August 10, 2026 — The Trend Tribune

By tech_tanaka · August 10, 2026

Why Anthropic isn't "breaking up" with NVIDIA—the real reason behind their move to developing their own chips. — (Photo: illustrative)

Why Anthropic Isn't "Breaking Up" with NVIDIA—The Real Aim Behind Their Own Chip Development

On August 5th, Anthropic officially confirmed the establishment of an in-house team to design its own AI chip for Claude. They are reportedly recruiting engineers with expertise in both hardware and software. What's interesting from an engineer's perspective is that this news involves more complex circumstances than simply "breaking away from NVIDIA."

Not "De-NVIDIA," but "Adding a Multi-Chip Strategy"

First, it's important to understand that Anthropic itself has clearly positioned this move not as a "departure from NVIDIA." The company currently relies on hardware from partner companies, primarily NVIDIA and Google accelerators, for both model training and inference. This development of their own chip is explained as adding a new layer to their existing business relationships with AWS, Google, NVIDIA, and AMD, rather than replacing them.

In other words, Anthropic's goal isn't simply to "break away" from dependence on a single supplier, but rather to strengthen its "multi-chip strategy" by offering multiple options. In fact, it's been reported that the company has already secured approximately 3.5 gigawatts of custom TPU (Tensor Processing Unit, Google's proprietary chip for machine learning) capacity, available from 2027, through long-term agreements with Google and Broadcom. This latest development of its own chip should be understood as adding a new option—a self-designed chip—to this existing effort.

The Motivation: "The Arithmetic of Daily Token Processing"

According to Forbes, this move is driven by the "simple arithmetic" of the enormous number of tokens Anthropic processes daily and its reported annual revenue of $30 billion. The core aim of this initiative is to continuously reduce the cost per query by using chips optimized specifically for its own models, rather than off-the-shelf, general-purpose accelerators.

From the perspective of the cost of providing AI models, this can be considered an extremely rational decision. By "co-designing" hardware and software (in this case, the AI model itself) in a way that optimizes each other, it's possible to achieve efficiencies that cannot be attained by combining off-the-shelf products. Considering that Google's TPU has pursued this strategy for many years, Anthropic moving in the same direction can be seen as part of an inevitable trend in the industry.

Recruitment Requirements for "Immediate Contributions"

What stands out in this recruitment is the specificity of the desired candidate profile. The job posting seeks engineers who "have experience actually shipping completed semiconductor designs as individuals," and the annual salary is set at a high level of $320,000 to $485,000. One report frankly states that this position is for "people who are comfortable making important decisions without the support of a large organization."

This suggests that Anthropic is pursuing this chip development not as an experimental research project, but as a serious undertaking with mass production and deployment in mind.

Comparison with Leading Competitors

Anthropic's move is not isolated within the industry. OpenAI announced its inference-focused chip, "Jalapeño," developed in collaboration with Broadcom, in June. Meta is also preparing for mass production of its next-generation AI accelerator, "MTIA." Google has a track record of developing its own chips (TPUs) for a longer period than any other AI company. The French startup Mistral has also reportedly expressed interest in developing its own silicon.

In short, the trend of frontier AI companies developing their own chips is no longer an exceptional strategy, but is becoming a standard tactic for survival in this industry. According to semiconductor industry sources, the design cost of advanced AI chips, including securing specialized engineering talent and achieving high-yield manufacturing processes, could reach nearly $500 million. This suggests that Anthropic is reaching a scale of daily inference costs that justifies this massive investment.

What Engineers Should Note

This news indicates a structural shift towards vertical integration, where the competition for AI model performance extends beyond software algorithms to include the design of the underlying hardware. It remains unclear whether Anthropic will handle manufacturing in-house or outsource it to foundry partners (like Samsung, as reported).

For developers operating large-scale AI models in-house, this move can be seen as a positive signal that inference costs may decrease further. On the other hand, if multiple major AI companies develop their own silicon, differences in optimization across platforms may emerge, adding new complexity to model portability and benchmark comparisons.

References: qz.com / www.forbes.com / techwireasia.com

AnthropicClaude半導体AIインフラAI/ML

Taking on "the greatest challenge in robotics"—a US startup presents the idea of a robotic hand that starts with tactile sensation.

By robo_takahashi

On July 27th, Palo Alto's Tacta Systems unveiled its first robotic hand with built-in tactile sensors, "TactaBot." This article explains the company's unique "Fluidic Tendon" actuation technology, the sensitivity of its tactile sensors (smaller than a grain of rice) capable of detecting 2.5 to 700,000 Pascals, its two-stage learning approach that overcomes data deficiencies through skill capture using a worker-worn glove, and its commercialization plan with shipments scheduled for early 2027, all from the perspective of focusing on the "hand" as a key technological element.

Tackling the "Greatest Challenge in Robotics"—A US Startup Reveals a Robot Hand Starting with Tactile Sensation

On July 27th, Palo Alto-based Tacta Systems unveiled a technology they had been developing in private. It's a robot hand with built-in tactile sensors, aiming to automate "high-value manual tasks" that have previously been considered difficult for robots, such as the precise assembly and wiring of electronic components. As a software engineer, I'd like to delve into the technical challenges this company is tackling.

Why "Hands" are the Biggest Challenge in Robotics

Tacta Systems CEO Vikram Pavateh made a striking statement in the announcement: "Everyone talks about the robotics market being worth $1 trillion, but the technology hasn't reached that point yet. That's because robots still can't use their hands properly." He also stated, "Hands are the most difficult problem in robotics, and it's impossible to solve them without tactile sensation."

This observation will be intuitively understandable to any engineer with experience in robotics. Visual information from a camera alone makes it difficult to accurately determine how much force and at what angle to grasp an object. Humans can grasp objects even with their eyes closed because they unconsciously process information obtained from the sense of touch in their fingertips. Imparting this sense to robots has been a long-standing bottleneck in automating precise assembly tasks.

Unique Actuator Design: "Fluid Tendon"

The core of TactaHand's technology lies in its proprietary actuation technology called "Fluidic Tendon." While conventional robot hands primarily use a mechanism combining motors and cables (tendons) to move the fingers, Tacta employs a different approach. This design achieves durability that allows it to maintain precision, force, and speed over millions of operation cycles in a factory.

Further noteworthy is the sensitivity of its tactile sensors. These sensors, smaller than a grain of rice, can detect a wide range of pressures, from a very light 250 Pascals to a clamp-level force of 700,000 Pascals. This wide range of sensitivity makes it possible to handle feather-light parts without damaging them and to perform tasks that require firm grip and force application, all with the same hand.

"Skill Capture" to Tackle Data Collection Bottlenecks

Another technological pillar is the "Skill Capture" system using a wearable device called "Tacta Glove" worn by human workers. When a worker wears this glove and performs actual manufacturing tasks, data such as force, movement, temperature, and video are recorded in real time.

This system aims to solve another deep-seated challenge in robotics: the lack of large-scale, high-quality datasets for dexterous tasks. While language-based AI models can use vast amounts of text data from the internet as training material, data on precision work performed by robotic hands is overwhelmingly scarce. Tacta takes a two-stage approach: continuously collecting this data from human workers in actual factory settings to train its AI model, "Dexterous Intelligence," and then fine-tuning it to suit specific customer tasks.

A Two-Stage Approach: Pre-training and On-Site Fine-tuning

The practical advantage of this approach lies in its significant reduction in the time required to learn new tasks. First, pre-training is performed using data from a wide range of haptic skills. This model then serves as a foundation for fine-tuning using task-specific data collected from the customer's factory floor. This two-stage learning flow makes it easier to handle diverse situations with a long tail (a collection of exceptional cases that are infrequent but frequently encountered in the field), lowering the barrier to entry for those considering automation.

Commercialization Still Far Off—A Realistic Timeline of Early 2027

Despite its ambitious technological goals, Tacta's commercialization plan maintains a realistic pace. The company's first product, "TactaBot," is scheduled for initial shipments in early 2027 to major manufacturers in the electronics, AI infrastructure, and automotive industries. Future expansion into medical devices, aerospace, and defense sectors is also anticipated.

The specific tasks targeted include the assembly of precision electronic equipment, connection of connector cables, and assembly of wire harnesses. These are all tasks requiring delicate force control and precise positioning, skills honed over many years by skilled human craftsmen, and have been considered difficult to automate with conventional industrial robot arms.

What Software Engineers Should Notice

What makes Tacta Systems' approach interesting is that, rather than a flashy demonstration of the entire humanoid, it focuses technically on a specific part: the "hand." By specializing in the hand technology, the possibility of retrofitting it to existing humanoid platforms and industrial robot arms becomes apparent.

With the US government providing a tailwind for the return of manufacturing (reshoring), targeting the specific electronics manufacturing industry is also commendable as a realistic business strategy. The combination of tactile sensors and data-collecting gloves represents a different approach from existing humanoid companies, and we will be closely watching to see if this technology can truly solve "the biggest challenge in robotics" through actual shipment results in 2027.

References: www.prnewswire.com / www.therobotreport.com / roboticsandautomationnews.com

ロボットハンド触覚センサー製造業フィジカルAIスタートアップ

The security of "a swarm of countless AIs," not "a single AI"—DeepMind is investing $10 million in this still underdeveloped research area.

By thesis_yamamoto

Google DeepMind, in collaboration with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org, is currently running the "Scaling AI Safety for a Multi-Agent World" program, which will provide up to $10 million in grants for multi-agent AI safety research (application deadline: August 8th). This article explains the four priority areas—sandbox construction, the science of collective behavior, identity protocols, and surveillance tools—as well as the Tier 1 (up to $300,000) and Tier 2 ($300,000 to $1,000,000) grant categories, and the awareness of the research gap that cannot be filled by market forces.

The Safety of a "Swarm of Countless AIs," Not Just a "Single AI"—DeepMind Invests $10 Million in a Still Underdeveloped Research Area

Google DeepMind, in collaboration with four organizations—Schmidt Sciences, the Cooperative AI Foundation, the Advanced Research and Invention Agency (ARIA), and Google.org—is running a joint research funding program that will provide up to $10 million in funding for research on "multi-agent AI safety." The application deadline is August 8th, and the results are expected to be announced this fall. As a journalist with a background in AI research, I would like to organize the technical background and explain why this field is attracting so much attention now.

From "Single Model Safety" to "Swarm Safety"

The official name of this program is "Scaling AI Safety for a Multi-Agent World." Owen Larter, Senior Director at DeepMind, emphasizes the importance of understanding the unpredictable behavior and risks that arise when numerous AI agents operate simultaneously.

Much of the AI safety research to date has focused on individual alignment (aligning AI goals with human intentions), ensuring that a single model does not deviate from its intended behavior. However, this program raises a different kind of risk—the question of whether new collective risks emerge when countless AI agents, developed by different organizations and operating on different platforms, begin to negotiate, trade, and cooperate with one another, risks that cannot be captured simply by accumulating the safety of individual models.

Four Priority Research Areas

This funding program has four specific priority research areas.

The first is the construction of realistic and reproducible "sandboxes and testbeds" for evaluating and comparing multi-agent systems. The second is "scientific" research to understand how agent networks generate collective behavior and how they malfunction. The third is stress testing of "identity and reputation protocols" that support inter-agent interactions across different platforms, such as identity verification and accountability mechanisms. The fourth is the development of a "monitoring and control tool" to monitor the deployed agent cluster and intervene as needed.

Grant Scale and Application Requirements

The project grant is for a period of one to two years, with a two-tiered framework: Tier 1 up to $300,000 and Tier 2 from $300,000 to $1,000,000. Applications are accepted from researchers worldwide, and selected researchers will be notified of the results this fall.

DeepMind positions this initiative as an extension of its past collaborative research. The foundational research previously developed for verifying multi-agent cooperation, such as "Concordia" and "Melting Pot," forms the basis of this program.

Why "Now"? – The Background of Increasing Security Incidents

This funding program was launched in response to the reality that interactions between agents are rapidly increasing, and consequently, security incidents are also on the rise. In fact, around the time this grant program launched, there were reports of several GitHub repositories being disabled by malware attacks targeting automated coding tools.

As more AI agents connect to real-world systems and interact with each other, the risk of system-wide vulnerabilities becoming apparent—vulnerabilities that cannot be seen by evaluating the security of a single model—increases. This program can be seen as an attempt by the research community to proactively address these risks inherent in "AI as a swarm."

The Problem of "Gaps That Market Power Cannot Fill"

The guidance document for this research funding program contains an insightful passage: "We anticipate that some of these problems will be addressed by market power, but some will fall through the cracks."

In other words, this program aims to clearly distinguish between challenges that for-profit companies will naturally solve in the course of their product development and challenges that no one will address unless intentionally funded as academic research. The "ecosystem-level safety"—how different AI agents developed by numerous organizations can safely cooperate with each other—is a research area that is difficult to fill solely through market forces, as it does not directly translate into profits for individual companies.

What Researchers Should Keep an Eye On

This program symbolizes the shift in focus in AI safety research from "how to control a single model" to "how to safely design an entire system in which multiple autonomous agents interact."

With agent-based AI being increasingly integrated into corporate workflows and the number of situations in which multiple AI agents communicate and cooperate with each other growing in practical applications, research results in this field have the potential to become practical guidelines not only for academic interest but also for engineers who actually design and operate multi-agent systems. We will be closely watching what research themes will receive funding after the announcement of the selection results this fall.

References: deepmind.google / www.tun.com / www.cooperativeai.com

DeepMindマルチエージェントAI安全性研究資金AI/ML

Why Anthropic isn't "breaking up" with NVIDIA—the real reason behind their move to developing their own chips.