Sections
00:00 What we're covering today
00:24 1. US proposes sharing AI incident alerts with China
01:22 2. CXMT says its new memory platform produces 50% more chips per wafer
02:21 3. SoftBank launches dollar and euro bonds to fund its next OpenAI payment
03:18 4. Anthropic is testing AI ideas in a Bay Area biology lab
04:13 5. GitHub will retire six Copilot models on October 19
05:01 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Monday, September 21, 2026. Today, we're looking at a proposed safety channel between Washington and Beijing, a Chinese memory maker's production claims, and SoftBank's borrowing plans.

We'll also cover Anthropic's biology lab and an October deadline for teams using GitHub Copilot. Narration uses an A I generated voice.

U S proposes sharing A I incident alerts with China. Treasury Secretary Scott Bessent said on September 20 that the U S had proposed an A I incident notification system to China. Reuters and AP reported his remarks after he met Chinese Vice Premier He Lifeng in New York. The proposal would cover incidents serious enough to affect national security. Bessent said the two countries also discussed more A I talks before the planned Trump-Xi summit. For now, it's still a proposal. The reports don't show that either country has started using the system or agreed to binding reporting rules. A summit statement could show whether both governments support it. Written rules could explain which incidents qualify, how quickly each country must send an alert, and who receives it. A joint exercise or a documented incident would show whether the system works. Another promise to keep talking wouldn't.

CXMT says its new memory platform produces 50% more chips per wafer. The Chinese memory maker announced G5 mass production on September 20. CXMT says the platform produces at least 50% more gross chips per wafer than its previous platform, based on an 8-gigabit comparison. It also introduced two 24-gigabit LPDDR5X products for phones and portable devices. Reuters reports that CXMT is seeking a stronger position against established global memory suppliers. But CXMT is counting chips that may fail testing. The company didn't say how many usable chips come from each wafer, and it provided no independent evidence for its cost claims. These products are phone memory. Their release doesn't prove that CXMT can supply the high-bandwidth memory used with leading A I accelerators. Customer shipments, independent tests, and usable yields could show whether buyers have a reliable alternative. More chips on a wafer matter only if enough of them work.

SoftBank launches dollar and euro bonds to fund its next OpenAI payment. SoftBank launched $10 billion in dollar bonds and 1 billion euros in bonds. Reuters reported the launch on September 21 in Asia, when it was still September 20 Pacific time. A term sheet says SoftBank will use the money for a $10 billion payment to OpenAI and general corporate purposes. The payment is expected to close on October 1. The bonds would replace a bridge loan facility. SoftBank has launched the offering, but investors have not priced or settled the bonds. Reuters says pricing is expected on September 24 and settlement on September 29. SoftBank must repay the debt even if its investment disappoints. The final interest rates and maturity terms will show what investors charge SoftBank to borrow. Settlement would confirm the bond sale. The October payment would confirm that SoftBank paid OpenAI.

Anthropic is testing A I ideas in a Bay Area biology lab. Anthropic's head of life sciences confirmed the company's Bay Area wet lab in a Reuters report published September 18. Anthropic runs physical experiments there and through outside partners. A spokesperson said the lab isn't specifically for drug discovery. Anthropic also hasn't publicly described all the research it plans to run there. The lab can test whether an A I prediction holds up in a real experiment. But its existence alone doesn't show that an A I system discovered a treatment or improved patient outcomes. Anthropic has confirmed the facility, but its scientific results still need their own evidence. Outside researchers could assess the work if Anthropic publishes its methods and results. Other labs would still need to reproduce the findings. That would tell us more than the existence of the lab.

GitHub will retire six Copilot models on October 19. The company announced the change on September 18. The models are GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash, and Grok 4.5. GitHub lists newer alternatives. Business and Enterprise accounts will get access under the default model settings unless an administrator has disabled it. This is GitHub's announcement about its own service, and the suggested replacements don't promise identical behavior. Teams with restricted model policies may need an administrator to enable a replacement before users can switch. Teams can test an available replacement on their normal coding work before October 19. That test works only if their account allows access to the model.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.