This Week in AI: Anthropic's Revenue Skyrockets as Tech Giants Intensify Their Battle

Over the past week, the AI industry maintained its dizzying pace of iteration. From Anthropic's rocket-like revenue growth to intensive moves across coding, Agent, and payment gateway verticals, to the mega AI factory being built by NVIDIA and OpenAI — every headline reveals a dramatic reshaping of the competitive landscape. This article provides an overview and analysis of the week's core developments.

Anthropic's Annualized Revenue Hits $6.5 Billion

According to investor data cited by multiple media outlets, Anthropic's annualized revenue (ARR) surpassed $6.5 billion by the end of July. Annual Recurring Revenue (ARR) is one of the most critical financial metrics in the SaaS and cloud services industry — it takes the monthly recurring revenue at a given point in time and multiplies it by 12 to produce an annualized revenue estimate. It's important to note that ARR is not the same as actual annual revenue; it's a forward-looking metric based on the current revenue run rate. $6.5 billion means their monthly revenue at the end of July was approximately $540 million.

This figure represents astronomical growth by SaaS industry standards. For comparison, Salesforce took 20 years to reach a similar annual revenue scale, and Snowflake's ARR at IPO was about $590 million. Even OpenAI's ARR was reportedly around $3.4 billion by the end of 2024. The fact that Anthropic reached this number in less than three years of commercial operations reflects how the generative AI market is exploding at a pace far exceeding traditional SaaS growth curves — the market is rapidly transitioning from a "trial" phase to a "large-scale procurement" phase.

Looking back at the growth trajectory is staggering: this number was about $900 million at the end of last year, grew to $4.7 billion by May, and leaped to $6.5 billion by the end of July. In just over half a year, revenue grew nearly 7x.

The core drivers behind this explosive growth come from two directions: the developer-facing coding tool Claude Code, and enterprise API usage. Claude Code is Anthropic's command-line coding tool, fundamentally different from traditional code completion plugins. Developers can describe requirements in natural language directly in the terminal, and Claude Code understands the entire codebase's context, autonomously completing code writing, debugging, refactoring, and even cross-file modifications. Its core advantage lies in "codebase-level understanding" — it can not only handle individual functions but also grasp a project's overall architecture and dependency relationships. This capability allows enterprise developers to offload large amounts of repetitive programming work to AI, significantly boosting development efficiency.

From a technical roadmap perspective, programming assistance tools have gone through several generations of evolution: from early IDE auto-completion, to syntax-tree-based code suggestions, to the LLM-based line-level/function-level code completion introduced by GitHub Copilot — each generation dramatically improved developer productivity. Claude Code represents a new paradigm — "Agentic Coding" — where AI is no longer passively completing code snippets but autonomously understands entire projects as an Agent, proactively executing multi-step programming tasks. This requires the triple support of ultra-long context windows (to load entire codebases), powerful reasoning capabilities (to understand architecture and dependencies), and tool-calling abilities (to execute terminal commands, read and write files).

This demonstrates that enterprise willingness to pay for high-quality large models is rapidly materializing, and Anthropic has successfully captured this wave of enterprise procurement by building a strong reputation in coding capabilities.

For the industry as a whole, this data also means that large model commercialization is no longer just a promise — it's real revenue. Coding is becoming one of the most solid and highest-monetizing application scenarios for large models.

Coding and Agent Tracks in Full Bloom

This week, multiple companies made intensive moves in coding and Agent directions, with competition reaching a fever pitch.

Grok Build Opens to All Plans

Grok's Grok Build feature officially launched on web and mobile this week, opening to users on all subscription plans. Users simply describe a website, app, or game in natural language, and Grok generates a working version. Even better, finished creations can be published directly to a standalone domain on Grok.me, or exported to GitHub.

The "describe it and it's built" philosophy that Grok Build represents is the ultimate evolution of the low-code/no-code movement in the AI era. Traditional low-code platforms (like Mendix and OutSystems) lower the development barrier through visual drag-and-drop, but still require users to understand component logic and data binding. AI generation tools like Grok Build skip this layer entirely — users only need to describe the final product, and AI handles everything from interface design to code implementation. The closed loop from creation to publishing on a standalone domain is particularly noteworthy — it means the time from idea to going live can be compressed to minutes, which is disruptive for indie developers, startup teams, and rapid prototyping scenarios. This capability further lowers the barrier to AI coding while completing the full loop from generation to deployment.

GLM-4.6 Launches, Aiming for Top Open-Source Model

Chinese large model GLM released a new version this week, with key improvements in complex coding, cybersecurity, and long-duration task processing. The officially published intelligence index reached 60 points, tying with Kimi for first place among open models. API pricing remains unchanged, though currently only the API is available — model weights are expected to be released later. This reflects Chinese companies' determination to keep closing the gap with the international first tier in open-source model capabilities.

Alibaba Launches GUI Agent

Alibaba released an Agent model specifically designed to operate graphical user interfaces, covering mobile phones, computers, and web browsers. Its core capability is being able to "see" screen content and then determine where to click and what to type.

This marks an important shift in Agent technology approaches. Traditional AI Agents execute operations by calling software-provided APIs (Application Programming Interfaces) — for example, sending emails or querying databases via API. But in the real world, a large number of software applications, especially legacy enterprise software, don't offer open APIs, or their API functionality is limited. GUI Agents use computer vision technology to "see" screenshots, identify buttons, text boxes, menus, and other interface elements, then perform click, drag, and input operations just like a human. This approach is technically more challenging because it requires simultaneous visual understanding, spatial reasoning, and operation planning capabilities, but it's also more versatile — theoretically, any software with an interface can be controlled.

From a technical implementation perspective, GUI Agents typically involve a multi-modal large model's Vision-Language-Action three-layer architecture. The model first performs visual understanding through screenshots, using OCR, object detection, and layout analysis to identify interface elements; then combines user instructions with the current state for reasoning and planning; and finally outputs precise operation commands (such as coordinate clicks and keyboard inputs). Pioneers in this field include Microsoft's UFO and Google's WebAgent research projects. Alibaba's launch of a GUI Agent covering all three major platforms suggests a breakthrough in cross-platform generalization capabilities. The impact on enterprise automation (RPA, Robotic Process Automation) is particularly direct — traditional RPA products like UiPath and Automation Anywhere rely on predefined operation scripts and element selectors, while GUI Agents can dynamically understand and adapt to interface changes, significantly reducing the development and maintenance costs of automated workflows.

This approach is closer to human operating habits and can theoretically cover software that doesn't have open APIs. However, the specific deployment methods and actual operation success rates still need more real-world testing to verify.

Tech Giants Positioning at Key Entry Points

Two more important stories this week reflect how major companies are competing for critical nodes in the AI value chain.

Stripe Acquires OpenRouter

OpenRouter announced this week that it's joining payment giant Stripe; the transaction amount has not been disclosed.

OpenRouter currently connects over 400 models, processes more than 10 trillion tokens per day, and serves over 10 million developers and enterprises. It is essentially an "aggregation gateway" for model calls — think of it as a "router for AI models." In current AI application development, different tasks may be best suited to different models — for example, using Claude for long documents, GPT-4o for multimodal tasks, and open-source models for cost-sensitive batch requests. OpenRouter provides a unified API interface so developers only need to integrate once to flexibly switch and load-balance across 400+ models. Processing over 10 trillion tokens daily means it has become an indispensable middleware layer between the AI application layer and the model layer.

Since its founding in 2010, Stripe has positioned itself strategically as "the infrastructure of the internet economy," with a valuation of approximately $50 billion in its 2023 funding round. Acquiring OpenRouter is a key step for Stripe to redefine its role in the AI era. In the traditional internet economy, Stripe processed relatively standardized payment scenarios like e-commerce transactions and SaaS subscriptions. But AI model billing is completely different — per-token billing, per-call billing, massive price differences between models, and real-time dynamic routing make payments and metering in the AI era unprecedentedly complex.

With this gateway in hand, Stripe could enable "model selection, cost control, and payment completion" all within a single system. Stripe's strategic intent is clear: integrate AI model selection, scheduling, billing, and payment into a closed loop, upgrading itself from a pure payment tool to commercial infrastructure for the AI era. It's essentially building a bridge between AI applications and model providers that serves as both router and cashier. This is an important extension by a payment giant into the AI infrastructure layer.

Cursor Launches Origin to Challenge GitHub

The company behind AI coding star product Cursor launched a code hosting service called Origin, with paying users already able to join early testing.

Origin supports code repositories, Pull Requests, and two-way synchronization with GitHub. To understand the significance of this move, consider the central role of code hosting services in software development: GitHub currently holds an absolute dominant position with over 100 million developers. Code hosting isn't just about storing code — it also encompasses code review (Pull Requests), CI/CD (Continuous Integration/Continuous Deployment), project management, and an entire suite of development collaboration workflows.

The code hosting market has long been highly concentrated. After GitHub was acquired by Microsoft for $7.5 billion in 2018, it further consolidated its dominance leveraging Microsoft's resources and Copilot's first-mover advantage. GitLab holds some enterprise market share as an open-source alternative, while Bitbucket (under Atlassian) has advantages in project management integration through the Jira ecosystem. The core challenge Cursor faces with Origin is developer migration cost — code repositories contain not just code itself, but complete commit history, Issue discussions, CI/CD configurations, Webhook integrations, and massive amounts of ancillary data and workflow configurations. Origin's strategy of supporting two-way sync with GitHub is precisely aimed at reducing this migration friction, allowing developers to gradually experience AI-native code hosting workflows without having to sever ties with GitHub all at once.

Cursor's strategic thinking behind Origin is this: if an AI coding tool can deeply integrate code hosting, it can provide an AI-native experience across the entire workflow of code generation, review, and merging — not just assistance at the editor stage. This is the exact reverse of GitHub's logic with Copilot — GitHub extended from hosting platform to AI editing, while Cursor is extending from AI editing to hosting platform. This means Cursor is no longer content being just an editor tool; it aims to extend upstream and build a complete development workflow ecosystem, entering direct competition with GitHub.

Safety and Compute: Two Sides of the Same Coin

OpenAI Pauses Some Model Training

Following concerns over cyber-attack capabilities, the deployment of a certain OpenAI model had been delayed. This week, OpenAI disclosed more specific actions: pausing reinforcement learning on the latest deployed model for two weeks, while its largest frontier reinforcement learning experiments were also temporarily shelved.

The background to this event involves a core topic in AI safety — "Dangerous Capability Evaluations." As large models grow more capable, they may exhibit unexpected abilities in areas like cyber attacks, bioweapon design assistance, and autonomous replication. Frontier labs including OpenAI, Anthropic, and DeepMind have each established their own safety frameworks (such as OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy), stipulating that when a model exceeds thresholds in specific dangerous capability evaluations, corresponding safety measures must be implemented before deployment can continue.

Specifically, OpenAI's Preparedness Framework categorizes risks into four areas: cybersecurity (model's ability to assist cyber attacks), CBRN (chemical, biological, radiological, and nuclear weapons-related knowledge), persuasion (large-scale social manipulation capabilities), and model autonomy (self-replication, resource acquisition capabilities), with each category having low/medium/high/critical risk levels. Anthropic's Responsible Scaling Policy (RSP) uses a similar tiered system called AI Safety Levels (ASL), currently defined up to ASL-4. When a model demonstrates capabilities exceeding preset thresholds during red-team testing, labs must deploy corresponding safety mitigations before continuing training or deployment. These frameworks currently rely primarily on voluntary commitments by each company and lack a unified external oversight mechanism.

Safety remediation directly interrupts training progress — something uncommon in the speed-obsessed AI race. This pause is a landmark event of these safety commitments moving from paper to actual enforcement, indicating that frontier model capabilities are approaching previously established safety red lines, and safety risks are beginning to have material impact on R&D timelines.

4.25GW Super AI Factory Takes Shape

Meanwhile, the compute arms race continues to escalate. NVIDIA, in partnership with SB Energy, is building an AI factory in Ohio, USA, with OpenAI as a tenant. The initial planned capacity is 4.25GW, with each system potentially deploying approximately 1.5 million GPUs.

To understand what 4.25GW means, some reference points are helpful: a typical large nuclear reactor unit (such as China's Hualong One) has a rated capacity of about 1.15GW, meaning this AI factory's power demand is roughly equivalent to the full-load output of 3.7 large nuclear reactor units. For comparison, the Three Gorges Dam's total installed capacity is 22.5GW, and New York City's average power load is about 11GW. Each system deploying approximately 1.5 million GPUs (likely NVIDIA's latest Blackwell architecture) consumes enormous amounts of power not only during training but also requires continuous power and cooling during inference.

In terms of specific power consumption, the NVIDIA Blackwell B200 GPU has a thermal design power (TDP) of approximately 1000 watts per card. At the chip level alone, 1.5 million GPUs would have a peak power draw of 1.5GW. Adding networking equipment, storage systems, cooling systems (which typically account for 30-40% of total energy consumption), and power distribution losses, the 4.25GW total capacity planning is reasonable. The International Energy Agency (IEA) noted in its 2024 report that global data center electricity consumption is expected to double from approximately 460 TWh in 2022 to over 1,000 TWh by 2026, with AI workloads being the primary growth driver.

This power scale approaches the output of several large nuclear power plants, vividly illustrating the enormous energy demands of AI training and inference. This also explains why AI companies are beginning to partner directly with energy companies, and even invest in nuclear and renewable energy projects — Microsoft restarting the Three Mile Island nuclear plant, Amazon acquiring nuclear-powered data centers, Google signing the world's first fusion energy purchase agreement. AI companies are evolving from electricity consumers into significant participants and drivers of the energy industry. The competition for compute is becoming a competition for energy supply, and future AI infrastructure development will be increasingly deeply intertwined with the energy sector.

Conclusion

From revenue explosions to gateway battles, from Agent paradigm shifts to gigawatt-scale compute factories, this week's AI industry developments clearly outline several major themes: coding remains the strongest monetization scenario, Agents are moving from APIs to interface operations, tech giants are accelerating their positioning at key value chain nodes, and compute and energy are becoming the foundational pillars of competition. The industry's boundaries continue to expand, and the competition is far from over.