Anthropic Aims to Surpass SpaceX's Historic IPO

Anthropic aims to surpass SpaceX’s $75 billion IPO record

Anthropic is preparing the public filing of its IPO prospectus documents, possibly as early as the end of August. According to people familiar with the plans, the company aims to match or exceed the offering size of SpaceX. SpaceX had raised $75 billion in June and increased this amount to $86.2 billion via an overallotment option. Anthropic had already confidentially submitted a draft registration to the SEC on June 1. Discussions about the volume and details are ongoing and may still change.

The public filing would make the company’s audited financial data available for the first time. So far, only individual figures from investor talks are circulating: In the second quarter, Anthropic generated more than $11.5 billion in revenue, compared to $787 million in the same quarter last year. The run rate was over $65 billion at the end of July. For 2025, however, a net loss of nearly $42 billion is on the books, after $8.3 billion the previous year. Leaked figures reportedly show a small operating profit in the second quarter.

The valuation has jumped within a few months: In February, Anthropic raised $30 billion in a Series G at a $380 billion valuation, and in May, $65 billion in a Series H at a $965 billion post-money valuation. This places the company above OpenAI’s $852 billion valuation from March. Citigroup is expected to join Morgan Stanley, Goldman Sachs, and JPMorgan in the banking syndicate. CFO Krisha Rao recently led investor briefings but left questions about a specific valuation unanswered.

The growth is driven by the enterprise business surrounding the coding-agent Claude Code, as well as other offerings like Claude Cowork and integrations into Microsoft 365. At the same time, both frontier labs are committing to enormous infrastructure pledges: Anthropic has signed a three-year contract for compute capacity with SpaceX, which could reach double-digit billion-dollar figures.

California startups have raised around $366 billion in venture capital this year, with half of that going to Anthropic and OpenAI. The US IPO volume stood at $160.6 billion on August 19, with the 2021 record at $195.2 billion. Both companies have filed confidentially, with Anthropic pushing the pace, while an IPO from OpenAI is more likely expected towards the end of the year or early 2027. → Yahoo Finance, PYMNTS, SiliconANGLE, Bloomberg Law, Bitcoin Insider

Synthszr Take: A company with a $965 billion valuation and a net loss of nearly $42 billion in 2025 has yet to present an audited balance sheet where these two figures meet. What’s circulating are metered snippets from investor briefings: $11.5 billion in quarterly revenue and a $65 billion run rate, garnished with a rumored small operating profit. Meanwhile, four Wall Street firms are arranging their spots in the syndicate, while the company’s own CFO leaves any valuation questions unanswered in these very briefings. The record Anthropic wants to break belongs to a company from which Anthropic is simultaneously buying tens of billions of dollars worth of computing power; the capital is circulating within a very small circle. The S-1 at the end of August will be this company’s first document whose content it doesn’t control, and until then, the valuation is running several laps ahead of any regulatory review.

ChatGPT can now read and send Apple Messages

OpenAI has released a plug-in that connects ChatGPT to the Apple Messages inbox. According to the company, it can be used to sort, analyze, and edit messages, as well as draft and send replies on behalf of the user. Deleting individual messages and searching through long-past conversation histories are also part of the feature set. A promotional video shows a user getting suggestions for follow-up messages to their contacts based on the messages received the previous day. According to TechCrunch, the plug-in also works with Codex and ChatGPT Work, so it’s not limited to private use. Regarding data privacy, OpenAI told Bloomberg that the feature runs locally on the user’s computer and does not create an index of all messages. What this means in practice remained unclear; TechCrunch has inquired further. The company also advises its users against enabling permanent permission, as this removes the final opportunity for review before ChatGPT sends a message under someone else’s name. → TechCrunch

Synthszr Take: The plug-in reads your history, formulates replies, and sends them out under your name. It runs locally on the machine, so you need the desktop app, and because it integrates with Codex and ChatGPT Work, the IT department—not the individual employee—decides on its approval in a corporate context. The real operational decision lies in a single switch for permanent permission: Once activated, the final glance at the text before it lands with a customer or your boss disappears. OpenAI itself warns against this, unusually blunt for a product announcement, yet the warning will likely fade away because any confirmation dialog is perceived as a nuisance after three days. The tool is already useful for search and drafting; it’s better to keep the sending part to yourself until permission can be set per contact instead of wholesale for the entire inbox.

Ex-manager blames Zuckerberg’s culture for youth protection failures

Arturo Béjar testified on Wednesday as the first witness in the case against Meta in Oakland, stating that the corporate culture shaped by Mark Zuckerberg made it practically impossible to ship features for safety and well-being. The trial has been underway since Tuesday before Judge Yvonne Gonzalez Rogers at the U.S. District Court for the Northern District of California; the plaintiffs are the Attorneys General of California, Colorado, Kentucky, and New Jersey. Béjar worked at Meta from 2009 to 2015 and again from 2019 to 2021 on the very issues now at stake, and later took his internal warning about harm to teenagers to the U.S. Senate. The states base their argument on design decisions like endless scrolling, auto-playing videos, beauty filters, and the like button; Béjar’s contribution is the testimony that countermeasures were delivered as optional settings rather than defaults. In the afternoon, Meta’s lawyer Brian Stekloff asked him if he or his teams had solved any of the problems described. → Techpresso

Synthszr Take: Building protective features as an opt-in is a product decision, not a technical limitation. Anyone who has ever designed a settings page knows that defaults almost completely determine behavior and that a deeply buried option statistically does not exist. The price for this was paid by thirteen-year-olds dealing with unwanted messages and comparison pressure in their feed, while the engagement curve on the dashboard continued its clean upward trend.

OpenAI CFO Friar internally calls the IPO just 'another funding round'

OpenAI’s CFO Sarah Friar has downplayed a potential IPO to her own staff: “The IPO is not a finish line, it’s a milestone, another funding round”, CNBC quotes her from an all-hands meeting on Wednesday, citing two anonymous insiders. She pointed to the $122 billion raised in March, which gives the company flexibility. CNBC had already reported in June that OpenAI had confidentially filed an IPO prospectus with the SEC (Confidential Filing), as had Anthropic. Friar reportedly said that Anthropic could lift its confidentiality in the coming weeks and go public in September: “That’s okay, we’re running our own race.” The appearance comes at a time when OpenAI, according to its own blog post on Tuesday, is slowing down the pace of new models and delaying the release of Astra to meet, by its own account, stricter safety standards. → Gizmodo

Synthszr Take: When a CFO declares an IPO a routine matter in front of her own team, she’s talking to people whose stock packages are tied to this exact event, and they listen differently than an analyst. The sentence about the 122 billion from March is the real internal message: The money is there, no one has to wait for a stock ticker. The reference to Anthropic, which Friar says could go public as early as September, inoculates the workforce against a headline before it even appears.

Cerebras delivers the CS-4 and promises 30x faster inference than GPUs

Cerebras has introduced the CS-4, a new inference system, that, according to the manufacturer, delivers up to 30 times more tokens per second and user than GPU-based systems. The system is built from three chips of the Wafer Scale Engine 3 Turbo generation. Cerebras specifies the computing power at 750 PFLOPS and the memory bandwidth at 129.6 petabytes per second. The throughput per watt is said to be up to ten times higher than its predecessor, the CS-3. The company states the latency between wafers is as low as two microseconds. → AlphaSignal

Synthszr Take: Cerebras is selling speed, but what’s being bought is operational leeway for agents. If a user gets a thirty-fold token rate, an agent can think, check, and call tools thirty times more often in the same waiting period. This is exactly the leeway needed by Cursor’s new architecture, where agents wake up on their own in response to Slack messages or PR updates and launch sub-agents on their own virtual machines. This aligns with the Stanford paper on 10,000 language model agents, which shows that such populations either find consensus or tip into polarization.

Box CEO Aaron Levie sees the value of enterprise AI between the model and the workflow

Aaron Levie, CEO of the software service Box, contradicts the common assumption that more powerful frontier models devalue the application layer. In his view, the biggest opportunity in the enterprise use of artificial intelligence lies in the space between the model and the real-world workflow. Specifically, he names six layers that companies will still need to provide themselves: domain-specific interfaces, data integration, change management, model routing, evaluations, and industry-specific pricing models. According to Levie, these layers create a differentiation that no single model can replace. → MyClaw Newsletter

Synthszr Take: Levie is defending his own business here, and yet the argument holds up. Of the six layers he lists, not a single one is a model capability: data integration and industry-specific pricing logic arise from legacy systems and customer contracts, not from a training run. The bottleneck in companies has been in the same place for years, namely the question of who is allowed to act on an agent’s suggestion and who provides the countersignature; a more powerful model doesn’t move this question a single millimeter. However, this moat is only solid as long as the application layer is built to be model-agnostic, with interchangeable routing instead of prompts hard-wired to a specific model’s behavior.

Australia will now tax Meta and Google even if they don’t show any news at all

Australia’s parliament on Thursday passed the News Bargaining Incentive, which imposes a 2.5 percent levy on the local advertising revenue of platforms with a significant search or social media presence in the country and more than 250 million Australian dollars (about 178 million US dollars) in local ad revenue, regardless of whether they display any journalistic content. The scope includes Meta, Alphabet’s Google, TikTok, and Microsoft’s LinkedIn; LinkedIn was added in early August along with the increase from the original 2.25 percent. The levy can be reduced or eliminated entirely through agreements with at least eight Australian news organizations, with deals with large publishers credited at 150 percent of their value and those with small and medium-sized providers at 200 percent. No single contract may offset more than a quarter of a company’s total liability. The crucial factor for credit is that the deal is concluded before the end of the platform’s respective fiscal period. → Techpresso

Synthszr Take: In 2024, Meta exited the old code in compliance with the rules, because a news law doesn’t apply where there is no news. Canberra has learned from this and shifted the tax base: away from content and toward local advertising revenue, at a rate of 2.5 percent starting at 250 million Australian dollars. The real craftsmanship is in the settlement rules: the 200 percent credit for agreements with small publishers and the cap ensuring no single contract clears more than a quarter of the debt render the convenient large check to a major publisher worthless.

Slack opens agent installation to ten third-party platforms, with partners handling OAuth

On August 20, 2026, Slack rolled out two agent features, the more important of which is called “Add to Slack”: Users can bring agents they’ve built on ten external platforms into their workspace without having to program a custom Slack integration. The installation starts in the partner’s tool, and the partner handles OAuth, app configuration, and permission scoping. The runtime environment remains outside of Slack, either with the provider or, for tools like NanoClaw, on the user’s computer or in their cloud account. The Slack app then receives messages and mentions and can join channels. The launch partners include Hyperagent, LangChain, Lovable, n8n, NanoClaw, OpenAI, Runlayer, Skydive, Superhuman, and Vercel.

According to Slack, installed agents inherit the workspace’s data boundaries and permission controls; the organization’s app approval policy still applies, and the OAuth scopes plus channel membership determine access. Previously, this required creating a custom Slack app and clicking through configuration, scopes, and authentication. With NanoClaw, the marketplace app becomes a manager app that creates a separate managed Slack app for each agent, each with its own WebSocket connection. In a demo, NanoClaw developer Gavriel Cohen had one agent create an entire team of other agents, which appeared in the channel under their own names and could be addressed individually.

The second new feature, Slack Code, is separate from this. When a coding agent is tagged in a conversation, it opens a project-specific code channel where the plan, code diffs, pull request details, and a live preview of the HTML output are visible to everyone. The team can provide feedback, pause or redirect the agent, and approve the work; according to the provider, high-risk steps like merging into production require human approval. Upon completion, the channel archives itself, leaving a log of who commissioned, changed, and approved what. Slack Code is available from day one on all plans, including free workspaces, and works with agents from the Slack Marketplace; the company names Claude Code, Devin, Vercel Agent, and GitHub Copilot as founding partners, along with ChatGPT.

Marc Benioff presented the whole thing under the label “multiplayer coding” at Dreamforce 2026. The four partners had already built parts of this workflow before: GitHub brought its Copilot coding agent to Slack in October 2025, Anthropic introduced Claude Tag as a shared identity for selected channels in June 2026, Cognition has long treated Slack as the primary interface for Devin, and Vercel provides the connection from Slack events through approval steps to deployment. This does not involve a common execution environment or a shared model layer; only the channel is shared. Additionally, there are Agent DMs and an Agents tab, where sessions can be found and running agents can be stopped. Slack announced it will open up the APIs more broadly later, including for agents outside of software development. → Salesforce, The New Stack, The Verge, The Register, Gizmodo, Salesforce, SiliconANGLE, WeRSM, unite, RuntimeWire

Synthszr Take: Slack is building a junction box here instead of its own agent ecosystem, and that’s the smarter decision: Ten external platforms bring their own agents, the runtime remains with them, and Slack only keeps what no one else owns anyway, namely identity, channel membership, and the workspace’s approval policy. The real progress lies in an unspectacular shift: OAuth, app configuration, and permission scoping are now handled by the partner, no longer by an admin who clicks through scope lists and hopes for the best. The NanoClaw pattern shows where this is heading, when a manager app creates its own managed Slack app with its own connection for each agent, and a single click suddenly turns into half a dozen bot identities. The approval policy was designed for a few dozen apps per year, not for agents that create each other; that’s precisely where the unresolved issue lies. The interesting number in the coming months will be the number of installed agent apps per workspace, and at what point the administration capitulates.