses to reinstate Huang
Nvidia reported $115.84 billion in revenue for its third quarter, 103 percent more than a year earlier and 20 percent above the previous quarter. Earnings came in at $2.68 per share against a consensus of $2.47. It is the first quarterly report since founder and CEO Jensen Huang resigned last week after reports about several affairs. Board chair Harvey Jones said the board does not intend to reinstate Huang as CEO. The share is down 10 percent since his departure, leaving a valuation of $3.7 trillion against a peak of $6 trillion in October. Huang appears before three House committees in early December, and the SEC is examining the China revenue.
Nvidia ends H200 sales in China and sets up export committee
Acting CEO Colette Kress said Tuesday that Nvidia no longer sells H200 chips in China, ending the exception Huang had negotiated with President Trump: delivery through the US at a 25 percent tariff. On the same day the company set up a board-level committee for national security, export control and trade risk. New rules took effect Monday that keep advanced US chips out of foreign data centres serving Chinese customers remotely. Jefferies estimates those rules could cost $44 to $62 billion of data centre revenue next year.
OpenAI launches GPT-6.2 Astra and Sol to harsh reviews
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday. Astra costs $8 per million input tokens and $40 per million output tokens, 20 percent below its predecessor, and ranks second on the Artificial Analysis Intelligence Index; Sol stays at $2 and $10. Investor Matt Shumer calls the models essentially unusable. Developer Theo Browne says Astra is the worst model OpenAI has ever released. Ethan Mollick found it very uneven and weird, and hard to steer. Dan Shipper of Every reports that Astra is weak on coding benchmarks.
Microsoft calls Copilot the new OS for work at Ignite
At Ignite 2026 Satya Nadella announced general availability of Word, Excel and PowerPoint inside Copilot, which he calls the new OS for work: agents are steered from one place and applications run as plug-ins on a knowledge layer over mail, meetings, chat and files. Copilot Autopilot is a persistent cloud agent that gets roles instead of separate tasks and is billed through Copilot Credits. Copilot Workspaces makes AI interaction multiplayer. Every agent gets a unique ID with permissions and logging, and Nadella said trust cannot be assumed by definition.
Microsoft claims 96% of Opus 5.5 performance at 5% of the price
Microsoft introduced two coding models at Ignite, MAI Code 2 and MAI Code 2 Flash, plus MAI Work Flash for Excel formulas, Word documents and Teams. The claim is that MAI Code 2 reaches 96 percent of Claude Opus 5.5 at 5 percent of its price, measured on Microsoft’s own benchmarks with a 50 percent introductory discount. The models are available as an option in GitHub Copilot, Microsoft 365 and Fabric. ZDNET calculates that a party spending $1 million a day on inference would land at $100,000, and notes that the figures come from Microsoft itself.
Amazon says stolen AI keys power agents that hunt for more
Records from Amazon’s MadPot sensors between 3 October and 5 November show AI agents in a cyber campaign used stolen API keys to reach the models that helped them find more keys. Amazon estimates with moderate confidence that the agents ran on ARTEX, an open-source AI pentest framework that turns natural language into commands through external LLMs. The agents split planning, execution and verification and shared results in a knowledge graph. Targets were internet-exposed LLM gateways including LiteLLM, Dify and Langflow. The FBI reports privileged access within three hours, and South Korea counted 67,000 victims.
Anthropic crypto futures imply $1.7 trillion valuation
Crypto futures tied to the Anthropic listing fell along with AI shares but still imply a valuation of about $1.7 trillion. That is well above the $1.4 trillion the company would be worth at the top of its proposed IPO range. Bloomberg readers get only the summary of the article.
Perplexity adds a mode where sixteen agents check each other
Perplexity added a Multiagent mode to Computer, an effort level above Ultra in which a lead agent directs up to sixteen agents running on Claude, GPT and Gemini models. The agents share findings and check each other’s work before the result is delivered. The article page itself was behind a cookie wall, so this is the RSS summary.
Gemini now shows and edits what it remembers about you
Google gives Gemini users an overview of the facts the assistant has stored about them and the option to edit those entries. Turning off the Memory switch in the Personal Intelligence menu stops new facts from being recorded but does not delete what is already stored. Old facts have to be removed through the overview or by deleting conversations in the Activity menu. Management runs from the Gemini web app under Settings.
EU lawmaker weighs deleting the highest cloud sovereignty tier
Rapporteur reports that a key lawmaker on the EU cloud file is considering deleting the highest sovereignty tier from the rules. In the subtitle, ECR MEP Diego Solier warns that sustainability data is also needed to increase public trust. The article is Pro-only: headline, subtitle, author and date are verifiable, the body is not.
Key Trends
- • Nvidia posts a record quarter while its board fights off its own founder and the China business closes.
- • Microsoft moves from reselling OpenAI models to selling its own, at a fifth of the price it compares against.
- • Agent infrastructure becomes the target: gateways, keys and permissions are the new perimeter.
- • Three frontier labs ship flagship models within ten days of each other.
- • Memory turns into a visible and editable object in consumer assistants.
Wat vond je van dit bericht?