Bedrock · 138

Anthropic 40% of ARR is sold by cloud vendors: $65 billion in annualized revenue is not that easy to earn

Comparative news, AI news, Anthropic's annualized revenue just reached $65 billion, and SemiAnalysis then split this revenue. According to its model estimates, more than 40% of ARR in the second quarter came from indirect channels such as AWS Bedrock, Microsoft Foundry, and Google's enterprise AI platform. What is worth paying attention to is how much profit these revenues can leave behind. In Bedrock, for example, Claude was sold by Anthropic. Anthropic will count the total amount of tokens sold into ARR, and then pay AWS for computing power and channel sharing. In other words, it's also a $1 ARR. If you sell it through a cloud platform, Anthropic will end up leaving less money than direct sales. So while $65 billion ARR isn't fake income, the revenue structure is clearly not that healthy. The higher the share of channels, the less direct equations between revenue growth and profit growth. If you only look at ARR, you might be overestimating the contribution of these revenues to Anthropic's final profit. Of course, there are benefits to the channel model. AWS, Microsoft, and Google already have a large number of enterprise customers and procurement contracts, and can directly cram Claude into existing cloud bills. Anthropic is now trading some of its profits for scale and customer acquisition efficiency.

3d ago

SemiAnalysis: Anthropic Q2 over 40% ARR comes from indirect channels

In comparison, according to SemiAnalysis, Anthropic's ARR from indirect channels (Bedrock, Foundry, Gemini Agent Enterprise) in Q2 2026 accounted for more than 40%, and API and B2B contributed most of the new revenue. Indirect channels differ from direct revenue monetization models. Cloud service providers usually charge IaaS fees or revenue shares. Laboratories include gross ARR, but related costs are reflected in sales and marketing expenses, and indirect channels require computing power allocation analysis accurate to the level of a single transaction.

3d ago

AI infrastructure upstart Nscale is preparing to go public in the US. It will be an IPO in September as soon as possible. Goldman Sachs and J.P. Morgan Chase act as advisors

Comparatively, according to TechFundingNews, AI infrastructure company Nscale is preparing to go public in the US and may launch an initial public offering (IPO) as soon as September this year. The company has disclosed to potential investors that cumulative contract revenue amounts to approximately US$51 billion. Currently, Goldman Sachs and J.P. Morgan are advisors for Nscale's potential IPO, but related discussions are ongoing, and the listing may still be delayed. Nscale is a spin-off company from crypto mining company Arkon Energy, which mainly revolves around AI data centers, GPU computing power, and energy infrastructure. In July of this year, it was announced that it would acquire distributed AI software company Anyscale for approximately $1.65 billion, whose customers include Coinbase, Runway, and Bedrock Robotics. Earlier, AI infrastructure Nscale completed a $2 billion Series C round with a valuation of $14.6 billion, led by Aker ASA and 8090 Industries, with Nvidia, Lenovo, and Nokia participating.

12d ago

Crypto Agent commercialization is accelerating, why are stablecoins the most critical part?

Core view: For AI agents to become real economic agents, the core obstacle is that traditional payment systems cannot support their autonomous payments. Stablecoins represented by USDC, along with dedicated infrastructure launched by companies such as Coinbase, Circle, and Stripe, are building a native programmable, all-weather, small, high-frequency “currency layer” for AI agents, spawning a program-driven on-chain microeconomy. Key elements: 1. Four major barriers to traditional payments: Agents cannot pass the identity barrier (no ID card), authorization (verification code required), time (not 7 x 24 hours), and cost (high fixed processing fee), and cannot perform small-amount high-frequency transactions. 2. Native advantages of stablecoins: programmable (automatic code execution), no license (self-generated wallet), 7 x 24 hours, transparent accounts and stable value, perfect for agent payment needs. 3. Implementation practices of leading companies: Coinbase launched AgentKit and X402 protocols (more than 50 million transactions have been processed); Circle launched the CCTP cross-chain protocol and AgentStack; Stripe launched a stablecoin API and supported USDC subscription payments. 4. Typical application scenario 1 (ultra-small payment): The x402 protocol and Circle's Gateway Nanopayments achieve $0.000001 micropayments, unlocking the long-term economy of pay-per-use billing for API calls, data access, etc. 5. Typical application scenario 2 (automatic generation): AI agents can achieve “self-hematopoiesis” through yield-bearing stablecoins (such as aUSDC), cover operating costs with interest, and platforms such as Ymax can achieve 8-12% annual stablecoin returns. 6. Large-scale implementation challenges: Private key management is vulnerable to attacks (such as the Owockibot incident), gaps in compliance (agents cannot be identified), and inaccurate AI intentions may lead to irreversible financial losses. Generative AI is changing from a “chatbot” to an AI agent (AI agent) that can do things by itself. A real question then popped up: How do these silicon-based “employees” receive money and how do they pay? Traditional banking stuff — real-name authentication, manual authorization, public accounts — inherently disapproves of AI agents. One answer that is rapidly evolving is to use stablecoins (USDC, USDT, and stablecoins with interest) to create a native “currency layer” for AI. This article will break down the implementation of leading companies such as Coinbase, Circle, and Stripe in this field, while also discussing compliance and security risks. The technical infrastructure is ready, but how to drive it is still a big problem. 1. The “payment breakpoint” encountered in the commercialization of AI agents Today's AI agents are already very capable: book air tickets, write codes, adjust interfaces... but they get stuck as soon as they get to the “payment” step. Traditional payment systems are designed for humans — you have to have an ID card, enter a verification code, operate on weekdays, and have a low processing fee for each transaction. These are all barriers for agents. Specifically, traditional payment systems set up four hurdles for agents: identity barriers: opening a bank account or credit card requires an ID card, face recognition, or even bank transactions, and agents can't even pull it out. Authorization: SMS verification codes, manual confirmation, and 3D security authentication are often required during payment, and agents cannot click buttons even if they cannot receive SMS. Time limit: Banks only process transfers on weekdays and business hours, while agents work 7×24 hours. Cost barrier: Each transaction has a fixed processing fee, such as starting at 30 cents for credit cards, so the pay-per-use model of $0.001 doesn't work at all. However, the financial behavior of agents requires exactly this kind of small, high-frequency charge (such as per number of API calls, per usage). The more fundamental problem is that the entire payment system has never considered direct “program to program” transfers. Even between two technology companies, the process is often: the agent generates an order → sends an email → person approves → person logs in to online banking to transfer money → each other's financial reconciliation. The agent can only do the first two steps and the final record. The most important step, “money from A to B”, must be done by hand. Current experiments: they are all modelling...

18d ago22#AI #stablecoins #wallets

IOSG: Why are Wall Streeters saying “no” to ChatGPT and Claude?

Author: IOSG Ventures Original title: IOSG Weekly Brief｜AI's Crossroads: Why Is Wall Street Saying “No” to ChatGPT and Claude? #336为什么需要私有 AI On July 1, Palantir CEO Alex Karp contributed 20 minutes of an interview on CNBC called a “mental breakdown” by some media. According to Karp, the company is paying a token premium to Frontier Labs while watching its IP flow to model vendors. He called this leak an alpha transfer, and the transfer is happening at the architecture layer: every request sent to the closed source model arrives at the service provider's server in plain text. Just a few days before the broadcast of the program, Palantir just announced a partnership with NVIDIA to run an open Nemotron model in a customer-controlled environment, along with a nine AI sovereignty declaration. PLTR jumped 8% after the CNBC show aired. Over the past 20 years, enterprises have relied on agreement level trust to adopt cloud software, and it works. Every SaaS vendor sees only slices of enterprise data, and most have little incentive to feed back to core products with customer data. Salesforce sees sales channels, Workday sees personnel, Jira sees development iterations, and AWS provides the foundation for storage and computing. Today's AI workflow, however, advocates uploading all the household items at once, and stringing together the structured context of each department to maximize productivity. Goodwill aside, upstream service providers can now use this data for new functions instead of leaving them lying in the server eating dust. No one is slowing down. Anthropic's annualized revenue reached $47 billion in May, a sharp jump from $9 billion at the end of 2025, while OpenAI surpassed 900 million weekly active users in February. Both companies completed a new round of financing this spring, are valued at close to $1 trillion, and are expected to IPO at higher market capitalization. Years of privacy and IP accusations haven't caused the two companies to lose any momentum. Some companies have already taken action. In February 2023, less than three months before ChatGPT was released, major Wall Street banks restricted its use. In May 2023, after Samsung engineers leaked the chip source code to ChatGPT, the company banned generative AI across the network. In response, OpenAI launched ChatGPT Enterprise in August of that year, promising not to use commercial data training, plus a zero-data-retention (ZDR) n agreement, which has since become a standard requirement for corporate procurement. However, the contract only locked the company account. IBM found that by 2025, shadow AI (employees feed company data into unapproved AI tools through personal accounts) was involved in one-fifth of data breaches, and heavy shadow AI use added an average of 670,000 dollars to the cost of the breach. In a 2025 survey by safety training company Anagram, four workers said they were willing to violate AI usage policies in order to complete tasks faster. Businesses can at least spend money to buy roads, ZDR contracts, untrained service files, if you're a government or Palantir customer and sovereign deployment. However, for ordinary users like you and me, the importance of AI privacy is still debated until the court subpoena was found. A court order in May 2025 forced OpenAI to keep even consumer chats that users had deleted. In November, the judge also ordered 20 million of these to be handed over to the “New York Times” lawyers as evidence disclosure materials. Then the criminal case: the ChatGPT records of the defendant in the Palisades arson case entered the evidence, and the affidavit for the murder of two dead in Florida cites questions from the suspect about how to dispose of the bodies. Sam Altman also admitted in an interview in July 2025 that ChatGPT conversations are not protected by legal privileges, and OpenAI “may be asked to hand over” user chat records in lawsuits. The point is not that only criminals need intimate conversations. People's conversations with AI are archived and can be summoned, and most users don't know...

38d agoburnking#AI #Claude #GPT #IOSG

SemiAnalysis: Meta will accelerate the procurement of computing power rather than slow it down, and is in negotiations with Anthropic to build its own AI model service platform

Comparing news, SemiAnalysis said in its latest report that after the news reports that Meta may become a new Neocloud, the market's first reaction was to sell off computing power cloud companies such as CoreWeave and Nebius, and once again worried about excessive AI computing power. But the agency's judgment is just the opposite: this fear is probably wrong. Meta's data center and computing power purchases will not slow down; on the contrary, they will continue to accelerate. The article mentioned that in the first half of this year alone, Meta has signed more than 5GW of capacity in the field of cloud services and hosted data centers, and this does not count as a self-built project that it is accelerating. SemiAnalysis says Meta is in final negotiations with Anthropic to gain access to Claude's privatized instance. If this is true, the point is not only that Meta is buying more computing power, but that Meta is probably building its own AI model service platform. This model is a bit like AWS's Bedrock, Microsoft's Foundry, Google's Vertex. Meta can either use Claude internally first, or package model capabilities into token-as-a-service in the future to provide external services. In the short term, it may be that its own model is used externally and that of the Anthropic model; in the long run, Meta may even incorporate Anthropic and OpenAI models into its own external service system. According to SemiAnalysis, the logic behind it is that Meta has computing power, advertisers, social network distribution capabilities, and a consumer portal. If it can combine cutting-edge models, agents, sales, and marketing SaaS, it won't just be a company that buys GPUs, but will go above and beyond AI applications and model distribution in the past.

50d ago

Meta is selling computing power, why is the market panicking

On Wednesday, a Bloomberg news blew up the US stock market — Meta is building a cloud business and plans to sell excess AI computing power to external customers. The stock price immediately reacted: Meta surged more than 10% intraday, the biggest one-day increase since January 29, 2026. However, hardware vendors such as Micron, Corning, and SanDisk collectively plummeted, reaching 6%, 11%, and 8%, respectively, while “computing power intermediaries” such as CoreWeave and Nebius fell by more than 10%. What exactly is Meta going to do? In fact, Meta's plans to enter cloud computing are not temporary. As early as October 2023, Zuckerberg proposed this “alternative.” In April of this year, Meta raised its 2026 AI-related capital expenditure forecast to 125 billion to 145 billion US dollars. After the financial report, the stock price plummeted by more than 7%. The market's patience with Xiao Za's “smash the money first, then talk about it” approach has been exhausted. At the annual shareholders' meeting in May, Zuckerberg clearly stated that entering the cloud computing market is “definitely within the scope of consideration.” He also revealed a key message: “Almost every week, different external companies come to us hoping we can build an API service, or ask if they have computing power to sell, and are willing to buy it at a premium above our procurement costs.” According to information obtained by Bloomberg, currently Meta is mainly considering two models in the direction of commercialization: First, selling “model access rights” — similar to Amazon AWS, developers pay to call AI models deployed on the Meta infrastructure (including the self-developed Muse Spark model), and Meta charges customers based on the number of API calls. Second, selling “raw computing power” — leasing raw computing power directly to external customers. The model is similar to the approach of “new cloud” service providers such as CoreWeave. The plan was incorporated into an internal top-level program called “Meta Compute,” which is co-led by three key players: Santosh Janardhan, head of Meta infrastructure, Daniel Gross, head of the AI department, and Dina Powell McCormick, president of Meta. It's worth mentioning that Meta moves are not an exception. Musk's SpaceX has launched a similar operation this year, selling idle computing power to Anthropic and Google. The two pessimistic interpretations believe that this is the first domino card for “oversupply” of AI computing power. Their reasoning is simple: if big model training really “eats up computing power,” why doesn't Meta keep it for itself and instead take it out? This shows that Meta's own R&D needs are no longer enough to fill its computing power warehouse, so they may have actually bought too many chips in the past two years. What makes the market even more tense is that Meta, as one of the “weather vane” of the industry, if it takes the lead in bringing excess computing power to the market, it means that the tight logic that the entire AI hardware is “always out of stock” will have to be re-examined. If other major manufacturers (such as Google and Microsoft) follow suit, the order growth rate of upstream suppliers such as Samsung, SK Hynix, and TSMC will probably put the brakes on, and even face the risk of cutting orders. This is why hardware stocks such as Micron and SanDisk have been hit hard — the market is digesting the expectation that “peak demand has passed” ahead of schedule. The optimists, on the other hand, think this is simply an overinterpretation. The logic is exactly the opposite: Meta spent hundreds of billions of dollars to build a computing power pool, and now “return blood” with idle resources that cannot be used for a while. This is just a normal asset management operation; it has nothing to do with “shrinking demand.” On the contrary, if the path of cloud leasing passes and can generate stable cash flow, Meta will be more motivated to continue to make major purchases of next-generation GPUs, optical modules, switches, and cooling systems — because the money earned from selling computing power can in turn support larger capital expenses. To put it bluntly, leasing is for better procurement, not the end point of procurement. If this logic works, then the decline in hardware stocks is an emotional “misslaughter”; in turn, it may be an opportunity to pick up bargains. What do analysts think? Judging from research reports, most institutional analysts are more inclined to make “optimistic” judgments. Bank of America (BofA) reiterated Meta's “buy” rating in a Wednesday briefing and gave a very sophisticated perspective: enterprise-grade AI cloud services are equivalent to investing 100 billion dollars in insurance for Meta — even if the consumer AI business (such as advertising) falls short of expectations, corporate computing power leasing can cover the bottom, preventing profit margins from collapsing. Bo...

52d agoWendy#AI topics #Meta #original #Arithmetic power #US stocks

Gross profit margin plummeted by 10 percentage points. How much did CBRS's business model transformation cost?

Author: David, Tide Research Original title: First financial report on the CBRS listing: revenue doubled but gross margin guidance plummeted, OpenAI's big order fulfillment path was too long Tide guide: Cerebras (CBRS) handed over its first quarterly report after the IPO. Q1 core revenue was US$191 million, up 92% year over year, exceeding market expectations. However, the Q2 core gross margin guideline plummeted from 46.5% to 36%-38%, and the stock price fell more than 10% after the market. This company, which uses an entire wafer as a chip and bets on the AI inference circuit, has an OpenAI contract of over $20 billion and the AWS cooperation framework, leading to annual revenue of 855-865 million US dollars. The growth data is hard enough, and the valuation controversy is big enough. The core focus was that revenue exceeded expectations, and guidance even exceeded expectations. Q1 core revenue of $191.3 million (+92% YoY) was higher than the agreed estimate of approximately $181 million. Core revenue guidance for the full year was US$855-865 million (+69% YoY), higher than market expectations of US$828 million. Under GAAP, cloud and services revenue was US$82.8 million, up 178% year over year, making it the fastest growing sector. The sharp drop in gross margin guidance was the biggest negative of the season. The Q1 core gross profit margin was 47%, up nearly 5 percentage points year over year. However, the Q2 guidance fell to 36%-38%, down about 10 percentage points from Q1; the full-year guidance was 38%-41%. Management attributed the reason to insufficient data center capacity: the company was temporarily leasing back systems from existing customers who had sold hardware to deploy capacity, and short-term costs worsened as a result. Shares fell more than 10% after the market. There is a direction for improvement in customer concentration, but it is far from being solved. 86% of revenue for fiscal year 2025 came from two UAE related entities (MBZUAI 62% and G42 24%). OpenAI began contributing revenue in February 2026, and the AWS partnership is expected to be reflected financially only in 2027. True income diversification will not be verified until 2027. The valuation price is until 2028. Based on about $200 after the market, CBRS corresponds to about 90 times the revenue of the past 12 months; even with a median value of US$860 million using the full-year guidance, the long-term P/S is still more than 50 times higher. The 10 coverage analysts had a median target price of $300 (range of $250-340), implied that OpenAI's contract of over $20 billion and the AWS deployment were fulfilled on time and in volume. Short-term catalytic and repressive factors coexist. Catalysts: Accelerated deployment of OpenAI 750MW computing power, implementation of the AWS inference solution, and launch of new data center production capacity in the second half of the year. Constraining factors: The lockdown period included unconventional early lifting clauses (the market capitalization can be triggered when the market value exceeds 40 billion US dollars; the current market value has reached near this threshold), the gross margin recovery path is unclear, OpenAI itself is not profitable, and it has promised to reduce part of its computing power. Financial reports reveal the transformation of business models: from selling chips to selling computing power The most easily overlooked in Q1 earnings reports is the change in revenue structure. Under the core caliber, hardware revenue was US$116.6 million, accounting for 58% of total revenue; cloud and service revenue was US$79.8 million, accounting for 42%. In the same period a year ago, the ratio was roughly 70:30. Cloud service revenue increased 167% year over year, nearly three times that of hardware. Management made this trend more clear during the conference call: hardware revenue will decline in stages in the next few quarters, as the company will deploy more hardware production capacity to its own cloud to fulfill inference computing power contracts with OpenAI and AWS rather than selling it directly to customers. Cerebras is changing from a “company that sells chips” to a “company that sells computing power.” This transformation also directly explains why Q2 gross margin plummeted. On the phone call, an analyst inquired about the details of production capacity deployment. Management revealed that the company's current bottleneck is not TSMC's chip supply, but rather the physical space in the data center. In order to deliver computing power to OpenAI as soon as possible, Cerebras is “temporarily leasing back” the hardware system already sold from G42 (previously its largest customer and minority shareholder). If you rent a third-party facility to deploy your own system, the cost structure will deteriorate in the short term. This is the main reason for the decline in gross margin from the 47% guideline to 36%-38%. The schedule given by management is in the new data for the second half of the year...

59d agoburnking#CBRS #OpenAI

Anthropic opens Seoul office and announces new collaboration with Korean AI ecosystem partner

In comparison, Anthropic announced the official opening of its Seoul office and the establishment of new partnerships with companies, startups, and research institutes in the Korean AI ecosystem. NAVER has recently deployed Claude Code across the engineering organization, and currently thousands of engineers are using Claude Code to improve coding productivity; the engineering team at global online gaming company Nexon is also using Claude Code to write, review, and publish code for real-time service games. For large enterprises, LG CNS is rolling out Claude to thousands of employees and plans to deploy it further within the LG Group; Hanwha Solutions is providing Claude to global employees through AWS Bedrock to meet regional data residency and security requirements; and Samsung SDS is deploying Claude to Samsung Electronics employees for everyday knowledge work, intelligent workflows, and software development. Additionally, Claude for Startups has launched in Korea, and Anthropic is co-hosting Claude Build Day with BASS Ventures this week.

66d ago

RootData: BR to unlock around $5.54 million worth of tokens in a week

According to the Web3 asset data platform RootData token unlock data, Bedrock (BR) will unlock about 40.63 million tokens worth about US$5.54 million at 12:00 Beijing time on June 20.

70d ago