
Anthropic の1.4兆ドルIPO:人類史上最大の富の創出イベントか
Anthropic’s $1.4T IPO: The Biggest Wealth Event in Human History?

Anthropic の1.4兆ドルIPO:人類史上最大の富の創出イベントか
Moonshots with Peter Diamandis
要約
Peter Diamandis、Salim Ismail、Dave Blundin、Alex Wissner-Gross の4人が、Anthropic の IPO とサンフランシスコへの富の集中、NVIDIA の Jensen Huang 辞任騒動を議論した回。あわせて、米国の大規模なサイバー障害、Starship による初の機体キャッチ、対中チップ規制、OpenAI の新モデル、洋上都市構想なども取り上げた。収録は2026年11月18日で、ライブ配信である。
この回のポイント
- ●Anthropic は ANT のティッカーで上場予定。出席者の説明では、価格帯の上限で約640億ドルを調達し、評価額は約1.4兆ドル。需要は供給の約5倍とされた。
- ●平均的な従業員の持ち株価値は、Anthropic で約7200万ドル、OpenAI で約1.26億ドルと紹介された。サンフランシスコに富が集中し、ユニコーン到達までの期間も急速に短縮していると語られた。
- ●NVIDIA は決算が好調だったが、Jensen Huang の辞任と取締役会の対立が話題に。出席者は、後継計画の不在、活動家投資家の介入、CUDA の堀の揺らぎなど複数の見方を示した。
- ●一連の障害(Outlook、FAA の全米運航停止、大手銀行、Optimizely への攻撃)について、協調攻撃とは見ていない。老朽インフラの更新と、攻撃側がAIで自動化する非対称性が課題だと指摘した。
- ●Starship Flight 15 で機体が初めてキャッチされた。Alex は宇宙輸送を「パケット交換」にたとえ、Peter は軌道到達の低コスト化を試算した。
- ●対中では、ソフト(モデル)の規制は効きにくく、チップなどハードの規制が要になるという意見が出た。短期の競争では有効かもしれないという見方もあった。
章立て
導入:トランプ訪中とNVIDIA騒動
米中首脳会談の話題から、NVIDIA の決算と Jensen Huang 辞任をめぐる議論、Google の動向への期待までを語る。
Anthropic IPOとSFへの富の集中
IPO の条件、従業員の資産規模、ユニコーン到達の短期化、起業家への助言、SFの街の課題を議論する。
相次ぐシステム障害とサイバー攻撃
Outlook、FAA、銀行、Optimizely の事案を整理し、老朽インフラとAIによる攻防の非対称性を論じる。
Starship初の機体キャッチ
Flight 15 の成功と V4 の構想、打ち上げコストの試算、宇宙輸送を「パケット化」とみる見方を紹介する。
チップ規制と中国AIをめぐる政策
サンフランシスコ宣言、東南アジア向けチップ規制案、中国製モデルの利用開示案について議論する。
xAIのデータセンター建設とGPT-6.2
建設速度の秘密、AIによる管理、OpenAI の新モデルと Opus 5.5 との比較に触れる。
洋上都市(シーステッディング)
Palmer Luckey らの Slab 構想と、Starlink やデータセンターが追い風になるという議論を紹介する。
視聴者からの質問コーナー
エージェント運用、教育、重みの直接編集、ウイルス対策AI、長寿と独裁、リサイクルなどの質問に答える。
解説記事
AI・テクノロジー分野のポッドキャスト「Moonshots」の2026年11月18日収録回では、Peter Diamandis、Salim Ismail、Dave Blundin、Alex Wissner-Gross の4人が、Anthropic の IPO を軸に最近のニュースを語った。ライブ配信で、タイ・バンコクにいる Salim が深夜に参加したことも冒頭で語られている。なお本稿の数値や見方は、すべて出演者の発言に基づく。
Anthropic IPOが示す富の規模
Peter の説明では、Anthropic は翌朝にティッカー ANT で取引を始める予定だ。825百万株を70〜78ドル程度で売り出し、上限なら約640億ドルを調達して評価額は約1.4兆ドルになる。需要は供給の約5倍とされた。直近四半期の売上は170億ドル、年換算の run rate は約810億ドルと紹介された。
Bloomberg の報道として、サンフランシスコでは数千人が1000万ドル超の資産を得たこと、IPO価格帯の中間値で平均的な Anthropic 従業員の持ち株が約7200万ドルになることが挙げられた。Dave は「ペーパー上の富」という見方に反対し、ロックアップが短く、実質的には流動性のある資金だと述べた。この資金が元従業員によるエンジェル投資としてAIスタートアップに再投入され、起業エコシステムを加速するという見方も示された。
Alex は、フロンティアラボが約2兆ドル規模になってようやく上場するのは問題だと述べた。個人投資家が直接触れられる段階がもっと早く来るよう、IPO制度の改革を望むという立場だ。これに対し Salim は、創業5年ほどの会社である点を踏まえ、「知能が資産クラスになった」ことを前向きに捉えた。Salim はユニコーンまでの期間が2000年の21年から現在は約1年に縮んだというグラフも紹介した。
NVIDIAの混乱とGoogleへの期待
番組冒頭では、NVIDIA の決算が売上1158億ドルと好調だった一方、Jensen Huang の辞任で株価が下がったことが話題になった。Dave は、取締役会が長年求めていた後継計画を Jensen が進めなかったことが背景にあるという自身の調査結果を述べた。Alex は活動家投資家が取締役会の刷新に動くと予想し、Dave は対中チップ流出などの問題があるため復帰は難しいと見た。
また Dave は、CUDA の優位性はAIによって崩されると述べた。Alex は、ハイパースケーラーやフロンティアラボが自前チップを持つため、NVIDIA の設計力を必要としなくなる可能性に触れた。さらにエコシステム維持のための循環的な資金提供が細れば、需要が長期的に弱まりうるとの見方も示した。ただし Dave は、その資金の使い方を詐欺的とは見ていない。
Google については、Dave が今後90〜120日以内に大きな動きがあると予想した。Alex は、現状ではフロンティアラボの3番手にとどまるとして懐疑的だった。
障害は攻撃ではなく老朽化の問題か
Outlook の障害、FAA による4時間超の全米運航停止、大手銀行の障害、Optimizely を経由した悪意あるJavaScriptの混入が相次いだ。出席者の多くは、これらを組織的なキャンペーンとは見ていない。Alex は、パッチ適用自体が障害を起こした例もあると指摘し、本質は老朽インフラとAIによる脆弱性発見コストの低下だと述べた。
Dave は、攻撃側がAIで自動化される一方、防御側は人手に依存する非対称性を懸念し、今後半年から8か月は厳しいと述べた。また、見た目だけ整った「vibe coding」製のソフトが政府調達に入り込む「vaporware」問題にも触れた。一方で、更新コストはAIで大きく下がったため、レガシー改修は起業機会だという意見でも一致した。Alex は、障害がAI規制強化の口実に利用される恐れも指摘した。
Starshipのキャッチと宇宙輸送の変化
Peter が今週最大のニュースとして挙げたのが、Starship Flight 15 だ。Ship 42 が約22時間半後に帰還し、Mechazilla のアームで初めてキャッチされた。Elon が語った V4 の構想や、推進剤コストの低下から軌道到達が1000ドル未満になりうるという Peter の試算も紹介された。
Alex はこれをインターネットの「回線交換からパケット交換へ」の転換になぞらえ、宇宙ミッションが数百〜数千の標準化された飛行の組み合わせになるという見方を示した。
チップ規制と、規制の効きどころ
米国と9か国のサンフランシスコ宣言、東南アジア5か国などへのAIチップ輸出ライセンス案、中国製AI利用の開示案が取り上げられた。Alex はこれを「Pax Silica 対 Pax Sinica」という新たな冷戦的構図の一部と捉えた。
Dave は、モデルは蒸留などで再利用できるため、規制が効くのはチップ工場やデータセンターなどのハードだと述べ、フォトニクスなど新しい計算基盤も見据えるべきだとした。Alex は、シンギュラリティ直前の短期競争なら牽制に戦略的意味があるが、長期では逆効果になりうるとの両面を示した。
まとめ
今回の議論から読み取れるのは、AI企業の評価額と資本の集中が、起業、投資、地政学の前提を変えつつあるという出演者の認識だ。ただし、語られた数値や予測の多くは出演者の見解や報道の引用であり、確定した事実とは限らない点に注意したい。日本の読者にとっては、AIによるインフラ更新の機会、ハード規制をめぐる国際競争、そして資本がAI起業に再循環する流れが、今後を考える手がかりになるだろう。
文字起こし(英語・自動生成)
Mute in between. Really? Yeah. It sounds like a rodent on your microphone. Yeah, it just makes a lot of noise. I'm sure it's just a pending rant. If I do nothing, you're slow-hearing? No, it's gone now. No, it's when you're moving on. Okay. All right. It's when you're talking. Just don't talk. God damn it. Do you guys realize... It's annoying when you talk. What time is it? It's 1 a.m. here in Bangkok, right? Oh, you're the man. I'm staying up again through the middle of the night. I'm tired. Whose fault is that? Scheduling. You know, we did this 24 hours ago. Note I've got the same shirt on in the same hotel room. Holy crap. Same lighting. I'm tired. I'm hungry. I'm cranky. I'm jet lagged. And you guys are wanting to do this live with no editing possible. Hello. Yeah, I can tell already. It's like you're trying to hurt me.
Oh, my God. I love you, Salim. Thank you for your dedication. Best episode ever. Oh, my God. So, guys, a lot happening in the last 24 hours, right? So President Trump landed in China. President Xi met him on the tarmac with a handshake and a hug. A rare welcome for U.S. President. Besides President Xi and President Trump, Vladimir Putin is there. The trilateral meetings are going to be hot. I can't wait to see what comes out of that. And then the last 24 hours, what's that? What could go wrong? Hopefully, what could go right? Yeah, yeah. That's so funny. Did you see the stairway down from the airplane? And, like, the pomp and circumstance in China is insane. There must have been 1,000 people there waving their little flags. But then Xi Jinping walks up to the bottom of the staircase, and it was just like this awkward moment where, like, is Donald in the bathroom in Air Force One? or like, is he supposed to go up the stair? Like, what's going to happen here? And finally. It's all positioning. It's all positioning.
Yeah, coming down on the other world leaders is always a good look. Yeah. Did any of you have, like, 1984 flashing before your eyes? Remember, in 1984, there were only three countries. There was Oceania, which was basically the Americas and the British Isles. There was Eurasia, which is basically Europe plus Russian landmass. And then there was East Asia, those three countries. Did any of you see, oh, my goodness, 1984 politics? Yeah, right now they're like, who gets Malaysia? Well, do you get Saleem if you get Malaysia? Does he come with that or not? I don't know. This is a great game of risk. So you're in Thailand. Where are you right now? Bangkok? Bangkok. And when President Xi landed and they met in Alaska, of course, President Trump had all of the jets flying overhead. You know, show of power. Show of power. Yeah, China should launch one of those hypersonic missiles right over the meeting. No, no ideas. You're not head of state anymore, Dave.
That idea, okay. And then NVIDIA just reported on earnings last night in the first quarter since Jensen resigned. The numbers were huge. Revenue was $115.8 billion, up 103% from last year, 100% growth year after year. Data center revenue was up $107 billion. Net income, $65 billion. These are incredible numbers. And the stock still went down about 1.5% this morning. So interim CEO Colette Kress guided fourth quarter earnings at $130 billion. And then, of course, the whole Jensen soap opera is going on. So what do you guys think about this, right? So Jensen's saying, I'd come back if you asked me. The board's saying, no way. Everybody's weighing in. Thoughts? It's a disaster, man. It's a total disaster. I have a clear thought, so I did a little research. It turns out the board for 10 years has been trying to get him to do a succession plan,
and he's not been doing it, and he's surrounded himself. His kids are in the company, and at some point the board kind of just went, we can't run a $5 trillion company this way, and there's no coming back from that. That's my thesis as to why this is happening in the way that it is. No question that this will be a massive issue. They had to take the pain sooner or later to deal with this. Yeah. And then all the comments in the peanut gallery, right? We've got Jeff Bezos and Sam Altman saying NVIDIA will do fine without him. Jeff saying, you know, bring back founder power. Elon saying, bring back the founder. It's a little different from the Sam Altman, you know, weekend of being fired in that Sam was outright fired, blindsided by the board, and then created an employee and shareholder uprising and got himself reinstated, and then he just quietly eviscerated the board. But Sam didn't want to go. Jensen quit in a very emotional kind of state, judging from all the reports and what he was saying, because he's got the infidelity home life thing going on concurrently, but he quit, submitted a resignation.
So that's a little bit of a different dynamic. Also, Sam's a very young guy with a lot of future in front of him. And Jensen's been running the company for three decades or more with no succession plan. So it puts the board in a really awkward spot. And I guess it's divided now, too. There's three people on the board that want him back. And I don't know, God knows how many. What's really weird to me is that it's apparently a handpicked, called in the press, pushover kind of board. but they're saying majority of them are saying no don't come back which is kind of odd like you push overboard what's your prediction for each of you what's your prediction is he coming back alex oh god watch for these likely down the middle watch for the activist investors this would seem to be a ripe case where activist investors of the bill ackman genre should and Probably, I would expect, will step in and attempt to overhaul the board. This is just – it smacks of terrible governance on so many different levels.
And it's, I think, frankly, an embarrassment for America that our most valuable, at least until recently, by market cap company is in the middle of such an absurd governance failure. I expect the activist investors to circle and take aim and try to fix the situation. And do what? I'll take the other side of that because that would maybe work in a normal environment, but given all the espionage issues and leakage of chips to China and so on, I think that seals the deal. So you think there's no way he's coming back? I don't think he's coming back. Not that I don't want him to. I just don't think he's out. I was out for a weekend. He has Greg Brockman as a super loyal right-hand person, and you could come back quickly. Time is going to kill this thing. Like if you said, well, we'll figure it out in three weeks, four weeks, five weeks, the board is torn, every week that goes by, whether or not he comes back becomes less relevant because it's very, very hard to come back and be empowered again
after six weeks or whatever. Like getting it back to where it was a week ago is going to be damn near impossible if time goes by. So they better figure it out fast. Steve Jobs came back to Apple after years through a reverse acquisition. I think there are all sorts of scenarios where Jensen comes back. But I think what Steve Jobs came back to a company that was borderline bankrupt. So if you follow your analogy along, sure, Jensen could wait three years for this thing to completely fall apart and then come back like a white knight. Yeah, that could happen. Actually, that may be why Polymarket doesn't think he's coming back for this year. But what was the polymarket like by the end of next year? 30% chance, I think, is the last question. Things may have to fall apart a little bit before they want him back enough or before NVIDIA gets acquired maybe by SpaceX Tesla. You never know. But what struck me most in this is history tends to rhyme. And I wasn't around for the 70s, but I read history.
And the period of, call it the 70s plus or minus a decade, was a period during which the great powers were in many cases playing games in the Middle East for oil because oil was the great commodity that made the world go round at the time. Now the great commodity is GPUs 50 years later. And the commodity isn't coming from the Middle East anymore. It's coming from here and Taiwan largely. And I can't help but escape the idea that the great powers, including our great power, the United States government, can't resist the temptation, including factions within the U.S. government, to play games with leadership for the 2020s version of oil. In the 1970s, it would be you'd launch a decapitation strike against the leader of a Middle Eastern country that wasn't maybe friendly toward U.S. oil consumption. Now it's a decapitation strike against the largest U.S. company that maybe isn't adhering to certain desirable policies regarding import-export of GPUs.
Well, I'll tell you what I would do if I were, Jensen. I would say, look, I'm a 3.5% shareholder in this company. Board, I will go quietly if you go ahead and cash me out. I will take that many billions of dollars and I'll do something else for a year, year and a half. Because during this next year, CUDA is going to get obliterated by AI. The CUDA moat is absolutely going to get obliterated by AI. And all these competitive inference time chips, you know, from all these other companies, Positron and Xscale, and there's many of them now, they look really, really good. And so let that all bottom out and then come back like Steve Jobs, the way you said, Alex, you know, come back to turn it around because there's going to be a huge pile of cash no matter what. Yeah, and maybe, David, to add to your comments, the hyperscalers and the Frontier Labs, they all have their own NVIDIA, CUDA, and just NVIDIA in general chip competitors. At this point, whether it's partnering with Broadcom or building what they're building internally or just using Frontier models to design their chip architectures, in many cases, they may think they don't need NVIDIA's chip design prowess any longer.
The AI has superseded however many PhDs NVIDIA has on staff doing their chip design. They may feel that at this point they don't need NVIDIA's chip architecture anymore. Maybe they need indirectly the TSMC transistor supply, but they don't need NVIDIA per se. And the argument would go, if you follow that to its logical conclusion, that the only thing keeping sort of the finger in the hole in the dam is NVIDIA's alleged circular financing to prop up neoclouds and startups to purchase NVIDIA particular architecture versus alternative architectures from the hyperscalers, neoclouds, and frontier labs. Jensen leaves the circular financing scheme, as reported and or alleged, vanishes under the watchful gaze of a new interim CEO who's mindful of pressure not to engage in circular financing schemes, the thing collapses, not investment advice, but the scheme collapses,
in particular, the artificially, as reported, propped up demand for NVIDIA's chips collapses due to that circular financing collapsing. And there are many ways to collapse. I don't mean like literally overnight collapsing. There's a lot of demand, but over the longest term collapses. And then Jensen can say, you know, I was just... displaced and the thing fell apart. Post hoc ergo propter hoc. So, you know, you should bring me back. After this slash that, therefore, because this, in other words. Thank you. Confusing. You forgot the open paren and close paren. Open paren, close paren. But the point is a lot of people will be tempted to conflate causality or causation with correlation when, in fact, it might just have been the circular financing scheme that was propping everything up.
I think there are a lot of posts saying this circular financing propped up. It sounds like a big Ponzi scheme, but I actually don't see it that way at all. I think they have massive amounts of cash piling up. They're going to end up doing a dividend, which is insane. Like for an AI company in hyper growth to do a dividend because you can't figure out what to do with your money. That's nuts. But, you know, I think, Peter, you said $130 billion next year at 80% or 75% gross margin on that. There's never been a cash pile that big piling up that quickly before. And so building out your own ecosystem around you with that money is actually a very natural thing to do. I don't think it's fraudulent in any way. But I also completely agree without Jensen at the helm, the board and the CFO are not going to keep doing that anywhere nearly as aggressively as Jensen was doing it. And I really think he does need to act a lot or the company needs to act a lot like Elon in this moment to be relevant, because there's no way you can sit on your existing GPU architecture and say, yeah, that's going to be around five years from now, seven years from now. That's so important.
They need to build. You know, what Elon does extraordinarily well is build adjacent companies over and over again. He's going up and down the stack. And NVIDIA needs to do that. They started to do that with open source models. And I'm curious, you know, what Jensen's departure is going to do to the open source world. Oh, my God. I don't know if we have more time to talk about that. But I cannot believe how big a deal this is and the balance of power between Google and the rest of the world. And, you know, Elon immediately came out supportive of Jensen. And I was like, well, I know he needs the Vera Rubin chips, but why is he so supportive? And then realized, like, Elon is, he started OpenAI purely to counterbalance Google, taking over the entire world. And then he lost control of OpenAI to Sam, and now he's, you know, trying to counterbalance the world with XAI instead. But Google has all the data, more data than any CIA or any other foreign government. They have more data than any government on the planet. And now the TPUs are starting to really take off.
And so that's a vertical monopoly like the world has never seen that is on the cusp of total AI takeover. And you notice they're very quiet in the press. There's no drama type person. Have you seen the rumors about Google is going to pounce soon and sort of take the lead across everybody? There's a lot of expectation of what Google is going to do given the power. Yeah, I saw a posting. They've already released a bunch of new chips and so on. So they're going full out. Oh, yeah. Google should have pounced years ago. They haven't been especially competent, and I say that with full editorial independence of any relationship Google may or may not have with this podcast. I think Google can and should have acted years ago to take the lead as a frontier lab, and they've flubbed it. Google's Gemini 4. Not over yet, Alex. They're going to come back. Gemini 4 Carbon is an improvement, yes. I'd love to, and I have friends on the Gemini team,
and friends on the TPU design team. I would love to see them actually pulling ahead and not just doing whatever they can to tread water as a number two or number three. Mark my words, Alex. They're going to be. We're going to see this, I don't know, next 90 days, 120 days, max. What happens in the next 90 or 120 days? Google comes out with a plan that sort of shakes the Frontier Lab infrastructure. You have insider info on Google's 90 to 100? I have expectations of what Sundar can and will do. And what Sergei – Sergei is not going to sit back and let this be. I mean, he's way too silent. They're cooking. They're cooking in the background. Oh, my God, yeah. Yeah, the ultimate, if you want to go conspiracy theory on it, if you really wanted to sneak up and take over the entire world, you would not be on stage. Your models would not be constantly leapfrogging the other guy. They'd be just like one notch behind.
You'd stay out of the limelight. You would slow play your hand while you build out your TPUs. But if you look at the actual TPU compute, the amount of compute that's getting cranked out now, it's grown exponentially in just the last year, and nobody talks about it. It's not in the news. It's just sort of quietly ready to pounce on TSMC and take over this entire void that NVIDIA has left. I mean, you can argue who's the light side and dark side of the force here, but the balance in the force is hugely different with NVIDIA not counterbalancing Google. Yeah, if I had to make the bull case, the Steelman case for how Google pulls, sort of pulls ahead from behind where they're so obviously number three right now among the Frontier Labs, I suspect it would be via some XAI cursor-style acquisition or strategic partnership that surprises the market. It's difficult and not impossible to imagine them organically becoming number one or number two, just like GCP, still not number one or number two public cloud.
They're still a solid number three. Yes, it's good for their business and everything, but they're not number one or number two, similarly for their frontier model status. Yeah. I can't wait to send you a text. Called it. I can't wait. Look, Peter, I want the frontier to be competitive. Frontier liberation. Frontier we go. I want Google to leapfrog everyone else. I just haven't seen it happen yet. All right. All right. Well, let's get into the show. Welcome to Moonshots, everyone. Your number one podcast in all things AI and exponential. Your front row seat to the hypersonic tsunami that we call the singularity. You know, it feels like the singularity. I mean, the rate at which stuff is happening is through the roof. Today, we've got the Fantastic Four, AWG, our resident physicist and genius, Dave Blunden, our impresario of AI investing, Salim Ismail, our globetrotter. So, Salim, again, for the records, where on the planet are you and what time is it there? It is 1.25 a.m.
I'm in Bangkok. I came here for a leadership retreat for a big company. and I'm flying back in 10 hours. So how many hours on the ground are you there? You are a probability function. Yeah, I've probably got the worst carbon footprint in the world right now. And I'm going to say it again for the record. I'm tired, hungry, cranky, jet lagged, and you guys are doing it live. Like, literally, I think you're trying to hurt me. My brain is trying to, like, crank. It's hard enough keeping track with all you guys, but late at night and it's, you know. Anyway, we'll see what we can do. All right. I'm going to rely on the comments from people to get clever ideas to talk about. Yes, and we are live streaming this on X just for fun and just to really help people get a sense of where things are going. So I'm Peter Diamandis, your host. You're a data-driven optimist. Our mission, help you understand what happened the last few days, what it means for you, your family, your company, and keep you abundance-minded as we're going through this hypersonic tsunami.
Okay, the number for this week is 22 and a half hours. That's how long Ship 42 stayed up before the Mechazilla arms caught it back at Starbase. If you're new here, please hit subscribe. You know, our mission here is to get, you know, sort of the optimistic, abundance-minded news out to as many people as possible. We put this out twice a week. You're not going to get it via email notification. You're going to get it if you subscribe, so please do. Also, on X, our handle is at moonshots underscore pod. We have a clip channel there for the hottest takes, in particular what Alex's genius has put on the show. We cut it out and feed it to you. All right, let's jump into our first story. It's the Anthropic IPO and the growing wealth in San Francisco. So tomorrow morning, at the time we're recording this, Anthropic starts trading under the ticker ANT. Today, they're expected to set final price. The deal is for 825 million shares marketed somewhere between 70 to 78 dollars.
There is massive demand, right? There's 5x demand for the shares they're selling. At the top of the range, the offering would raise about 64 billion dollars, valuing Anthropik at 1.4 trillion, still below the 85.7 billion SpaceX raised back in June. Wall Street expects it to rise to about 80 bucks a share. I'm curious if any of you are buyers, but we can talk about that afterwards. The filing shows third quarter revenues at $17 billion, up $1.4 billion from a year earlier, and an annualized run rate of about $81 billion as of October 31st. Seven companies, including Google, NVIDIA, AMD, and Microsoft, agreed to buy $26 billion of non-voting shares in the IPO. Reminder, the seven co-founders have pledged to donate 80% of their personal equity, A huge amount of capital is going to be flowing in the philanthropic world out of OpenAI and Anthropic. This is the public market's first hard mark on the frontier AI labs to see how they dominate the stock market going forward.
It's going to be the S&P 3 between Anthropic, SpaceX, AI, and ultimately OpenAI. In a related story, Bloomberg Business says that in a seven-mile-wide city of San Francisco, barely 800,000 people, the AI boom has minted thousands of fortunes worth over $10 million. At the midpoint of Anthropics IPO range, the average employee would own shares worth $72 million. So if you can go back in time, get a job at Anthropics. The average open AI insider take is worth almost $126 million. for their employees based on the firm's most recent valuation. With only 0.1% of the world's population, San Francisco is now, by some counts, absorbing more than a third of the global venture capital marketplace. Dave, let's go to you first. Your thoughts on the IPO and on this concentration of wealth in San Francisco? Oh, my God. Those numbers, so 72 million, that was the average?
Just take insider holdings divided by headcount? Yeah. And then for OPI, it's $126 million at their current estimated valuation. I mean, it's crazy. Yeah, I know this is where people don't generally – they don't think a lot about trillion being that much different from billion. But then when you divide it out, it's like nothing the world has ever seen is of this scale. And I was looking at the data on the wealth in San Francisco, and it is so dwarfs any other event at any city in any point in human history. is just miles and miles and miles ahead. I think a lot of people also don't realize the SpaceX IPO, you kind of associate that with Texas now because Elon moved. But almost all of the shareholders and investors from the early days also come from San Francisco, Steve Jurvetson and a whole bunch of other friends of ours. And all that wealth already landed in San Francisco. So OpenAI is not liquid yet, but when you add it all together, it's so funny on that chart. The wealth from Anthropic is a big bubble.
The wealth from OpenAI is a big bubble. All other companies combined is an equal-sized bubble. That's how big these are. So it's just a massive concentration of wealth. It definitely moves the financial capital of the entire world to SF. And, you know, New York can complain about it, but it's just a mathematical fact. And also, when people look at global wealth, you know, they usually look at all assets combined. Most of that is real estate. But if you go and ask somebody, hey, I want you to invest in my company. Can you give me your real estate? That's not a conversation that ever flies. But after this IPO, if you go to somebody and say, hey, can you invest in my company with your new liquid cash that came from this IPO? Also, all the slides and all the postings say this is paper wealth. That's nonsense. This is trading wealth that will be fully liquid right after the lockup, which is a very short lockup. So this is real money. You can bar against it, yeah. What's interesting is all of the accelerants and the singularity. We talk about acceleration from the GPU expansion of the planet, but we also have an acceleration from the fact that these thousands of employees will individually become angels and super angels,
seeding the next companies coming out of the Bay Area and the rest of the world. So it's a hyper-acceleration that's going on. We're not seeing an exponential. We're seeing a hyper-exponential. Alex, your thoughts on this? Yeah, I think it's disappointing that it took this long to get to an Anthropic IPO in the first place. Individual retail investors, certainly via exposure through index funds, had exposure via Google, for example, to Anthropic and via maybe financial services institutions, but no direct exposure. And I think it's a travesty that it's so difficult to IPO in this country that frontier labs wait until they're worth almost $2 trillion to give retail direct exposure to them and the wealth creation. So, yes, it's great that, in my mind, that Anthropic eventually got around to IPO-ing. It's not great that it took until $2 trillion to get here, and I would love to see real reform of the way IPOs are handled in this country
to make it a true incentive for the next Anthropic to IPO at a much lower valuation at an earlier stage in their development to give retail investors direct indexed, close paren. There's no open paren there. Exposure to that equity. Yeah, and when we start trading on digital exchanges, it's going to accelerate things. Yeah, I want to push back just a little bit on this, Alex. The company was formed in 2021, right? It's barely five years old. Like, you know, I know it would be great to have an IPO soon, but you've got to get it to some level of maturity and stability. Not that this is a stable environment, but I think this is kind of the most unbelievable wealth creation we've ever seen. What I like about this is we're seeing intelligence become an asset class now, and that, I think, is a really good portent for the future.
I would maybe just comment the hypersonic, if you will, growth trajectory of Anthropic from a standing start from really a schism with OpenAI. Putting the schism aside, we want to see rapid growth for IPO-able companies. If anything, I would say – You do. So I would say, again, hopefully Anthropic fulfills their S1 vision of capturing $30 trillion. If so, there's a lot of delta left between where they are right now and saturating whatever their ultimate target market is. But on the off chance that they don't completely fulfill their vision, that is a lot of wealth creation. Maybe call it paper wealth creation. Arguably, real wealth comes from creating greater future freedom of action for humanity and not just from moving dollars around. But that is at least a lot of paper wealth creation that retail investors missed out on that certain private investors captured.
So even the fact, as you're pointing out, Salim, that it was just a few years from a standing start in schism to IPO, I would say, if anything, I want more for three to one year from start to IPO ability companies rather than saying this is an anomaly. Okay. So I have a question for you guys. At some point, is the holding period three months or six months in this case? I think it's six months, right? Normal is six months, but I think they have an early release on some of it. So assume it's six months. At the six-month mark, do we see a flood of exits as people liquidate and go do their own thing? Or is it so exciting to be there that people are like, damn, I don't dare to leave this place? Which way does it go? I think Anthropic, based on what I read, has what's perceived by markets as a relatively unusual schedule for employees taking their vested shares and liquidating them.
This may not be business as usual as you would see with a conventional IPO. Yeah and I think it a huge amount of rotation on that day but there be a ton of people taking probably half off the table But the demand is so high I don you know trading on what level it trading that day because it not going to be anywhere near the IPO price Usually it's either much higher or much lower six months later or three months later. But I don't think they'll have too much trouble rotating it out. So then that will all be fully liquid and reinvestable. And like Peter is saying, it's all going to get reinvested back into AI startups and AI Series A's. Dave, check out this graphic I made. I went and looked at time between founding and reaching unicorn status. So in 2000, it was 21 years, dropped down to 11 years in 2010, four and a half years in 2020, and a year now, right? I mean, the rate of wealth creation is, you know, it's insane.
It is insane. It's the best founder time I've ever seen by so wide of a margin. I can't even overstate how wide a margin it is. And I keep telling people, like, don't – AI could do anything. You know, two years to three years from today, you don't know what world we're going to be living in. But right here, right now, you know exactly what you should be doing. Sprint. Build your company and sprint in that direction. And this capital, you know, most of my life, the entire venture capital industry combined has been, you know, $20 to $40 billion. Now it's bigger than that. That's over $100, $150 billion plus now. This is $1.4 trillion in personal liquidity. So it's like a decade of the venture capital entire industry that's unleashed, all coming from people who made their money in AI. And so naturally they want to reinvest in their roots or what they know worked. I mean, just the scale of that reinvested capital, it's like nothing the world's ever seen. It would be amusing if time from founding to unicorn status actually went negative.
I could imagine crazy scenarios where it crosses the horizon. Back to your options? No, where the value exists before the company. Alex, you're absolutely right. So what happens is in the companies that are getting unicorn valuations out of the gate, it's dependent on the people they collect. So you've got some known founder who aggregates two or three other rock star founders, and just based on who they are, not even on their projections or their business plan, they're getting billion-dollar valuations out of the gate. We saw that already with some of the early folks leaving Anthropik and OpenAI. Sure, sure. I mean, I see this all the time, actually, now that I'm forcing myself to think about it. We are in narrow cases, call this spiky, spiky, anti-causal or a-causal unicorns, where safes are being written to teams at unicorn valuations before the company exists all the time.
So in a spiky sense, we're already there. Query whether we could see this more broadly in the market, where time to unicorn status goes broadly negative. So advice to entrepreneurs out there listening, your team matters more than anything else. Find some extraordinary people who have pulled it off before. Have a clear moonshot vision. That's the most important thing. When I'm investing, when Brett Adcock, we just released a podcast with Brett at Figure that Dave and I did. When Brett first came to me in the earliest days, I was an early investor through my fund, Bold. But, you know, what was amazing was his success having, you know, hit multiple exits, him putting his own money in. But the team he aggregated was through the roof. And it was like, if you can do that, we're going to write a check. Well, I'm telling all the teams to go to SF. You don't have to move there. In fact, if you're, you know, the schools and the ecosystems that are hugely untapped are, you know, Carnegie Mellon, Georgia Tech, you know, those areas, Villanova, Northwestern.
Stay close to wherever you're recruiting from, but spend a ton of time in SF. And the difference is so noticeable, just walking down the streets and talking to people. It's like first world difference, you know. You'll see it immediately. And that's where you're going to want to do all of your idea generation and your capital raising, but not your recruiting. If you hang out in the schools, I don't want to get in trouble with schools that I really love, but if you hang out in the major universities in SF, there's a very, very noticeable talent void. You're like, where did everybody go? But then you look at the amount of money that is trapped in their vesting schedule at Anthropic and OpenAI. they're not going to walk away from an average of 75 million or 140 million that's still got two years of vesting in front of it. And if you look at Anthropic, you know, like Selim just said, it's only a couple years old. So everybody's still early in their vesting cycle. So they can't leave. And so, okay, well, who's going to start the next wave of companies then? And it's like this huge void of talent with this massive surplus of capital.
So what you want to do is recruit somewhere else, but go to SF and spend as much time as you can there networking and getting to meet people. How many trillions of dollars of IPOs in SF do you guys think it takes before SF looks like it is a city that hosts trillions of dollars of IPOs? Until recently, I walk around Union Square and I see people in the fentanyl fold. At least that's been somewhat cleaned up recently. But how many IPOs does it take before SF looks like a city befitting all this newly created wealth? Do you mean how long it's cleaned up? Yes. and modernize. It's going to be a while. It's going to come long until it looks like the future that it's minting. I would say three years, I think, because the liquidity has to happen first, and that happens tomorrow. And then, you know, Mark Andreessen and a bunch of people were trying to build a new city, like just in East Bay, and I don't know how that's going, but it takes a while to build a brand-new city from scratch. That takes a few years. But I think AI as a management tool is going to start to penetrate SF.
If you look at Waymo as an example, it's like you could say three years ago, when am I going to see Waymo everywhere? What is taking so long? Now you're like, oh, wait, it's here. It just sort of happened. Just not in Cambridge. Just not in Cambridge or Boston. It's awful. We're going to have Optimus and figure robots walking around the streets of San Francisco, cleaning things up, taking people into the halfway house. Maybe that's the solution. It's not obvious to me. I mean, it's, again, sort of a best of times, worst of times in SF, where on the one hand, you have super intelligence. In some ways, SF feels post-singularity. One can look at parts of it and imagine Starfleet Academy or Starfleet headquarters being built there at some point in the near term future. And then you see people who are so obviously on fentanyl or, you know, other criminal problems that are just happening very visibly on the streets juxtaposed next to each other. Blade Runner. Okay, so you're saying this is the cyberpunk future, a.k.a. high-tech low-life.
But do we really want a high-tech low-life cyberpunk city being like our embassy to the future? We want Mecca to look like Starfleet Academy. We definitely do. We have predicted that we're going to be running all science fiction plots in real time, so we're seeing that happen. I don't like Blade Runner. okay i'm gonna move us on so you know let's go to our next story here so this week um you know started out with a number of what we call cyber attacks and system failures uh on our email our banks the entire u.s air system uh making it all look fragile all in six days so four blows came close together uh it's not a campaign but let's talk through it because there's a lot of fear circulating throughout the public discourse. Last Thursday, Microsoft Outlook and Exchange went down. Down Decoder logged nearly 130,000 Outlook reports across 74 countries. It hit me. Yes, I'm switching over to Gmail.
I have my Gmail. I just need to make it my number one email. Then on Friday afternoon, the FAA issued a nationwide – what's that? I just have some memories of you hitting – you know, every time you landed on a flight, I would get 20 emails whacked into my inbox, and I'd go, oh, boy, Peter just landed. And you could see you've been in Outlook on the whole flight, tapping away. We were just begging you to move on by when you're finally doing it. It took a global Microsoft hack to do it, but thank God. Those days were pre-StarLink on the airplanes, though. Okay, let's continue. So then on Friday afternoon, the F.A. issued a nationwide ground stop for more than four hours, the longest nationwide ground stop in 25 years. Transportation Secretary Sean Duffy said about 3,500 planes that were ready in the sky were driven to land, making it, you know, he may use this as a plea for his $50 billion revitalization of this ancient system we call air traffic control. And as a pilot, believe me, it's an ancient system.
You know, the fact that you're actually, you know, talking over a open channel to air traffic control, You're reading your comments to them. They're reading it back. You're writing it down. You're reconfirming it. Then you're entering into your flight control system. It's insane. About 12,462 flights were canceled. Homeland Security Secretary Mark Wayne Mullen said there was no indication of a cyber attack. Then on Monday, three major U.S. banks were facing reports of widespread service disruption, including Chase, Bank of America, and Wells Fargo. J.P. Morgan called it a software bug, an on-premise hardware with no evidence it was a cyber-related attack. It was resolved that evening. And then also on Monday, there was a confirmed cyber attack. Attackers used a flaw in Optimizely's experimentation sign-up flow and injected hostile JavaScript into 178 organizations. Companies like CNN, Zoom, HSBC, Equifax, Salesforce, and IKEA.
Apparently, the code tried to hijack crypto payments, skin cards, steel cookies, and screenshots and load iPhone exploits. Optimizely says it closed the hole the same day. So the question is, should the public be concerned? So, you know, I think they should be concerned about centralized systems that are old equipment and, you know, and trusted vendor scripts need to be rewritten. Right. This was not a coordinated mega hack for me. to my eyes at least the ground stopped and optimizely were the worst of them the fixes are all boring essentially we got to rewrite legacy hand-built software and modernize legacy hardware and interfaces so i went to you alex on this and asked what's your opinion uh you said no big deal yeah it was a nothing i mean unfortunately i was not in the position although sorry dave i i know you were in the air when this happened i was on the position where i needed to fly So I don't want to write it off completely. It was a terrible inconvenience for all concerned.
But at the purely technical level, I think, is there some grand scheme here? Probably not. If you want to squint at this and tell some sort of story, in some cases, it's been reported that some of these outages were the result of attempting to produce patches for software. Sometimes patching software can be just as interruptive as vulnerabilities being exploited. If I had to stitch together a conspiracy theory tying together all of the recent infra issues without reference to that, I think, pretty hilarious Saturday Night Live spoof, it would probably be something about how we have aging hardware, aging software infra on the one hand. On the other hand, AI has now made vulnerabilities and CPEs and exploits too cheap to meter. And there's just going to have to be a great overhaul of all of our hardware, telecoms, software infra, stack,
to recognize that there's now a new asymmetry between both the needs of modern software and superintelligence to consume bandwidth and compute, and also the capabilities of modern superintelligence to be able to discover flaws in the existing legacy stack. And if you look as just another barometer at the Linux kernel, Linus Torvalds thinks now we're actually past the peak of number of patches to the Linux kernel that AI is discovering. I don't know whether that will hold up or not, but to the extent that we're in some sense past the worst, in his view, of AI discovery of software issues. Maybe this will just take a couple of years to percolate throughout the rest of the economy to some of the slower lanes before finally we have an upgraded stack. It's amazing how concentrated our risks are getting over time. Everything is built on a few platforms. The entire U.S. aviation system can go down from a single point of failure.
That alone is remarkable. But I was on a plane coming from SF back to Boston, and we were over Minneapolis, and then everybody's iPhone at the same time gets the same news alert saying, something's wrong, all flights are grounded, nobody knows why. So there's immediate, like, you know, panic ripple that goes across the entire plane. Did you get diverted, or did you land in Boston? No, very few flights got diverted. It actually worked out fine for me, but what a shit show for everybody trying to leave. I couldn't believe when we got into the airport, the line filled the entire, there was a big snake, filled the entire terminal. I was estimating how many people are in this line and how quickly is the line moving. From the back of the line was eight hours to the front of the line, and more people were piled into it. And so, you know, everybody's like, I miss my connection, and, you know, I can't see Salim in Bangkok. And it was just horrific for everybody outbound, And it was totally fine for flights coming inbound because they generally went to where they were originally destined to go.
But, man, talk about, like, single points of failure pop up all over the place. And I think, you know, there's a cybersecurity issue that's percolating, but there's also a vaporware issue that's percolating where a lot of really kind of cheesy startups have popped into the world with vibe-coded front ends. Right. And when they teach them to a government. Yeah, half-big software. You know, in the old days, like four years ago, if you built something that graphically looked like it worked and you click on the buttons and it worked, you must have done some coding. So there must be some competence in your company. Now, if you're selling to the FAA, you could literally be totally fake and vibe code something using AI in about an hour. And from the FAA's point of view, it's like, wow, that looks like really good code. You guys must know what you're doing. And so I worry a lot about this vaporware window we're in right now where it's transitional because the AI is getting so good that it actually can build, you know, solid server-side code now too.
But there's a whole window of time here where there's just so much garbage, so much fake garbage flying around. So I wonder to what degree, you know, some of these issues are not cybersecurity. They're just vaporware. Yeah. You know, we've got a whole civilization now that's so totally connected but not resilient. And that's a dangerous combination. That's a good point. You know, I'm the eternal optimist, but just to be clear, we are going to see major cyber attacks. We are going to see systems coming down. I don't know if it's next week or a couple months from now. But, I mean, people need to be prepared for that. and that will be a transition period until we go and patch up all of the old C and C++ handwritten code that's still in existence in background. It's not just that, Peter. It's all of our cyber offense is now human out of the loop with agents attacking multiple vulnerabilities in parallel at scale.
The cyber defense is still human in the loop, patching one port. That asymmetry is going to be ugly until they reorganize the entire cyber department of every major company in the world and government. It's going to be a very brutal six, eight months. Every time I see a cyber disaster, I'm actually happy in a way that it's not a catastrophic nuclear meltdown type thing, because we need an angry voter base to get the government moving. And there's so little motion and nothing. We talked to Eric Schmidt about this four times. Nothing is going to move the government until it affects votes. And so we need those votes to happen with as little tragedy as possible because we need post-AGI governance. You know, it's going to be totally different than pre-AGI governance. But they haven't even started. It's like if you look at what's even being proposed, it's so trivial compared to what needs to be done. So these, you know, like an FAA outage is actually a really good one because nobody died. But it was massively inconvenient. And that's probably enough to get voters to get active in the, you know, we need massive amounts of change and much faster reaction.
I would worry, though, it's such an opportunity to false flag, if there's an outage of some important system, point a finger at AI, point a finger at some politically expedient target, when in fact it could just be a bunch of copper wires that haven't been replaced with fiber optics. I worry in an era when AISI is getting so politicized that outages far from being prophylactic and encouraging cyber vulnerability patching will instead get weaponized against broader progress with superintelligence. That would be my worry. I totally agree about that constantly, too. How long before the systems are hardened, you think? How long before we have vulnerabilities still in our infrastructure? Oh, gosh. I mean, I can see scenarios where humans will be visiting other star systems and there will be legacy code stuck in the starship or whatever it is, and it's unpatched.
I think there's a long tail, just like we were talking about earlier with SF hosting multi-trillion dollar IPOs on the one hand and having a fentanyl epidemic on the other. I think we'll have interstellar humans riding on, and Werner Vingy dwelled on this just a little bit in his novel, A Deepness in the Sky, riding on legacy code bases that still are unpatched. I think there is going to be like a, we wrote about this a bit in Solve Everything. I think there is, we're in the middle of a great patching moment, and it goes by a few different names, the great refactoring, the great factoring, when superintelligence is going to be broadly used to fix every vulnerability, the software equivalent of curing the top 5,000 diseases. But just because you've cured the top 5,000 diseases doesn't mean that it's been implemented in every possible patient. And same idea with software and especially embedded software that's just being used in the physical world without any plausible, convenient, ergonomic way to update it or patch it.
Yeah. There's a great comment from one of the viewers here called Tiffany. Inconvenience is the best case scenario. Catastrophic failure is the worst. We need leaders who don't leverage the mindless insecurity and fear of the masses. We need folks who can lead for real. I think there's a leadership failure in some of this also. And also the inconvenience of replacing your 1990s hardware systems, all connected to the web, all sitting pretty for cyber hackers. There's a lot of embedded software out there. The cost of upgrade is so much lower now, though. It's inexcusable to not upgrade ancient garbage given how low the costs have become in just the last year. like if you're a year from today if you haven't patched your 1990 software it's inertia it's inertia people not you know people don't pay attention until it happens to them i do think it's a market opportunity at the same time so maybe rather than just lamenting purported failures of leadership let's all adopt our own purported moonshot mentality and call out to entrepreneurs go
build the next trillion dollar company solving all of these embedded hardware software issues and make embedded hardware and software that never succumbs to all of these upgrade-itis issues. Especially if you're semi-technical. You don't have to be super technical, but you're semi-technical, but you're very, very articulate and you have good social skills. One of the highest callings right now is go into J.P. Morgan, go into Goldman Sachs, go into Moderna, go into some of these big, very profitable companies and work on the where's the data problem, how do I connect AI to it, how can we do this securely. You know, a lot of that is just social engineering, trying to figure out what is the legacy system. And, you know, the people who wrote it are usually long gone. There's a lot of sleuthing to do. But that's kind of a missing ingredient right now. Dave, going to SMEs, there are, you know, hundreds of thousands of small and medium-sized businesses that need, you know, need their systems overhauled. Need someone to walk in and say, I'm going to redo your security. I'm going to rehaul your system.
and any business owner who doesn't take advantage of that offering is waiting to be hacked. Yeah, that's right. And you'd be surprised how big that void is. You'd think, well, hasn't somebody already tried and stepped in and had that conversation? But AI came upon us so much more quickly than the Internet that that community just hasn't developed quite yet. It's just so laggy. But if you go to any random company that you happen to know, go to anybody who works, any friend of yours who works at some midsize company and say, has a credible entity come to you with an AI upgrade plan? And nine times out of ten, they'll say, not a single meeting yet. Yeah. Like, okay, that's your opportunity. Step in there and fill that void. All right. I'm going to move us to our next story. For me, the big story of this week, in my personal opinion, was the success of Starship Flight 15. So on Saturday, McZilla caught a Starship for the very first time. Ship 42 lifted off on Friday evening from Pad 2 at Starbase, reached orbit, deployed 26 Starlink V3 satellites.
SpaceX said it made contact with every one of them. They were working. After a long coast, after 22 and a half hours, it deorbited, reentered, lit three Raptors on the landing flip, and dropped into the tower's arms and made history. You know, forever since I've been in the space business for the last 40 years, fully reusable spacecraft have been the golden, you know, the target. It changes everything. This is a big deal. Let's watch an epic video. This is a video of Starship landing filmed from the Maxilla Tower. All right. Take a look-see. Thank you.
Oh, man, I wish I was there. I couldn't make it. But watching the crowd, just the roar of excitement, it was like, you know, the great believers of humanity's future were celebrating. The super heavy booster aboard of the tower catch diverged to an offshore landing in the Gulf. Remember, on the first attempt of the super heavy capture, it worked. And also on the first attempt of the ship capture, it worked. On Sunday, Elon shares his vision for Starship V4, which he says is going to be 70 meters long, roughly 40% increase in length and 60% increase in propellant. Starship engine count will also increase to nine with three more high-thrust Raptor vacuum engines. I posted my own math of this on X. So if you look at a mature, reusable, refuel-and-go transport system like a car, a ship, a plane, fuel typically represents only 25% of the operating cost. So when Starship goes high volume, fully reusable, with propellant cost initially estimated $1 million to $2 million per launch, that implies a total cost of $4 to $8 million per flight or $27 to $53 per kilogram.
Just to remind you, the space shuttle was operating at about $54,000 per kilogram. This is a thousand-fold reduction in cost. What it means is that the average 150-pound American, which is how much I roughly weigh, can go to orbit for roughly $2,000 to $4,000. It's the same as a first-class airline ticket. It's epic. Hours later, Elon tightened the fuel number. He said propellant costs per flight will drop to $500,000 and payload will increase to 200 tons, which means a per person ticket price could drop to below $1,000 per person to orbit. I find that awesome. Let's take a look at a quick image of what Steve Jervison posted. and we're going to be doing a pod with Steve this coming week. So here you see the cost per kilogram on the left axis running from $100 to $51,000.
And on the bottom, this is the time frame. And you can see the improvements in the old style systems, which were used once and throw away. They were basically ICBMs. You pull the warhead off the top and you fly them were pretty flat. Then you saw Falcon 1, Falcon 9, Falcon 9 Heavy, and now Starship, and we see this plummeting. And this is a curve that shows genius, and it's extraordinary. Alex, over to you. Well, my first reaction I wrote on X, the solar system is cooked. And I meant it. I think this is – one can analogize this easily to so many important moments in human technological history. But what I think strikes me the hardest is the transition that computer networks made from circuit switching to packet switching. Some of the folks here may remember circuit switch days when telephones and computer networks used to actually have a physical, require a physical circuit in order to pass data.
And there were, at the time that this was happening in the 70s, 80s, and 90s, almost until 2000, there were wars over what the best way to packetize information would be so that information bits could be fully decoupled from atoms and you can send one bit of information in a way that's agnostic to its physical carrier. The first packet, first Internet packet, or I should say the first ARPANET packet was sent in the late 1960s. Some would suggest that is when the singularity happens and not now. But I think this is the packet switched moment for space in the sense that you remember how Jeff Bezos and Blue Origin were basically on paper laughing at Elon and SpaceX for requiring so many Starship flights in order to achieve a moon mission. versus more monolithic proposals. And I think the recovery, the successful recovery of the Starship upper stage,
Starship with these tweezers, with these chopsticks and Mechazilla, but in a reusable way, this, I think, underlines that we're going to have a, this is an analogy, we're going to have a packet-switched solar system where just like with cargo containers, with standardized shipping containers, you can packetize matter and ship it all over the world. The fact that it's now possible to do this with whatever you can cram into the current version of Starship or the next version of Starship, these are going to be individual packets, and we need to stop thinking about monolithic missions like the way we used to assign names to, say, the space shuttle back during the era when NASA dominated human spaceflight We moving past the named space mission era to an era when a space mission is going to be composed of hundreds or thousands of individual packetized flights
That then self-assemble where they are to meet the mission. Correct. That refuel on demand and the packets meet up in LEO and can refuel and they come and land. and it's no longer about the individual craft or rocket or engine or vehicle anymore. It's about just they're interchangeable, and they're all constantly going up and down and sideways. That's new. I think shipping containers is the perfect analogy. Yeah. Agreed. Salim or Dave, any thoughts? Well, just so you can remember, it's a dollar a pound. If you say, hey, this weighs about a pound. I want to put this in orbit. How much fuel do I need to get this thing into orbit? It's a dollar a pound. There's actually a used clothing store outside here in Cambridge called Dollar a Pound, so you can remember it that way. Dollar a Pound, that's where Yelan's going in that next bubble, which is only here. Salim? I think I literally get chills watching that thing come down and get caught.
It's just unbelievable. I look forward to the day that it becomes so routine that I don't get the chills. Yeah. That would be a cool day. Yeah. You know, the lore has it that when Elon sort of set out the Mechazilla strategy, because why are they doing that? So they can catch it, refuel it, remount it on a booster, and go again, theoretically in hours. You know, people thought he was insane. But it's a beautiful architecture. That's a great example of the future is faster than you think, which is somewhere up here. Yeah. the convergence of technologies. There are a couple of good examples in that book, but now what we're actually living in should rewrite the book, actually. But if you think about the process of landing, the process of landing this massive thing in exactly this tiny little landing zone, the confluence of technologies, including GPS and everything else, the feedback and control, how many other things can we invent now
that take advantage of that level of feedback and control precision There must be thousands of things that we're not even thinking of right now that are suddenly possible. I had a conversation this morning with the head of moonshot thinking at LILA, LILA Biosciences. And we're talking about the fact that you really have to delve into the area of crazy to try and really focus on moonshots. Because, you know, the Mechazilla mechanism is kind of insane. And, you know, we talk about this all the time. The day before something is truly a breakthrough, it's a crazy idea. If it's not a crazy idea, it wouldn't be a breakthrough. It would be an incremental improvement. So where inside your organization are you thinking about crazy ideas? And it's really hard. You know, we're so landlocked in our belief of what we can do, what is possible, and there's self-imposed limitations. And during this period of superintelligence, we need to literally shatter those expectations. And one of the best ways to do that is, you know, brainstorm with your favorite LLM. like, okay, what's a crazy idea that would be 10 times bigger?
And then go and pursue it. I read a really cool analysis of, you know, to land the way that Elon is landing, you've got to carry the fuel for the retro rocket. It's not a tiny amount of mass that you have to carry. Why don't we land like the space shuttle did, you know, just glide in, and then you have to have this huge, you know, undercarriage weight, but you don't have to have all that retro fuel. And the answer is like, yeah, dude, we're going to be landing on other planets that don't have runways. I want to be going everywhere. Like, what about that? Like, oh, yeah. I mean, one can steel man this also. And if you squint at the grid fins on the most recent Starship version, one could easily imagine a multimodal Starship 5, 6, 7 that doubles as what the space shuttle should have eventually been that can do maybe not necessarily even needing propulsive landings on Earth but is able to do propulsive landings on the moon or Mars where atmospheres are thinner or non-existent while still being able to do a glide.
And I could imagine now that SpaceX has developed AI religion and is using AI design that we eventually do get sort of the perfect unified synthesis of the original spatial concept and the Starship. Yeah. I'm excited. You know, I sort of consoled myself of not going to launch by saying we're going to be seeing thousands of these per year. So they're going to get old eventually. But man, oh, man, what a beautiful moment. Congratulations to everyone at SpaceX. All right, let's move to – Somebody put – Rick Chatz says self-docking technology has come a long way since the Roomba. well well said it's better than the jet bridges on most airplanes uh all right i'm gonna move us to the whole chips and china conversation uh so let's weave three stories together uh on chips in china one's diplomatic one's regulatory one's policy
all aimed at who gets access to build super intelligence so on november 15th at the pack Silica Founders Retreat in San Francisco, the U.S. and nine partners signed the San Francisco Declaration on the Secure Foundations of Enduring Prosperity. I love this name. It feels Asian in some way. I don't know why. The signatories are the U.S., Australia, India, Israel, Japan, the Republic of Korea, Qatar, Singapore, the UAE, and the U.K. They are treating economic security as national security. And I think that's a really critical point. There are five working groups that have been stood up covering policy, adoption, supply chain, safe harbor, and security. Then comes the hard edge. Politico reported on November 16th that Commerce Bureau of Industry and Standards has an interim final rule expected in the federal registry that would require licenses before selling or shipping advanced AI chips to Indonesia, Malaysia, Singapore, Thailand, and Brazil.
These were gateway countries for China. The aim is the remote compute backdoor being closed so that China doesn't get access to these advanced chips. Politico said it would hit companies including ByteDance and Alibaba. The third story, and this is one, Alex, that you flagged this morning, is out of the White House, reporting that the Trump administration is weighing a rule that would make U.S. companies tell you when they're using Chinese AI. Reuters reported that the Treasury Department and the White House Office of Science and Technology Policy, this is Michael Kratios' gig, have discussed requiring companies to post their use of Chinese AI models on their websites. Airbnb, for example, uses Alibaba's Qen to power its AI customer service agent. The logic is simple. Chinese open-weight models like Kimi, DeepSeek, and Qen are cheap and good enough. And the three most used models on Open Router in October of this year, they were all Chinese.
Hawks fear that this hooks U.S. companies on Chinese AI. Airbnb says that AI runs overwhelmingly on U.S. origin models, and Chinese models are used and run only through approved U.S. providers. So, I mean, my take is I'd rather see a label than a ban. You know, clear disclosure, I think, is important to have. We've heard conversations that, you know, people are stockpiling the open weight models because they're fearful they're going to be taken down. So, gents, Salim, you first. What do you think about this? I thought there was a couple of interesting things here. I thought that the way they've done the subgroups is very interesting because you've got these five working groups, which is a very interesting organizational design, right? It's like a network of countries trying to coordinate. I fear that it's a bit bolting the door after the horse has gone aspect to this, so there's not to thrash an old metaphor.
But I think this seems a little bit too little too late. And I'm concerned that a blanket approach is going to risk kind of harming some of the ecosystems we want to build. I'm particularly interested in what's happening with India where they're a signatory, but they desperately need compute. They have no compute of any kind. Indian startups are really, really struggling on this one. So it's going to be interesting to see how this plays out. Did you bring some GPUs in your suitcase to Thailand? Yeah, I've been carrying them along in my carry-on. Alex you know Paxilica this potential ban on use of opiate models thoughts the silicon curtain is descending it would seem so we see on the one hand the Paxilica and on the other hand the Pax Sinica which is in historiography as the term for periods of relative stability in East Asia that are China-centric. These two could not be more visibly in opposition at this point. This is like Cold War
2.0 with two basically carving up the world, or at least attempting to carve up the world into two different spheres of influence in the form of different chip stacks, different models, import-export restrictions. I'm even reminded the notion of labeling. It perhaps goes without saying that we label our physical goods in this country. They're actually required under interpretation of 1930s era statutes, those like made in China, made in Vietnam labels. Those don't get applied at the moment to digital goods like models. So I don't think it's a terrible idea. Just like we would say, you know, pick up a plate and say you look under it's made in China. You pick up a model, you look under and say this was trained in China. I don't think that's necessarily a terrible idea by itself. I'm also not sure it moves the needle that much. I'm not so convinced that enterprises or individuals, if they have to put up the equivalent of California's
Prop 65 warnings that, oh, we've used somewhere deep in our tech stack a Chinese-trained model, whether that will be any influence at all on consumption by users, I tend to doubt it. But I do think it's symptomatic of a broader issue. It goes back to the Jensen debacle, which is battle lines are being carved right now. And there is there is a move to to do the great decoupling between America and its allies on the one hand and China on the other. And to just to encircle them in some sense with our tech stack versus their tech stack, at least for the moment. And all of these, the Jensen drama and the made or trained in China labels and the San Francisco Declaration, all of these are just facets of the great game, which has been played throughout the centuries between all the great powers. Right now, the great game is fixated on superintelligence, and it's being played between largely the U.S. on the one hand and China on the other.
And a whole bunch of bystanders and enterprises are getting caught in between. Are we strengthening American leadership through these restrictions or incenting the rest of the world to develop alternative stacks? Of course we're incenting the world. This seems really dumb policy from my perspective. Because the open source and open weight models are so good right now, you're going to drive everybody to that model. We did this in the aerospace industry. We put export restrictions on U.S. satellites. It just meant that the European and Asian satellite manufacturers became more capable and competed with us. It happens over and over again across the board. I think the analogy, though, has a huge flaw in it. I think it all comes down to hardware and not – like the open source models don't matter a whit. And I'll tell you why. Because you can take one of these models and you can transfer learn from it. You can use the thought traces to create a new clone off of it. You can disassemble the layers now.
We're doing a lot of research with that in our quantization group. It's just it's this really amorphous product that it just doesn't matter. What matters is where are the freaking chip fabs and what hardware is this all running on? That's the only thing you can measure, control, contain. Trying to do anything related to the software is just a joke. You know, it's like you can ban it if you want. It doesn't matter. You can retrain off it so quickly anyway. You know, the Chinese proved that if you take a foundation model, even if you don't have the weights, you can use it to bootstrap your new model for one-tenth the cost. If you do have the weights, it's more like one-one-hundredth the cost. And that's getting better, you know, with AI assistants helping with that, that's getting better and better by the day. So software is not containable in the age of AI. Hardware and data centers are the only thing that are containable. And then you have to be careful on the hardware side, too, because the computing substrates are going to start shifting very quickly now. New physics will be discovered very quickly. Photonic computing will come online.
So you have to really race ahead. It's like, where is it going next? Where is it going next? It's where I'm really, really excited about potentially talking to the White House about this. You know, what is the three-year view of photonic computing? Because it's all fine to ban the current GPU way late. Like you said, the horse is way out of the barn. Let's ban the current GPU. But the new computing substrates, well, what about them? This won't even matter. in three years. What about the next thing? So what matters is regulation at the hardware level, not at the software level. Interesting. Alex, do you agree? Yeah, I mean, there's, I think, an implicit trade-off between short-term benefit versus long-term benefit of, call it, technical protectionism. In the short term, if you think that we're in a race to, I mean, I think we're already in the singularity, but call it a max Q event of the singularity that could happen next year in 2027, then if that's your mentality, then it might actually make strategic sense to try to throw sand in the gears of China,
even if it means in the long term it just spurs Huawei to deploy more of their chips to Chinese frontier labs, even if it's not a long-term deterrent or a slowdown for them. If we're only months away from a max Q moment of the singularity, then every minute might actually count in the strategic calculus. If, on the other hand, you think this is like years or decades away, then it doesn't matter at all. It's just bad policy. Alex, to play this out for those listening, it's because if you when you hit that max Q moment, any advantage you have is massively accelerated. Yeah, this is like, you know, to two racers on a I don't know, a marathon or something. pick your sports metaphor, and the finish line is potentially within sight, and now they're shoving each other, jostling, doing all sorts of things to throw each other off their game because the finish line is so close and every second counts. So what does that look like, by the way, in 27? What does it look like if we're, you know, is it hard takeoff as we've described it?
Surely you're not asking me to spoil the predictions episode that everyone's going to want to do for New Year's. It's going to be fun. Yes, we do our predictions episode every year. It's coming up soon, yeah. Alex is like 10 for 10, I think. This year, we should do a retrodiction, though, judging ourselves. I will say I feel pretty good about my prediction for 2026, and I have some spicy ones for next year. I've missed Optimus on Mars big time. Anyway. I don't even remember what I said, so we have to go back. We'll dig it up. We'll dig it up. I'm sure you guys have kept receipts somewhere. I'm going to go and edit the videos so I come out better. All right, our next story. Well, Starship was making history at Starbase. The same company, SpaceX AI, and again, full disclosure, it's my largest holding, was breaking records on the ground in data centers. So Dodge Designer highlighted a semi-analysis report that SpaceX had finished mini-hard,
part of the Colossus II campus in the Memphis area in about five months. A new site called Gun Range, just south of the Tennessee-Mississippi line, is targeted for a roughly four-month build. Semi-analysis says that most comparable projects are still fighting for an 8- to 12-month timeline. So one of the questions is why is Elon able to build things two to three times faster than everybody else? He's got projects like MacroHard is the largest existing hall in Colossus 2, about 700 megawatts of requested power. MiniHard is the new building next to it, about 450 megawatts. MacroHarder is the next expansion. You've got to remember, Elon's a teenage kid when he names his projects this way. Jesus Christ with the naming God. You know, the speed is not a one-off. Semi-Analysis says Colossus 1 retrofit put 150 megawatts online in 122 days, and a later Colossus 2 retrofit put 210 megawatts online in 91 days.
So this is extraordinary. So why so fast? Semi-Analysis points to a dense hardware culture. You know, Tesla builds modular electrical and cooling gear, so they're not stuck with vendor lines. You know, this is Elon's playbook. You know, verticalize, own every component of your system. And then construction steps run in parallel. Those build times and the 2.4 gigawatt figure that are semi-analysis estimates, you know, end of the day, it's pretty extraordinary the speed that he's building. Salim, four-month Greenfield data center. If they hit it, it's a new tempo. What happens to everyone else when the build time, you know, is moving faster? I mean, you know, what's your point for? Well, this kind of speed, yeah, this kind of speed creates such a massive, like an exponential learning advantage, right? Because every time, every cycle, you're seeing things that other people aren't even getting to, and therefore you're able to create learning loops around this.
I mean, this is the organizational similarity made tangible, right? This competitive advantage is not from better technologies coming from the organizational velocity he's able to achieve, which is now into physical. and that's kind of an amazing comment to make. This is such a huge vector he's on where he can take any domain with hardware in it and just completely change the cycle time. Now, let's see the machines that builds the machines, and this is another argument for the SpaceX-Tesla merger, isn't it? It gets better. One of his data center sites is being named Spam, spam and it was the warehouse where Hormel was literally manufacturing spam products. We've come full circle in this economy from a warehouse from a spam factory to a spam factory. It's funny you say that because last night Rob Fisher texted and said, hey, we can get half a gigawatt of power
over, sorry, half a megawatt of power over at the yogurt factory. They've abandoned that space. So I might grab that, actually. We can put about 500 GPUs over there, but it would be the yogurt equivalent of the spam factory. When we first came from India to Canada, my mother gave me lunch of fried spam because she read somewhere it was really good. And I was like, what are you reading? It is good. It was terrible. What are you doing to your kids? What are you doing to your kids? I want to say something about beating swords into plowshares as applied to yogurt and spam. It's an interesting set of rezonings that we're going through. Oh, my God. Culture GPUs. Oh, my God. Okay. I think one thing that's missing from the story here is that AI is the greatest management tool in the history of all management.
And one of the challenges, actually, I put in front of our team is for our interns, when they come in, have them buy a piece of land somewhere in the country and build a house on it and then sell it for a profit. And we'll split the game 50-50 with them. And they go, these are like high school seniors. What do they know about building a house? I'm like, exactly. This will teach them to use AI. AI knows everything about how to build a house. And, you know, what happened historically is you go talk to a contractor. The contractor tells you a million things you can't do. Now you go to AI and you say, why can't I do this concurrently? Why can't the framing and the flooring be done at the same time? Why can't the foundation be done sooner? And the AI just gives you a straight answer. And you use that to manage the construction. Also, the other really underappreciated thing is video capture AI. If you're building something in the physical world, every action by everybody working on it can be very easily translated from video feed into event or action. Like, what is this person doing? right now. And that is the greatest management tool in the world.
It's usually underutilized. One of the coolest things I've seen is they're taking robots onto a building site and they're painting where the pillars and stuff will go from the drawing so that they have absolute millimeter precision as to where they're making their cuts and things. It's incredible to see the progress that's happening around all that. And you're going to see this accelerate I don't think 10x is going to be possible but 2 or 3x when you put robots in the learning loop. where the AI is simply asking the robot to do the thing, and the robot is giving feedback, and it's just optimized in that fashion. Jacques, Peter, I think you're thinking too linearly. 10x should be minimum in what's going to come from this. Maybe. Because it's all the systemic compounding of everything. Yeah, we saw Sonny Altman talk about this in terms of robotics building next data centers, But I guess OpenAI also shut down their robotics lab, didn't they? Well, they shut down their team dedicated to a dedicated robotics model because Astra, GP6.x, is such a good generalist model, passes Driving Bench.
You don't need a specialist robotics team or model anymore. I think they also shut down their humanoid robot effort. It will be interesting to see. I think it was actually their team focused on dedicated robotics. Now it's part of a generalist effort. And Anthropic has a generalist effort, too. They're all competing against each other to see which of their mainline generalist models can do really well at all these robotic manipulation tasks. Speaking about generalist models, yesterday OpenAI released GPT 6.2 SOL and GPT 6.2 Astra. OpenAI says Astra is its first release model to meet the, quote, high threshold for AI self-improvement. Altman posted, quote, these models can do amazing things and are very token efficient. The Frontier Lab is now formally rating its own product as a meaningful accelerant for AI research. Alex, over to you. What do you think of GPT 6.2? Yeah, my preference is still Opus 5.5 and Fable 5.2.
But I am very glad that OpenAI, at least according to Noam Brown from OpenAI, is now releasing at least one of the previously internal models responsible for their major progress, a.k.a. slop drop in math. That is a very good sign. There are a number of leaderboards out there that track AI for math capabilities until this new model, until the 6.2 series was publicly available. The number one model, Peter, this will make you and probably Google happy, was actually for a few days. It was Gemini for Carbon, I think. And that is no longer at the top of the applicable math leaderboard. It seems like progress for math. It seems like progress for the benchmarks that OpenAI cares about. We talked on the pod about LifeVal. It seems like progress for doing well at a given price point on the ARC AGI 3 challenge.
On the other hand, there are a whole bunch of people complaining that the model just smells wrong, that it gets itself caught in doom loops if asked to perform certain activities. I would say anecdotally, and again, the model is newly out, so take this with a huge grain of salt, But anecdotally, my go-to is still Opus 5.5 and Fable 5.2 for their overall balance. It's incredible how long Opus 5.5 has held that crowning reign. It's been many months. It's been all of like, what, a month and a half now. That's a long time. We get bored so quickly. My goodness, it's ancient. Long in the tooth. That's funny. Opus TV. I mean, Alex, sorry, not Opus. Alex, do you – What's the difference? Opus 5.5 is about a quarter of the price of Fable 5.2. Is it better at anything, though?
I've moved everything over to Fable 5.2, but I'm paying through the nose. But is Opus actually better at anything? I mean, it depends on the task. So if you look at the broader leaderboards, then you'll conclude Fable 5.2 is marginally better than Opus 5.5 at a much higher price point. So if what you want is fast reactions, tight feedback loops, and or lower cost, at the moment, for everything that I care about, Opus 5.5 is just chef's kiss. It's an amazing – it's a good model, sir, as the popular expression goes. And it would take a meaningful step up from Opus 5.5 right now to get me to switch to something else. Interesting. Because with 5.2, you're right, it's slow as shit. But then you can put it in fast mode for only twice the price again. It's actually pretty responsive, but it's like, wow, this is getting to insane price points. Yeah.
There are upsides and downsides to each. And you can always pay from time to time. I'll pay for the privilege of just using what according to the frontier charts, regardless of the horizontal axis, what according to the vertical axis is the best model. But in terms of overall balance of responsiveness and big model smell and capabilities across a broad range of tasks, Opus 5.5 right now for me is the utter sweet spot. It is an amazing model. And then after 6.2, they brought in at about half the price. So over slide five, is that right? Yeah, I haven't looked at the direct price comparisons, but they're both – last I checked on the AAII frontier chart, they're both on the frontier. It depends also on what reasoning effort you set for each. And OpenAI has not been, in my view, hot take, especially good about when they present their comparisons against anthropics models about making totally clear what reasoning level effort is being set.
Well, we're going to need to get much more scientific about this topic soon, only because right now AI is so mind-blowingly capable and cheap. It's just like this is just a miracle. Everybody's overjoyed with the miracle that it exists. But very soon people will want to deploy hundreds or thousands of concurrent agents, and then they'll say, oh, wait a minute. Now I've got to be thinking, like, if it's half the price but one notch lower, is that a good buy or is that not a good buy? And, you know, am I routing the tasks to the right models? And all that complexity right now is just kind of a zoo. Like I joke, if only there were a company, let's call it hypothetically Open Router, and if only there were a popular way to route and benchmark on a price performance basis across a generic API surface, if only. If only. All right, I'm going to move us to our final story, and it's a fun one. This is part industrial moonshot, part floating city, in other words, seasteading. Yesterday, friend of the pod, Palmer Lucky, posted this render, showing in the slide right now,
a massive concrete block in the open water with smokestacks, towers, and gardens on top. The seasteading company is called Slab, and it's designed with a thousand-year design life. Love it. The goal was to be cheaper than land and absent quote absent state and local laws that cripple critical American industries including metal processing medical research and treatment desalination oil refining, and chemical engineering. He and a few friends, including Joe Lonsdale, wrote the check, cashed it up to a billion-dollar valuation back in 2017. He says there's been significant progress, but there's a long way to go. Reviving manifest destiny and westward expansion. I love that. Having new lands to go conquer has become difficult. You need an army, but building new lands is something that is now possible. He calls it a project designed to make America stronger. Quote, a relief valve for ill-considered or outdated rules and regulations.
Not a way to escape the United States governance. He's not named a location yet, a size, a budget, a builder, or a date. Sitting outside state and local rules would not automatically put a platform outside of federal and maritime law. Seasteading is not new. Alex, you and I have talked about this. You know, it's been something we both have written about and dreamed about. The Sea Setting Institute was founded back in 2008 by Patrick Friedman and Wayne Gramlich with early backing from, of course, Peter Thiel. So, gents, you know, a fun story, a thousand year life. You know, I was flying back with Palmer on his plane from when he won the Wildfire X Prize. And he was telling me about this. You know, there are many South Pacific plateaus where the water is only, you know, hundreds of feet deep. You can build pilings and you can stabilize this platform out there.
You basically build it in the large ports in South and Central America or in the U.S. and then you float it out and then you stabilize it and you've got, you know, a launch port or you've got a, you know, regulatory. I love this longevity research facility where if you want, you're the wealthy or national leaders. You can go there and get access to experimental treatments that are not allowed in the U.S. yet. Alex, you've been tracking this for a while. Your thoughts? Yeah, I was a big supporter of the Seasteading Institute back in the Pat Free Friedman early days. Joe Lonsdale was involved around that time as well, and then more recently under Joe Quirk. I think we made this point on the plot in the past. I think humanity thrives when we have frontiers, and I think we're starved of frontiers right now. The Internet, not so much a frontier these days. Superintelligence, arguably a frontier. Space, arguably still a frontier. The oceans could be a frontier if we get our act together and start building seasteads and ocean cities,
which is why I'm generally in favor of it and doing what I can to support it. But I think we need more frontiers. And I think there is an unlock now that we didn't have maybe in the early days of the Seasteading Institute, which is Starlink. We now have the ability to get high bandwidth to almost anywhere of interest on the Earth. That simply wasn't an early option in the early days of the Seasteading movement. It is now. There's another factor, which is data centers not seeming to be very popular terrestrially. So the same impetus that could and seems at least on a trajectory to drive data centers to orbit and build out our Dyson swarm could potentially also be the missing driver, the missing killer app for seasteading, namely ocean-borne data centers. And those data centers could become the nuclei for ocean cities or seasteads. I think this is within the realm of possibility. So I would say there were, in summary, two missing factors for why now and why we didn't get the dream of ocean colonization decades ago.
One was on the supply side, that you need a way to connect it up to the broader economy. Starlink solves that. And two was on the demand side. You need an actual economically valuable purpose for being out there other than just like oil drilling or something equivalent. And now maybe if data centers are persona non grata, grata, excuse me, on land, keep my Latin correct, then maybe oceanic data centers will be the demand side killer app and we can finally get ocean colonization. He makes the point in his in his ex post that this was pre data centers. That was not the driver here. It was simply, you know, it was liberation. um salim your thoughts two things i think uh i love the moonshot i mean it's great to have more than just elon doing completely and utterly crazy things it's it's we need a thousand people like this thinking we're the only limit is your imagination right i think that's incredible
it's great also it's been going on for a while uh but for me the interesting thing is not about putting like a factory in the ocean, but you're trading an operating environment for industrial innovation that is the product. Like there's a whole foundation there of really interesting futures. And I think that's really cool. So I'm really, really looking forward to seeing what happens here. I'm not really a big fan of sea setting. I feel like we evolved for billions of years to get out of the ocean, why the hell are you going back in? But industrial – You're not living in the ocean, Salim. You're living in – I know, I know. What do you have against the ocean, Salim? But if we're building stuff on the ocean, I'm a huge fan of ocean-based data centers and so on. The sea is just really corrosive and it's really hard to navigate. You know, all those world islands in the UAE never worked out. Erosion is a very tough model to beat there. But we've been running oil rigs for decades perfectly fine.
So in theory, this should work fine. And there's an infinite amount of space. You want abundance. There you go. There you have it. Yeah. You know, I think if this works, we're going to see hundreds of these, not just one. It's two-thirds of the Earth's surface, and we're just wasting it, right? And we haven't even properly mapped it, let alone colonized it. And it seems now is – It's a long way to go. Agree. Yeah, now is as good a time as any to speak of manifest destiny in the U.S., but there is a certain sense, I think, in which the oceans are at ocean level and subsurface, I should add, as well. The oceans are a final frontier that we really should be aggressively developing. They have energy. They have cooling. They have lots of different attributes we can use. They have space. They have mystery. They have – there are many, many opportunities there, I think. For sure. All right, gentlemen, you up for some questions? Absolutely.
All right. Let's go to AMA with the mates. Okay. Let's see here. Salim, you get first choice. Oh, I've got to go with number one there. Okay. Let me make this bigger so I can read it. when agents work for weeks, which business functions shift first from supervising tasks to supervising outcomes? And that's from Solly D7809. You know, I think there's going to be huge batches of business functions that shift. I think first would be customer service and customer operations. The outcomes are very measurable and it's bounded. So you could deal with resolution rates and customer satisfaction. all sorts of retention measurements. And so there's a huge opportunity there. But the opportunity, the shift will go way past just customer service. And in this whole organizational singularity,
we've been talking about coordination activities, which were typically done by human beings, now gets done by AI agents. And doing that, you could coordinate for a long time very effectively. By the way, just to put a placeholder here, we're just finishing up the pilot that we've been running for three, four months. The results that are coming in are unbelievable. The first company we've looked at to see where they ended up, they took a process that took 20 days, and they're now doing it in 38 minutes. So I'm really – I can't believe how excited I'm feeling about what's coming out of that. So I'll report – I hope you'll share that with us, Mike. Yeah, we'll get much more data. We're just aggregating it all right now. This is incredibly exciting to see. But I think the organizational redesign is a massive shift towards how we're going to use agents agents to run all of our organizations, our government departments, our impact projects, et cetera. And all the orchestration and harnesses are just about there for that. So super exciting. Nice. If you guys don't mind, I take number two. What should education look like in a world
where preparing for a job is no longer its main purpose? And that's from at Stephan-SWE. Stephan, you know, I've talked about this extensively and it's my next chapter of my life, reinventing education here. I think what it looks like is using education for a few main things. Number one, shaping your mindset, right? The way you see the world is critically important. You know, your mindset, the way your neural net works as a kid or as an adult is how you deal with opportunities and threats. What's your immediate reaction? How do you parse that information? Number two, helping, you know, somebody through educational process find a purpose in life some you know a passion they love that helps other people find a purpose because when you've given super capable you know superpowers from ai and robotics where do you point it you know what do you love what do you wake up every day with a hunger in your heart that you want to go after most kids and most adults don't know their purpose my favorite quote from mark twain
is there are two important days in your life the day you were born the day you find out why So helping that kid find their purpose is a massive unlock. Education should be about how to use the top AI tools and then learning entrepreneurship. And instead of getting a job, which is giving agency to someone else to give you a means of living, find your purpose. Find something you want to solve. Use entrepreneurial schools. Use the latest AI models and build a solution. Become an entrepreneur. And is entrepreneurship for everybody? probably not for everybody, but it's for a much larger percentage of the population than currently are using it because it's become possible. So I think that's what education becomes, giving kids agency to control their own future. And, you know, the more entrepreneurs on the planet, an entrepreneur is someone who finds a problem and solves a problem, the better the world is. So that's my rant on the future of education. Sure, sure. Okay. Dave, over to you.
All right, I get number three. Could one AI directly adjust another AI's weights while teaching a robot a new skill? From FrankNeath2148. Yes, absolutely. I'm super excited about that. Thank you for asking. There's a new paper, actually, out of MIT CSAIL. It's called Weight Space Programming. So you're adjusting the parameters directly with a neural net. So one neural net is putting thoughts right into the brain of the other neural net. There are also a lot of stepping sites. It's a very hard thing to do, but there's a lot of stepping stones along the way. And so I might roll this into a prediction for next year, but the progress on this is going to be incredible. I know that already, actually, just from our internal work here. And you can do it the hard way by trying to adjust individual parameters, but you can also do it modularly by changing just the attention side or by adding entire new expert modules. And that actually is an easier stepping stone to what you're talking about here. But this is going to be a huge breakthrough for AI because what we've done historically in AI is retrain from scratch every single time, go back to bare metal, build the whole thing over again.
And nobody ever cared about better ways to transfer knowledge until the training costs got really big. Now it's $100 million to train a frontier model from scratch. So all of a sudden it's worth recycling. And so there's going to be huge progress on this. But the short answer is yes, it's 100% possible. tons of research going on in that area right now. Super exciting. All right. Number four is for you, Alex. Number four asks, where is AI making real progress against viruses from long COVID to HIV? And this is from TRP C8N. I think it's pretty obvious. Moderna and OpenAI announced a number of years ago that they were, well, the partnership has been in place for, I think, three years. They announced more recently that it even existed. It is, I think, borderline obvious at this point that all of the major frontier labs are training off of genomic, probably multiomic and proteomic data as well. There's quite a large body of evidence. Folks go into the frontier models from anthropic and open AI, prompting them
with genome sequences. I wouldn't necessarily recommend prompting them with viral genome sequences, but prompting them with genome sequences and measuring the perplexity of the extrapolations. And it's the same mechanism that the authors and their guilds when they're suing the frontier labs, arguing that they've violated copyright laws for training off of their books. Same concept. You prompt off of known text, in this case, biological sequences, and you find that they're able to predict the remainder. They've been trained off of real biology. And so to the broader question of how is AI making real progress against viruses, I mean, I could flippantly cite a bunch of studies in the literature of particular point solutions to particular antivirals. But I think actually the better answer is much broader. And so I'll open parentheses just for Saleem. Mechanize, which is a spinoff, quasi spinoff of Epic AI folks, when they first started up, they posted a really interesting essay arguing that for some problems, not necessarily specifically antivirals, but for some problems, including problems in medicine,
it's almost easier for the broad capacity of civilization to increase versus narrow capability increases. And I do think that is likely to be in substantial part the case for AI progress against viruses. Yes, you can do point solutions for particular viruses, but I tend to think the solve everything mentality of increasing our civilizational capacity by training off of basically all biology, all biological sequences anywhere, building those into our generalist frontier models to the point where the models speak biology and think biology is actually where most of the net long-term progress against virus is going to come from. In the short term, it's probably Moderna using chat GPT tools to speed line all of their regulatory approvals and other paperwork. There is, you know, to give a A quick citation, there is an AI-discovered broad-spectrum antiviral that's called MDL001 that looks strong in animal models right now.
And so I think this will come, and Moderna is very much focused on this area, as you so noted. All right, let's move on. Number five, six, seven, and eight. Salim, over to you. I think I've got to take number seven. What should an AI entrepreneur in India build when local infrastructure lags Silicon Valley? And that's from Karan Jain091. So, you know, the obvious answer here is to you don't try and compete with Silicon Valley because you don't have the capital to build large foundational models, and you've got huge expertise, dedicated researchers to do this. But what you can do is focus on applications and local problems where India has a massive advantage, right? So consider you help a small business operate, use a voice-only agent to operate in multiple Indian languages.
You'll suddenly expand your business in the scope because people forget India has 20 different major languages with different scripts for each language. It's a complete patchwork of different cultures, et cetera, et cetera. It's more like Europe than the U.S. And so the opportunities to manage that complexity with AI are absolutely huge. So I would just as an example build a voice-first agent that would help businesses interface with different customers, different languages. That would create unbelievable economic value. There's already huge opportunities in taking very specialized workflows, training with your hands and mapping that, et cetera. So I think there's incredible opportunities. You don't need to copy Silicon Valley. You can build on it because the money is usually made in the application layer, not the infrastructure layer or the model layer over time. Great. Alex? Can I do number six? I was going to grab that, but go ahead.
Go ahead. It's fine. Off-brand, but I can't resist. So number six asks, what happens to democracy if longevity breakthroughs let dictators stay in power for centuries? And this is from William L. Ramsey. So I'm sure you guys have all seen the Charlie Chaplin 1940 movie, The Great Dictator, where Charlie Chaplin is pretty bold for the time. Basically, it's a parody of Hitler. and at the end of the speech, Charlie Chaplin's character delivers a pretty, I think, moving speech that includes the line, and so long as men die, liberty will never perish. And this question, I think, speaks to that Charlie Chaplin line, which is, if people can live forever, then dictators, human dictators, can potentially live forever as well. But I don't think actually that concern will stand the test of time, even though it was an inspiring line in the Charlie Chaplin movie.
I don't think it will stand the test of time. Why? Because people as late as the early 1990s used to talk about software as immortal. And what actually happened is software itself, which doesn't have all of the necessarily biological frailties of human meat bodies, itself is very much not immortal. Software has vulnerabilities and bugs, and it operates in a complex ecosystem. And so software for it to survive has to be patched. It has to continue to evolve. Similarly with humans, I'm less concerned about a god emperor in the Frank Herbert style, maintaining power for centuries, because I think the actual real world universe that we live in is so complicated that living for centuries is probably going to necessitate transforming oneself, reinventing oneself, if you will, at every level of the sack, in the same way as, say,
if you wanted a piece of software to operate for decades successfully without falling over, that software is going to, ship a Theseus style, need to be patched beyond recognition. Same too for any would-be dictators. You know, to steel man this, you know, I've guys remember the hot mic moment. I think it was last year when President Xi and Putin were discussing organ transplants and Xi saying, you know, some predict that they could live to 150 years old. So there's no question that these individuals are looking towards longevity to maintain themselves. But I think at the same time, you know, AI or SI is empowering the public. and, you know, dictators only stay in power so long as they're able to oppress their people. And I think the people are becoming more powerful than ever before. Dave, you want to take number five? That looks like a good one for you. Sure, I can take five. Could startups trade control of their businesses for access to a lab's most capable internal models?
And that's from Wandering Joy. Yeah, this is actually a frustration of mine. OpenAI already has a program for that, and Sam Altman is great as an entrepreneur who's turned into a Foundation Lab CEO. He gets it. So he's the only one I know of that has an active program. It's relatively new. But inevitably, this has to be done by all the labs. Rather than saying everyone's going to lose their job, why don't you say here's the roadmap of the future where everyone can contribute, and here are the rules. and that's exactly what this question is calling for. So I really hope Anthropic gets its act together and does a program like this where you know they're not going to give their best models to someone who could use it for an evil, nefarious purpose. But an application program where you're vetted and you're let under the hood and you don't have to give your control of your company to them, you just have to share the equity upside, I think that's an inevitably great outcome from the foundation models and then labs, but OpenAI is already there.
Google is not, as far as I know. Anthropoc is not, but they should be. So, yes, I think that's a very likely outcome. All right, number eight, can robotics finally make recycling work at the scale our waste problem demands? And that's from at MJBitcoin. And I think the answer is yes. It's not just robotics. It's exponential technologies. I can't wait for robots to be cleaning up the beaches here in Santa Monica, going along the 10 and the 405 freeway and picking up all the garbage. I mean, it's pretty incredible. But, you know, one of the things, we had an X-Prize for this years ago that never made it. I still love the idea. You know, a dump or a, you know, a waste management site filled with everybody's crap is resource refined elements and capabilities. So I always thought it'd be interesting to have a machine that basically goes into a dump. It grinds everything up into a slurry. It spits it out.
And then, you know, electron beam charges the particles as they come out. And it sorts it into organic matter, into copper, into silver, into gold, into plastics. And basically take a waste dump and make it a resource refinement facility. I think that is something which is very capable in the future and in some way for us to recycle the world that we are consuming and tossing out in our trash. It's funny, Peter, out in California where you are, everything is always three bins. All the restaurants and everything have three bins. You've got compost, recycling, and trash. And then you look into the bins, and everything is in every bin. Like the compost, you could never possibly compost it. It's got plastic in it. It's got it. It's just, you know, everything's everywhere. But it'd be so easy to build an AI capability that sorts after the fact anyway now.
So, you know, it's got to happen. MIT does have that, though, Dave. You know, walk around the MIT campus, MIT trash bins, the three stream recycling, have little cameras with an AI attached to them that will offer you the option of looking with computer vision at whatever you have and recommending a stream. Yeah, that's pretty cool. or catching you if you put organic matter into the recycled bin. Big brother. Well, you know, if you make it social stigma, that's no fun. Just sort it after the facts. But the microplastics issue is a major crisis, actually. I didn't realize how bad it was, but a very good... Seven grams of plastic in your brain, a credit card worth in the average human brain. It is crazy. Jesus. All right, everybody, please send us your... Wait, wait, before you get there. Yeah, please. Before you get there, I've got two or three quick quotes I want to give from the comments. Yeah. Mr. Duckett's 5818 says, need AI to tell me why I'm so addicted to this show. LOL. Appreciate the work you put in, fellas. Then B. Greer 4752 says, I always feel fired up after an episode, as do we, Mr. Greer.
Chris Bell says, I want Peter to live longer. He can steer human agency in space. You got this. Thank you. I'm working on it. So we're all working on it. And by the way, we have an epic episode coming out with David Sinclair on the future of longevity. So that's coming out. Oh, my God, it's crazy. It was so good. It was so good. All right. Please send us your outro videos. Here is one from our dear friend CJ Trueheart. CJ, thank you for posting. Everybody else, please catch up. We want your outro videos. Send them to us. You can send them to media at diamandis.com. All right, everybody, enjoy.
while we're here to say the future starts today. Keep your mind from the moment and your moonshot in your mind. When you figure out your gift, let go and let it shine. So why not take the shot? Put yourself on the spot. Wanna be a moonshot late? Then you gotta be moonshot Monday. Moonshot Monday, moonshot Monday. In due time it all comes to pass Nothing here on earth will forever come to last So what legacy will you craft? What impossibility will you make? Your unique ability This is bigger than disruption Bigger than the next I feel This is human evolution This is how far we can go When we aim beyond the ceiling Beyond what's been done before
Every moonshot that we're taking Opens up another door For the kid who's watching Somewhere thinking dreams are just pretend We're the proof that crazy ideas Are just where breakthroughs begin Keep your mind on the moment And your moonshot burning bright When you harness your gifts You become the speed of life So go and take that shot Put in everything you've got, cause if you wanna change the world, maybe you gotta be moonshot minded Moonshot minded, we are moonshot minded We are moonshot minded Moonshot minded, yeah, yeah Moonshot minded From this moment forward, you're not the same as before You've got a moonshot call, and what are you waiting for? The impossible is possible, and minds like ours unite.
All right, moonshot-minded. Gentlemen, as always, a great one. Salim, safe travels home, pal. Thank you very much. I'm looking forward to getting out of this hotel room bubble that I'm in. Are you going to work the whole way home, or are you going to get a sleep? I'll probably do both. because once you have internet on a flight with AI, you can kind of start doing all sorts of stuff over time. You know, you have time to kind of really get into it. So I'm looking forward to that. I've got a bunch of episodes of For All Mankind to catch up on and the Rings of Power Season 3 dropped. So I want to track that out. I have no time to watch anything or read anything these days because I spend all my goddamn spare time researching these episodes or being with you got a lunatic. We love you for it. Alex, thank you for your genius. As always, grateful for you. Happy to be here. Thank you, Peter. And Dave, same, right back at you. All right, gentlemen.
Until our next episode, it's going to be with Steve Jurvetson. That'll be a fun one. That will be awesome. Yeah. Yeah. I get to bring up my parallel universe question. And I'm going to talk to him about, is intellectual property still a thing? Should you file patents? He said no last time. This was like 10 years ago. Let's see what he believes today. All right. Until next time. All right. Take care.
番組の概要欄(原文)
The mates sit down live to discuss the latest developments in AI, technology, and innovation, from cybersecurity challenges and SpaceX breakthroughs to the global AI race and the future of industry. Get access to metatrends 10+ years before anyone else - https://qr.diamandis.com/metatrends Peter H. Diamandis, MD, is the Founder of XPRIZE, Singularity University, ZeroG, and A360 Salim Ismail is the founder of Open ExO, a GP at Exponential Venture Capital/The Organizational Singularity Fund and a sought after global speaker and thought leader. Dave Blundin is the founder & GP of Link Ventures Dr. Alexander Wissner-Gross is a computer scientist and founder of Reified – This episode is brought to you by: Get the blueprint for generative media https://goo.gle/startupgenmedia Go to Blitzy to book a free demo and start building today: https://qr.diamandis.com/blitzy Your body is incredibly good at hiding disease. Schedule a call with Fountain Life to add healthy decades to your life, and to learn more about their Memberships: https://www.fountainlife.com/peter _ Connect with Peter: X Instagram Substack Website Xprize A360 Connect with Dave: Web X LinkedIn Instagram TikTok Connect with Salim: LinkedIn X Join Salim’s 10X Shift Subscribe to Salim’s YouTube channel Exponential Venture Capital Connect with Alex Website LinkedIn X Email Substack Spotify Threads Listen to MOONSHOTS: Apple YouTube Follow MOONSHOTS: Instagram TikTok X Threads – *Recorded on November 18th, 2026 *The views expressed by me and all guests are personal opinions and do not constitute Financial, Medical, or Legal advice. Learn more about your ad choices. Visit megaphone.fm/adchoices
出典・この記事について
- 番組
- Moonshots with Peter Diamandis(PHD Ventures)
- 元の回
- Anthropic’s $1.4T IPO: The Biggest Wealth Event in Human History?
- 配信日
- 要約作成日
- 公式サイト
- www.diamandis.com/podcast
この要約・解説記事は、元の音声を文字起こしした内容をもとに AI(Claude)が日本語で作成したものです。 固有名詞や数値に誤りを含む場合があるため、正確な内容は元の音声でご確認ください。制作方法について
関連エピソード

なぜ全部が壊れるのか? Anthropic の1.5兆ドル IPO も解説
Uncanny Valley | WIRED

Nvidiaの危機はAI崩壊の前触れか ― Ed Zitronが語るジェンセン・フアン退任と密輸疑惑
Tech Won't Save Us

Jensen Huang なきNvidia:復帰の意向と取締役会の拒否、Anthropic IPO など
AI Inside

ジェンスン・フアンがNvidia CEOを辞任、AnthropicのIPO、AI 2027の前倒し予測、HubSpotのAI再創業
The Artificial Intelligence Show

AI業界の行方を左右する5つの問い:Anthropic IPO、OpenAI新モデル、ジェンスン・フアンの去就
The AI Daily Brief: Artificial Intelligence News and Analysis

Figure AI はなぜ中国・Tesla に勝てるのか:Brett Adcock とロボティクス徹底討論
Moonshots with Peter Diamandis