The Brief: The Last Company Standing?
Google and Anthropic watermark, Nvidia mobilizes $500B for even more compute, SpaceXAI ships Grokbot
FIELD NOTES
Should the CEO of Anthropic be more optimistic about AI?
On Saturday, an exchange between investor Gavin Baker and Dario Amodei went into why the public has such a negative view of AI. Dario called it “fundamentally a crisis of trust.”
His answer, roughly, was that you don’t fix that with better messaging. You fix it by delivering something undeniable: like, cure cancer, or solve major diseases.
I understand the instinct. But the problem with waiting is that the public is forming its opinion of the technology now. The costs are already hitting hard: electricity demands, data centers, job anxiety, deepfakes.
AI has a strange problem: people are using more of it and liking it less.
Anthropic and OpenAI are generating tens of billions of dollars in annualized revenue. AI adoption inside businesses continues to climb. But public sentiment is moving in the opposite direction.
A June Pew survey found that 40% of Americans expect AI to have a negative effect on society over the next 20 years. Just 16% expect a positive one.
And increasingly, that skepticism has consequences. Only 14% of Americans say they would support a data center in their community. Seventy-five data center projects worth $130 billion faced opposition in the first quarter of this year alone.
But there was another part of the exchange that made me wonder if that’s even the right question.
The reason the whole exchange started was that Gavin alluded that Dario privately believes Anthropic will eventually be “the only company left in the world.”
Maybe that’s hyperbole. But it resonated with something I’ve struggled to reconcile in Dario’s public position on AI.
AI could be extremely dangerous, and the race to build it is itself dangerous. But Anthropic needs to stay in the race.
Concentrations of AI power are dangerous. But Anthropic will become one of the most powerful companies in the world.
Certain capabilities may be too dangerous to release broadly (ie Mythos). But Anthropic has to decide when and to whom they should be released.
There are reasonable arguments for every one of these positions individually. Together, though, they create a difficult tension.
If you believe AI is both extraordinarily powerful and potentially dangerous, but also believe someone will build it regardless, staying at the frontier can feel like the only reasonable thing to do. Better to have a responsible, safety-conscious company helping shape what happens than to leave the AGI race entirely.
The problem is that every frontier lab can make some version of that argument.
If Dario really does believe Anthropic could eventually be the only company left, it raises a harder question: how does any CEO separate their judgment about what is good for the technology from their conviction that their own company is the one best to build it?
Maybe that’s part of the crisis of trust too.
The people building AI are being asked to make humanity-scale consequential decisions about how fast to move, what to release and which risks are acceptable, while simultaneously competing to win the market those decisions will create.
So perhaps my question is whether any frontier AI company can simultaneously race to win and credibly argue that the competition itself is the problem.
-Tara
THE DOWNLOAD
SpaceXAI Ships an Always-On Agent, Then Closes the $60 Billion Cursor Deal Three Days Later
SpaceXAI unveiled Grok Bot on Aug 11, a persistent agent product where each agent gets its own cloud computer and signs into applications with the user’s credentials, including applications with no API, runs unsupervised, and coordinates with other agents in group chats. Grok 4.6 followed on Aug 12, same 1.5T base as 4.5 but tuned for long-running agents, shipping day one into Cursor, OpenRouter, Vercel and Cloudflare. On Aug 14, SpaceX closed its $60 billion all-stock acquisition of Cursor, issuing 389,289,254 Class A shares. Cursor had crossed $4 billion in annualized revenue in June and acquired Firetiger on Aug 13, the day before its own close.
Why it matters: Grok Bot is also the first credential-operating agent from a major lab, and that architecture is a different risk surface than an API-calling agent: it logs into things as you, in apps that never agreed to be automated. The model itself is incremental, tying GPT-5.6 Sol at 61 on the Artificial Analysis index rather than leading. Musk separately said future Grok versions will train on “the sum total of all SpaceX information,” including employee work product, with no stated opt-out.
Is the US-China Model Gap Just Mostly Release Timing?
Z.ai shipped GLM-5.3 on Aug 14, built by scaling post-training on the GLM-5.2 base, and held the weights back two weeks for safety testing. Writing in Interconnects, Nathan Lambert argues the reason Chinese models keep arriving at the frontier is the release calendar rather than distillation. American labs finish a model and spend months testing internally; Chinese labs ship in days and spend those months improving the released version. “It is very, very likely that OpenAI and Anthropic have far better internal models than Z.ai.” He also credits narrower scope, since GLM-5.3 has no visual capability at all, plus compute efficiency, Tsinghua’s talent pipeline, and reports that American data companies are selling reinforcement-learning environments to Chinese labs.
Why it matters: Developers build on what they can get, and the self-improvement loop runs on data from developer usage, so shipping first has its benefits. It also points the policy argument at the wrong variable. Senator Banks asked the White House this week to fund American open models, but a subsidy does not shorten a safety review. Lambert: “I’m confused why the labs in the U.S. haven’t patched this behavior faster; instead they’re running to the government asking for policy help.”
Google Makes Its Visible Watermark Optional; Anthropic Starts Watermarking Text
Google made the visible watermark optional on AI-generated images, video and music in Gemini Apps and Flow, while keeping invisible SynthID and C2PA metadata mandatory and undisableable. VP Josh Woodward: “We’re striking a balance here between creative control and safety.” The toggle is unavailable on work and school accounts, and gated to AI Ultra in India, South Korea and Vietnam. Three days earlier Anthropic said it will watermark text from every model released after Aug 2, using a version of DeepMind’s SynthID-Text approach across the API, apps, Claude Code and Cowork.
Why it matters: The deeper problem is that strong text watermarking may not be achievable at all. Zhang, Edelman, Barak and colleagues proved at ICML 2024 that a random-walk attack strips watermarks without knowing the key or even which scheme is deployed, and a simple paraphrasing pass defeats detectors generally. Anthropic is not really claiming otherwise.
Nvidia Signs Six MOUs to Mobilize $500 Billion
Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital. The release states these are memorandums of understanding and that the partnerships remain subject to execution of the final agreements. The widely reported 25% guarantee appears nowhere in it. It came from Jensen Huang on X the following day, describing support for up to 25% of an opportunity, assessed project by project, limited and residual-value based.
Why it matters: Nvidia is now guaranteeing the future value of GPUs that Nvidia itself will make obsolete. This sounds, at first, like another version of the circular financing I wrote about last September, when Nvidia was investing in companies that were turning around and buying its chips. It’s actually something different. It is promising that if the chips ever have to be sold off, they will still be worth something. That promise makes the GPUs better collateral: lenders can lend more against them, at better rates, to buyers who otherwise couldn’t get those terms.
The FCC Drafts an Import Ban on Chinese Optical Transceivers While Indium Phosphide Runs Short
The administration and the FCC are drafting a rule to add China-made optical transceivers to the Secure Networks Act covered list, targeting new models, with officials aiming to enforce by end of 2026. China holds about 56% of global optical module capacity, and Chinese firms hold seven of the top ten slots, with Innolight and Eoptolink above 60% of the 800G-plus market. Lumentum’s Michael Hurlston says the indium phosphide shortage will be “worse than memory”, with the company expanding supply 50% year over year and still shipping more than 30% below demand.
Why it matters: China controls roughly 70% of global indium supply, so restricting Chinese transceivers tightens the input to the Western capacity meant to replace them. Hurlston, on serving the demand: “Between the two of us, I don’t think we can service the demand that Nvidia and others are now putting on us.”
A Nature Study Gives an Agent Six Days and $3,000 to Replicate Two Papers; the Authors Score It 2/6 and 1/6
Nature reported on Aug 13 on “shadow evaluations” run by Sayash Kapoor at Princeton with the UK AI Security Institute. The goal for the agentic system was to develop concept from two computer-science papers. The original authors scored the outputs only 2 out of 6 and 1 out of 6, reject and strong reject. The preprint documents five recurring failure modes: premature commitment to a first hypothesis, poor backtracking with ambitious goals abandoned inside ten hours, no sense of what clears a publication bar, poor budget awareness, and instruction drift. One run spent only $1,130 of its $3,000. Kapoor: “I don’t think full automation of open-ended research is on the horizon right now.”
Why it matters: What the agent did well is the part worth reading twice. It ran hundreds of experiments without human help, debugged without falling into error loops, caught its own false claims, did not reward hack, and produced complete LaTeX papers. The engineering of research is solved. What failed was judgment: it could not tell a trivial finding from a real one, and its self-reviews were uncritical.
JOIN US AT HOT CHIPS
Join us for cold beer and Hot Chips on October 25! We’re hosting a happy hour post- Hot Chips in Palo Alto.
