Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and today we've got a menu heavy on money, machines, and a phone that costs more than a used car but can't book a flight to save its life. Welcome to Barely Possible. Buckle up.
Let me start with the story that made Wall Street flinch this week, because it's the one that actually matters for anybody trying to build something.
Chinese company Moonshot AI dropped a new version of its Kimi model, and the reaction was immediate. According to the report, the announcement coincided with a speech from Chinese president Xi Jinping at the World AI Conference in Shanghai, and it seems to have spooked Wall Street. The Nasdaq dropped about 1% on Friday as investors sold off chip stocks like Nvidia. Now, Moonshot's own framing is honest, which I appreciate. They said Kimi K3 "still trails the most powerful proprietary models," but the new open source model "demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models." Independent analyses from Arena.ai and Vals AI backed that up, saying Kimi is competitive with flagship frontier models.
Now, if you've been listening to this show, you know we've been circling the China open-weights theme for a couple weeks. So I'm not going to give you the same lecture again. I want to zoom into the actual argument that broke out, because that's where the real signal is.
Here's the tell. The reactions in the piece are all over the map, and the people reacting are the ones who actually set policy and set narratives. David Sacks — the Trump administration's former AI czar, now co-chair of the President's Council of Advisors on Science and Technology — used Kimi to bang the deregulation drum. He said the United States is "tying itself in knots: politicians and bureaucrats are banning new data centers, piling on state regulations, and pushing for new federal agencies to pre-approve frontier models. This is how you lose the AI race." He also took a shot at Anthropic, calling Claude a "woke lobotomized" model. That's the political frame: China's shipping, we're regulating, get out of the way.
Then there's the distillation crowd. Former Uber CEO Travis Kalanick echoed the complaint that the Chinese are "distilling off" American models — meaning training on the outputs of American AI. His line was, "If distillation isn't enforced against, then everyone should be able to distill from everyone else." And the piece drops a very dry parenthetical right after: American models have also been built on top of Chinese ones, specifically Kimi. So the distillation grievance cuts both ways, and everybody in the room knows it.
But the part I want you to sit with is what OpenAI's head of strategic futures Dean Ball said, because it's the most revealing thing in the whole story. Ball admitted Kimi is "a very good model" whose performance probably can't be "explained away by distillation or anything like that." Then he went somewhere else entirely. He said the "probable outcome of an open-weight-model-dominant world is full AI communism," where AI becomes "a public good which will ultimately be provided by the state as a kind of digital public infrastructure." He called that a "dystopian hellscape." And then — this is the important bit for you as a builder — he laid out the playbook for how you kill open Chinese models without ever banning them. His words: "You don't need to ban open source. You just need to direct every agency to issue soft law that creates FUD" — fear, uncertainty, and doubt. He literally invented a fake headline as an example: "A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models." His point was it doesn't need to be well justified. You just create enough regulatory risk that every regulated enterprise backs off.
Let me translate that from the executive suite into the kitchen table. If you're a founder building on top of a cheap, capable, open Chinese model to keep your costs down — and a lot of you are, we'll get to Databricks in a minute — the risk to your business may not come from a law. It may come from a memo. From a whisper. From a "bulletin" that makes your enterprise customers nervous enough to say, "let's not, just to be safe." That's a different kind of risk than a ban. A ban you can fight in court. A cloud of doubt you just have to live under.
Now, to be fair, not everybody thinks the sky is falling. Shakeel Hashim, editor of the AI publication Transformer, argued much of the worry is overblown — both because Kimi "likely does not have dangerous cyber capabilities," and because the Chinese government will face "extremely similar incentives" to restrict its own open models once they get more dangerous. In other words, the incentives to lock things down are universal. Nobody stays fully open forever once the capabilities get scary enough.
So where does that leave you? The capability gap between open and closed, between China and the US, keeps getting narrower. That's real. But the story this week isn't really about benchmarks. It's about the fact that the people who lost the technical argument are already reaching for the regulatory one. Watch that. If you're building a product whose cost structure depends on open weights, your job now includes reading the political weather, not just the model cards.
That thread — cost structure depending on cheap open models — is exactly the bridge into the biggest money story of the week.
Databricks. On Thursday the company announced a new round of funding that values it at $188 billion. The round was led by Coatue. And here's the funny part — Databricks didn't even disclose how much it raised, said the money isn't in its hands yet, and the round won't close until later this summer. Other outlets reported the raise at roughly $3 billion. Announcing your valuation before the check clears is unusual, but a VC told TechCrunch the deal is solid, with so many firms wanting in that the company had no reason to keep the shiny number a secret.
And look at the tempo here, because this is the part that should make your head spin. Only five months ago, in February, Databricks closed a $5 billion round at a $134 billion valuation. Five months before that, September 2025, a billion dollars at $100 billion. And roughly nine months before that, December 2024, ten billion dollars at $62 billion. They've raised so many rounds that people are running jokes about running out of letters in the alphabet. Somebody posted, "Turning on alerts for when we get a Series AA." Sixty-two billion to a hundred and eighty-eight billion in about a year and a half. That's not a fundraise, that's an elevator.
Now here's why I think this is genuinely instructive and not just a big number. Databricks wasn't born an AI lab. It was founded in 2013 in the big-data era — software for storing enormous amounts of enterprise data in the cloud and running fast analytics on it. Boring, useful, profitable. And because it was already sitting on troves of enterprise data, it was perfectly positioned when companies started wanting AI with the same security and governance they expect from grown-up enterprise software. So they rolled out product after product — a database built for AI agents called Lakebase, an AI gateway called Unity, and a "meta-harness" called Omnigent that manages multiple agents.
But here's the connective tissue to the Kimi conversation. Databricks became one of the poster children for enterprises adopting cheaper Chinese open-weight models for cost control. They're a particular champion of Z.ai's GLM 5.2 for coding. Last week CEO Ali Ghodsi shared internal benchmarking he did to manage his own AI bill for his 3,000 software engineers. They tested models on the actual tasks their programmers do. And the finding was — open models, GLM 5.2 in particular, could now handle even the highest-difficulty coding tasks at a lower total cost than proprietary models from Anthropic and OpenAI.
But the surprise wasn't the model. The surprise was the wrapper. They found that the choice of harness — the agentic coding tool like Codex or Claude Code that wraps around the model and manages its context and instructions — mattered just as much for cost. They found an open-source harness called Pi was one of the best at managing context around each prompt, which made it one of the lowest-cost choices without sacrificing quality. Their own conclusion: "The lesson here isn't that one harness is always cheaper. Instead, model choice is only one piece of the puzzle."
Now, this is a real, useful operational insight for anybody running a big engineering org — the tooling around the model can move your bill as much as the model itself. Chew on that. But I also want you to hold the two stories side by side. Databricks just got rewarded with an AI-halo valuation, in part, for being loud about using cheap Chinese open weights to control costs. And in the very same week, an OpenAI exec is publicly workshopping how to make regulated enterprises too scared to use exactly those models. Databricks is a regulated-enterprise darling. If Dean Ball's soft-law FUD strategy ever actually gets deployed, the poster child for the open-weights cost play is standing in the blast radius. I'm not predicting anything. I'm just saying — these two headlines are not living in separate universes.
And the AI-halo thing is real to the point of comedy. The piece notes that the effect is so strong these days that even sandwich shop Jersey Mike's mentioned AI 22 times in its S-1 documents. Twenty-two times. For a place that sells you a sub. If your deli is name-checking large language models, we have officially reached the "put a bird on it" phase of this cycle.
Alright, from the sublime to the ridiculous. Let's talk about a phone.
Vertu — the UK-founded luxury phone maker, the folks who sell hand-finished devices that can run tens of thousands of dollars — has a foldable called the Alphafold. It starts at $6,880. And the whole pitch, the entire reason for the price, is that it comes with an AI agent built for executives. Not benchmark scores, not cameras. An agent called Hermes Agent, built on top of the open-source Hermes project, that's supposed to analyze your files, automate tasks across apps, remember your conversations, and hand off to a human concierge when it can't cope.
TechCrunch's Jagmeet Singh actually used it for days as his primary phone and ran it head-to-head against a Samsung Galaxy Z Fold 7 running Google's Gemini. And I love this review because it's the most honest thing I've read all week about where agents actually are versus where the marketing says they are.
First, the awkward part. Under the leather, the reviewer noticed the Alphafold looked an awful lot like the $1,100 ZTE Nubia Fold — same hinge, same dimensions, same placement of speakers and mics and the fingerprint reader. Even the system information showed ZTE identifiers in the software. Vertu confirmed the device was developed through a supply-chain partnership using ZTE and Nubia's hardware platform, but said Vertu was responsible for the luxury materials, the software, quality control, and after-sales service. So you're paying about six times the price for calfskin leather, a titanium frame, packaging that opens like a jewelry case, and an AI agent. The leather and the box, fine, that's what luxury is. The AI agent is the actual bet.
So how'd the agent do? Here's where it gets useful. The reviewer ran a classic executive scenario: message a contact that I'm running twenty minutes late, navigate to the airport, flip the phone to Do Not Disturb, and remind me to call the hotel in fifteen minutes. Hermes sent the message, enabled Do Not Disturb, opened Google Maps — but didn't actually start navigation, and set the reminder for 9:08 p.m. despite the request being made at 2:32 a.m. for a reminder fifteen minutes out. So it was eager, it grabbed a bunch of tasks, and it fumbled the details.
Gemini on the Samsung took the opposite approach. It asked follow-up questions — which airport, which reminder app — and once the reviewer answered, it got the reminder right. Hermes did more of the workflow. Gemini produced the more accurate outcome. That is the entire agent debate in one anecdote. Do you want the assistant that charges ahead and does eighty percent of it wrong-ish, or the one that stops and asks and does the smaller thing correctly?
The trip-planning test was worse. Asked to book a business trip from Mumbai to Pune with a morning flight and a hotel, Hermes said no direct morning flights, offered a "Contact Butler" button to punt to the human concierge — and then created a calendar entry for the wrong dates, July 7th instead of the 18th to 19th, and left the whole thing half-finished. Gemini, when it hit the no-direct-flight wall, kept going and suggested alternatives instead of giving up.
And the one that really stuck with me: memory. Hermes analyzed a financial spreadsheet, correctly summarized the Q2 figures. Then the reviewer came back to the same conversation days later, and the agent had completely forgotten the document. It said, and I quote, "I cannot access files stored directly on your local device. Please upload or attach the Sales spreadsheet here in the chat." The thing that "remembers your conversations" didn't remember the conversation. Gemini, meanwhile, held the context and was still answering follow-up questions about that spreadsheet days later.
Here's the founder lesson buried in a luxury phone review, and it's got nothing to do with the leather. The reviewer's verdict was that Hermes is "an ambitious AI assistant rather than a finished one." That word — finished — is the whole thing. Vertu is selling a $6,880 promise of agentic autonomy, and what the buyer actually gets, today, is a work-in-progress that got server-side fixes mid-review. The reviewer even flagged that the experience buyers get in a few months may not be the one he tested. So you're paying frontier-luxury money for something that's still cooking in the kitchen.
If you're building agents, the takeaway is blunt: autonomy without reliable memory and without knowing when to ask a clarifying question isn't a premium feature. It's a liability you charge extra for. The Samsung-plus-Gemini combo, at a fraction of the price, won on the thing that actually matters — getting it right. Confidence is cheap. Correctness is the product.
Now let's shift from what the agents can't do yet to who's going to own all the money if they ever can.
There's a piece by Connie Loizos built around a conversation she had back in the spring with Neil Rimer, a co-founder of Index Ventures — one of the most successful venture firms of the last thirty years. And Rimer said something that stuck with her, and now with me. Talking about the wealth piling up around AI, he said he has "a strong sense that there will be some sort of a redistribution." And then: "It'll either be voluntary or it'll be involuntary, but it'll happen, and I hope it's voluntary." He added that tech leaders "can play a leading role in seeing that through."
Now, coming from a random person, that's a bumper sticker. Coming from a guy whose firm has raised roughly $15 billion and whose exits last year — including Figma's IPO and Google's purchase of the security firm Wiz — reportedly netted Index around $9 billion, it lands different. This is a man talking about giving back the very windfall he's collecting.
And the numbers in the piece are the part I can't shake. The Giving Pledge — the Buffett and Gates thing from 2010 to get billionaires to commit half their fortunes to charity — is fading. A hundred and thirteen families signed in the first five years, then 72, then 43, then just four in all of 2024. And it's not just billionaires. Total American charitable giving hit a record $592.5 billion in 2024, but the number of Americans actually giving has fallen five straight years. Two-thirds of households gave in 2000. Roughly half do now. Even affluent-household giving slipped from 90% in 2017 to 81% last year. Record dollars, fewer givers. The money's concentrating and the generosity is thinning at the same time.
And this shows up right inside Index's own portfolio, which includes Anthropic. The piece cites a financial planner, Alex Caswell, who was asked whether his newly-wealthy clients — a lot of them Anthropic employees tied to effective altruism — were planning to give away the bulk of their fortunes. Anthropic even matches employee donations of up to 25% of their equity. And some used it. But most weren't building philanthropy into their plans at all. They were focused on angel investing or starting their own companies. Caswell's line: "That's what I'm seeing more than the desire to become philanthropic." Effective altruists, redirecting the altruism into founding more startups. Make of that what you will.
Then the scale of it. Musk is worth just over a trillion dollars after SpaceX's IPO last month made him the first person to hit that mark. Forbes counted 45 new AI billionaires in its 2026 rankings, worth a combined $2.9 trillion — and that's before Anthropic or OpenAI have even gone public. The piece notes that once those two do IPO, their combined employees will hold enough wealth to buy nearly a third of all the homes in the San Francisco metro area. A third of the housing stock, held by the staff of two companies.
Rimer's history lesson is the kicker. Last time American wealth concentrated like this was the Gilded Age. In 1889, Andrew Carnegie published "The Gospel of Wealth," arguing a rich man should give his fortune away in his own lifetime, calling it a disgrace to die wealthy. That became the intellectual grandfather of the Giving Pledge. But it didn't hold the other path off for long. By the mid-1930s, Senator Huey Long had a national movement called Share Our Wealth demanding steep taxes on the rich, and a nervous Franklin Roosevelt pushed through what the press called the "soak-the-rich tax," raising the top rate as high as 79%. Rimer's whole bet is that his peers pick the easy way — voluntary — before history picks the hard way for them.
Why does this matter to you, a builder? Because it's not abstract anymore. California voters will decide this year on a 5% one-time wealth tax on the state's billionaires. Google founders Sergey Brin and Larry Page have already moved their primary residences to South Florida to be safe. And the piece floats a genuinely cynical wrinkle — one reason OpenAI is reportedly eyeing a public offering in 2027 is that the California tax, if it passes, would calculate net worth on worldwide assets as of the end of this calendar year. There's also OpenAI reportedly discussing handing the federal government a 5% equity stake, which Sam Altman frames as sharing the upside and critics see as buying political cover. Roelof Botha's dry line from a separate interview: some of the most dangerous words in the world are, "I'm from the government, and I'm here to help." The point is, the environment you're building your company's value in is about to have redistribution baked into it, one way or another. Rimer's just saying the quiet part out loud.
Okay. Let me clear the desk with a few faster ones, because there's good stuff in the back of the fridge.
The Zoom hack that quietly tells you something real about where we are. There's a piece noting VC Jeremy Levine changed his Zoom display name. He's not "Jeremy Levine" anymore. He's "Jeremy Levine I do not consent to transcribing or recording." Petty or brilliant, your call. But the reason is that always-on recording has become ubiquitous thanks to the flood of AI note-taking apps. Another VC, Eric Bahn, says he now just assumes every founder meeting is being recorded before he even sees a phone slide across the table. And there's a founder in the piece who records most of her first dates with the Granola app and feeds the transcript to Claude afterward to see if she could've been more "engaging or empathetic" — and to check who did most of the talking. Dating in San Francisco is apparently a performance review.
But the smart question in that piece is the one for builders: if every meeting and watercooler chat and dinner gets transcribed and summarized, who's actually reading any of it? At what point does the audio landfill of every conversation you've ever had stop being useful and just become another recording nobody has time to play back? The capture problem is solved. The "so what" problem is wide open. If you're building in this space, the money isn't in recording more. It's in making the mountain of recordings actually mean something.
Quick one on policy: the Department of Justice says federal employees can download TikTok on their government phones again. A 2022 law banned it, but the DOJ says the law no longer applies thanks to a deal transferring TikTok's US operations to a joint venture backed by Oracle, Silver Lake, and MGX — with Oracle as security partner and old owner ByteDance keeping a 19.9% stake. The saga that would not end has, apparently, ended, at least on this front.
And on the crypto-security beat, because a lot of you build near this: the FBI arrested a 21-year-old Florida man, Zyaire Wilkins, accused of uploading fake video games loaded with malware to Steam. Once people downloaded and played them, the malware stole passwords and drained crypto wallets — around 8,000 victims, roughly 80 wallets hit, at least $220,000 in crypto stolen. Games with names like BlockBlasters and PirateFi, all designed to look and play legitimately. Valve, Steam's maker, has been pulling these. But here's the detail I love, because it's the opposite of high-tech: the feds traced crypto payments used to buy gift cards, including Uber Eats. They subpoenaed Uber, saw the gift cards were tied to an account making deliveries to Wilkins, got a warrant, and grabbed his MacBook. Guy allegedly drained crypto wallets for a living and got caught by his dinner delivery. Operational security ends where the burrito arrives.
Now let me tie the week together, because there's a spine running through all of it and I don't want you to miss it.
This whole edition is a story about the distance between the promise and the plumbing. Kimi is a frontier-class open model that spooked the market, and the loudest response wasn't a better model — it was a plan to scare people away from it. Databricks got a $188 billion valuation partly for using cheap open models well, while an exec at a rival is drawing up how to make that exact strategy radioactive. Vertu will sell you a $6,880 phone whose AI agent can't remember a spreadsheet from Tuesday. And Neil Rimer, a man who just made billions on the promise, is telling his own peers to start giving it back before somebody makes them.
The connective thread isn't "AI is overhyped." It's more useful than that. It's that in every one of these stories, the technical reality and the story being sold have come apart, and the interesting action is happening in that gap. The FUD memo instead of the ban. The AI-halo valuation on top of boring data software. The autonomy pitch on top of a forgetful agent. The philanthropy pledge that quietly turns into more angel checks. As a builder, your edge this year is being the person in the room who can tell the difference between what a thing does and what it's being sold as. Everybody else is reading the marketing. You read the review, you read the benchmark caveats, you read the display name that says "do not record me."
As we covered earlier this week with the Hyundai walkout and the whole robots-versus-workers fight, the machines keep arriving faster than the institutions know how to price them — the wages, the taxes, the trust. This week just showed you the money side of that same lag. The capability is here or nearly here. The wisdom about who benefits and who pays is running way, way behind.
So watch the soft law, not just the hard law. Watch whether your cost structure depends on a model somebody wants to make scary. And if you're building an agent, for the love of everything, make it remember the spreadsheet before you make it book the trip.
That's the menu today. I'm Tony DeLuca, this has been Barely Possible, and I'll be right here tomorrow. Take care of each other out there.