GOVT STAKE IN OAI | ANTHROPIC X SAMSUNG | NVDA REVENUE SHARING x.com/i/broadcasts/1dJrPPvMe…

SITUATION REMEMBERED: 250 years ago today, the USA became an independent country.

people online love to talk about whether dario amodei or sam altman holds the mandate of heaven. but over the past few months it’s felt like several disparate threads of dissatisfaction with the frontier labs have started to come together. their tokens are expensive, they might drop your app as a feature, and they might have rather strong opinions about what the DoW should use their models for. they might retain your data (for safety purposes), they might bump you down a model (for safety purposes) and they might give your competitor access to frontier models before you (also for safety purposes). at the same time, there are now more alternatives than there have ever been. open source models, many from the fast-growing chinese AI ecosystem, are still several months behind the frontier. but that now covers many capabilities that are sufficient for many different applications, from consumer apps to internal enterprise tools. tools for comparing or fine-tuning and post training open weight models have also matured—i spoke to @samhogan last week, and they have a very neat feature that lets you compare a live service with an open source model. thinking machines announced a paper with bridgewater that saw them fine tune a qwen model to better-than-frontier performance at 10x cheaper price on a specific task set. but, until last week, what the open alternatives lacked was a figurehead to rival dario and sam. alex karp, ceo of palantir and absolute live wire, made a bid for that role live on CNBC. i wrote about it for the MTS newsletter.

DAILY SITUATION RECAP (via @theojaffee): Sam Altman explains his vision for AI governance. He proposes giving the US government a 5% stake in OpenAI (and presumably other AI labs), and creating a US-led international forum to establish standards, assess capabilities and risks, and “[make] the technology available to nations and companies that participate and follow the rules”: essentially a global governance regime similar to the IAEA. Anthropic is in talks to make an AI chip with Samsung. Much like OpenAI and Broadcom’s Jalapeño or Google’s TPUs, it will likely be a chip designed for and around Anthropic’s models, and will likely focus on inference rather than training. Unlike most other AI labs and hyperscalers, Anthropic has not yet made their own chips directly. Nvidia is launching the AI Compute Partnership. Neoclouds like Coreweave and Lambda buy Nvidia GPUs in bulk to rent out to AI labs, but they often have weak credit ratings, which makes acquiring GPUs expensive. Nvidia solves for this by guaranteeing to rent back any unused GPU capacity, and taking a share of the neocloud’s revenue in exchange. This helps labs finance their GPUs while turning Nvidia into a sort of central bank for the cloud industry. Crusoe is raising $3 billion at $30 billion, tripling their valuation in nine months. The company builds data centers, including OpenAI’s Stargate Abilene, and has nearly 5 gigawatts (GW) under contract and over 40 GW in the pipeline. (For reference, the largest data centers today use about 1 GW of power). Kling AI raises $3 billion at an $18 billion valuation led by Chinese internet giant Tencent. Kling is a Chinese video gen company competing with ByteDance’s Seedance and (before its death) OpenAI’s Sora. It recently hit an ARR of $500 million, and expects to go public in around a year. Zuckerberg says AI agent progress hasn’t “accelerated in the way we expected”. Meta has had a difficult run in AI in the last couple years, spending tens of billions on hyperscale data centers and talent, undergoing a major reorg, and now having to rent out excess compute as its models remain behind schedule and below the frontier. Incoming UK Prime Minister Andy Burnham is less pro-AI. Burnham will pivot away from deals with US-based AI labs (essentially, all of them), pursuit of self-driving cars in London, AI Growth Zones, and “unfettered tech boosterism” in favor of British AI sovereignty and greater review from local governments. Google loses its fight over a $4.7 billion EU antitrust fine for requiring developers to pre-install Google Search, Chrome, and the Play Store on their Android devices. The fine was levied in 2018 and upheld in the EU’s highest court today, signaling the EU’s gradual legal escalation against American tech. Microsoft is overhauling Copilot and cutting unused features in response to feature bloat and low usage compared to competitors like ChatGPT and Claude. SoftBank is launching a neocloud, SB Neo, to rent AI compute to US companies. The company also plans to build data centers in Japan as soon as possible. Tesla is capping employee AI spend at $200 a week. That’s 8M tokens of Opus or 4M tokens of Fable. The tokenmaxxing era is truly over.

SITUATION UPDATE: Alibaba has banned employees from using Anthropic’s Claude Code at work after the tool drew scrutiny for features that can identify China-linked users, per Reuters. The ban follows Anthropic’s accusation that Alibaba was distilling Claude.

SITUATION EXPLAINED: Why is China building a sovereign AI stack to compete with Nvidia? We asked @StevenGlinert and @MitchsBrew, co-founders of Sphere Semi. "What I see as the only real viable crack in Nvidia's armor is what's coming out of China, which is ironic given Jensen's famously kind of weird stance on China." "If you think about the things that Nvidia can do, the way that neural nets are trained into CUDA, the way they've built the data center rack into Nvidia, the way they've built memory infrastructure into Nvidia. The only thing that is incredible about China is they've built this sovereign stack for all of these things." "They are now in many ways seemingly able to actually have independent training capabilities for their own models that are entirely outside of Nvidia." "There is no threat to him but Xi Jinping. That he should take serious."

SITUATION EXPLAINED: Why did token maxing happen and why is it ending? @kmad, AI Engineer and Developer: "It's funny how quickly we turned... it was like three months or something of token maxing. Had all these leaderboards and everything." "A lot of folks have kind of just defaulted to that without the governance, without thinking about the broader implications of using Opus for everything, for drafting an email." "We're rapidly approaching a phase now where there will be some cost controls in place, a little bit more intentionality for matching the tasks with the appropriate sized model or cost of model. That's where the FDEs and the consultancies of the world are starting to step in." "There'll be a bifurcation of undefined knowledge work and then these well-defined tasks. The well-defined tasks, I think we'll see a trend of using cheaper open source models. It's almost like a portfolio of different AI services or AI APIs within the company that are company specific and not generic."

SITUATION EXPLAINED: What would it take to make Americans feel like they have skin in the game for the future of AI? We asked @stephenkentx, media director at @ConsumerChoiceC "I am not offended by the idea of a sovereign wealth fund. I think that we are in a populist moment. It's only going to get more populist, and we have to do something substantive that makes people feel like they have some skin in the game for the future of the country." "Just a couple generations ago, what you're talking about is every person who has a house has a stake in the future of the country. Every person who has some money in the stock market has a stake in the future of the country." "Are we talking about a direct stipend that goes down to every American household to spend as they see fit? Which I think might be what I favor... because you don't know what one household values over another." "I do think of data as a natural resource that should be making people rich."

SITUATION EXPLAINED: Why doesn't the Alaska sovereign wealth fund analogy actually work for AI? We asked @stephenkentx, media director at @ConsumerChoiceC "Bernie Sanders has made his first opening offer in the American AI Sovereign Wealth Fund Act to have the US government acquire 50% of these companies, basically have the United States government with seats on the board of every one of the major AI firms. In my view that would be a seizure of the equity of these companies." "Sam Altman is offering 5% to the US government. It's a pretty bold strategy. I almost admire it." "They're modeling this off the Alaskan dividend fund. But you're basically talking about an asset that belongs to the people of Alaska, the land, and the natural resources beneath the land, which in theory is a limited resource." "The analogy just doesn't hold up super well. I'm not sure how you're going to make the case that these AI firms just sort of belong to the people, unless you're talking about the intellectual property that they studied in order to train these models."

SITUATION EXPLAINED: Why does the KIDS Act pretend to give parents power they already have? @stephenkentx, media director at @ConsumerChoiceC "I still happen to believe that parents have full control of their households. The problem is that they just don't like to say no. The problem is parents who don't wanna say no." "I've never met a teenager who pays their phone bill. I've never met a teenager who's helping to pay for the Wi-Fi. They are being given these devices and paid for every month by their parents, and the parents are never going, 'Eh, you're cut off. You're not doing this anymore.'" "But I think that worm is turning. I think young people right now and a generation away from now are going to embrace a more offline lifestyle." "It will be elder millennials like my generation who forced younger people onto social media before they were ready who will be held in contempt for that."

SITUATION EXPLAINED: Why does duckweed produce 10 to 15 times more ethanol per acre per year than corn? We asked @jared_western, founder and CEO of Western Chemicals. "The great thing about duckweed, and kind of the superpower of duckweed beyond the wastewater cleaning, is just the speed at which it reproduces." "It doubles its mass once every two to three days. Literally within two to three days you have twice the mass that you started with." "In terms of ethanol conversion, it's probably on the order of 10 to 15 times more throughput per year comparative to corn." "So if we're looking at gallons of fuel that we can get out per acre, per year, kind of in our mid-range estimates, we're looking at like 5,000 gallons per acre, per year. Which is quite good." "Comparative to corn, you typically get like 700 or 800 gallons per acre, per year."

SITUATION EXPLAINED: How do you turn wastewater into fuel? We asked @jared_western, founder and CEO of Western Chemicals. "The good news is biology basically does all the work. We just need to create the ideal conditions for the biology to thrive and do its thing." "The duckweed, as it grows on the wastewater, is gonna pull nutrients out of the wastewater to clean it. Primarily nitrogen and phosphorus. It pulls that out of the water, fixes those into proteins, and uses that as fertilizer to fuel the growth process." "Those carbohydrates are exactly what we need to do fuel production processes. It's got our basic building blocks we need to make hydrocarbon fuels." "We harvest the duckweed, do some mechanical processing on it, and then we use a fermentation process to turn those starches first into sugars, and then we put in yeast and basically brew beer and turn the sugar into alcohol." "You do distillation and dehydration to turn it into basically anhydrous ethanol, pure concentration 99.5%. And then it's fuel grade. It's ready to blend in with gasoline."

SITUATION DETECTED: Microsoft has launched Frontier Company, a new $2.5B operating business embedding 6,000 engineers into customer organizations to deploy AI systems.

SITUATION EXPLAINED: Why was legal one of the most surprising early AI adopters? We asked @kmad: "I've worked with a bunch of attorneys, I've testified in court, so I can appreciate the level of nuance and specificity that you need. When you get hammered on a cross, you need to know the numbers or you need to know where your information came from." "But we've seen multiple examples of attorneys or expert witnesses using AI to write their reports or write their briefs, and they get nailed for it because there's hallucinations and whatever else." "I would've expected them to be much more conservative. But legal has been one of the front runners in this space." "If you're training on a bunch of contracts, a lot of that's going to be in distribution. This is an NDA format. This is an MSA format. It's very easy for the model to put that together."

SITUATION EXPLAINED: What is token capital efficiency? We asked @kmad, AI Engineer and Developer "Instead of token maxing and using Opus and GPT 5.5 for everything, can you have a strategy as a business to measure the effectiveness of your spend on tokens? I call this token capital efficiency." "The way I would argue you do that is you inventory your tasks, you prove out what works, what is automatable, what you want to automate. You define your evals for measuring how successful that will be." "Once you have that kind of flywheel, then you can walk down that cost curve. 'Here's my acceptable level of performance. I want 97% accuracy or better on whatever this task is.' Once you can measure that, then you can try a bunch of different models under the hood." "What is that optimal point between retaining your level of performance and the lowest cost possible?"

LIVE NOW W/ @brexton talking compute

SITUATION EXPLAINED: Why is consumer robotics really about giving you your time back? @evan_wineland and @kaandogrusoz, co-founders of @weaverobotics "Consumer robotics are ultimately about giving you optionality in the limit. You always have the option to do any of these things, but you'll then have choice about whether to offload that." "Consumer robotics done right, if they're accepted by people, they actually drive value in the home. They give you lots of discretionary time back that you can use for literally whatever you deem worth your time." "That could be spending time with your loved ones or on your craft, or even just sleeping or relaxing." "It's going to be really nice to have a helper in your home that you can offload these things to when you want and only when you want."

SITUATION EXPLAINED: Why is cooking the ultimate humanoid AGI test path? We asked @evan_wineland and @kaandogrusoz, co-founders of @weaverobotics "Cooking is, I think, one of the hardest tasks you can build up to in home robotics. It's extremely multimodal. If you think about how we cook, we rely on scent, we rely on touch, we rely on timing, we rely on recipes, we rely on very fine-grained details and visuals." "It's a task where timing really matters. Do this thing for 30 seconds only. You have no... it's like, I failed at flipping the steak, let me try again for another minute. Everything has to be in sync. That's a really good humanoid AGI test path." "The reason we went for laundry folding first is it's something people do a lot and generally do not want to do. There are very few people who would say, 'I view it as my art.' Cooking is a form of art for some people. Laundry folding is way more unobjectionable."

SITUATION EXPLAINED: How do we build democratic institutions that stay robust in a post AGI society? @jasonhausenloy, AI policy researcher at @CAIS: "What is the foundation of democracy? Why is it that we have the democratic institutions that we do? It is ultimately because humans provide value and democracy is a way to organize those humans so they continue to provide value." "Humans can opt out of society and that would be terrible for the governments that manage them. They provide their labor to the economy and their physical strength towards the military." "If you're able to have these things split apart in pretty rapid succession, then what you're hoping on is the reliability of these institutions that have lasted 250 years in the US to continue. And I actually don't think that they are so robust." "Whether that be in five years or twenty years, both of which seem totally reasonable, and in the grand scheme of things a very, very short amount of time, we will reach an end state where this technology does allow for the split of where humans' value comes from and how they can contribute in democratic systems."

SITUATION EXPLAINED: The automation discourse assumes automating people end to end is how AI must be developed. @jasonhausenloy, AI policy researcher at @CAIS: "The core problem with a lot of the discourse around automation is that it assumes this path that we've taken along the tech tree to automate people kind of end to end with long-horizon tasks is the way that AI must be developed." "It is simply the first way, the easiest way, that we've been able to turn compute into human lives getting better." "If we could pour all of our compute into training larger AlphaFold-type systems, go and kill cancer, go and do the things that really treat AI as an extension of what humans want to do. Like tools." "That to me is a much more compelling vision than the ones that have been laid out otherwise."