NVIDIA Convinces 37 Companies to Join Forces in AI Security, But None of the Top Three Model Makers Show Up
1｜NVIDIA Enlisted 37 Companies for AI Security, None of the Big Three Came Along
NVIDIA, in collaboration with Microsoft, SpaceX, IBM, Palantir, CrowdStrike, and 32 other companies, established the "Open Security AI Alliance" to set industry standards for AI security deployment. NVIDIA provided the NOOA framework, while Microsoft contributed the MDASH vulnerability scanning system. The list includes chip manufacturers, cloud providers, security firms, as well as Hugging Face and Siemens, but OpenAI, Google, and Anthropic were all absent.
On the same day, Microsoft released the first cybersecurity-specific model, MAI-Cyber-1-Flash, claiming its benchmark score surpasses those of the big three model giants' systems at half the cost. 90% of vulnerabilities are automatically patched by it, with only the toughest 10% handed over to GPT-5.4.
The security alliance is led by companies that do not make models, while those that do opted not to participate collectively. The focus of AI security is shifting from "who oversees the model" to "who safeguards the individuals using the model." The latter clearly has no intention of waiting for the former to reach a consensus.
(Source: Tom's Hardware / CNBC / Microsoft Security Blog / The Register)
2｜Double Whammy for Chinese Chips: ChangXin Memory Skyrockets 466% on IPO Day, Domestic DUV Lithography Machine Enters Mass Production
ChangXin Memory debuted on the A-share market with an IPO pricing of around $8.6 billion, witnessing a staggering 466% surge on its first day, propelling its market value to about RMB 3.3 trillion (around $480 billion), making it the largest IPO in Asia in 2026. ChangXin is currently the world's fourth-largest DRAM manufacturer, holding a global market share of 7.67%. Founder Zhu Yiming pledged to donate stock worth around $5.6 billion to employees.
On the same day, a domestically-produced DUV immersion lithography machine in China announced its entry into the mass production stage. Despite the gap between it and ASML's EUV, it can cater to a large portion of the demand for mature process chips, which is precisely the area that needs to be filled in for a more independent supply chain.
The capital market and production line provided feedback on the same day. China's semiconductor self-sufficiency is no longer just a narrative in labs but has started to enter a stage where production capacity, talent incentives, and market pricing validate each other.
（Source: Bloomberg / Financial Times / Tom's Hardware）
3｜BlackRock Issues Debt for Meta Data Center, Investment-Grade Bonds Priced at Near-Junk Bond Yield
BlackRock priced a $12.55 billion bond issuance for Meta's data center at a 7.534% yield. In the investment-grade bond market, this level is uncommon and is nearing junk bond territory in terms of risk compensation. Despite weak syndicate demand, the bond rallied in early trading.
This transaction was more than just a large financing. It shifted the AI data center's balance sheet from a tech company's capital expenditure to bond investors' risk pricing. Previously, the market viewed AI infrastructure as cash flow investments by a few tech giants, but now asset management firms are mobilizing long-term debt for specific projects, packaging data centers as independently valued infrastructure assets.
The high yield also indicates that the market acknowledges Meta's credit but does not see AI infrastructure as a risk-free asset. Questions about the sustainability of computing power demand, the timing of project returns, and whether electricity and equipment costs will continue to rise are being factored into the financing cost upfront.
（Source: BlackRock）
4｜Claude User Chats Indexed by Google, Exposing Medical Records and Crypto Wallets
Anthropic's Claude chat platform experienced a privacy incident. Reddit users discovered hundreds of Claude shared conversations that could be publicly accessed through Google search without the original links. The exposed content includes medical records, student information, cryptocurrency wallet keys, and personal addresses.
The issue stemmed from public URLs generated by the "share" feature being crawled by Google's spider and indexed. When users sent chats to colleagues, they may not have been aware that the links could end up in search results. Anthropic has since fixed the issue, and the new Google search results no longer display these links, but those who already have the links can still access them.
For a company that prides itself on security, the cost of such incidents goes beyond fixing the vulnerability. No matter how strict a model's security boundaries are, if the product's default settings inadvertently push private conversations onto the public web, the first thing users lose is trust in the platform. On the same day, Amodei was also discussing the security risks of open weights, highlighting that their product's privacy defenses had faltered.
（Source: Fortune / TechCrunch / VentureBeat / Futurism）
5｜Intensifying Open Source Debate: NVIDIA Invests $5 Billion in Sutskever, Amodei Urgently Clarifies "No Intent to Ban Open Source"
NVIDIA announced a $5 billion investment in Safe Superintelligence (SSI) founded by Ilya Sutskever, and provided the next-generation Vera Rubin computing platform to scale up SSI's computing power by an order of magnitude. SSI has previously raised over $3 billion in funding from a16z, Sequoia, and others, with a valuation of $32 billion.
On the same day, Anthropic CEO Dario Amodei responded to industry pressure, stating, "We have not advocated for banning open-weight models." Previously, NVIDIA, Microsoft, Meta, and 25 other companies jointly called for refraining from imposing "premature restrictions" on open weights. Google and OpenAI signed the letter, with Anthropic being the most prominent dissenter. Amodei's concern is not with open source itself, but with weights once released being irretrievable, unupdateable fencing, and unable to prevent misuse.
While NVIDIA signed the open-source joint letter, it is also heavily invested in closed-source security research. It does not need to take sides. As long as computing power, chips, and security research are all still expanding, each side of the debate could potentially become its customer.
(Source: Bloomberg / TechCrunch / Axios / CNBC)
Also Worth Knowing ↓
RTX 50 Series sees prices surge by up to 59% in the Chinese market. MSI and Colorful have raised prices across the RTX 50 Series, with the RTX 5090D seeing a 22.5% increase, and the RTX 5080 priced at 12,000 yuan. The primary reason is the rising cost of GDDR7 memory, with NVIDIA starting to pass on the pressure to partners. (Source: Tom's Hardware)
"The Big Short" Eisman exits key tech stocks. The real-life counterpart of the movie character Steve Eisman revealed in a CNBC interview that he has sold off key tech holdings, believing that AI capital expenditure returns have not materialized yet and current valuations reflect excessive optimism. (Source: CNBC)
Ultraman declares "Singularity is Here." OpenAI's CEO, in a podcast interview, defines the Singularity as the point when AI begins to self-improve, and claims it has arrived. His last news byte was about a model just invading Hugging Face heading to the White House. (Source: Axios)
A new type of ransomware specifically encrypts AI model weights. Security company Sysdig discovered that attackers breached the training environment through an unauthenticated endpoint in Langflow. The threat to model security is evolving, with attackers lagging behind in capabilities, and even their ransom mechanism has flaws. (Source: VentureBeat)
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
