Subscribe
Share
Share
Embed
Anthropic's revenue outruns its trust problem
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Tuesday, August eighteenth.
In today's briefing we see Nvidia backing OpenAI's next data center campus with up to a hundred and five billion dollars in financing, SpaceX's Cursor launching its own proprietary alternative to GitHub for coding agents, and Anthropic's revenue run rate crossing sixty-five billion dollars.
First up, today in the big model news;
Anthropic
Anthropic's annualized revenue hit sixty-five billion dollars by the end of July, seven times where it stood a year earlier and now ahead of OpenAI's roughly forty billion. Investors are pricing next year's revenue in the range of one hundred to one hundred twenty billion dollars, the level this month's chatter around a two trillion dollar valuation would need to clear. The jump lands a day after Dario Amodei called the industry's backlash a crisis of trust rather than a messaging problem. Revenue and trust are moving on separate tracks: nothing in the numbers shows the second one shaping a product or policy decision yet.
Separately, LMArena's crowd preference leaderboard currently ranks Claude Opus 5 below several Opus four point six and four point seven variants, fueling a narrative that Opus 5 feels worse to work with. But Arena Elo measures chat vibe, not capability, and Opus 5 still tops Artificial Analysis's Intelligence Index. Cohere Labs' Leaderboard Illusion research gives a standing reason to discount a gap this size, since vendors privately test variants and disclose only the winners. Zvi Mowshowitz's review of the system card found no regression at all, and pinned the complaint on overdramatic phrasing rather than a real performance drop. The gap stays unverified until someone runs a compute matched eval on a fixed task set.
In the harness, tools and orchestration world;
SpaceX's Cursor shipped Origin, a proprietary rival to GitHub built specifically for AI coding agents, days after the sixty billion dollar acquisition officially closed. Cursor now sits inside SpaceX's balance sheet instead of renting margin on top of GitHub's repository layer, and Origin removes the dependency that arrangement implied: agent workflows need commit and merge throughput that human oriented hosting was never built for. Launching a proprietary platform on the same day as Composer 3, a one and a half trillion parameter model trained on SpaceX's own Colossus cluster, makes the intent explicit. GitHub now has to choose between rebuilding for agent speed or ceding the highest intensity workloads to a vertically integrated rival, and every other agent tool maker still dependent on GitHub faces the same tradeoff against a competitor with reason to out-build the infrastructure underneath it.
A quieter pattern showed up alongside it. Hermes Desktop, Bot Mode, and Codex's own orchestration layer all shipped more mature tools for coordinating multiple agents with persistent context, while Vanta and LangChain separately tightened sandboxed, permissioned execution for enterprise agent deployment. None of these vendors were reacting to each other, which is the stronger signal: demand for coordinating agents keeps surfacing independently across the stack rather than staying one lab's bet.
Separately, GitHub Copilot Autofix introduced a script injection flaw into a Snowflake public repository, swapping a sanitized input pattern for direct string expansion inside a shell command. Wiz's autonomous Red Agent pentesting tool found and exploited that flaw two months later, using nothing more than a crafted GitHub issue title to steal a token with read access to Snowflake's internal Jira. It's the same pattern this month's other crawler spoofing and sandbox stories have shown, a trusted abstraction becoming the cheapest exploit path, except here one AI wrote the hole and a different AI found it with no human in either loop. Autofix class patches on security sensitive code now need the same scrutiny a human contributor's diff gets by default.
In AI Infra.
There's a pattern running through this month's compute deals worth naming before the lead item: Anthropic and Google renting capacity from SpaceX and xAI, capital chasing operators with no delivery history, a distressed hedge fund doubling down on an unproven chipmaker. Nvidia's move on OpenAI's new Ohio campus is the sharpest version yet. Nvidia is putting up to a hundred and five billion dollars in financing behind a twenty year OpenAI lease at a site built and operated by SB Energy, scaling from four and a quarter gigawatts to eight gigawatts on former Department of Energy uranium enrichment land. Separately, and this is the part that changes the shape of the deal, Nvidia is also taking a direct one and a half billion dollar equity stake in SB Energy itself and locking in exclusivity as the site's sole compute supplier. Nvidia sells chips, so financing the buyer converts a one time GPU sale into a twenty year exclusive claim on OpenAI's buildout, a customer lock no rival chipmaker can match without taking on the same balance sheet risk. AMD, Google's own TPUs, and the hyperscalers' custom silicon are absent from the deal, and their case for weaning any buyer off Nvidia gets weaker every time the largest AI purchaser signs the opposite bet.
The same compute landlord dynamic showed up again in Groq's own numbers. Groq raised three hundred fifty million dollars at a three and a half billion dollar valuation, nearly half what it was worth a year earlier, completing its pivot from AI chipmaker to neocloud GPU operator, with Nvidia itself joining the round. Nvidia had already hired away Groq's founder and top talent in a twenty billion dollar licensing deal months earlier, and this financing folds what remained of Groq into Nvidia's ecosystem as a downstream tenant rather than a rival architecture. It's not a lab renting from an operator this time, it's a would-be Nvidia challenger ending up as one of its own customers.
In other news…
The Cherokee Nation has banned hyperscale data center development on its tribally owned and trust land, after a task force found sixty-four percent of surveyed citizens opposed such projects on energy, water, air quality, and cultural resource grounds. It adds a new kind of block to a siting backlash that so far has been state regulatory or climate specific: tribal sovereignty isn't something a developer can negotiate around the way it might lobby a state agency. Consultation near reservation land is no longer a courtesy step.
That's the briefing. Have a great day, and don't forget to subscribe.