読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
すべて技術製品投資株式市場その他

読書

2026-11-139 投稿

更新

vLLM Announces Day-0 Support for Thinking Machines' Inkling 2 Small Multimodal MoE Model, Achieving 630 Tokens/s Per User on GB200

重大
2026-11-12

Open-source inference engine vLLM announced Day-0 support for Thinking Machines' newly released Inkling 2 Small model on November 12, 2026. Inkling 2 Small is a multimodal (text and vision) Mixture-of-Experts (MoE) model with 594 billion total parameters and 17 billion active parameters, supporting a context window of up to 1 million tokens. The model also comes with a native DSpark speculator.

The vLLM team performed deep optimizations targeting NVIDIA Blackwell architecture GPUs (such as the GB200). Through techniques including kernel fusion, programmatic dependent launch (PDL), highly tuned Gluon attention kernels, and sequence parallelism, vLLM achieved significant performance improvements. According to official tests, single-user decoding speed reaches approximately 630 tokens/s at TP4 with an acceptance length of 3.5. Additionally, sequence parallelism optimization reduces the 8K-token prefill step by 22% at TP4 and 26% at TP8, freeing 1.3–1.9 GiB per GPU for KV cache.

Inkling 2 Small adopts new architectural features, including block attention residuals, gated attention, and mixed-precision MoE (NVFP4 for the first 40 layers, MXFP8 for the last 8). vLLM maintains the reference implementation's numerical precision in its integration and verified that accuracy remains consistent with non-speculative runs on benchmarks like NIAH, MMMU-Pro, and HLE when DSpark is enabled. Currently, vLLM also supports LoRA fine-tuning and reinforcement learning training for this model.

ソース:vllm.ai

リサーチ

Tomasz Tunguz: Nvidia's Three AI Capital Channels and the 112% Cash Flow Overdraft in the Last Quarter

トピック · 英伟达AI资本管道構造的
2026-11-12 08:00 GMT+8

Tomasz Tunguz, Partner at Theory Ventures and former Google product manager, published a structured industry analysis on the day Jensen Huang resigned. His core argument is that over the past four years, Nvidia has acted as the 'Central Bank of AI,' and Huang's departure exposes this role to systemic tightening risks. This analogy is not mere rhetorical exaggeration; he provides specific figures and compositions for three capital channels, along with a triggering concern: last quarter, Nvidia deployed 112% of its operating cash flow into equity and M&A.

The scale of the three channels is as follows: $42 billion in M&A and technology acquisitions over the past two years (including Groq's $20 billion LPU technology license and engineering acqui-hire, Hugging Face's $12.93 billion acquisition, Poolside's $6 billion Model Factory license, etc.); $61 billion in cash-cost-based equity investments covering 229 private companies and 13 public companies (with disclosed stakes in OpenAI reaching $30 billion); and $164.5 billion in lease guarantees and cloud capacity backstop commitments (including up to $105 billion in guarantees for OpenAI’s Ohio data center leases and $6.3 billion in capacity guarantees for CoreWeave). Additionally, there are $63.1 billion in customer accounts receivable, with payment terms extended to as long as one full year.

Tunguz points out that in the last quarter (FY27 Q2), Nvidia's operating cash flow was $24.08 billion, while the cash deployed into equity and M&A was $26.98 billion, a ratio of 112%. He traces data across eight quarters: rising gradually from 4% in FY25 Q3 to 72% in FY26 Q4, and reaching 112% in FY27 Q2. Over the past four quarters combined, Nvidia generated $134.4 billion in operating cash flow but invested $75.8 billion into equity and M&A. This means the deployment pace has exceeded the organic cash generation capacity of a single quarter, requiring the consumption of existing cash reserves or debt financing to maintain. Under Huang's leadership, this unconventional deployment was viewed as a strategic bet driven by personal conviction and relationships; however, Tunguz believes that facing stricter board scrutiny, a more conservative successor might compress these expenditures.

This is precisely where the 'Central Bank' analogy lies in its mechanism. Tunguz writes that if Nvidia tightens equity checks, payment terms, or lease guarantees, AI construction will slow down, just as the broader economy slows when the Federal Reserve tightens credit. He specifically emphasizes that these deals run on Huang’s personal conviction and relationships—Nvidia simultaneously acts as chip supplier, lead venture investor, anchor cloud customer, and lease guarantor across more than 200 startups. What sustains these multiple roles is not contract terms, but trust among founders. A successor must maintain these founder partnerships without the co-founder who built them.

Tunguz’s argument has a notable boundary. The 112% figure is a single-quarter peak rather than a sustained rhythm—the average deployment rate over four quarters is approximately 56%, and of Groq's $20 billion, only $2.94 billion fell in Q2 FY27 (the remaining $13 billion was paid in the 72% quarter of Q4 FY26). More importantly, Tunguz himself lists two alternative paths: First, custom AI ASICs already accounted for 28% of AI server shipments and 15% to 20% of accelerator revenue in 2026, with Google TPUs holding 65% of the custom silicon market, and 'leadership turbulence at a single-source supplier accelerates this shift.' Second, Meta, Google, and at least eight US startups that have raised over $100 million continue to train open-weight models, and they will now have to carry more of the American open-source mantle themselves. Furthermore, the $164.5 billion in guarantees represents 'contingent liabilities that do not appear on the balance sheet unless triggered,' but require maintaining substantial liquidity reserves. Therefore, a more accurate phrasing of the 'Central Bank tightening' narrative is: Nvidia’s capital channels serve as a liquidity supplement mechanism for the AI ecosystem; tightening them will raise financing costs for mid-sized operators dependent on its credit endorsement and slow project progress, but it will not eliminate the capital expenditure needs of hyperscale cloud service providers themselves.

ソース:Tomasz Tunguz

リサーチ

Noah Smith and Brad DeLong: The U.S. Needs Land and Loans, Not Workers, for Reindustrialization

トピック · 美国再工业化瓶颈重大
2026-11-13 15:10 GMT+8

Economist Noah Smith and University of California, Berkeley economics professor Brad DeLong discussed whether and how the United States should advance reindustrialization in the latest episode of the Hexapodia podcast. Their conclusion defied most expectations: the bottleneck lies not with workers, but with two institutional constraints—land use permitting and bank financing.

They first dismissed the path of economic equality among three common justifications for reindustrialization. Noah Smith and Brad DeLong argued that revitalizing the American middle class by creating high-paying blue-collar manufacturing jobs is unworkable because AI-driven robotics will make most manufacturing highly automated; competing with robot hands using human hands in factories is no longer viable. This means the era of mass semi-skilled blue-collar factory employment has definitively ended.

However, this does not mean manufacturing cannot create jobs. They pointed out that income generated by automated factories creates a multiplier effect locally, driving local service sector employment. This requires that the operators of these factories actually reside near the factories so they can spend their money locally.

With the path of economic equality ruled out, geopolitics and military security become the primary drivers of reindustrialization. They cited an example that is not necessarily optimistic but is significant: Ukraine is the only known country to have successfully achieved reindustrialization, having built a strong drone and robotics industry out of necessity to resist the Russian invasion.

Noah Smith and Brad DeLong ultimately attributed the lack of reindustrialization in the United States to two missing elements: land use permits and bank loans. Regarding land, a permitting reform bill currently moving through the U.S. Congress attempts to address this issue. Regarding banking, they consider this an area where the United States truly lags behind other countries.

Noah Smith proposed a specific plan: utilize America’s sluggish megabanks as pools of capital for reindustrialization by granting them selective exemptions from the Dodd-Frank restrictions enacted after the 2008 financial crisis. He stated that bank financing appears important for manufacturing, but promised to elaborate on the specific mechanisms in subsequent articles.

It is worth noting that Noah Smith’s proposal is currently a suggestion rather than implemented policy, and he personally acknowledged that detailed arguments regarding bank financing mechanisms are incomplete. The logical chain linking Dodd-Frank exemptions to manufacturing financing, along with the potential systemic risks involved, still requires more comprehensive argumentation to support it.

ソース:noahpinion.blog

リサーチ

TSMC Confirms Employee Data Leaked to External Site via Cloud Provider, States No Evidence of Internal System Breach

トピック · 台积电云端数据泄露重大
2026-11-13 09:04 GMT+8

On November 12, TSMC stated that personal information of some current and former employees was unauthorizedly posted on an external website. The company emphasized that there is currently no evidence showing its internal systems were breached, and it has notified affected individuals to provide necessary assistance.

According to INSIDE, TSMC explained that the leaked data matched records within a certain cloud service provider's system and was unrelated to the company's production or customer data. TSMC is cooperating with this cloud provider, investigative authorities, and a third-party cybersecurity firm to clarify the incident, though it did not disclose the supplier's name or the specific fields of the leaked information.

Prior reports indicated that hackers named TSMC in underground forums selling employee personal data. The seller claimed to have 6.9 GB of data covering approximately 46,000 employees, priced at 3 BTC, and threatened to release all data if no buyer was found within a week. These figures are solely the seller's claims and unconfirmed by TSMC; the actual number of affected individuals and scope of data remain pending further investigation.

The key aspect of this incident is that the data leak occurred not within TSMC's own systems but through outsourced cloud services. When enterprises entrust HR or employee data to third parties for storage, they effectively transfer part of their protective responsibility; if a supplier fails, it is still the enterprise's own employees who suffer.

ソース:INSIDE

リサーチ

Former Nvidia CEO Jensen Huang Agrees to Testify Before US Congress

トピック · 黄仁勋间谍指控案構造的
2026-11-13 00:36 GMT+8

According to TechCrunch, former Nvidia CEO Jensen Huang has agreed to testify before the US Congress by Thanksgiving. The announcement was made on Thursday by House Foreign Affairs Committee Chairman Brian Mast, citing reports from Bloomberg and NBC News.

Huang stepped down as CEO following allegations of an affair with a woman suspected of spying for China. He agreed to testify on Wednesday after several senior lawmakers threatened to issue subpoenas.

The testimony involves three key committees: the House Foreign Affairs Committee, the House Select Committee on the Chinese Communist Party, and the House Intelligence Committee. Their respective chairs, Brian Mast, John Moolenaar, and Rick Crawford, had sent a letter to Huang on November 4 requesting his appearance.

"Jensen and I have been very at odds... on export controls, but Nvidia is a very important company to the United States of America," Mast told reporters. "I don’t want to see anything hurt in Nvidia, but he has some very serious questions to answer."

The hearing will take place behind closed doors, though Mast stated he expects the briefing to be recorded on video and transcribed. This development marks an escalation in legislative oversight regarding AI chip giants and their ties to China.

ソース:TechCrunch

リサーチ

US Democratic Lawmakers Launch Probe, Questioning 'Illegal Export Tax' on Nvidia by Trump Administration

トピック · 英伟达美出口税争议重大
2026-11-13 06:42 GMT+8

On Thursday, US Congressional Democrats formally launched an inquiry into how the Trump administration handled Nvidia's chip sales to China. This follows the company's recent ouster of its founder and CEO, Jensen Huang.

The investigation is led by Rep. Richard Neal, the top Democrat on the House Ways and Means Committee, and Sen. Ron Wyden, the top Democrat on the Senate Finance Committee. In a letter to US Commerce Secretary Howard Lutnick, they requested the department preserve records regarding an arrangement they deem "questionable" and allegedly "illegal." The letter describes a mechanism criticized as an "unconstitutional export tax system": President Trump agreed to allow sales of Nvidia's H200 chips to China, provided that these Taiwan-manufactured chips were first shipped to the United States. Upon arrival, Trump added a tariff, effectively taking a cut of Nvidia's sales to China. Subsequently, the Commerce Department issued the export license.

As reported by the AP, this revenue-sharing deal was public. In August 2025, both Nvidia and AMD agreed to share 15% of their revenue from chip sales to China with the US government as part of a deal to secure export licenses. However, the Democrats' letter questions why Lutnick continued to allow Megaspeed, a Singapore-based company where an alleged Chinese spy worked, to buy Nvidia chips even after being reportedly informed by US intelligence about suspicions regarding Jensen Huang's relationship with her. It also questions why other Asian companies with no physical presence at their listed addresses were green-lit.

The inquiry also examines whether Nvidia's political donations influenced the decision. The letter notes that the export deal was approved after Trump learned of Huang's "compromised position" due to his affair with an alleged spy, and after Nvidia donated $26 million toward Trump's projects, including the White House ballroom. CNBC reported that Nvidia committed at least $26 million to Trump-backed projects, including $25 million for the ballroom and $1 million for the 2025 inaugural fund.

This is the second Congressional investigation into Nvidia's China sales and its dealings with Trump since news of the alleged spy affair broke. Previously, Republican-led committees demanded Huang testify in closed-door sessions. Although Democrats, as the minority party, cannot compel answers, this move signals an escalation in legislative oversight over the interplay between executive power and tech giant interests.

ソース:TechCrunch

リサーチ

Higress Integrates Qwen3Guard for Zero-Code AI Safety Gateway

重大
検証済み 2026-11-13 03:24 GMT+8

Alibaba Cloud's open-source cloud-native API gateway, Higress, has recently integrated the Tongyi Qianwen safety model Qwen3Guard-Gen as a Wasm plugin to provide AI content security capabilities. This solution allows enterprises to implement unified input and output safety moderation at the gateway layer without modifying application code or upstream model services.

In traditional AI application development, security policies are often scattered across individual application SDKs, making it difficult to manage risk thresholds and refusal logic centrally. Each model upgrade or policy adjustment requires redeploying applications. Higress places content extraction, safety calls, risk decisions, and streaming buffering logic within the data plane path, achieving decoupling and reuse of security capabilities.

The plugin defaults to calling the Qwen/Qwen3Guard-Gen-4B model and supports real-time moderation of non-streaming JSON responses and SSE streaming outputs. For streaming responses, the plugin uses a segmented buffering mechanism, triggering a moderation check every time a certain number of characters accumulate (default 1000 Unicode characters), balancing latency and security. If a risk is detected above the threshold, the gateway can directly return a refusal response, preventing the request from consuming backend inference resources.

Qwen3Guard is a family of safety moderation models launched by Alibaba's Tongyi Lab, trained on the Qwen3 architecture, covering risks such as violence, illegal acts, and privacy leaks. This integration emphasizes "self-hosting" features, allowing enterprises to independently deploy and scale safety models without relying on cloud-based content safety services, thereby better controlling data privacy and risk policies.

ソース:Alibaba Cloud Community

リサーチ

Indian Public Library Faces Anti-Corruption Probe Over Alleged AI-Generated Books

重大
2026-11-13 00:37 GMT+8

The Periyar Arivulagam library in the Indian state of Tamil Nadu has become the subject of an anti-corruption investigation just days after its opening. The probe was triggered by accusations from internationally recognized Tamil novelist Perumal Murugan, who claimed that up to 1,800 books purchased by the library may have been partially or wholly generated using artificial intelligence.

According to TechRadar, citing BBC reporting, the books were purchased from a single publisher at unusually high prices. Many reportedly contained false information, suspicious author names, and questionable International Standard Book Numbers (ISBN). The library, built at a cost of 3 billion rupees ($34 million), opened on October 16.

Tamil Nadu's School Education Minister A Rajmohan stated that if the accusations are true, it could be "India's first AI scam." The state government has blamed the previous administration for the purchases, an allegation denied by the Dravida Munnetra Kazhagam party. Over 160 writers, researchers, artists, and publishers have signed a statement calling for an independent inquiry and a public release of procurement records.

ソース:TechRadar

リサーチ

Graph Attention Networks: Solving Inductive Learning Challenges in Graph-Structured Data with Self-Attention Mechanisms

検証済み 2026-11-13 22:12 GMT+8

長期読解 · 《Graph Attention Networks》(2017)

In the field of deep learning, Convolutional Neural Networks (CNNs) have achieved great success in processing grid-structured data such as images. However, many real-world datasets, such as social networks, biological networks, or communication networks, do not possess regular grid structures but instead exist in graph form. Early graph neural network methods were primarily divided into spectral and non-spectral approaches. While effective, spectral methods typically rely on eigendecomposition of the Laplacian operator, meaning that trained models cannot be directly applied to new graphs with different structures, limiting their flexibility in inductive learning tasks.

To address this challenge, Petar Veličković et al. proposed Graph Attention Networks (GATs). The core innovation of this architecture is the introduction of masked self-attention layers. Unlike traditional methods that depend on fixed neighborhood sizes or predefined graph structures, GAT allows nodes to "attend" to features of their neighbors and implicitly assign different weights to each neighbor. This mechanism not only avoids expensive matrix operations (such as inversion) but also does not require prior knowledge of the complete graph structure. By stacking these attention layers, the model can efficiently process node pairs in parallel and flexibly adapt to nodes with varying degrees.

A key advantage of GAT is its inductive learning capability. Since attention weights are dynamically computed based on node features rather than fixed graph topology, the model can be directly applied to entirely new graph structures unseen during training. The paper notes that this approach simultaneously addresses several critical challenges faced by spectral-based graph neural networks and achieves state-of-the-art or comparable results on citation network datasets such as Cora, Citeseer, and Pubmed, as well as protein interaction datasets. For researchers interested in understanding how to extend Transformer-based attention mechanisms to non-Euclidean data, this is a foundational seminal work. It is recommended to read "Graph Attention Networks" (submitted 2017, published at ICLR 2018), focusing on how it solves inductive learning problems in graph-structured data through self-attention mechanisms.

ソース:《Graph Attention Networks》(2017)

リサーチ

このビューで最新の記事まで読みました

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク