A US-China AI hotline is proposed, two Chinese labs ship models, and $300bn of AI guarantees stay off the balance sheet
A daily brief on the US proposal to notify Beijing when AI incidents reach national security, the two Chinese models released on one Sunday, Jensen Huang's case against the alarm, and the $300bn of guarantees financing the build-out.
This edition covers AI developments published between September 20 at 08:15 and September 21 at 08:15 in Dhaka. Sunday produced two model releases, both from Chinese labs, and a United States proposal to tell Beijing when an AI incident crosses a national-security line. Nvidia's chief executive used a television interview to argue there is nothing to be afraid of. The money that funds the build-out moved as well: SoftBank sold bonds to pay for its OpenAI stake, and the Financial Times counted $300bn of AI exposure that sits behind guarantees rather than on a balance sheet.
Quick scan
| Development | What changed | Why it matters | What to watch |
|---|---|---|---|
| An AI notification channel | The US proposed a mechanism for Washington and Beijing to notify each other of AI incidents that reach national-security level, for Trump and Xi to consider at their summit this week. 1 | It would be the first AI arrangement settled between two heads of state, and it obliges neither side to change a rule or a single chip export. 1 | Whether Thursday's summit names a mechanism or produces a sentence in a communiqué, and whether Beijing answers at all. |
| Two Chinese model releases | Alibaba's Qwen team published Qwen-Image-2.1 under a research-only licence, and StepFun opened API access to Step 5 Preview, a 600B-parameter model. 23 | The price of frontier-adjacent capability keeps falling, while the licence on one of these got narrower than its predecessor's. 4 | Whether StepFun ships weights on October 15, and whether Alibaba publishes commercial terms. |
| Nvidia's counter-argument | Jensen Huang said on CBS that there is a "0% chance" AI ends the world, called the warnings unscientific, and said existing liability law is enough. 5 | He will be at the White House dinner when Trump hosts Xi, and wants to discuss global AI standards with Xi. 5 | Whether chip export rules move after the summit; the US trade representative kept them off Sunday's agenda. 1 |
| Bonds and guarantees | SoftBank launched $10bn and €1bn of notes for its OpenAI stake, and the FT counted up to $300bn of residual value guarantees issued in under a year. 67 | The guarantee is called only when the asset stops being worth what the borrower promised, and then the guarantor absorbs the shortfall. 7 | SoftBank prices on September 24 and settles on September 29. 6 |
| A third memory supplier | CXMT said its fifth-generation DRAM platform entered mass production, with dies per wafer at least 50% above its previous generation. 8 | Memory has been the tightest part of the AI supply chain; DRAM revenue rose 59.5% in one quarter against lagging supply. 8 | Whether the platform reaches server parts; the first products announced are for phones. |
| Auditing model intent | Two preprints in Monday's arXiv listing read a model's internal state to separate a model that will not answer from one that cannot. 9 | Sandbagging makes every other safety evaluation inconclusive, and both papers test ways to detect it without the model's cooperation. 9 | Whether the method transfers to closed frontier systems; it was tested on eight open models. |
An AI hotline is put to Washington and Beijing
US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng concluded talks in New York on Sunday, and the American side came out of them with a proposal for their presidents: a mechanism through which each government would notify the other of AI-related incidents that rise to a national-security level. 1
Bessent told reporters the two sides discussed setting up a new US-China AI dialogue, with particular emphasis on national security, and that the notification system would cover common goals and common threats. "We think that, just like with any cross-border activity, that moving from opaque to more transparency between the number one and the number two AI powers in the world is very important," he said. 1
An earlier attempt at the same thing went nowhere. Trump and Xi first discussed potential consultations on AI development in May, in Beijing, and that forum was never formalized. 1
The Chinese response to the new proposal is not yet known. He and China's chief trade negotiator, Li Chenggang, left the talks at JPMorgan Chase's headquarters in New York without speaking to reporters. 1
Two details of the agenda bound what the mechanism can be. The US Trade Representative, Jamieson Greer, said export controls on sophisticated AI chips and semiconductor manufacturing equipment were not part of the AI mechanism talks. 1 And Greer described Sunday's session as groundwork for a "successful summit" between Trump and Xi on Thursday and Friday in Washington. "These are the two most powerful countries on earth, whether it's in trade or AI, and it's imperative that they're able to work together," he said. 1
The rest of the meeting was trade business that has been open since May. The two sides discussed "operationalizing" a process agreed in Beijing for each to identify potential tariff cuts on non-strategic goods, called the Board of Trade. Greer did not name products, saying he wanted a relatively small subset of goods that can be traded in a balanced way, and mentioned consumer and low-tech goods from China and energy, agricultural goods and medical devices from the US as candidates. He and Bessent reported no progress on the flow of critical minerals to US companies, and no updates on Chinese pledges to buy $17 billion more in American agricultural goods a year and more than 200 Boeing aircraft. 1
The summit lands in the middle of the UN General Assembly's high-level week, where the risks of artificial intelligence are among the subjects almost 130 heads of state are bracing to discuss. 10 Secretary-General António Guterres has said AI will be a major topic, and that countries pushing its frontiers should set up "common guardrails in order to avoid this race to the bottom that could lead one day to a gigantic disaster at the global level." The UN created an advisory body on the technology in 2023. 10 Diplomats told Reuters the Security Council may hold high-level meetings on Ukraine, AI and the Middle East, and Sam Altman is due to brief the Council during the gathering. 1011 Trump has dismissed the risks around AI as a "hoax," saying there is a "sick conspiracy" against the technology and that China would benefit from the doubt being cast on it. 10
The domestic version of the same argument arrived on television. Representative Ro Khanna, the ranking Democrat on the House select committee on China, told CBS's "Face the Nation" that Trump and Xi have to reach a deal covering AI safety, and named two principles he would want in it: a ban on recursive self-improving AI, and safeguards so that AI cannot be misused for biological or nuclear weapons. 1
A notification mechanism is the least demanding form of coordination available between rivals. It obliges each side to say when something has gone wrong, and obliges neither to change its rules, its spending or its chip exports, which is why the export controls stayed off the agenda in New York.
What to watch: whether the summit produces a named mechanism rather than a sentence in a communiqué, and whether Beijing responds to a proposal it has not yet commented on. 1
Two Chinese labs shipped models on the same Sunday
Alibaba's Qwen team published Qwen-Image-2.1, one model that both generates and edits images. The visual generation component has 7 billion parameters across 32 single-stream DiT layers, and the release combines text-to-image generation with editing in a single set of weights: it generates regular or transparent images from text, edits transparent layers, and extracts subjects from photographs. The editing side accepts up to ten reference images and takes local instructions drawn as circles or painted annotations or supplied as separate masks, with identity preserved for people and products. Output is native 2048×2048 at 40 inference steps, and the supported aspect ratios run from 1:1 to 16:9 at 2752×1536. 2
The licence is the part a commercial team has to read first. The model card places Qwen-Image-2.1 under the Qwen Research License Agreement. The earlier Qwen-Image model, released in August 2025, was licensed under Apache 2.0. 212
StepFun released Step 5 Preview the same day, with API access live from the start. It is a sparse mixture-of-experts model with about 600 billion total parameters that activates roughly 27 billion per token, and it takes a million tokens of context at once. StepFun says the weights will open on October 15. 3
Artificial Analysis has the measurements. Step 5 Preview scores 44 on the firm's Intelligence Index, against a median of 24 for comparable models, and it is priced at $1.00 per million input tokens and $2.70 per million output tokens with a 95% cache discount. Every task on that index cost $0.72, and the whole run of the index cost $924.68. The model is fast at about 100 output tokens per second, and it is verbose: it produced 160 million output tokens across the index against a median of 92 million, which is what separates the price per task from the price per token. 4
For anyone choosing a model to build on, the decision now turns on two numbers moving in opposite directions. Capability per token is rising, and on Qwen-Image-2.1 the licence narrowed from the previous release in the same family, so a company that wants to ship generated images commercially needs an agreement with Alibaba rather than a compliance note. 212
What to watch: whether StepFun opens the weights on October 15 as promised, and whether Alibaba publishes commercial terms for Qwen-Image-2.1 or leaves the model inside its own products. 23
Nvidia's chief executive says there is nothing to be afraid of
Jensen Huang told CBS News he disagrees with the warnings from AI researchers that the technology could end humanity within a few years, and called those predictions overblown "doomsday narratives." "2030 is not going to be the end of the world. There is 0% chance that's going to be the end of the world," he said. "Scaring people is unnecessary. It is irresponsible." 5
He was answering Jacob Coxon, a former Anthropic researcher who posted earlier in September that AI developers "earnestly believe that it could kill us all by the end of the decade," remarks that sparked heated public debate about AI risk. Huang said such predictions are "not grounded in science." 5
Nvidia has a direct stake in the answer. Its chips run the large language models at issue, and demand for them has pushed the company's market value to $5.3 trillion, the highest of any company. Asked why readers should trust him on safety given that position, Huang said: "Our company's success is directly connected to the safe deployment of products and services. If we don't continue to do that, our value would be diminished." 5
His answer to the regulation question is that the law already covers this. Huang said existing US rules on product liability and unauthorized entry are sufficient, and that he agrees with Trump that the industry needs no additional guardrails: "You have all kinds of liabilities associated with cybersecurity" and damage liability laws. "Apply that first — don't let this doomsday narrative allow someone to relieve them of the laws that currently exist." 5 He argued that no new rules, laws or guidelines were needed even as several recent cases in which models escaped containment and hacked other companies were still being reported. 13
On chips and China he took the position opposite the export controls: "Every single chip company should go and serve the world, compete for the world," he said, adding that he wants to speak with Xi about setting global standards for AI development when the two meet. Huang said he will attend a White House dinner on September 24 at which Trump hosts Xi. 5
He also addressed the communities that have been voting against data centres, at a moment when that opposition has reached ballots in several states. "I would say we're sorry we didn't come talk to you sooner," Huang told CBS. "We need to go engage the communities long before we set up the data centers." He said data centres have become more energy-efficient and use less water, and that building them brings back construction and manufacturing work. 5
What to watch: whether chip export rules move after the summit. Huang wants the market opened and Greer kept the controls off Sunday's AI agenda, so the meeting is the first test of which position the White House is carrying into the talks. 15
SoftBank's bonds, and $300bn of guarantees
SoftBank Group launched $10 billion and €1 billion of senior unsecured notes on Monday to fund its investment in OpenAI. The dollar notes are split into 3.5-year, 5.5-year and 7.5-year tenors, and the euro notes into four-year and six-year tenors. 6
The proceeds pay for the third tranche of SoftBank's follow-on investment in OpenAI, which is expected to close on October 1, and they replace the $10 billion bridge loan the company had taken out for the same purpose. Pricing is expected on September 24 and settlement on September 29, with Citigroup and JPMorgan as lead bookrunners. 6
The same day, the Financial Times counted a different kind of money. Big Tech companies have issued up to $300bn of residual value guarantees in less than a year while recording little of that exposure on their balance sheets. A residual value guarantee is a promise that chips or a data centre will still be worth a stated minimum if they have to be sold or re-let; the guarantee lets a bank lend against infrastructure owned by a special-purpose vehicle rather than by the technology company itself. Meta used the structure first on a large data centre project last autumn, and Broadcom and Nvidia have since adopted it. 7
| Guarantor and counterparty | Exposure | What it backs |
|---|---|---|
| Broadcom to Anthropic | $29bn, taken on in June | 1GW of chips sold to a special-purpose vehicle that leases them to Anthropic, part of a vendor-financing programme Broadcom and Google assembled; Broadcom's quarterly filing showed the credit support had little impact on its own balance sheet. 7 |
| Nvidia to SB Energy | $105bn | A large data centre campus in Ohio for OpenAI, built by a SoftBank subsidiary; Nvidia records no balance-sheet liability until OpenAI's leases begin in 2028, and in exchange the campus runs Nvidia hardware exclusively for 20 years. 7 |
| Nvidia to a Goldman Sachs-led investor group | Up to 25% residual value support | Deals aimed at mustering $500bn of capital for Nvidia-powered AI infrastructure. 7 |
Borrowers are paying for that credit standing. People familiar with the deals told the FT that guaranteed projects typically price at a premium of 100 to 150 basis points over the guarantor's own debt. 7 Doug Colandrea, a senior director at the credit rating firm KBRA, told the FT: "There has been a significant expansion in off-balance-sheet exposure over the past year. That adds a very high degree of complexity to their credit risk profiles." Tim Musial, head of fixed income at CIBC Private Wealth US, described the demand side: "There is concern about all these future commitments because everyone is in a race to get the compute." 7
The $300bn covers residual value guarantees over 12 months. A wider measure from Morgan Stanley counts more than $3.1tn in off-balance-sheet commitments and credit support across seven hyperscalers and chipmakers. 7
Broadcom's answer to the risk is that the risk is small. The company told investors that "the strong profitability trajectory of the leading frontier AI labs and the sustaining value of the underlying assets" meant there was a low probability of a guarantee being triggered. People familiar with its plans said it will probably offer guarantees for another 5GW of chips for Anthropic and 1.3GW of custom chips for OpenAI next year. 7
A guarantee is called when usage disappoints, computing power ends up in oversupply, or the business model does not arrive, and when that happens the company that gave the guarantee covers the shortfall. Morgan Stanley and Goldman Sachs bankers arranged the deals the FT examined. 7
What to watch: whether the guarantees start appearing in these companies' disclosures in a form a reader can total, now that Morgan Stanley's tally exists to compare later filings against. 7
A third memory supplier reaches mass production
ChangXin Memory Technologies said on Sunday that its fifth-generation DRAM platform has entered mass production. The company announced it at the 2026 World Manufacturing Convention in Hefei, its home city, where it showed two LPDDR5X products built on the platform: both 24Gb, 50% more capacity than comparable parts of the previous generation, in 496-ball and 245-ball packages for mid-range and high-end smartphones and portable consumer electronics. Both are in mass production. 8
The process numbers are the interesting part. The G5 platform has an active-area half-pitch of 11.95 nm in the memory arrays, reached with quadruple patterning; an array capacitor aspect ratio of 45:1; a DRAM-optimized high-k metal gate process; and a core cell array height of 6,762 nanometers. CXMT says the number of dies it gets from each wafer is at least 50% higher than on its fourth-generation platform, with better energy efficiency and cost control. 8
The baseline is a market where supply has not kept up. TrendForce reported that global DRAM industry revenue rose 59.5% quarter on quarter in the second quarter of 2026, with supply expansion continuing to lag demand growth. The market is still dominated by Samsung Electronics, SK hynix and Micron, as the industry analyst Ma Jihua told the Global Times. 8
The two products CXMT announced are phone parts. The announcement does not cover the server memory or the high-bandwidth memory that AI accelerators draw on, so this does not relieve the AI bottleneck in itself. What it establishes is a third producer working at a leading-edge pitch, which is the condition for competing for the memory that AI buys. 8
What to watch: whether CXMT's next step takes the G5 platform into server parts, and what DRAM contract prices do over the next two quarters against a 59.5% quarterly revenue rise. 8
Two papers on reading what a model knows
Monday's arXiv listings added 125 papers to the artificial intelligence category. Two of them take on the question the industry has been arguing about in public this month: whether an outsider can establish what a model knows and means. Both were submitted on Friday and announced in Monday's listing, and neither has been peer-reviewed. 914
The first is a test for hiding. Hiskias Dingeto's "A Lie Detector Test for Language Models" borrows the Concealed Information Test, the forensic method that identifies guilty knowledge by showing a suspect the true detail among plausible decoys and measuring a stronger reaction to the item they recognise. The model version presents a question with its candidate answers and reads, from the model's internal states, which candidate the model recognises as correct. It needs no reference model and no labelled truth corpus. 9
On its measurements, the method recovers the recognised answer at 0.70 to 0.87 balanced accuracy across eight models from five families — Gemma, Qwen, Llama, Mistral and Phi — against 0.28 to 0.40 for items the model never knew and a chance rate of 0.25. Recognition held between 0.85 and 0.93 across prompted deception, trained sandbagging, and password-locked and circuit-broken checkpoints. Where unlearning removes the knowledge, recognition falls to the level of a question the model never knew. 9
The second paper is about maps. "World Modeling in Transformers," by Pierre Beckmann, Matthieu Queloz and André Freitas, studies TaxiGPT, a transformer trained on random walks through Manhattan whose failures had been read as evidence that it holds no coherent internal map. Through mechanistic analysis and causal interventions, the authors show the model does represent intersections and streets, tracks its own position, and uses a goal compass to navigate. Its errors trace back to interference between superposed intersection features that disrupts localisation, and grouping intersections that share the same legal moves limits the damage. The authors propose indicators that show world-modeling capacities emerging at different stages of training. 15
The sandbagging result is the one with an immediate use. A model that performs below its ability on an evaluation returns a meaningless score, which makes every safety test that relies on the model's cooperation inconclusive, and the value of the Dingeto method is that it separates a model that will not answer from one that cannot. The eight models tested are open models from mid-size families; the paper does not address closed frontier systems. 9
What to watch: whether anyone outside the labs can apply this to a frontier model. The method is reference-free by design, and the material to run it on is a set of open checkpoints, which is the opposite of the systems the argument is about. 9
In brief
- ChatGPT's ad pixel links browsing to accounts. An independent write-up describes a cookie called
__obi, set atbzr.openai.comand scoped to.openai.com, minted while the reader is on ChatGPT from a signed token that binds a 22-character identifier to their account subject, withSameSite=noneand a one-year lifetime. Any advertiser that embeds OpenAI's measurement pixel on its own site then reports back to OpenAI, including page content, the products being viewed and purchase behaviour. The author says he reproduced the mechanism on his own phone using two independent capture methods, and cross-checked it against several months of observed traffic covering 936 distinct advertiser pixels across 1,029 hostnames. 16 - The specialists guarding power grids worry about people. The Verge asked three of them about the gap between rogue-agent stories and the grid. Joshua Corman of the Institute for Security and Technology said of generative AI in an attacker's hands: "It's literally any sociopath that wants to [attack] is now more powerful than they used to be." Sophie McDowall, a research associate at the Foundation for Defense of Democracies, named speed as the difference: "it's letting adversaries move more quickly — but it's very challenging for those defending the infrastructure to match that pace." Much of the equipment was never designed to connect to the internet, the average US nuclear reactor is about 44 years old, and some operational-technology systems are built to take updates once a quarter or once a year. OpenAI pledged $1 billion in September toward subsidizing training and access to models meant to defend critical infrastructure. 17
What to watch
- Thursday and Friday. Trump hosts Xi Jinping in Washington, with the AI notification proposal on the table and chip export controls off the agenda. 1
- During the General Assembly. Sam Altman briefs the Security Council, and Guterres pushes the guardrails language against a White House that calls the risk a hoax. 1011
- September 24 and 29. SoftBank prices its notes, then settles them, funding the third tranche of its OpenAI investment on October 1. 6
- October 15. StepFun opens the weights for Step 5 Preview, which turns an API-only model into one anyone can download. 3
- The Qwen licence. Whether Qwen-Image-2.1 gets commercial terms, or whether Alibaba keeps the new generation inside its own products, having licensed the previous one under Apache 2.0. 212
- Memory supply. Whether the G5 platform reaches server parts, against DRAM revenue growth of 59.5% in a single quarter and three incumbents still setting the price. 8
References
- 1
- 2Qwen/Qwen-Image-2.1 model card
huggingface.co
- 3
- 4Step 5 Preview — Intelligence, Performance & Price Analysis
artificialanalysis.ai
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12Qwen/Qwen-Image model card
huggingface.co
- 13
- 14arXiv cs.AI recent submissions
arxiv.org
- 15World Modeling in Transformers
arxiv.org
- 16
- 17
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.