Anthropic wants to become the new Accenture

Anthropic enters the consulting business, competing with Accenture & Co

A new company called Ode with Anthropic launched this week under its full name, set up by Anthropic, Blackstone, and Hellman & Friedman as a $1.5 billion venture, as first reported by TechCrunch. Ode sells AI implementation: small teams of senior engineers go into a company, find the areas where AI can help, and then build the systems to do it. The approach is 'Claude-first,' meaning it preferentially uses Anthropic's models and competitors' where necessary. CEO Chris Taylor told TechCrunch it's 'pretty easy to imagine this as a trillion-dollar company one day.' Ode operates with around 100 engineers, more than half of whom have founded startups themselves; a Blackstone manager called them 'special forces.' The basis is the applied AI boutique Fractional AI, acquired in May, which ended an eleven-month partnership with OpenAI for this purpose. Co-founder Eddie Siegel downplays the model question: The choice matters, but that's not where most of the energy is burned. → Techpresso

Synthszr Take: Siegel is right, and the $1.5 billion is the price for him being right. Most enterprise pilots never reach production because no one has clarified beforehand what 'solving the problem' actually means. This is precisely the new scarcity: the intent to know exactly what should be built in operations and where it connects. Ode sells this clarification work as a service, and the fact that Anthropic is downgrading its own model to an interchangeable ingredient ('matters, but it does not decide the project') is the most honest market analysis a model manufacturer can currently provide. The 100 expensive engineers are proof that this work doesn't scale, at least not yet: you can't roll it out via API; you have to carry it into each company individually. It will be interesting to see if OpenAI's The Deployment Company, Deloitte, and Accenture can industrialize the same craft more cheaply before Ode can maintain its special forces margin. The HubSpot data uproar already shows where the real test lies: in the permission hooks and the question of whether the customer trusts the whole thing.

Xi warns of 'new historical injustices' at Shanghai AI conference

President Xi Jinping warned at the World Artificial Intelligence Conference 2026 in Shanghai against creating 'new historical injustices' in the age of artificial intelligence. In his speech, he called for more support for the Global South's access to AI and a more open, inclusive development approach. It was Xi's first personal appearance at the conference since its launch in 2018, which, according to the South China Morning Post, is seen as a signal of Beijing's ambition to lead global AI governance. Xi urged that AI must remain 'safe and controllable' and spoke out against the increasing securitization of the technology: No country should place its own security above that of others or overstretch the concept of national security in the AI field. The speech came shortly after US President Trump attacked China in a prime-time address, accusing it of the 'largest compromise of election data in history.' Beijing thus positioned itself as an advocate for an open approach at a time when export controls and blacklists restrict access to advanced technology. → Techpresso

Synthszr Take: This speech is a rhetorical maneuver, packaged as a moral offer. 'Open, inclusive, for the Global South' sounds like fairness, but it means market access: where the West locks down access to chips and models with export controls and blacklists, China sells its open weights as the more generous operating system for everyone who can't get Nvidia. Xi turns the West's concept of security against itself when he warns against the 'overstretching of national security,' while reserving the same vocabulary ('safe and controllable') for his own claim to control. The timing, right after Trump's election data accusation, is too precise to be a coincidence; this is a gesture of sovereignty responding to a gesture of sovereignty. The real audience is in Jakarta, Nairobi, and São Paulo, where decisions are currently being made about whose AI stack the next economic generation will be built on. Rhetoric doesn't win races, but it occupies the terms by which scores will be settled later, and Beijing has just secured the framing of fairness with a single word.

Almost every second token in US developer tools now comes from China

According to Hello China Tech, Chinese AI models have gone from being a cheap alternative to productive infrastructure in American developer tools within just a few months. On OpenRouter, a routing platform for developer APIs, Chinese models have accounted for more than 30 percent of the weekly token volume from US companies since February 8, with a peak of 46 percent. The speed of adoption is seen as a visible driver. Z.ai's GLM-5.2, released in June, saw the fastest initial adoption of any model in 2026 on Vercel's platform. The daily token volume there grew by about 27-fold, and the number of customers increased by about 80 percent. The report interprets these figures as proof that the models are being used in live operations. → Hello China Tech

Synthszr Take: A 46 percent peak share among US companies is a statement, and one without flags. No developer team routes almost half of its token volume to China because it has a geopolitical opinion. It routes there because GLM-5.2 gets the job done at a fraction of the price, and the switch is a single line of configuration. In June, the headline was 'GLM-5.2 beats GPT-5.5 at a fraction of the price'; now it's a production reality in American toolchains. Origin has become invisible at this API interface, and invisibility is the most powerful salesperson there is. If Washington wants to regulate the origin of models, it will have to contend with a demand that already manifests in the token log of every second request. It's only a matter of time before someone tries to turn back the traffic and realizes that the switching costs are now on them.

OpenAI burns $21 billion and spends almost three dollars for every dollar of revenue

Leaked financial figures cited by Scott Galloway show: OpenAI posted a loss of $21 billion in 2025. According to Galloway, the cost structure is disproportionate to the revenue, as for every dollar a subscriber spends on ChatGPT, the company spends almost three dollars. Added to this are the departure of several top managers, a lawsuit from Apple, and reports that OpenAI might postpone its IPO until 2027. Galloway explicitly draws a parallel to the collapse of the dot-com and telecom bubbles around 1999. As evidence for what he sees as unrealistic planning, he cites the forecast of generating $100 billion from advertising. His verdict: The business model is reminiscent of a large language model's hallucination. → Scott Galloway

Synthszr Take: A model that spends three dollars to make one has a math problem. With physical products, unit costs decrease with volume; here, they increase because every new user triggers real compute costs. The $100 billion advertising forecast is the point where the plan falls apart: A company that lures its users today with ad-free intelligence wants to turn them into advertising space tomorrow, on a scale that even Meta took twenty years to reach. This is the post-rationalization of a deficit, poured into a revenue slide. The 1999 parallel only holds halfway, because back then, companies were burning venture capital; today, an entire compute supply chain from Nvidia to data center operators is attached to OpenAI. If the discrepancy between one dollar of revenue and three dollars of cost doesn't fall to a sustainable ratio by 2027, the postponed IPO will become the second hallucination on the same balance sheet.

Meta wants to rent compute power to Anthropic for around $10 billion

According to the New York Times, Meta is in negotiations to rent out computing capacity from its own data centers to Anthropic. The deal could be worth around $10 billion over two years. Observers like Mike Isaac interpret the move as Meta currently overbuilding its data centers for future demand that doesn't yet exist internally, while Anthropic already has that demand. Fittingly, Dave Brown, previously SVP for Compute and AI at AWS, will be moving to Meta in the coming weeks to work on the expansion of the data centers (Wall Street Journal). Gary Marcus interprets the move as a shift from frontier model development to cloud provisioning, similar to xAI. Patrick Moorhead points to Meta's long list of failed B2B attempts and the resentment of many companies after the Llama course change. → us.list-manage.com

Synthszr Take: Meta is building capacity for a demand that the company itself is not yet producing, and is turning this gap into a rental business. That's the real trick in this news: overcapacity becomes a product as long as its own utilization is lacking. Zuckerberg has sunk billions into compute without having a frontier model to fill these data centers, so he's turning the empty racks into an anchor tenant named Anthropic. It's still risky, because whoever takes in $10 billion over two years is marrying their infrastructure bet to the growth of a direct AI competitor, and the capacity that's rented out today will be precisely what's missing when Meta's own demand finally picks up. SpaceX is currently pulling the same pattern with the DOD, and xAI is anyway. The interesting metric will only come in 2027: whether Meta can reclaim the rented clusters or if the landlord ends up becoming a prisoner of its best customers.

EU forces Google to open Android and Search to ChatGPT, Claude, and Perplexity

On Thursday, the European Commission, via a DMA decision, mandated Google to grant competing AI assistants and search engines significantly more access to Android and Google Search. Specifically, Google must give third-party assistants the same system functions and data access as its own Gemini: users will be able to decide whether ChatGPT, Claude, or Perplexity can access app interactions, voice commands like 'Hey Google,' and device hardware. A second proceeding forces Google to share search data with competitors, with the Commission explicitly including AI chatbots that function as de facto search engines. Google has until January 2027 for the data sharing and until July 2027 for the Android changes; non-compliance could result in fines of up to 10 percent of its global annual revenue. Google executive Kent Walker accused Brussels of undermining data protection and security guardrails for millions of Europeans with these decisions. The rulings are also seen as a signal for similar issues with Apple, which is holding back Siri AI in Europe, citing the DMA. → AI Secret

Synthszr Take: Brussels says 'competition and diversity' but in practice mainly means a company that already has 800 million weekly users. Through the opened gate march ChatGPT, Claude, and Perplexity, all from Silicon Valley. There is hardly a European search startup that could benefit from this. The Commission mandates openness, but the only player with the brand, reach, and capital to embed itself as the system assistant on every Android device tomorrow is located in San Francisco. This is industrial policy for OpenAI, packaged as antitrust against Google. And the old contradiction remains unresolved: The same regulator that tightens data protection to the maximum simultaneously demands that third-party assistants get deep access to hardware and user data. Google is bound by security audits and bears the liability; the access is provided by competitors. By 2027, this will primarily decide which US assistant gets the next default slot on the home screen. The power struggle between Europe and GAFA fades into the background.

Philipp Pausder pitches his startup Arken Labs as the 'German Palantir'

Former Thermondo founder Philipp Pausder is back in startup mode and working on Arken Labs, a software for analyzing infrastructure data. According to joint research by Sarah Heuberger and Caspar Schlenk in manager magazin's Tech Update, the application is intended to make impending damage to bridges, roads, and energy grids visible earlier, a kind of early warning system for authorities and operators. According to the report, Pausder is peddling the pitch 'German Palantir' to investors. Pausder is very well-connected in the digital scene and politics; his wife, Verena Pausder, leads the German Startup Association and is repeatedly linked with political ambitions. Details about the product idea, team, and financing are behind manager magazin's paywall. The report leaves open how the targeted investors are reacting to the comparison. → Tech Update – manager magazin

Synthszr Take: 'German Palantir' is a clever labeling trick because it flashes a billion-dollar reference before the first line of code is written. The problem: A name is a promise, not a value proposition. Palantir's real moat lies not in its data model, but in twenty years of embedded government and security knowledge that no investor can buy overnight. But therein also lies the real opportunity for Arken Labs: bridges, roads, and energy grids are a deeply regulated domain market that is too small and too German for Silicon Valley to penetrate. If Pausder truly understands the 'why' of German infrastructure—the inspection cycles, the liability logic, the regulatory rules—he won't need the Palantir comparison in two years. If, on the other hand, he still needs it to explain what he's building, it was more of a press photo than a product from the start. The market will decide this based on the first three government contracts, not the pitch deck.

xAI launches Grok Build, a terminal coding agent in Rust

xAI has released Grok Build, a terminal-based coding agent. The tool runs as a full-screen TUI, understands the codebase, edits files, executes shell commands, searches the web, and manages long-running tasks. It works interactively, headless for scripting and CI, or embedded in editors via the Agent Client Protocol (ACP). It is written in Rust; the source code is periodically synchronized from the SpaceXAI monorepo. Pre-built binaries are available for macOS, Linux, and Windows. Installation is done via a single curl command from x.ai/cli. Simon Willison has documented the release in his newsletter. → Simon Willison

Synthszr Take: Another agent harness, and from a developer's perspective, Grok Build looks like all the others: terminal TUI, file edits, shell access, ACP connection to the editor. Claude Code, Codex, and the Gemini CLI have had this form for a long time; the feature list is barely distinguishable between providers anymore. The real competition is decided elsewhere: by Grok's latency, token price, and the question of whether the model can last for hours in an agent loop without the bill exploding. The switching costs are practically zero—one curl command, and the fourth agent is running in parallel in the terminal. What xAI is missing here is a reason to stay: The Rust performance is nice, but it's not a moat, and the X web search will only become a unique advantage if the response quality per token beats the competition. The more interesting bet is on the ACP standard, because whoever hooks into editor integrations early on sits at the contact point where developers call their agents daily. Until then, Grok Build remains just another terminal icon in a field that is filling up by the week.

Anthropic gives Claude a web-fetch tool, but only for pre-known URLs

Anthropic has given the Claude API a new tool: With the beta header web-fetch-2025-09-10 a web_fetch-tool can be added to the tool list, and Claude can then fetch content from URLs itself during its response. According to Anthropic's documentation, the tool extracts the full text content of a page and also reads text from PDFs, limited to five fetches per request by default. The security logic is deliberately tight: Claude is not allowed to construct URLs dynamically, but can only fetch those that have already appeared in the conversation context, i.e., in user messages, client-side tool results, or from previous web-search and web-fetch results. URLs that Claude generates itself, or those from container-based tools like Code Execution and Bash, are excluded. Simon Willison classifies this as protection against prompt injection exfiltration but points out that URLs in user messages are still followed, which in many applications is where untrusted content lands. He mentions the parameter allowed_domains, which can be used to restrict access to an allow-list of fixed domains. → Simon Willison from Simon Willison's Newsletter

Synthszr Take: Three lines of JSON in the tool list, and the chat model becomes an agent that closes the loop between searching, reading, and responding on its own. This is what leaps in agentic capability look like: not a new model generation, but a small API extension that fundamentally changes behavior. What's interesting for developer practice is that Anthropic doesn't put the guardrail in the model, but in the harness: Claude can't invent URLs, which eliminates the classic exfiltration vector where you trick the model into sending private data encoded to evil.com/log?.... That's clean engineering. But the gap remains that user messages and tool results are considered trustworthy, even though that's exactly where external text comes in in real-world applications. The real leverage is in allowed_domains: only those who strictly restrict the allowed domains to targets from which nothing can demonstrably be exfiltrated can build production systems on this. This domain selection is trickier than it sounds, because even harmless documentation servers sometimes pass through query parameters. Build the allow-list tightly now and test it against real injection payloads before you let the tool access sensitive contexts.

In biology, computing power beats expert knowledge in drug discovery

The Semafor author describes, after a week with biotech thinkers in Boston, how the scientific method itself is shifting. The 'Bitter Lesson,' which AI research has learned since Richard Sutton's 2019 text—that scaling compute beats human expert knowledge—is now reaching biology. According to the report, a new generation of drugs is emerging from the brute-force analysis of huge datasets. Nanotechnology and AI can now measure thousands of proteins in human blood, and better computational methods are finding patterns in them without any human understanding the mechanism of action. Humanoid robots are expected to automate lab work like handling mice or cutting tissue in the future, allowing labs to run 24/7. In so-called cloud labs, an AI chatbot could design a study and have it executed for real at the push of a button, in agentic loops that conduct experiments, evaluate them, and derive new ones. → Semafor Technology

Synthszr Take: Sutton was right in 2019, and for language, it was a blessing. Scaling beats expert knowledge because a wrongly guessed token costs nothing but a correction in the next pass. In biology, the calculation is different. When machines screen thousands of blood proteins and derive active substances from them without anyone understanding why a pattern works, the end result is an intervention in a living body. Here, a wrongly guessed pattern becomes a side effect in a patient. The pace is tempting: cloud labs running 24/7, robotic hands on tissue, experiments in agentic loops, but the feedback loop runs through biology instead of compute time. The scarce resource remains the same as in any domain: the clarity of the question you ask the machine, and the discipline to validate its hits before they reach the patient. Brute force finds correlations in time-lapse; whether that becomes causality is still decided by biology.