LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-10-28154 publications

Actualiser

Anyscale Field Engineer Naila Khan: KubeRay Default Configurations Fail at Five Boundaries in Multi-Team GPU Clusters

Sujet · KubeRayMatériel
Vérifié 2026-10-28 06:59 GMT+8

When RayJobs scale from a single successful run to hundreds of jobs across multiple teams and clusters, default configurations that worked well at small scale begin to fail under production parallelism and contention. This is the core observation from Part 2 of the KubeRay series. She emphasizes that these are not defects in KubeRay itself, but boundary issues between Ray, Kubernetes, and the organization running both.

She describes two types of teams from her perspective as a field engineer: enterprises fitting Ray into clusters built for other purposes, and digital natives pushing Ray beyond its previous limits. She notes they repeatedly encounter the same five scenarios. The first scenario is debugging data tied to the cluster lifecycle: a RayJob configured with shutdownAfterJobFinishes: true deletes its Ray dashboard while cleaning up after itself. The hardest case to troubleshoot is when the driver process exits within the first few minutes—GPU worker nodes start, receive no tasks, and the cluster is torn down before anyone can capture the driver logs.

The second scenario is hardware faults surfacing as application errors. PCIe or NVLink faults, or hung NFS mounts, can interrupt a training run while the Pod still reports Ready status. Ray continues to schedule work to that node, so the failure presents as an application error. Some teams have built modified operators and custom resources to detect failing hardware.

The third scenario is that default settings require tuning at production scale. Khan gave specific examples: when 20 clusters compete for one GPU pool, idle GPU worker nodes may be reaped while their placement group is still forming, preventing training from starting; setting maxReplicas with no headroom can stall the autoscaler when Ray Train retries out-of-memory (OOM) tasks; each retry can also leave a placement group behind. She stated that field team feedback indicates weeks of platform engineering work are needed before reliability stabilizes.

The fourth scenario is head node disruptions affecting the entire cluster. Khan cited the experience of the Klaviyo engineering team, which has been running Ray Serve on KubeRay for a year: node drains for node-image updates and security patches can take down Ray nodes, and memory pressure can evict the head Pod. Without GCS fault tolerance, the whole cluster fails; with it enabled, workers continue running only if the head recovers within about ten minutes, which is KubeRay's default worker reconnect timeout.

The fifth scenario is coexistence of Ray with existing platform services. Service meshes, shared namespaces, quotas, and audit requirements predate Ray. Khan described an enterprise team serving thousands of models on Ray Serve that found Envoy sidecars creating a listener for every Ray port, which strained the mesh control plane and affected neighboring services.

For these five scenarios, Khan offers six practices, emphasizing that none require new tools but instead leverage capabilities already available in the Kubernetes ecosystem. Core recommendations include: enabling the History Server (introduced as alpha in KubeRay v1.6, promoted to beta in v1.7) to write events and logs to object storage; running the NVIDIA DCGM exporter and alerting on GPU Xid errors rather than just monitoring utilization; using a gang scheduler to ensure job worker nodes start together or not at all; leaving headroom of at least two additional worker nodes for to accommodate replacements after OOM retries; setting on the head node; and keeping the service mesh at the edge while excluding Ray's internal ports.

Sources :Anyscale

Recherche

AI Firm Anthropic Expands Partnership with GSA to Offer 40% Discount on Claude Products to All Levels of US Government

Sujet · Claude政府版Matériel
2026-10-27 08:00 GMT+8

AI company Anthropic announced on October 27, 2026, an expansion of its partnership with the U.S. General Services Administration (GSA). Through a new agreement, government partners will receive a 40% discount on Claude Enterprise, Claude for Government, and the Claude API. This marks the first time the discount applies across every product surface, including direct purchases today and future acquisitions through cloud providers, and extends beyond federal agencies to state, local, tribal, and territorial governments.

According to Anthropic's official announcement, approximately 500 U.S. federal, state, and local government organizations currently use Claude. Weekly active users of Claude for Government have increased about eightfold since February, while usage across federal civilian agencies has grown roughly 20-fold over the past year. The agreement aims to help these institutions, which often operate under tight budgets and long procurement cycles, apply AI to complex problems without large upfront commitments.

In addition to price reductions, Anthropic is launching a dedicated program to help agencies implement and optimize their use of Claude. The company states this initiative brings best practices from its private-sector customers to public-sector partners. On security and compliance, Claude for Government became generally available on September 30 at the FedRAMP High authorization level; Claude Enterprise is currently in process for FedRAMP Moderate authorization through the FedRAMP 20x program.

Anthropic estimates that the discounts offered through this agreement will save the government approximately $1 billion. Notably, agencies are not bound by minimum spend commitments and will pay only for the usage they consume. This model lowers the financial barrier for public sector adoption of advanced AI tools, potentially accelerating market penetration in the government segment.

Sources :anthropic.com

Recherche

NVIDIA Demonstrates Fine-Tuning Models with Synthetic Resident Data to Master Local Regulatory Processes: Linguistic Fluency Does Not Equal Procedural Correctness

Matériel
2026-10-28 00:00 GMT+8

Imagine you are developing an AI assistant for the French passport agency and receive a question from a citizen unable to travel due to health reasons: How do I apply for a passport? An unfine-tuned foundation model responds in fluent French: You can mail your application documents via registered post. This answer is grammatically correct and polite, but procedurally wrong—the official website of the Government of France explicitly states that in such cases, a civil servant must visit the applicant at home; mailing is not a legal option.

In a tutorial published on October 27, 2026, NVIDIA technical blog authors Sophia Abbassi and Shashank Verma use this case to illustrate a limitation of general large language models: they can master a language but may not master the specific administrative rules of the jurisdiction associated with that language.

The core of their demonstrated method lies in the Nemotron-Personas dataset. Each synthetic resident is not a generic 'French adult needing a passport' but a specific profile endowed with age, commune, occupation, and household structure based on French census data. This shapes what residents ask and how they ask it, and can also change the correct answer: the passport stamp duty for a resident of French Guiana is €43, whereas it is €86 in mainland France. The dataset was constructed by NVIDIA in collaboration with French AI startup Pleias, with the Belgian version also involving KU Leuven, covering Dutch, French, German, and English.

The fine-tuning process consists of five steps: first, define a behavior contract (target audience, task scope, language style, factual source limitations, safety red lines, output structure) while freezing a set of manually written test cases; second, use an LLM to convert each synthetic resident into a customer service dialogue, where every answer's facts must come from official procedure texts; third, fine-tune on Nemotron 3.5 Lightning using LoRA and deploy after merging the adapter; fourth, send the same group of frozen test prompts to both the base model and the fine-tuned model, comparing them item by item. The authors emphasize that a decrease in validation loss alone is insufficient to prove the model has improved; one must check item by item whether fabricated fees, deadlines, or document requirements appear.

Pleias applies this method to its partner's spine care assistant, Lamina: expanding a small number of real patient cases into diverse training dialogues using three Nemotron-Personas datasets, and testing the assistant's performance across different patient profiles. Pleias itself states that better grounded data makes small, efficient region-specific models viable. The authors also note that persona grounding improves training data coverage and reduces blind spots, but does not guarantee fairness or representativeness.

It should be noted that this is a first-party tutorial demonstration by NVIDIA on its own models, its own datasets, and its own hardware (DGX Station GB300 or DGX Spark). Evaluation results only show a comparison of one Q&A pair, without providing quantitative pass rates across multiple test cases, nor independent reproduction. The authors explicitly state that for tutorial purposes, a frozen test set is not strictly required—but it should be created for production deployment. The transferability of the method depends on the ability to obtain official procedure texts and demographic statistics for the corresponding jurisdiction.

Sources :NVIDIA Technical Blog

Recherche

Report: US FTC Staff Indicate Plan to Sue to Block $5 Billion Cintas-Unifirst Merger

Matériel
2026-10-27 23:14 GMT+8

According to a Semafor report citing people familiar with the matter, staff at the U.S. Federal Trade Commission (FTC) have privately indicated to Cintas Corporation that they intend to file a lawsuit to block its proposed $5 billion acquisition of rival uniform provider Unifirst Corporation. The regulator has not yet voted on the final decision, and a formal meeting to discuss the case has not been held.

This development challenges the prevailing assumption among dealmakers that the current U.S. administration would take a light-touch approach to mergers. Following the report, shares of Unifirst dropped by 12%, while Cintas shares fell by 6%. An FTC spokesperson stated that any final decision would not be communicated through the press.

Sources :semafor.com

Recherche

Hebbia’s Kavita Anand details Astra improvements in source selection and presentation generation for investment research

Sujet · AstraPrise rapide
2026-10-27 22:50 GMT+8

Hebbia’s Kavita Anand recently discussed specific improvements in Astra’s performance within the company’s internal evaluations. The focus was on two critical aspects of investment research workflows: the accuracy of source selection and the capability to generate PowerPoint presentations.

For professionals such as bankers and investors who routinely process thousands of documents, these updates are designed to streamline the transition from raw research data to actionable insights. The episode highlights how better source identification and automated presentation creation can enhance efficiency in high-volume information environments.

Sources :youtube.com

Recherche

Epoch AI Launches Chip Import Tracker: US Holds 72% of Global AI Compute, China's Chip Share Nearly Twice Its Compute Share

Matériel
2026-10-27

Chip designers and cloud service providers rarely disclose which countries their chips are deployed in. This means a fundamental question—where global AI compute is actually located—has long relied on scattered company announcements and industry estimates. Research organization Epoch AI’s AI Chip Imports explorer, released on October 27, 2026, attempts to fill this gap using customs records: starting with monthly import/export data from 16 major destinations, netting out parts transiting only through hubs like Singapore and Hong Kong, it estimates the share of remaining server and parts trade related to AI chips, then splits global chip sales estimates by country proportion.

Initial findings show that of the 27.4 million H100e (measured in Nvidia H100 equivalent compute units) arriving globally between 2022 and June 2026, 19.8 million, or 72%, arrived in the United States. The European Union accounted for 5% (1.4 million), Malaysia for 2% (0.7 million), and no other country exceeded 2%. Notably, the US lead has widened: the US received 60–63% of new AI compute annually from 2022 to 2024, rising to 72% in 2025 and 78% in the first half of 2026.

Another set of figures reveals a structural gap between chip volume and compute quality. Including domestic chips, China accounted for 18% of AI chip volumes in the dataset but only 10% of compute (2.6 million H100e). Epoch AI notes that a significant portion of China’s hardware consists of lower-performance chips, including Nvidia H20 chips designed to comply with US export controls, as well as domestic chips from Huawei and Cambricon, which together account for about 40% of China’s total. Counting imports alone, China received 12% of chips and 6% of compute.

This estimate carries significant uncertainty. Under different assumptions about trade composition, China’s imported compute estimates range from 1.9 million to 4.2 million H100e. The estimate includes chips brought into China in violation of US export controls, provided they appear in China’s own customs records; shipments that go undeclared leave no trace. Epoch AI previously used this same dataset in research on chip smuggling routes via Malaysia.

Methodologically, customs data cross-validates with Epoch AI’s AI Chip Sales data, built from company disclosures and analyst estimates. The two are not fully independent—estimating the value of chips within each server still uses Epoch AI’s own chip prices and product mix—but except for China, trade values come solely from customs data, with no mechanism forcing the two sets of numbers to align. Results show both series grew from under $10 billion per quarter in mid-2023 to nearly $80 billion by mid-2026. Cumulative customs-derived chip spending fell within 4% of the corresponding chip values in sales data, though individual quarters had noise of about 15%. Customs values ran mostly below sales estimates through 2024 and mostly above them after mid-2025, possibly reflecting changes in time from sale to arrival, or chip price fluctuations over a model’s lifecycle while estimates remained fixed.

Sources :Epoch AI

Recherche

Arize AI Open Source Lead Mikyo King Explains Code Mode for Agents: Sandbox Tool Composition and Efficiency Gains

Sujet · Arize AI代码模式Matériel
2026-10-28 03:00 GMT+8

In building AI agents, traditional methods often fill the context window with extensive tool definitions and raw outputs, leading to information overload. Mikyo King, Head of Open Source at Arize AI, proposes an alternative: 'Code Mode.' In this paradigm, the agent is given a sandbox environment where it can dynamically search, compose, and call necessary tools by writing and executing code, returning only the required results.

King details how this mechanism is implemented in Arize's open-source platform, Phoenix, supporting Bash, Python sandboxes, SQL, and even UI control. He also highlights hidden costs such as secrets management, permissions, and durability. Notably, evaluations using the Harbor framework indicated that agents using Code Mode reduced token usage by 43% and lowered operational costs by 31% compared to traditional approaches.

For developers focused on agent infrastructure and efficiency optimization, this episode offers specific engineering practice references. While a full transcript is not provided, the publisher's described quantitative results and technical details offer valuable clues for evaluating agent architecture choices.

Sources :youtube.com

Recherche

US Cloud Provider Oracle Announces Native Snowflake Service on OCI

Matériel
2026-10-28 04:08 GMT+8

U.S. cloud provider Oracle announced on October 27, 2026, at the "Oracle AI World" conference in Las Vegas that it will offer the Snowflake platform as a native service on Oracle Cloud Infrastructure (OCI). This allows enterprise customers to purchase, provision, and use Snowflake directly through Oracle, with simplified pricing and billing options, including the use of Oracle Universal Credits.

The partnership enables Oracle's Exadata data warehouse service and Snowflake's cloud platform to coexist within the same infrastructure. Customers can utilize data stored in Oracle AI Database, Exadata, and Oracle Fusion Applications within Snowflake without physically moving it. This architecture aims to reduce the need for custom data pipelines, accelerate reporting, data engineering, and AI insights, while maintaining control over data access and governance.

Oracle CEO Clay Magouyrk stated that this move gives customers more choices and simplifies the deployment and management of AI and data services. Christian Kleinerman, SVP of Product Management at Snowflake, emphasized the goal of removing friction by allowing joint customers to seamlessly integrate data managed by Oracle applications into Snowflake's AI Data Cloud. While specific performance metrics and detailed pricing terms were not fully disclosed, the move reflects a trend among cloud vendors to compete for enterprise data workloads in the AI era by opening up their ecosystems.

Sources :바이라인네트워크

Recherche

NVIDIA Solutions Architect Yang Liu: Missing Interface in DGX Spark MoE Inference Optimization Causes Performance Below Baseline

Prise rapide
2026-10-27 23:00 GMT+8

In edge inference scenarios with a batch size of 1, the performance bottleneck for Mixture-of-Experts (MoE) models is often not compute power but memory bandwidth. NVIDIA Solutions Architect Yang Liu recorded this assessment in a developer blog post dated October 27, 2026: profiling of Inclusion AI’s Ling-3.0-flash—a MoE model with 124 billion total parameters and 5.1 billion active parameters per token—on NVIDIA DGX Spark showed that the decoding phase was limited by weight reads, with two hotspot operations being MoE expert routing computation and output-layer vocabulary projection.

To address these two bottlenecks, Yang Liu tested three complementary optimizations: replacing the FlashInfer MXFP4 MoE backend with the Humming kernel library (leveraging NVRTC runtime compilation to adapt to specific dimensions, yielding a 9.3% throughput increase); applying online FP8 quantization to the LM head to halve its weight read volume (increasing throughput by 21.7% when combined with Humming); and enabling the model’s built-in NEXTN layer as a draft model for multi-token speculative decoding (achieving a 185.1% throughput increase under synthetic workloads). These three mechanisms do not overlap—the first two reduce the cost per step, while the third reduces the number of steps—and can theoretically be stacked.

However, stacking does not happen automatically. Yang Liu identified a critical integration trap: when FP8 LM head quantization and MTP speculative decoding are enabled together, the draft model must share the entire LM head module, not just the weight tensor. The original Ling NEXTN implementation did not expose this interface. Without it, the framework silently falls back—the draft model receives FP8 weights but interprets them as BF16, causing the accept length to drop to 1.00 and throughput to fall below the unoptimized baseline configuration. This failure mode produces no error messages and can only be detected by verifying at startup whether each optimization is truly active.

Another notable lesson is the significant gap between synthetic and real workloads. In synthetic tests (32K input, 2K output, greedy decoding), the MTP accept length was 3.90, with a throughput gain of 185.1%; however, in the full GSM8K evaluation (1,319 questions, temperature 0.6, top_p 0.95, median output approximately 310 tokens), the accept length dropped to 2.68, the combined throughput gain from the three optimizations was 42.5%, mean latency decreased by 24.5%, and accuracy showed no statistically significant change. Yang Liu explicitly warned that synthetic workloads with repetitive outputs and simple sampling systematically overestimate speculative decoding benefits, and deployments should use evaluations that reflect actual workloads.

All data are self-test results from NVIDIA on DGX Spark hardware, using a single model and benchmark, without independent third-party reproduction. For developers, two transferable insights emerge: first, when stacking quantization and speculative decoding, it is essential to verify module-level interface compatibility rather than merely confirming that configuration parameters are set; second, do not extrapolate percentage gains from synthetic benchmarks directly to real-world workloads.

Sources :NVIDIA Technical Blog

Recherche

Replit CEO Amjad Masad: AI Agents Will Drastically Compress Corporate Coordination Costs, Fortune 500 Companies to Be 'Self-Driving' by End of 2030

Sujet · Replit代理栖息地Matériel
2026-10-28 05:16 GMT+8

Amjad Masad, CEO and co-founder of Replit, posted a thread titled "The Rise of the Self-Driving Company" on X on Monday. His core argument is that AI agents are absorbing routine execution tasks within enterprises, eventually leading to an organizational model where "humans exercise judgment and agents execute most of the work." He anchors this view in economist Ronald Coase's 1937 classic question: Why do firms exist? Coase answered that firms survive because coordinating work internally is more efficient than through open markets. Masad argues that AI agents are poised to drastically compress the internal coordination overhead—management, meetings, and reporting—that firms have treated as fixed costs for nearly a century, thereby altering the optimal boundary of the firm. Consequently, he predicts that by the end of 2030, every company in the Fortune 500 will become a "self-driving company."

Masad refers to the underlying architecture of this transition as the "Agent Habitat"—a space where AI agents collaborate with employees and each other without having unrestricted access to company resources. Replit is advancing toward this vision through its commercial product line, Replit for Business. Agents can access company systems, data, and workflows, while permissions and approvals remain in place. On Tuesday, Replit further launched Cloud Agents, allowing employees to modify a company's codebase without setting up a development environment or understanding the application's internal structure. Engineers retain final decision-making power, but employees can submit tested changes rather than waiting for tickets to be picked up.

Replit reported some internal operational data to support the claim that this mechanism is functioning: after using shared instructions (called "skills"), token consumption dropped by 51%, and a routine data task that previously took 16 minutes was shortened to 1 minute without sacrificing output quality. Engineering teams saw code output nearly triple in the first half of 2026, while review times did not increase, nor did rollbacks or production incidents. Replit is also experimenting with agents learning from code reviews; when engineers flag issues, Cloud Agents can suggest updates to the team's shared instructions so future agents can avoid similar mistakes. However, engineers must approve these updates, and the feature remains in early access.

These numbers require careful attribution boundaries. Replit has not disclosed sufficient detail for independent verification of these gains, nor can it separate the impact of agents from other changes in how its teams work. As an AI-native company, replicating this path in large enterprises with decades of legacy systems and processes may be far more difficult. The Agent Habitat concept is currently closer to a blueprint than a published technical specification, and Replit has not yet demonstrated that agents can handle the kind of company-wide work Masad describes. Cloud Agents are now generally available to Pro and Enterprise customers, while the skills feature remains in early access.

Sources :The New Stack

Recherche

Anthropic Details Drug Development Progress with Six Partners and Acquires Solo Biotech Mithril

Sujet · AnthropicMatériel
2026-10-27 08:00 GMT+8

US-based artificial intelligence company Anthropic published a report on October 27, 2026, detailing early progress from its life sciences team's deep research partnerships with six organizations, including BigHat Biosciences and Bristol Myers Squibb, aimed at accelerating drug development. The company also announced the acquisition of Mithril Medicine, a one-person biotechnology firm focused on treating muscular dystrophy. Anthropic stated its strategy involves integrating AI into every stage of the process, from target identification to clinical trial management, to address distributed efficiency bottlenecks.

The report disclosed several specific laboratory-validated results. In cell therapy, Generate Biomedicines used Claude Code to redesign genetic instructions for CAR T-cell therapy. One new design was approximately 70% shorter than the standard version and, in lab tests, enabled immune cells to produce roughly 10 times as many receptors while killing tumor cells more efficiently. For antibody discovery targeting food allergies, Divergence Labs used Claude to autonomously generate about 35,000 designs. Early tests identified three designs capable of binding IgE and twenty capable of dislodging it, with three of the latter outperforming standard medications.

In lead optimization, BigHat Biosciences connected Claude to its automated lab system, Reccy, testing about 1,010 variants over ten weeks. By reanalyzing raw data and adjusting protocols, the best versions showed more than 10x tighter binding, an 18°C higher stability temperature for cancer-targeting leads, and a 100x slower clumping rate for long-lasting drug leads. Additionally, Enveda used Claude to review manufacturing documents, catching seven issues missed by human reviewers across 1,061 pages and reducing review time by 89%. Claude also drafted two key regulatory documents, with one taking four hours compared to the traditional three-month manual process.

Anthropic also confirmed the acquisition of Mithril Medicine, operated solely by founder Cody Tranbarger, which develops treatments for rare forms of muscular dystrophy. According to Anthropic, Tranbarger previously used Claude to manage contracts, logistics, and experiments with 11 external labs, advancing a potential treatment from concept to promising mouse study results in under 18 months. This acquisition marks Anthropic's transition from providing tools to directly running internal drug discovery programs.

Sources :anthropic.com

Recherche

Two Sigma Architect John Feminella: Five Composable Controls for AI Agent Autonomy Illustrated by Citibank's $900M Error

Sujet · AI代理可控自主权机制Matériel
2026-10-27 23:00 GMT+8

In 2020, Citibank accidentally sent about $900 million to Revlon’s lenders instead of an $8 million interest payment, even though the procedure, the warning pop-up, and three approvers all did what they were supposed to. This famous operational failure highlights the risks inherent in traditional human-in-the-loop processes.

John Feminella, who works on AI architecture and engineering at Two Sigma, uses that story to show how to give AI agents real autonomy without locking everything down. In his recent video episode, he argues that simple restriction is not the solution; instead, sophisticated control layers are required.

Feminella walks through five composable controls designed to address governance gaps in AI agent deployment:

  1. An agent registry: Clearly identifying which agents exist within the system.
  2. Risk-based review: Adjusting scrutiny levels based on the potential impact of actions.
  3. Tool access that follows an agent’s declared purpose: Ensuring agents can only use tools relevant to their stated tasks.
  4. Shared firm-wide evals: Establishing unified testing standards to validate agent behavior across the organization.
  5. Classifiers that send humans only the approvals that matter: Filtering routine requests to mitigate "approval fatigue" by escalating only critical decisions.

The episode was published on October 27, 2026, by the AI Engineer channel. Feminella explicitly notes that these views are his own and do not represent Two Sigma's official position. For engineers and managers exploring enterprise AI agent integration, this framework offers concrete architectural reference points.

Sources :youtube.com

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement
maxReplicas
num-cpus: "0"

At the end of the article, Khan proposed a decision framework: if you need to maintain a dedicated pipeline to capture logs before RayJob termination, spend hours troubleshooting whether a failure was due to the GPU, Kubernetes, or Ray, manually tune autoscaling defaults, run redundant clusters or custom health checkers, manually map workloads to teams for auditing or chargeback, or operate Ray across multiple clusters, clouds, and regions—if more than two of these describe your week, a platform layer may be worth considering. She then introduced Anyscale’s KubeRay Connect private preview as a concrete implementation of this layer.

It should be noted that this article is published by Anyscale, whose commercial product KubeRay Connect is precisely the solution recommended in the article. The six practices themselves are based on verifiable Kubernetes and Ray ecosystem tools, but the criteria for 'when to introduce a platform layer' align with Anyscale's commercial interests. Readers can treat the practices section as independent technical reference material, while viewing the product recommendation as a vendor claim.

Khan also mentioned a notable detail: seeing logs and dashboards for completed RayJobs was one of the earliest requests in the KubeRay community, with users raising it as early as 2023. The difficulty arises because ephemeral clusters are the very design intent of RayJob, meaning evidence must leave the cluster before the cluster disappears. The History Server is the community's response, built with contributors from several large platform teams.

The tool has clear boundaries. Customs codes generally mix AI and non-AI chips, as well as chips and non-chip components; data reflects when chips arrive rather than when they run, meaning some may not yet be online while others may later move to another country. Countries lacking detailed customs data, such as the UAE and Vietnam, are estimated only from export records of other nations. Chips made and used in Taiwan do not cross borders and are thus not captured; domestic production in China is assumed to remain entirely within the country, but no estimate exists for chips retained in Taiwan.

For readers tracking the geographic distribution of AI infrastructure, this work adds an incrementally updateable public estimation framework, shifting the question of where compute resides from reliance on selective corporate disclosure to independently verifiable observation. Its core findings—the expanding US share and the near-doubling gap between China’s chip volume and compute quality—provide a new quantitative benchmark for understanding AI industry competition, but readers should note all figures are estimates, with particularly wide confidence intervals for China.