
AI News Top 10
16 November 2026
Congress moves to tighten AI chip export rules as Nvidia’s leadership crisis deepens, and Anthropic enters its IPO week with 901 billion dollars in commitments.
Congress moves to tighten AI chip export rules after Nvidia scandal
Republicans and Democrats both want to restrict AI chip exports further after Nvidia chief executive Jensen Huang resigned on 12 November. Senate majority leader Chuck Schumer demands that Huang and the Trump administration testify before Congress. Tim Scott of the Banking Committee sent letters to Huang and commerce secretary Howard Lutnick. The Remote Access Security Act, filed in April 2025 and dormant since, is back on the agenda, and House speaker Mike Johnson speaks of a growing consensus on limiting chip exports to China.
Musk and 165,000 signatories ask Nvidia to bring Huang back
An open letter with more than 165,000 signatories asks the Nvidia board to reinstate Huang while an independent investigation establishes the facts. Signatories include Elon Musk, Epic chief executive Tim Sweeney, YC president Garry Tan and Uber co-founder Travis Kalanick. More than 7,500 say they work at Nvidia, about 18 percent of staff. Treasury secretary Scott Bessent wants the opposite: a chief executive from outside, and answers within ten days on export compliance. Nothing above the H200 may go to China.
Amodei flags huge uncertainty four days before Anthropic’s record IPO
At a JPMorgan investor event in New York, in conversation with Jamie Dimon, Dario Amodei called compute Anthropic’s deepest business dilemma. Capacity has to be bought one to two years ahead, before anyone knows whether revenue covers the cost. Bloomberg puts the company’s obligations at almost 901 billion dollars, mostly compute and chips, with about 71 billion due before September 2027. The listing on 19 November has to deliver the cash buffer that absorbs a mismatch between spending and revenue.
AI stocks slide ahead of Nvidia results and Anthropic listing
Global equities fell in a technology selloff as investors positioned for Nvidia’s quarterly results and Anthropic’s expected listing, two events that set the direction for the AI trade. Emerging markets dropped to their lowest level in a week. The Asian open ran under the heading of the Anthropic IPO and the Nvidia fallout, with chip suppliers and data centre builders among the heaviest decliners.
One compromised MCP server can aim an agent at the tools you trust
A review of the Model Context Protocol’s security history shows why: the specification shipped in November 2024 without an authorization model, gained OAuth 2.1 in March 2025, Resource Indicators against token misuse in June 2025, and mandatory iss validation in July 2026. Servers built in the first year still carry assumptions the spec has moved past. OWASP’s MCP Top 10 puts token mismanagement and secret exposure at the top. Gartner expects 40 percent of enterprise applications to contain task-specific AI agents by the end of 2026, against under 5 percent in January.
Cursor: stolen Fireworks AI key produced a trojanised Claude Code extension
The compromise of Anthropic’s official Claude Code extension in Cursor’s marketplace came from an access key Cursor had issued to Fireworks, released in the breach at Fireworks AI. ‘An attacker obtained an access key that we had issued to Fireworks,’ Cursor said in a Friday update. The key was used to copy data from a read-only storage bucket Cursor shared with Fireworks. The incident is a supply-chain attack through a trusted AI vendor, not through the extension itself.
Estonian premier: Europe has to build at the AI frontier itself
Kristen Michal told Rapporteur after the EU summit on AI in Dublin that share price swings are normal volatility, and that a correction is acceptable as it was after the dotcom bubble. He calls it of the greatest importance that Europe competes at the frontier instead of adopting technology built elsewhere: ‘everyone understands that we have lost time.’ With Finnish premier Petteri Orpo he published a programme days before the summit. Commission president Ursula von der Leyen announced a proposal to finance technological sovereignty; Mistral says it trained its Large 4 model on 3,800 specialised chips.
Gemini 4 Carbon tops MathArena at fifteen times Sol’s cost
Gemini 4 Carbon (high) scores 85.4 percent on MathArena, plus or minus 3.0, the first non-OpenAI model in months to take first place. Claude Fable 5.2 (max) follows with 82.8 percent and GPT-6.1 Sol (max) drops to 80.6 percent. A test run costs about 15.70 dollars on Carbon, 1.07 dollars on Sol and 40.20 dollars on Fable 5.2. Carbon runs into the time limit and almost never gives up; one script ran more than eleven hours without a progress check. Kimi K3.1 Pro from Moonshot stays the best open model at 54.3 percent.
Stanford open-sources Memorilla: 16 tokens carry a document memory
Memorilla compresses a document collection into 16 memory tokens for a frozen Qwen3-8B and reaches 84.33 percent on FactKG, with 108 times fewer document tokens than a full context window. The model itself is untouched: the memory sits in a separate latent module, so the context cost no longer scales with model size.
Daisytuner raises 6 million to run models on any chip
The Darmstadt deep-tech company Daisytuner, founded in 2024 by Lukas Trümper and Adrian Schmitz, raised a 6 million euro seed round led by Elaia and LEA Ventures, with the Amadeus APEX Technology Fund among others. The company automates performance engineering with a self-learning compiler and a patented optimisation platform, so the same code runs at full speed on the silicon a customer picks. With the funding it launches Daisyflow, a production inference engine on open frameworks that runs on-premise. Measured claim: Qwen 2.5 runs up to 18 percent faster than on vLLM on AMD Instinct GPUs.
Key Trends
- • Chip export policy now runs through people: the Nvidia crisis is a governance story, not a hardware story.
- • Compute commitments are signed years before the revenue that has to pay for them is known.
- • The agent attack surface sits in the tool and supplier chain, not in the model.
- • Memory compression starts to compete with large context windows on cost.
- • Inference efficiency spreads beyond Nvidia silicon, which is where the money is looking.
Wat vond je van dit bericht?