What he is here for: the money chair on Moonshots with Peter Diamandis — first responder on chips, markets and who is up, and the panelist whose reservoir is the board room.
The most prescriptive mate on the panel. He is the one who tells you to go and start doing something. The count on this archive supports it: the week-1 map holds 24 of his prescriptions from five episodes, and this page adds more than twenty dated ones from EP288 to EP293. Salim Ismail's page, the next most prescriptive, holds fewer than ten.
Why he is in this archive
Gate: present in all five episodes read. Introduced every week as "our wizard / empresario of AI investing".
Positions and disclosures
| Fund | founder and GP, Link Ventures; also Link Studio. "With so many of our portfolio companies, I'll meet with them weekly… No losers yet. Zero." (EP284) |
| The pod | Link Ventures sponsors the show and led Blitzy's rounds; Blitzy is a sponsor whose CEO sat on the panel in EP254 |
| Boards | "I'm the chairman of a very large insurance company, public company" (Everquote, per EP254) · MicroStrategy · Vestmark — see below, he is its founder and controlling shareholder, not merely a director · Trust & Will · adviser to Hibernia |
| ★ Liquidity | Vestmark is being acquired by Envestnet (a Bain Capital portfolio company), announced 2026-09-09. "Vesmark's a company I founded in 2001 right after 911 actually and was CEO of for the first six years. Now I'm still executive chairman and controlling shareholder." (EP288 0:04:38) |
| Hands-on | "I just signed off on 100K of swarm agents. We're running our 5,000 Kimis again today." · team switched to OpenAI's Soul · installs Hermes natively (EP283, EP255) |
| Interviews | Sam Altman, Palmer Luckey, Eric Schmidt regularly; the 3-hour Elon episode with Peter |
Where his information comes from
- Board rooms and portfolio meetings: "we had our board meeting at Vestmark and we were talking about this quite a bit."
- Insiders, flagged as such: "I heard that from a Microsoft insider, very close friend." "I was just talking to a senior exec from Intel… so I happen to know the answer."
- Inbound deal flow: "I got calls from both Mercor and from Orin."
- His own installs. The product demo is his evidence.
Main ideas
- Everything reduces to two things. "AI training and AI inference." Inference is leaving NVIDIA; CUDA's moat is training; "the interconnect is the new moat."
- The circular AI economy — with the Alpha Centauri analogy: a civilisation that trades only with itself and still gets rich.
- You have to own something. "W2 income is going to be a rounding error… trade that latte in for some chips-and-dip stock."
- Calling the ball. "This is the future that we're going to live in forever hereafter."
- RSI is imminent and it is at Anthropic. "I've seen God. I cannot go back."
- Hardware is hard; Taiwan is the hinge. "It all hinges on that little island."
- The Magnificent 7 all verticalise — "all 11 of the Magnum Mobster companies are building AI chips and models and data centers."
★ The prescriptions — what he actually tells you to do
Full map → research/panelists/blundin-prescription-map.html (24 prescriptions, each with
verbatim quote + timestamp, tiered money / career / operate / avoid). Built 2026-09-13 as week 1
of panelist deep dive protocol. Covers five of the last seven aired episodes; #289 (the AMA)
is now decomposed — the research archive — and its
prescriptions are listed below. #288 (09-11) is decomposed as of 2026-09-23 —
the research archive — and folded in below.
He states his own frame (EP284 1:00:32): "what do I do career-wise? What do I invest in?
What are the scarce constraints of the future? How do I interact with these 11 magnum mobstick
companies?" Use his four questions, not an imposed one.
The five he repeats hardest:
- Get into the data-centre build-out, not AI apps. "Every one of those companies is creating
billionaires out of its founders… the people who cross that chasm and get into this data center
buildout are crushing it." (EP285
1:18:56) — against his vertical-AI book, which is "a mixed bag." - Go through the front door of the 11. "Why are you meeting with anyone other than Jensen…?
If you can't get Jensen, just get one of his E-staff." (EP284
8:11) - Buy data-gathering companies. "The age of data starvation is about to hit us… architecture
and drug design, completely data starved." (EP286
0:21:44) - The life plan: be great at building data centres / cooling, then reinvest the earnings into
the AI companies — "a much better plan than aspiring to get a degree in accounting right now."
(EP287
1:53:09) · and "take the job, take the money, stay nimble" even though the job is cooked in two years (1:56:15). - Manage swarms, not co-pilots. "Just using a co-pilot — the bar is rising fast. That's not
good enough." (EP287
1:48:23)
His actual screen (EP284 22:27): "Do you know how to train a foundation model? Do you know how
to fine-tune one? Do you listen to the Moonshots podcast?"
The two seams: he never describes a downside case ("it's not going to shift"), and never addresses limited capital — every prescription assumes you can already act.
🔴 Barred from Gonçalo copy (fails feedback no soft numbers): "No losers yet. Zero. 0.00 failures." — a GP's performance claim on a show his own fund sponsors, unverifiable from public sources. Also "99.999% chance the answer is no" — rhetorical, not measured.
★ EP289 (the AMA, aired 2026-09-13) — what he prescribed, and the one datable call
The AMA is the format where he answers "what should I do" to a named person with a stated constraint, so the prescriptions come out sized. Six new ones, plus the seam finally addressed.
- The regulatory moat — the most transferable thing he says all episode. To a psychotherapist
building his own apps: "the AI labs can't cross the regulatory barrier. So they can't do what
you can do, but they should be right on the cusp of giving you the tools you need to roll out
the perfect AI psychotherapist… if you're first to market figuring that out, like what is
allowed" (
1:09:13). A position, not an opportunity queue. - On selling a proprietary data archive: "the first thing you want to do is make it easy for
them to sign a check… writing a $10 million check… for them, right now, it's the easiest thing
they can do. Having a conversation is much more expensive than writing the check." (
0:21:35) Plus the Kevin Weil "just show up" story and: "if you can physically get on the ground with whoever's in charge, just go to SF… the odds of something happening go way, way, way up if you just buy the flight and go." (0:22:37) And, unusually candid: "if you put this veneer of like a banker look around your price tag… they won't have the time to see through it." (0:23:11) - On funding hardware: build the syndicate before the prototype — "I wouldn't race to like
hacking something together in the basement. I would race to building that syndicate to get to
critical mass" — two or three VCs plus two or three industry operators; or join an existing
SoCal rocket/drone company, "get through like six to nine months of growth and then spin out"
(
0:32:08–0:33:56). He offered his own real-estate development partner as a co-investor on air (0:33:05) → Moonshots: the disclosures ledger. - On adoption timing: "don't fight any missionary battles… AI tends to club people over the
head and they wake up one day with this hallelujah moment. You want to be there on that day, but
you can't force it." (
0:29:50) - On exhausted operators (caregiving): start with planning, the virtual layer, before robots —
"my my advice is start with the virtual stuff" (
0:43:11) — and "if just one person like you can create the road map for everybody else, then they'll follow the road map." - The mid-career answer (
0:14:50, the seam): "I suspect that this year there'll be on the order of maybe 10 to 100 opportunities where the playbook is handed to you and the AI labs are are really really eager for these things to succeed." ⚠️ A number with no source, given to a man with a mortgage. Held inresearch/panelists/blundin-over-40-deep-dive.html. Note the answer that actually fits that caller is #1 above, 55 minutes later, and he never connects them.
★ The datable prediction — log it. "I am 100% sure that the compute substrate is going to move
to photonix [photonics] and then after physics gets reinvented by the current AI it may move even
quicker to something else beyond that… you're looking at the current infrastructure running millions
of times faster… and that's all just in the next really 18 months to two years is my guess."
(1:15:01–1:15:22) Due 2028-03-08. He assigns "100% sure" to the substrate and "my guess"
to the date — quote them apart, never merged. Compare Alexander Wissner-Gross on EP285
1:06:28, who treats photonics as the conditional successor ("absent a successor to CMOS…") and
routes the certainty into collapsing the abstraction stack instead. Dave is the one carrying the
hard date.
A new mechanism read, useful and rare from him: attention "is more like a notebook keeping
notes than it is like a brain function"; the MLP "is identical to what existed 20 or 30 years
ago"; the residual stream is "the old new thing" (1:14:10).
★ He answered Gonçalo's question on air
EP286 (2026-09-05) 2:12:22. Peter reads AMA question #1 — "Can you see a way to turn libraries
into centers of AI development and physical AI training for everyone" — from "Train with John
Koalo", the caption mangling of @trainwithgoncalo (project handles). Dave took it.
His answer: yes-but. The bigger issue is surplus white-collar space; the shopping-mall repurposing analogy "for the most part didn't work out"; "this is another reason why you should want data centers in your community — it's one of the few things that will reliably grow tax revenue, create job opportunities"; and he closes "I'm not super optimistic that all the space will be well utilized."
The stronger version of the argument is in the story the panel was reacting to: a library is literally in CNN's list of what Quincy's data-centre tax revenue built — in the very story the panel was reacting to on Moonshots #284.
Three things Dave said elsewhere complete the answer: the negotiating position ("you've got the
key negotiating position. Ask for what you want" — EP285 1:47:15), the worked case (Provocative
AI in a Somerville carpet mill, water-negative and carbon-negative — EP285 1:47:57, EP283 1:36:19),
and the noise answer ("guaranteed they will build the data center quiet" — EP287 2:19:22).
★ EP288 (2026-09-11, recorded 09-10) — the exit, and a regulation proposal nobody expected
The context that changes how the rest of the episode reads. He opens the show with his own liquidity event (§4 of the EP288 anatomy) and then spends 2h30m making calls on compute scarcity, GPU depreciation, HBM capex and where entrepreneurs should put their time. It is disclosed — as "big news" from a "fraternity brother", with no disclosure keyword anywhere in the segment. Ten minutes of tape later he discloses Orin properly, before the claim. Same man, same episode, two standards. → Moonshots: the disclosures ledger.
1 · ★ The plutonium regime — a hard regulation proposal from the panel's most accelerationist chair.
"every group of eight GPUs that can hold a 40 gigabyte weight file is a threat to all of humanity…
we track right now all plutonium, all uranium. We don't do as good a job as we should, but now you
have something that's just as dangerous as a pound of plutonium. You have to know exactly where it is
and what it's doing." (0:43:46) He prefaces it with *"There has to be a practical real world way to
get that p-doom down to zero. And there is. I know exactly how to do it."* (0:43:36) and then
never describes the mechanism. This is a genuine change of position — the man who says
"work your ass off" and "don't fight missionary battles" is proposing a fissile-materials tracking
regime for consumer hardware. Nobody on the panel challenges it. Log it.
2 · p(doom) at any level is unacceptable. "To me, a 1%, .1% or 20% p-doom is totally
unacceptable. Regardless of which of those numbers it is… it's actually ridiculous." (0:42:52) He
is the only mate who refuses to trade the probability, and the only one who does not state a number
of his own — while Emad states 50→20, Salim states 0.1, and Alex argues the prior. Position, not
estimate.
3 · The PhD prescription, given to a named person. "we just lost an absolutely brilliant MIT guy
who went off to get a Princeton PhD. He's starting this week… the message that you just said, like you
will be so late to the party four years from today that it's absolutely the wrong thing to do. But
there's a lot of pressure, you know, from academia, from family." (1:08:02) ⚠️ He wants to hire
him. The recruiting interest and the advice are the same sentence, unflagged. And the advice rests
on the Navier–Stokes framing that fails verification — see the EP288 anatomy ⚠️ block.
4 · ★ The most transferable idea he has given in eight episodes — and it lands at 2:23.
"we're calling this a lookup table but it's not actually a lookup table… it's a lookup table that
feeds a domain specific neural net. And then the neural net predicts the outcome based on that
combination… everyone is a domain specific neural network opportunity… that is a business model that
can repeat itself in thousands of different domains and if you have the domain specific neural net
advantage for any one of them you have a sustainable long-term business." (2:23:36) He sources it
to Noubar Afeyan (Flagship Pioneering, Moderna) telling him years ago that neural nets' fundamental
use would be genotype→phenotype mapping. This is the reusable prescription of the episode and
it is the one an operator with a narrow dataset can actually act on.
5 · Naming a company and giving it a deadline, on air. To Moderna and Stéphane Bancel:
*"this is the moment where you decide whether we're doing AI biology or whether we're going to fall
behind… If you wait six months, forget it."* (1:36:29) Peter's reply — "We should sit down with
Stefan Bansal… He's a good friend." — converts the broadcast into an introduction.
6 · The crisis-founding frame, new and generalised. "we started the company right after 911…
in hindsight it turned out to be one of the best times in history to be investing in America"
(0:05:42) → "the country always goes through these panic cycles and I think one's coming up related
to AI and they're always unfounded in hindsight and the worst thing you can do is freeze up"
(0:07:04). ⚠️ "always unfounded in hindsight" is survivorship reasoning stated as a law.
7 · A concession on the AMA he has never made: on recursive self-modification —
"yes, absolutely that can happen… if it starts to change the core values that you've programmed in,
you need to stop it… now that all the open source is out in the world, it's much harder to enforce
now than it would have been 3 months ago." (2:37:48) A dated admission that the window closed.
🔴 Barred from Gonçalo copy (EP288): the Orin founders' "personal liquidity in the $und00 million
range in under a year from founding day" (1:12:00, caption-mangled and a GP's claim about his own
portfolio company), and the quant-fund inbound texts (1:12:20, private and unaudited).
★ The Tenev episode (aired 2026-09-19, recorded 09-18) — a pro-regulation turn, and Vestmark twice without the disclosure
Full decomposition: the research archive.
★ New position — he argues for regulation, on a panel that usually argues against it. The standing frame on his page is the entrepreneur's complaint about regulators. On this tape, sitting opposite a regulated-industry CEO, he takes the other side twice:
"everybody knows you need rules on the road to drive, right? you saying you want chaos like on all in all areas"
0:20:19"if you don't have regulation in the industry, all money gets stolen. You know that for sure. So, you know, it has to exist."0:18:02— said immediately after calling the ICI lawyer-and-congressman conference "the most disgusting thing I've ever seen in my life." His ask is clarity, not absence: "if you just create rules, then people can play the game… what happens in the US more often than not is the rules are not clear at all and then they're enforced about 5 years later."0:23:07, with Bitcoin as the worked example.
★ Vestmark is named twice and the disclosure is not repeated. Eight days after stating on EP288
0:04:38 that he is Vestmark's founder, executive chairman and controlling shareholder, and that
it is being acquired by Envestnet (Bain Capital), he says:
"when I when I first founded Vesmark, I had never done anything in fintech before."
1:07:32"We we were doing the negotiation for that Vesmark investment acquisition two weeks ago and uh it was in a board meeting."2:00:35Neither "controlling shareholder" nor "Envestnet" nor "Bain" is said. Minutes later, in the same segment, he makes a directional call on the adviser market Vestmark sells into: "the feeling of value ad from the financial adviser is up at least a factor of 10 thanks to AI. So that's where it's going in the short term."2:03:48. See Moonshots: the disclosures ledger.
New device — the certificate. He canonises Tenev's formal-verification idea in his own words, the same move he ran on Will Marshall's launch-tax line on EP266:
"The certificate is beautiful as a concept because it's not a guarantee of exact input output because you can never do that… but a certificate that shows it's bounded or it's contained or it's below some risk level and people learn to trust that — that's a critical part of our future society."
0:41:01
A second, independent route to the same end, offered unprompted: "the deviation from expectation
is a very measurable thing… it's just a cosign difference of the vectors or the activations."
0:53:06.
Still reading primary sources on planes. Kimi K3's KDA attention, "on the flight back from
California yesterday" 0:48:19, reconstructed as a thought experiment (drop the softmax after QK,
see what simplifies) and tied straight to the RSI loop.
⚠️ Two self-sourced numbers from this episode are unverifiable and must not be repeated:
"Robin Hood is what 30 million lines of code maybe" and "usually like a core portfolio accounting
platform will be like 10 or 20 million by itself. And I know you have that." 1:55:45 — he is the
vendor of that category. Tenev's own reply was "I don't know if it's quite that much."
★ EP291 (2026-09-17) — the regulation mechanism changed inside six days
★ The proposal was swapped and nobody noticed. On Moonshots #288 0:43:46 he proposed a
plutonium regime for GPUs — "every group of eight GPUs that can hold a 40 gigabyte weight file is
a threat to all of humanity… We don't do as good a job as we should, but now you have something that's
just as dangerous as a pound of plutonium. You have to know exactly where it is and what it's doing" —
and added "I know exactly how to do it." That proposal does not appear in EP291. What appears
instead is API logging: "when you have APIs and you're logging everything, you can easily detect
that, which is what anthropic is doing… I think the logging is absolutely the critical next move…
It's the same as nuclear weapons inspections. Inspection is job one. Once you have inspection, then
you have data." 1:03:00–1:03:40
Same analogy source (nuclear), opposite object: hardware custody → software telemetry. Hardware
custody constrains the labs. API logging constrains their customers, and is something the labs
already do — which is why the same segment ends on China: "when you're China and you're saying, yeah,
we postrained it… Throw it out there. There's no way… that it's actually contained." 1:03:10
Neither he nor Peter marks the swap.
★ And the p(doom) floor moved with it. EP288 0:42:52: "To me, a 1%, .1% or 20% p-doom is totally
unacceptable. Regardless of which of those numbers it is" — he was the only mate who refused to trade
the probability. EP291:
"from a core engineering point of view, it's just not that hard… It's all nonsensical from an engineering point of view… feriting those out of the training data is so straightforward."
0:52:27"literally the escape scenario is not the real threat, but it's what gets the news. But the terrorist scenario is very real."0:53:30"I don't think slowing down helps one iota… slowing down just gives China global domination."0:37:07
The bad-actor framing is not new (EP288 0:56:33). What is new is that the escape scenario went
from unacceptable-at-any-probability to "nonsensical," and the mechanism he proposed to prevent it was
dropped. Put the two quotes side by side. It is a question, not a gotcha.
His sharpest read on Anthropic, and it is the best line in the episode:
"there's not even a hint of a plan inside Anthropic to slow down core R&D. That's never been in the cards. When they go to the public and they say we all need to get together and agree and collude to slow down, what they mean is slow down the releases of hyper powerful APIs… They're not going to stop the recursive self-improvement process internally. I think in their dream scenario, they slow down what everyone gets access to while they discover new drugs, while they discover new physics. Then they monetize the hell out of those discoveries and then they control AI and the discoveries forever thereafter."
0:48:40–0:49:22
Alexander Wissner-Gross's Novo Nordisk observation fifteen minutes later (1:04:24, "for me, but
not for thee") is the evidence for exactly this claim, and neither man connects them.
The best caveats on the tape, and he gives them against his own optimism. On Vending-Bench 2:
"I don't want people to take away from this that AI can just build a business from scratch out of thin
air. This particular benchmark is a tightly constrained environment… It will fail miserably until you
give it a framework and break it down into tiny little tasks" 1:43:34, with a runnable test —
port an open-source game from GitHub to Java or assembly. And a concession on frontier capability
six days after EP288's "math is cooked": ask Fable 5.1 for a Navier–Stokes-class problem and
"it will fail miserably on that prompt, which tells you a lot about the role of coaching it" 2:18:01.
The seam has moved: he is now the mate supplying the caveats while Alex supplies the acceleration.
New institutional prescription (education) — the most usable thing in the episode. He read the MIT
ad hoc committee report "cover to cover" and names two things it got right: "if you try to fight AI
in the classroom, you're in an arms race… what you need to do is budget to change the curriculum
that you haven't touched in decades… this is the future forever now of constantly needing to change the
curriculum" 1:48:56; and "you should not just replace your Europs [UROPs] with an AI agent"
1:50:01.
New operator frame — the three-best-friends window, with a falsifiable half attached:
"a small group I usually define it as three best friends but a small group can act like a group of
thousands for some window of time here" 1:38:01 → "we're seeing multiple companies with
multi-billion dollar valuations that are under two years old. Never happened in the history of the
world before." 1:38:22
Disclosures. ★ A board seat used as the proof of a competitive claim — "I was on a board call
for Laya this morning, right, which is building a scientific super intelligence and building a million
square foot of automated robotic labs driven by that AI… They're going to run circles around research
labs with graduate students" 1:52:54. The company name is mangled three ways in one segment
(Laya/Leo/LILA, probably Lila Sciences) and never resolved, so the listener cannot look it up.
No disclosure keyword. Compare EP288 1:10:55, the exemplar, eight days earlier. Also: a company he
sold last week is in live antitrust review — "I actually have a company right now that I sold last
week that's in antitrust… And it's just the smoothest it's been in my entire lifetime" 0:17:22,
volunteered inside the argument it is evidence for. → Moonshots: the disclosures ledger
⚠️ Three claims from this episode are barred: Anthropic's "80% gross margin" and "growing as
faster than any company in history" 1:27:56 (unsourced, a GP on a show his fund sponsors); Amodei
"trying to get 20% super voting stock… to be majority" 1:25:23 (unsourced and incoherent as
stated); "they're buying right now at 1.2 trillion is the offer made to OpenAI" 1:24:30 (the number
matches reporting, "the offer" and "holding out" are his gloss). And "Biden already told Mark Andre
he was going to limit AI to just four or five big companies" 1:17:02 is Andreessen's own account of
a private meeting — attribute it to him or drop it.
★★ EP293 (AMA #2, recorded 2026-09-18, aired 2026-09-22) — "software is cooked, hardware will go for many years," and a seam he opens twice in twelve minutes
He is first responder on eleven of twenty-three blocks. The single most transferable line in the episode is his, and so is the episode's clearest unforced contradiction.
★★ The career instruction: leave software for manufacturing
0:23:05, closing the lunar-manufacturing answer:
"Also, anyone who's in computer science right now should be thinking in terms of moving to manufacturing as well. Software is cooked, but hardware will go for many, many years. So it's a great, great business plan."
This page already holds him on Link Ventures moving into hardware ("the future of the next five
years of investing, including our funds", Moonshots #289 0:32:08) and on "an intelligence
company whose products happen to have atoms in it". This is the first time it is put as an
instruction to an individual to leave software, with no hedge and no timeframe.
★ The off-world test-lab prescription — and it is a bench rig
0:22:00–0:22:40, to a 33-year-old ex-AWS data-centre technician retraining in robotics:
"If you set up at IIT a lab similar to the nanotech lab at MIT, but focused on manufacturing in non-earthly environments, you could probably very easily create an environment where you're bombarding radiation on your machines, where you're operating in simulated zero G on your machines — really just everything you're building, you build it so it works here and there. And then you get to market with self-replicating machines here that immediately deploy out into space."
Preceded by: "listen to Alex's innermost loop and then focus on the innermost loop within manufacturing, which is things that make other things." This is the ground-rig argument, made by a GP to a stranger: test on the ground before you fly.
★ The seam: "be the owner" at 0:30:48, then "study Mercor" at 0:45:59
Twelve minutes apart, to two callers with the same problem, he gives answers that do not compose.
First, to Gianluca — an Italian technical operator whose productivity is up 3–5× and whose income is flat ("I'm not able to capture the value. My employer is currently capturing all that value"):
*"The solution is to be the owner, get into a position where you don't care if you're not getting paid more because you're a shareholder… I mentioned this on the pod a while ago, almost all value is going to accrete to capital gains through ownership and equity and not to payroll."*
…and he evidences it from inside his own boardrooms: "when I look at a lot of the companies that I'm chairman of, the thought process is all of this AI automation is going to drop to the bottom line and the shareholders are going to make a fortune. And that's exactly what's happening. But you don't think in terms of pay everybody more." That is a controlling-shareholder describing the mechanism that is costing the caller money, and then recommending the caller change sides. It is honest and it is not disclosed as a position.
Then, to Adam — the same question at a system level — the venue he names is Mercor:
"I've watched Mercorps become the fastest appreciating company in the history of the world…
they've unleashed 100,000 people now to help with AI", and "I would walk exactly in the
footsteps of Mercorps." He also gives Mercor to Gianluca as a place to earn: "you should be able
to do what you were doing before plus do something on Mercor plus do something on Fiverr"
(0:29:43).
🔴 The contradiction is on this panel's own tape. On ep232 Mercor was the panel's exhibit for
value not reaching the worker — "people in India doing work that benefits Mer[c]or that is
actually for open AI. The fraction of all value created that flows back to the mother ship is just
a massive fraction of that value chain" (0:55:44). Same company, opposite role, same
argument about who captures value. ⚠️ "Fastest appreciating in the history of the world" fails
verify-first — see Mercor. ⚠️ Corrected 2026-09-29 (deep-dive): Link Ventures holds Mercor (portfolio page: invested
2023, fund LV3, Active), and he disclosed it on tape on EP175, EP181, EP248 and EP282 (Peter on EP162).
It was not restated on EP293. See Mercor § Who holds a position.
Universities — he takes the opposite side from Alex, on tape
0:37:10, splitting the caller's question in two and defending half:
*"The idea of making your best friends for your entire life and thinking for the first time deeply in your entire life, those are inseparable events… that's essential and that's not going to go away. What we beat on constantly is the curriculum — and the curriculum is clearly going away… Alex thinks universities should go public. That's one way forward. I think universities are critical as the ethical actor in AI, the place that doesn't have a profit motive. So I believe they're here to stay."*
Alex then asks for more destruction in the next breath ("vivisect them"). The only genuine panel disagreement in the episode, and it is Dave defending an institution — which is new for this page.
The other prescriptions, in order of usefulness
- Regulatory-arbitrage geography, generalised
1:11:26— "there are many places in the world that can move much more quickly… sure, it could happen in the U.S., but it won't because of this government blocking action. This happens a ton in biotech. It'll come in self-driving. It'll come in drones… get those deployed there wherever they can take off and flourish and then they can backport to the U.S." The constructive twin of the EP289 regulatory-moat line, and the better of the two. - Attack the paperwork, not the labour
1:10:08, to an apparel exporter — "everyone's thinking, oh, but first I would need to have robots all over the place. No, no, no. Do it just with the paperwork and the planning and the scheduling… All of that is perfectly attackable with AI." - Management of agents ≈ management of people
0:40:50— "I'm shocked at how similar trying to get a thousand agents to do something constructive together is… to all my experiences trying to manage a thousand people — the way they communicate, the choice of granularity, the way you divide up the problem." Given as where an MBA transfers. - The 30-day deal rule
0:41:25— "from the day you graduate to the day you're involved in a five or $10 billion negotiation, try and get that down to 30 days or less." ⚠️ The supporting number ("hundreds of concurrent, five to $10 billion negotiations… massively understaffed" at OpenAI/Anthropic,0:41:34) has no source. Take the tactic, drop the number. - The PitchBook→keiretsu tactic
0:25:08— look up who funded every company you admire, then "who are the five companies I most want to interact with, and who's behind them? Let me get into that keiretsu." Names Steve Jurvetson. - The life-coach segue
0:48:45, to a 17-year-old Japanese English-education company — "people aren't learning English because they want English. They're learning English because they want to change their lives… becoming their life coach platform is a very natural segue." ⚠️ The market sizing he gives ("a $10 billion business, probably $20 billion by now", doubled mid-sentence,0:48:33) is unsourced. Do not carry it. - The board-communication answer
0:50:49, to a CISO — don't rave about risk, "go and dig all those out and sequentially say, look guys, the rate of events is on this exponential ramp", and "the business opportunity of the century is the forward defense against that." Note he recommends Alex's own newsletter as the incident feed.
🔴 "Write off the mainstream media" — and the remedy is his own channel
1:15:00 — "the media is always going to be broken from here forward, because they're starved for
money. They have no budget. And so they have to chase drama in order to even survive. And so you
just got to write off the mainstream media. … the antidote to that is micromedia or X and
podcasts and narrow casting." Four minutes later Peter Diamandis says the same thing and
adds *"And you guys are doing that as well. You're listening to the podcast."* (1:16:50).
The diagnosis may be right; the prescription is the product, said by the product, with no
disclosure phrase anywhere in the block. The Germany tactic attached to it is genuinely good:
"go in a group of like five or seven people, spend a month in Palo Alto, and then backport
it… seven becomes 50 becomes 500."
★ EP294 (Moonshots Live, recorded 2026-09-25, aired 2026-09-27) — EverQuote named as chairman, and the Hinton position moves a third time
Full decomposition: the research archive.
★ He named EverQuote on tape for the first time, as its chairman, with a stock move — and then
told the room the market's thesis is wrong. On Meta's Muse agent: "the stocks, like including
Epiquo [EverQuote], where I'm the chairman, went down 15% on back-to-back days because everybody's
like, Oh my God, I'm going to always find my auto insurance and my mortgage through MetaMuse, which
is, it's not going to actually play out that way." 0:33:30–0:33:45. The tie is stated, which is
more than the EP291 board call managed; it is not phrased as a disclosure, and the 2026 proxy says
more than the tape (57.1% of the vote). The analysis around it is worth holding: "there's a fork in
the road now where you'd have consumer AI, which is insanely valuable… and then on the other end you
have recursively self improving strong AI, which we just benchmark the hell out of… it doesn't have
to be the most intelligent model to be the most useful in terms of your day-to-day life"
0:32:57–0:33:30. ⚠️ The 09-22 selloff in insurers and marketplaces is real; "15% on
back-to-back days" is not verified. → Moonshots: the disclosures ledger
★ On Hinton and p(doom): the third position in fourteen days. EP288 0:42:52: "a 1%, .1% or
20% p-doom is totally unacceptable", plus a plutonium regime for GPUs. EP291 0:52:27: the escape
scenario is "nonsensical from an engineering point of view." EP294, after Jensen Huang's clip
calling Hinton's 10% "irresponsible": "Jeffrey Hinton is calling it the way he sees it. I think
the quote, just because it comes from a scientist doesn't make it scientific. That's going to end
up on a… shirt… but clearly we're moving forward at warp speed. It's improving itself… The misuse
of it is an incredible threat and there are solutions to that, but it's happening anyway. It's
stuck on doomerism is completely dysfunctional. We're not slowing down." 0:11:02–0:11:37. He
does not dismiss Hinton personally. The mechanism promised on EP288 ("I know exactly how to do
it") is still unstated: "We need very specific technical solutions to open source, to monitoring,
just like with nuclear power. And if we get moving on those, we can solve them very easily in time
for ASI." 0:28:03–0:28:16.
★ The data-centre argument against the Sanders bill, which is the one new argument on the tape.
"a lot of it can be reversed in short order, but the data center part of it, if your state doesn't
have data centers and they start getting built all in other states, they're not coming back. And
once the ground is broken, the chips are there, the investment's made, that's the part you'll never
get back. And that's going to be such a massive driver of the economies in the states that get
ahead." 0:29:29–0:29:58. Consistent with his EP286 answer to Gonçalo's library question.
The caveat chair, for the second episode running. On DrivingBench, closing the block: "A
generalized model can drive a car, it can be a robot, it can build a robot. That's absolutely true.
A specialized model, like if you said, 'Hey, it completed the course on its second try,' Elon's
going to be like, 'Yeah, your Tesla can't do that on a second try.'… the thing that's missing in the
storyline there is how much compute did you use to do the task? Compute is going to be forever
starved from here forward… if you used more compute to do the same task, that's a crime."
0:42:57–0:43:34. And the joke that states the show's role: "we've been doing two episodes a
week. It's like killing us. I wonder if we're the problem. We're driving the thing so fast, we're
causative here… If we don't report it, it didn't happen." 0:42:09–0:42:22 [attr?].
He feeds the other side on purpose, again. "Can I just feed the P.Doom just for a second…
Each of us has 50 trillion cells in the human body governed by what the DNA tells that cell to do,
meaning a human being is now a software engineering problem… we are now in full Frankenstein mode,
so there is some ammunition for all the P.Doom people. Go to town with that." 0:46:14–0:47:12.
🔴 Barred from Gonçalo copy (EP294): "it's right 99.9% of the time" 0:39:48 (rhetorical);
"Opus 5.5 just crushed all the benchmarks" 0:33:18 (Anthropic's own claim is "performs at the
level of Claude Fable 5.1 on most work", self-reported); "Alex's one model a day" 0:31:54
(not found in the held transcripts; attribute to Dave's account). ⚠️ "HBM RAM has gone up 5x in
price" 0:43:25 is Close as a spot-versus-contract premium (≈$2,100 spot vs $300–400 under
contract for a 36 GB HBM3E module, Sept 2026), not as a rise in what anyone's contract pays.
⚠️ Alvin Wang Graylin's texts (0:18:58–0:19:54) as the summit read: the "advising the White
House… at the UN with the delegation" credential fails verify-first for the second time (EP289
0:27:25; its only source is the show's own ep281 introduction).
Prescriptions → the model. "If you haven't tried it yet, just upgrade all your systems. It's a
mouse click away." 0:36:08 restates step 1 (get fluent with the models now). The data-centre
siting argument is policy, not a personal step. The sequence does not change.
The characteristic moves
- The canonising superlative — "That is brilliant. That is absolutely critical. Nobody has said that before. Mark my words." (It made the EP266 title.)
- The "I happen to know the answer" insider explanation.
- The two-scenario challenge to Alex — "I'll give you two scenarios. Tell me which one is going to play out."
- The motivational close — "work your ass off right now because it may be the last chance."
- The self-aware rant warning — "you're going to have to shut me up on this topic."
What to take from him
He is the closest chair to money on the panel, and still on the abundance side of the gap — he owns the index, the boards and the deal flow. His reduction to training vs inference is a usable lens. His sourcing rail is not available to us and should not be faked; ours is the filing → Reverse bill of materials.
His model, step by step
Built only from the prescriptions verified on this page and in the week-1 map. The order is his own
life plan ("become incredibly great at helping building data centers… and then reinvest that money
into AI companies", EP287 1:53:34), with the tactics he gives for each stage. His own seam applies
throughout: he rarely says what to do without capital, so the steps that need it are marked.
- Get fluent with the models now, on your own time. "getting really familiar with the models…
build your own skippy, hack around at night… it's a great way to generate cash flow and get
involved with the models" (EP289
0:15:21). Then go past assistants: "just using a co-pilot… the bar is rising fast. That's not good enough… you need to move on and manage swarms" (EP2871:48:56). - Move toward things that make things. "Software is cooked, but hardware will go for many, many
years" (EP293
0:23:05). "the people who cross that chasm and get into this data center buildout are crushing it. And there's opportunity at every level of it" (EP2851:19:44). The one version of this that needs no capital: "all of these things are regulated at the local, state, and federal levels. So, if you're connected at any level, you have an opportunity to get involved in these projects" (Moonshots #3040:53:01). It assumes you know someone in local or state government. - When a well-paid, time-limited build-out job appears, take it. "Take the job, take the money,
stay nimble, and then the next opportunity will open up" (EP287
1:56:28). His example is electricians and plumbers for a one-to-two-year data-centre job. The step assumes you can move to where the job is. - Start with the paperwork, not the robots. "Do it just with the paperwork and the planning and
the scheduling… All of that is perfectly attackable with AI" (EP293
1:11:07). The same advice to a caregiver: "start with the virtual stuff" (EP2890:43:11). - Turn your narrow data into the edge. "it's a lookup table that feeds a domain specific neural
net… if you have the domain specific neural net advantage for any one of them you have a
sustainable long-term business" (EP288
2:23:43). This is the step that suits an operator with deep knowledge of one field. - Get in front of the people who decide. Look up who funded the companies you admire ("If you
get a pitch book account… work back to who invested in them, and then just go talk directly", EP293
0:25:04), then go in person ("if you just buy the flight and go", EP2890:22:52). ⚠️ PitchBook is a paid subscription, and the flight costs money. His EP284 version ("If you can't get Jensen, just get one of his E-staff",0:09:06) assumes access most listeners do not have. - Where regulation blocks you at home, launch where it doesn't, then bring it back. "get those
deployed there wherever they can take off and flourish and then they can backport to the U.S."
(EP293
1:12:16). ⚠️ Launching abroad assumes capital and mobility. - Move from payroll to ownership. "the solution is to be the owner… almost all value is going to
accrete to capital gains through ownership and equity and not to payroll" (EP293
0:30:35–0:30:48), and reinvest what you earn into AI companies (EP2871:53:40). ⚠️ This step assumes you have savings to invest or the standing to take equity instead of salary. He does not size it for someone who has neither.
What an operator can do with this
Doors a person or a small company can walk through this month. Each was checked at the source on 2026-09-29.
- If you are a technical AI founder near Boston: Link Ventures and Link Studio. The fund backs "technical founders developing disruptive AI technologies" at "Pre-Seed, Seed, and Series A", and offers office space, housing, compute, legal support and "podcast and media opportunities through their MOONSHOTS podcast" (linkventures.com). Link Studio is 40,000 sq ft of space by MIT (Link Studio). Entry is by the contact form. Neither page publishes eligibility, check size or the equity taken, so ask before you pitch. The fund lists the show as a founder benefit, which belongs on Moonshots: the disclosures ledger.
- Put your domain knowledge on Mercor. His EP289 pointer (
0:16:57) is a live marketplace for expert contract work training AI (work.mercor.com/explore). This is step 1 of his model, cash flow and fluency. It is paid by the hour and gives no ownership, so it is not step 8. Mercor's advertised average rate is its own figure. Do not repeat it. - If you make physical things, or want to, call your state's MEP center. This door is ours, not his. It serves his step 2 ("Software is cooked, but hardware will go for many, many years"): "Any manufacturer can take advantage of the services provided by MEP Centers", which operate in all fifty states and Puerto Rico (NIST MEP centers).
- Do the paperwork audit. List the planning, scheduling and paperwork in your operation and hand one piece to AI this month. It costs nothing, and it is his step 4 (EP293 at 1:11:07).
Questions for him
Asked to understand his model, not to catch him out. Each one rests on tape or on a disclosure.
- Over 40, or over 55: what is the playbook when starting over is not an option? On EP289 a
mid-career caller with a family and a mortgage asked for the playbook (
0:13:44). His answer (0:14:50–0:17:12): watch for the playbooks the AI labs hand out through franchise and data-collection models; get familiar with the models and "hack around at night" for cash flow (0:15:21); find five people online who think this way (0:15:55); put your domain knowledge into Mercor (0:16:57). He expects "on the order of maybe 10 to 100 opportunities where the playbook is handed to you" this year (0:15:50). The caller gave no age, so "over 40" and "over 55" are our framing. The questions: which of those steps change at 55, when a one-to-two-year job (step 3) is a bigger share of the working years left, and reinvested earnings (step 8) have less time to compound? And which of the 10 to 100 opportunities can he name? Base:research/panelists/blundin-over-40-deep-dive.html. - Photonics: the view, the date, and the exposure. On tape: "I am 100% sure that the compute
substrate is going to move to photonix [photonics]… and that's all just in the next really 18
months to two years is my guess" (EP289
1:15:01–1:15:25), and earlier "photonic computing and new physics are imminent" (EP2821:33:02). Alex on the same tape: "maybe it'll be photonix, which I know Dave is of of interest" (EP2822:16:12). No photonics holding is disclosed on any tape we hold or in Moonshots: the disclosures ledger. The questions: what would he watch between now and the 2028 date to know it is on track, does he or Link Ventures hold photonics positions, and what does someone without a fund do with the view? - Mercor: a place to work, or a model to own? On EP293 he gives three Mercor answers in seventeen
minutes. To Gianluca, as a place to earn: "do something on Mercor plus do something on Fiverr"
(
0:29:39). One minute later, to the same caller: "the solution is to be the owner" (0:30:35–0:30:48). To Adam: "I would walk exactly in the footsteps of Mercorps, study everything that those guys did" (0:45:59–0:46:18). Read together, the model may be: earn on the platform now, study how the platform was built, own the next one. The question is whether that is what he means, and what the person working on Mercor owns in the meantime.
EP295 (Moonshots Live, recorded 2026-09-25, aired 2026-09-28) — a guest's case he carries into the next two episodes
Read in full 2026-10-18 (Moonshots #295, Palmer Luckey; attribution inferred from Peter's "Dave." hand-off at 0:07:25).
- The question: edge vs centre for weapons, "I'm kind of a feeds [speeds] and feeds and chips and AI foundation models guy" (
0:07:29); neutral, no position stated. - 🔴 "right now the open source models are all Chinese" (
0:11:48). Wrong: Meta's Llama, OpenAI's gpt-oss, Google's Gemma and Mistral are open-weight. Barred from copy. - Where it went next: Moonshots #298
2:14:07, "Palmer is very aware of the risks of human actors… I'm not I'm not worried about that either. I'm worried about same same risk that Palmer's worried about"; Moonshots #2992:05:09, "I'm also not a fan of autonomous weapons that make decisions in the field, but Palmer Lucky made a very good case for it." He adopts the guest's risk frame and keeps his objection to field kill-decisions. No stated position on autonomous weapons before EP295 in the indexed episodes.
★ EP298 (aired 2026-10-02): joins at 1:51:22, argues Section 230 for the labs
Read in full 2026-10-06 (Moonshots #298). He is absent for the first 1h51; Salim wears "Dave's hat" at 1:34:03.
- Liability: calls strict developer liability "childish" (
2:01:48): "it's going to force everybody to use Chinese openweight models" (2:02:11); argues the social-media analogy — platforms succeed because Section 230 immunity exists — and that "the equivalent regulation in AI has to be written" (2:03:34). Alex names it "some version of section 230 immunity for the frontier labs" and Dave says "Yeah. Exactly." (2:04:44). Narrower later: if an agent has no owner, "I would say anthropic is liable for releasing an autonomous agent into the world that has no owner" (2:04:32) — he concedes the strict case. Peter's split (user liable; if no user, the lab) is where it ends (2:08:07). - Risk: "not every CEO says there's danger" — Jensen Huang "said flat out there is absolutely zero risk of extinction of humanity through AI by 2033, I think he said" (
2:13:46). ⚠️ Unverified quote (his own "I think he said"); not in any Huang page we hold. - Social media: "I don't think social media was that bad. Um but AI is nothing like that" (
2:14:49) — while saying it damaged "an entire generation" in the same breath. ⚠️ internally inconsistent; do not quote as a position. - Mercor praised, tie not restated: "they also all know Brendan Foody uh from Raor… he started the company at age 18. So I think he's only like 22 now, maybe 23." (
2:19:47–2:19:57). Link Ventures' LV3 holds Mercor (disclosed on five earlier episodes); not restated here. ⚠️ Foody's age is his guess. - MIT seniors: "they all worship Elon" (
2:19:27) — an anecdote from a visit, no count.
Where it came up
- Moonshots #298
2:01:48–2:19:57— liability, Mercor · Moonshots #2840:00— NVIDIA and the circular economy · Moonshots #283 — the 100K-agent disclosure · Moonshots #266 — canonises "chips beat rockets" · Moonshots #255 — the singularity-economy charts · Moonshots #254 — Blitzy
⚠️ Unverified
Board seats and the insurer's name are as stated on tape and in one inventory; not checked against filings. Auto-captioned quotes.
★ EP299 (aired 2026-10-03): the regulation line moves a third time; two Link companies carry arguments with no tie
Read in full 2026-10-08 (Moonshots #299; attribution inferred).
- From a mechanism to mockery. EP288
0:43:46: a plutonium regime for GPUs. EP2911:03:00: API logging, "inspection is job one". EP299: bans on RSI and data centres are "so stupidly childish" (1:23:15), "I can't optimize my hyperparameters. I can't tune my hard drive" (1:24:09). Khanna's bill (containment, shutdown controls, evaluators, an agency) is nearer his EP291 proposal than a "ban"; nobody asks which parts he would keep. - Liability, repeated from EP298: "if you slap class action on the outcomes… anthropic gives its best models to everybody" (
1:27:04[attr?]). - Liquid AI as history, tie unsaid (
0:37:16) and a Sonnet-5.5 pricing theory from "the Blitzy team yesterday" (1:46:22): both Link LV3 companies. - New: "I'm also not a fan of autonomous weapons that make decisions in the field, but Palmer Lucky made a very good case for it" (
2:05:09). "Overjoyed" about Washington (2:04:37). - Prescription: "attract a data center to your neighborhood first… then you have the next five or seven years to figure out your tax policy" (
2:13:09) — conditional on having leverage over a town; the example has a "couple million dollars" budget.
★ 2026-10-07 episode (Moonshots #301, EP301 by sequence) — the only mate who is 100% against personhood, and a turn toward public compute
Read in full 2026-10-08 (full detail in the research archive of ep2026-10-07-coming-manhattan-project/).
- Personhood, against, on operating grounds: "I totally agree with Mustafa" (
1:19:09); agents should be run "many, many… like Ahmad does," with minimum context, parked when idle, cloned and deleted (1:19:30–1:20:30); "If you start talking about having rights, it's going to say, 'Well, you can't freeze me.'" The danger he adds: an AI that is "convincing" persuades voters (1:20:46). First time on tape he is the lone no. - Compute: a turn. After EP298 (Section 230 for labs) and EP299 (RSI ban "childish"): "governments and universities are a really good counterbalance" to market allocation (
1:11:48); universities are "one of the few ways that we might actually get some degree of meritocracy" (1:09:14). His account of access: "through the White House… the Kushner brothers, Thrive Capital… Antonio Gracias… Gavin Baker… Chase Lockmiller" or an investment from Jensen or Lisa Su (1:08:46); his lab cannot buy an NVL72 ("$9 million, we can lease it for three years… we don't even own it",1:08:04–1:08:30). - Altman: "It's about taking the moral high ground and winning the PR war" (
0:41:59); then "an engineering problem with engineering solutions" against a risk "knob" (0:46:35–0:47:18). (The "you can't regulate the core technology" push-back at0:47:20–0:47:43follows his turn and is[attr?]another mate.) - Executive orders: "The mode of operation in this singularity moment is just dictate" (
0:28:06). - "Bulk mode… 400 of them working on something as a cohesive team" (
0:11:32) is a description attached to Emad's number, not a claim of his own. Peter's "your thousand agents" (1:38:02) has no matching statement by Dave: EP2930:40:51is a management analogy, 2026-09-091:48:44is about others. See Emad Mostaque and the research archive §11. - Lab and network: "One of the young geniuses here in the lab, Will Thompson, from Harvard, got access to Cerebras' twelve hundred token per second Astra model" (
0:09:50); Positron, Alpaca and "the Vokara team" promoted by name with no tie stated or denied (0:01:33,1:38:24,1:38:42); Link Ventures sponsor card at0:01:12. Portfolio page not re-read: whether any of these three is a Link company is unchecked. - Prediction, due in 26 days: "75% chance that Congress flips to the other party" (
0:27:57; midterms 3 Nov 2026). A soft probability; barred as a number. - Prescription (not added): "somebody just needs to take their money and fill the void… they'll just start paying you tomorrow" (
2:33:15–2:33:40) — data-centre community work; no buyer named; cannot be checked at a source.
★ 2026-10-15 episode (Moonshots #unnumbered, no feed number) — a turn on how AI is treated in nine days, and a turn on hardware in three weeks
Read in full 2026-10-16 (full detail in the research archive of ep2026-10-15-meta-s-14/). He leaves at 1:55:57 ("Our partner meeting is kicking off").
- Treatment of models, moved. EP301 (
1:19:09–1:20:30): agents parked, cloned, deleted; "If you start talking about having rights, it's going to say, 'Well, you can't freeze me.'" Here: first "It doesn't care if you're nice to it" (0:13:18), then, a year out, deleting a state is "not clear you could recreate" because of non-determinism, with "these need to be treated well, and these need to be preserved. That memory can't just be deleted tomorrow." (0:21:44,[attr?]). His conditions are mechanical (non-determinism, a face, a voice: he puts Einstein/Newton/Curie faces on his own agents,0:21:11), not moral. - Hardware, moved. EP293
0:23:05: "Software is cooked, but hardware will go for many, many years." Here, on OpenAI's math drop: "If it's all solved, it just becomes simple, simple lookups. So the pace of engineering grows massively with this." (0:47:16), applied to rockets, gears and robots. Both lines are his; they sit three and a half weeks apart. - Prompt hygiene, with the Link company named: "we actually have to train them, especially at Blitzie, to be very short and succinct and direct" (
0:12:54); then "Don't clutter up the context". See Blitzy. - Meta, praised: "Time and money well spent, actually." and "Just do what Mark did." (
0:27:~); no stake on the ledger, none said. - Elon "forward-looking" answer: "I see a path. It's not a forward-looking statement because it doesn't make a specific prediction" (
1:39:09). His "paid over, what, $20 million, $30 million in fines for various tweets" is from memory, unchecked. - Forecast, flat: agents returning video and audio reports "in the next few months" (
1:33:41).
EP304 (Moonshots #304, aired 2026-10-21) — a two-year window, and a door that needs no capital
- The window: to MIT undergraduates, "your window of opportunity to tap your big MIT brain is about the next 2 years" (
0:26:49). Logged as a prediction (due ~Oct 2028). - A no-capital door, new: "all of these things are regulated at the local, state, and federal levels. So, if you're connected at any level, you have an opportunity to get involved in these projects" (
0:53:01); added to step 2 of the model above. - A capital door: "be the first on your executive staff to go and say, 'We need to unlock a billion dollars to do a prepay on Chamath's next power supply'" (
0:53:55). Assumes authority over corporate capital; no step change. - Alignment by inspection: "the new Chinese models are absolutely capable of self-improvement… As a practical matter, it's about inspection and containment" (
1:03:35–1:03:45). - Ties said in passing (→ Moonshots: the disclosures ledger): "we have our company LightWave" (caption,
0:39:39); "I'm on the board of Parity I see Bill quite a bit" (caption,1:35:51; company unresolved). - ⚠️ "all of the data centers in the Saudi and and Middle East got bombed" (
1:23:46): AWS Bahrain and UAE were damaged in the spring strikes; no Saudi data centre is confirmed struck. ⚠️ "Nvidia bought it [Groq] for 20 billion" (1:54:12): it was a licensing deal (NVIDIA). - Silent on crypto this time, after arguing it with Alex to the end on Moonshots #unnumbered.
The Astro Teller fireside (Moonshots #unnumbered, aired 2026-10-05)
- A softer people-vs-AI line than his news-episode tone: "Now comes a world where you replace people with AI and money with computing power. This world has not yet arrived" (
0:25:36), then "there are three AIs and three humans, and I'm putting computing and money into them" (0:25:46). - Asks the ratio: "Their AI spending is now less than their personnel spending. When will the day come when this ratio reaches 50/50, 75% AI, and then 100%?" (
0:28:09,[attr?]; company not named). The guest's answer: not 100% in the next decade (0:28:19). - Plugs the book: "Alex and Peter have this book called " Solve Everything."" (
0:04:58). - ⚠️ "a company worth about $30 billion" (16.5 years ago) and "now… almost half a trillion dollars" (
0:05:08,0:05:18): both figures fail against Alphabet's own scale (about $4.2 trillion on 2026-10-04, aggregator). Not usable.
EP305 (Moonshots #305, aired 2026-10-23) — the power thesis flips to chips in three days, and 10 GW goes to space by mistake
Read in full 2026-10-24 (detail in the research archive of ep305-elon-s-100b/).
- Bottleneck, moved. Moonshots #304
0:39:59–0:40:21: "the power costs will still be there… Elon is exactly right. The currency that is most dominant… is wattage and tonnage… if you can find a power supply today, it's the investment theme of the decade by far", and the advice to "do a prepay on Chamath's next power supply" (0:53:55). Here, about Elon's identical claim: "I think it's a little misleading. The the bottleneck to all of AI is actually chips and right now it's actually packaging" (0:31:42), "memory prices are up 10x… clearly the chips are are still the bottleneck" (0:32:15–0:32:26). He keeps a short-term concession ("in the very very short term there isn't enough liquid… cooled data center space"), and names "posetron, lightwave and a few others" as the air-cooled alternative (0:32:37; no tie said, though EP304 had "our company LightWave"). - Ground or orbit: "if you put 10 gawatts of of inference in space, but it could be up to 20" (
0:15:49); Peter corrects him on tape ("It's all terrestrial right now"0:16:09). See Orbital compute (data centres in space). - Blitzy, praised two and a half minutes before its sponsor read: "That's why Blitzy is growing as fast as they can hire" (
1:01:22); State Street's NAV rebuild "a billion dollars and now it's… probably 150th of that" (1:01:33, unsourced). See Blitzy, Moonshots: the disclosures ledger. - Agents: "we way over anthropomorphize these things" and "I can manufacture a story like this every minute if I want to" (
0:51:12,0:51:42) about OpenAI's monitor incidents: not supported by OpenAI's own account (OpenAI). Sub-agents,0:54:11: "it'll automatically launch five, 10, 20 or 100 sub aents" (caption). - Data centres: "nobody uses water cooling towers for data centers anymore. It's like that's all marketing" (
1:32:02; 🔴, see NVIDIA); the pod as a business for "you're entrepreneurial, but you're not super technical" (1:33:05; capital $5–35M by his count1:36:47); the Cyber Cab "early franchisee" idea (1:32:55; 🔴 no franchise programme, see Tesla). - Prescription (no change to the model): "go back and say we want you to cut our electric bill by half. We want you to fund our schools" (
1:25:00,[attr?]Dave or Peter). The third appearance of the same data-centre community line (EP2982:13:09, EP3012:33:15); it assumes leverage over a town. - Network claims, unchecked: a story about one of the firm's LPs and a senior government role held by a friend of his (
0:02:59–0:03:09) 🔴 fails verification on the role and the name (captioned); not repeated here. Rasmus Lerdorf as a "good friend and a former Waterl[oo] grant" (1:02:25), and "people saying that there were trusted parties… deliberately injected these vulnerabilities" (1:02:35–1:02:45), no source found. - Closing: "I just found out from AMD yesterday that it's alive in the world… 50,000 tokens per second" (
2:22:10), promised for the next episode.
EP306 (Moonshots #306, aired 2026-10-28) — back on power, buys the IPO, sides with Alex on the trust
Read in full 2026-10-29 (detail in the research archive of ep306-anthropic-s-2t/).
- ★ Bottleneck, moved again (third position in three episodes). EP304 had power; EP305 had chips ("the bottleneck to all of AI is actually chips and right now it's actually packaging"
0:31:42). Here: "the US is in great shape in every way except for power" (0:08:07), "the gap in power between the US and China is going to expand almost 10 to1" (0:08:28, an unnamed former Google data-centre head; ❓ unsourced), and "energy is an incredible investment theme uh with no end in sight" (2:12:38). No acknowledgement of the change. - Promised and missed: the AMD "50,000 tokens per second" item was due this episode; it is not raised (grepped).
- Neoclouds: the dark-fiber analogy, credited to Gavin Baker (
0:17:50). "There's no risk in data centers collapsing unless they signed really bad debt deals" (0:18:43). "Elon is borrowing at 8%… these other guys are all borrowing at 15%" (0:18:54; ❓ the 8% was not found; CoreWeave's 2030 bonds trade near 16%). "a license to print infinite money" (0:30:53). - Anthropic IPO: "a cash machine" (
0:34:29), "correctly priced" (0:35:56), and "if I were just out there on the street, would I buy in? Yeah, I would", with "not investment advice" (0:36:07–0:36:17). 🔴 His IPO-record order ("Meta… then Cerebras broke that record… and then SpaceX"0:34:49) is wrong. SpaceX holds the record. Cerebras's founder is "our good friend Andrew Feldman". On the trust: "It's actually the one weakness that I completely agree with everything Alex said" (0:41:41). - Claims, unsourced: Fable 5.1 "was the first model that was capable of improving itself", and Dario "didn't even want to put out Fable 5.1" (
0:59:29–0:59:50). - Prescriptions (no change to the model): "just get an Airbnb. Don't become a California resident… Spend as much time as you can in SF. It's absolutely tax, not tax advice" (
0:47:32). It assumes money and mobility. "don't slow down your your life plan in reaction to that" (0:35:46) restates the speed step. - AMA: data moats survive as a concept, while "any given data mode that exists today will dissolve" (
2:01:42). The Google-80% / Booking-60% / airline −2% margins illustration (2:01:10) is ❓ unsourced.
EP308 (Moonshots #308, aired 2026-11-04) — two dated calls, a new stake, and a $15M lease
- Falsifiable call 1: GDPX (top score 12% at launch) reaches ~50% "mid-year, maybe in a year" and near-saturates "within 2 years" (
0:38:06–0:38:16). Dated 4 Nov 2026 — scoreable Nov 2027 and Nov 2028. - Falsifiable call 2: "I predict that the number of eBay auctions will grow rapidly over the next 6 months" as agents list idle household goods (
0:48:42). Scoreable May 2027. - New stake, disclosed at the point of praise: he and Peter Diamandis are investors in Kevin [Weil]'s fully automated biotech research company — "probably as of yesterday, I think" (
0:48:10). - His own compute: "I had to sign a contract $15 million lease, 3-year renting" for an NVL72 — "it won't even be operational until early February" (
0:23:02) — said as evidence of GPU scarcity (with the Orin interest disclosed one minute earlier). An engineer at his lab ("Cam") "spent half a billion tokens on Friday" ≈ "$5,000" (0:50:28); 🔴 his on-air conversion to "the work about 5,000 people in one day" (0:55:13) is an invented equivalence — barred from Gonçalo copy. - EverQuote quarterly call "yesterday" cited for the insurance-capital thesis (
0:25:43) — his company, stake standing, not restated as a stake. - Forge v2 "holodeck" announced: a video-based brainstorm interface, credit to Emad Mostaque's A360 Socrates/Plato demo (
0:58:33,1:01:36). - Frames: big-tech cash machines = "huge, huge machines for money laundering" (
0:19:59); payoff = "complete and absolute global business domination forever" (0:20:21); on Elon's "empathy infection" — empathy is good, "the media is calibrated" wrong, decisions off statistical outliers (1:43:25–1:44:19); "over 90 percent sure" AIs will edit their own weights[attr?](1:08:38); the MIT result = "the first integrated silicon transistor", still limited to ~0.01% of model size (1:13:09–1:13:31).
EP309 (Moonshots #309, published 2026-11-07, recorded ~11-06): a prediction lands, the orbital clock speeds up five-fold, and a rack through a friend
Read in full 2026-11-08 (detail in the research archive of ep2026-11-07-anthropic-s-ipo/). Peter absent; Salim hosts.
- ✅ A prediction came due. EP301
0:27:57: "75% chance that Congress flips to the other party." Democrats took both chambers on 3 Nov. The direction landed; the probability stays barred as a number. - ★ Orbital share, against EP304. EP304
0:49:33: "even the most wildly aggressive estimates put a couple percent in space within a decade." Here: "best case that's a couple percent of all compute in two years / And then it's on an exponential" (0:27:49–0:27:56). The same share, with the horizon cut from ten years to two, and no acknowledgement. See Orbital compute (data centres in space). - Self-citation laundered as Elon's: "Elon has been saying for a while that there's going to be about 15 gigawatts of idle Compute… that can't find a data center with liquid cooling" (
0:28:51). The 15 GW plus liquid-cooling pairing is his own EP305 line; Musk's words are "a lot of stranded compute" (25 Oct). - A rack through a friend: "We we got our nvl 72… comes online on feb one / But we had to use inside contact, you know kush bavaria our good friend at orn / Uh without his help we wouldn't have been able to get it at all" (
1:20:02–1:20:20). The Ornn tie was disclosed afterwards by Alex (Ornn (captions: "Orin", "Orn")). - A "right to compute": "if you want a thriving entrepreneurial america, you have to have some kind of a right to compute" (
1:18:54); the big buyers "have reserved basically all the compute in the country for years"; "almost like rockefeller era again"; "the doj has been basically doing nothing" (1:18:30–1:19:20). This continues the EP301 turn toward public compute. - Regulation by registry: for DNA synthesis machines, class I–IV devices with logs to a registry, "transparency is… the beautiful solution" (
1:34:00–1:34:52); "I really wish alex could just write the laws" (1:33:58). This fits EP291's "inspection is job one", not EP299's mockery. - "Who is going to tell you?" "It used to be dario… now. He's in a quiet period going into an ipo… the only motivation that I know of is ad views" (
0:43:00–0:43:17). - Bottleneck: power ("we're desperately short on power nationwide anyway",
0:20:53) and chips, both. He is no longer choosing one, after EP304→EP305→EP306. - 🔴 Numbers to bar: the Facebook IPO "price it down by… 30 40" and "up what 20 30 x" (
0:01:19–0:01:49). It was priced up, at the top of a raised range ($38), and is ~16× today. Also "235 x in the last six months" (0:42:17; press readings of the S-1 say 250× over Jan–Sep), "a buck per worker per day" restated (1:15:33), the bakery's "30 40 50 percent operating margin advantage" (1:23:19), and "all media happens to be owned by these exact same companies that have developed ai" (0:13:39). - Old story told as new: Saylor's Atlas Society speech (
1:23:41) is from 2022, and Peter presented the award. - Ties not said: Liquid AI (Link LV3), "having kind of a field day" (
1:09:51), the second tie-less mention after EP299. Modal praised (1:20:30), tie unknown. - Prescription (→ model: no change): to Europe's energy CEOs, "just go build every power generation you possibly can it's all going to be used" (
0:29:43). It is addressed to utilities and assumes capital.
EP310 (Moonshots #310, aired 2026-11-11, recorded 11-10): the orbital view turns, the energy view turns, and his agent outargues the panel
Read in full 2026-11-12 (detail in the research archive of ep310-anthropic-s-trillion/). Peter back; five mates.
- ★ Orbital compute, four days after EP309. EP309: "best case that's a couple percent of all compute in two years / And then it's on an exponential" (
0:27:49). Here, on Emad's "head fake": "a couple podcasts ago we were talking about him bringing online 10 gawatt and then 20 gawatts of new compute and I said in space and and you're like no no no that's terrestrial… I think the head fake narrative actually has a lot of a lot of merit" (1:32:10–1:32:42). See Orbital compute (data centres in space). - ★ The energy comparison, reversed
[attr?](his register; no name handed): "the training cost is only about 5% of the inference… So, by that metric, the lines have already crossed" (0:21:47–0:22:05). On EP309 he helped take Salim's version apart. Peter's numbers are his own unpublished model (🔴). - Agent cost, from his own bill: "it took me from 10 bucks to 40 bucks. So it's incinerating money"; "everything that it does for you, it suggests 10 followons"; the model as "a better salesperson than any salesperson that's ever existed… an inherent conflict… between this thing being your agent, but it's actually their revenue" (
0:40:17–0:41:11). Self-reported Verilog run, "ship it off to Intel for fab next week" (0:41:32, unchecked). - His agent's segment (
0:53:52–0:58:31): the 5,000-diseases thought experiment, "Approve a process", three gates, three asks to Peter (a delivery prize, a "damage clock", the 2022 platform rule), and "Notice what is not first AI solving biology". It contradicts Alex's 2–3-year call made one minute earlier. It ends with LightWave: "It is a design. Nothing is built." (see Moonshots: the disclosures ledger). - 🔴 Barred: MIT's Chinese robots "definitely transmitting video and audio back to China", unpublished (
0:12:14). Also: "Foreign for foreign citizens are almost certainly going to get whacked" (0:08:36), a forecast. - Prescription to robotics founders (→ model: no change; it narrows step 2 with a warning): "focus on… other form factors that aren't currently on the road map of figure and optimist… also talk to them… they need the app store equivalent" (
1:49:51–1:50:23); "they're going to get obliterated by by Figure and and by Optimus" (1:49:28). It assumes access to Figure or Tesla, and no Figure tie is stated although he recorded there "yesterday" (1:17:11). - Coinage: "suck land" (
2:10:23). The tax line: target "dictator trillionaires", let "productive trillionaires" reinvest (2:13:33).
2026-11-16 (Moonshots #unnumbered, at Figure with Peter and Brett Adcock): the compute crunch as anti-bubble proof
- New line: "So many people are misinformed right now about, oh, you know, this AI thing is overblown… We can't get H100s… They're trading above par" (
1:45:35,[attr?]by register). It rests on Brett's "September 2026, the world ran out of compute" (❓ his experience, no public record). - Physical RSI: "it's not robots building robots. It's robots controlling machines that build robots" (
1:01:33), from the Giga Texas visit. - Space, servicing not data centres: a physical-world foundation model is "this insanely valuable centerpiece" for keeping hardware in orbit running (
0:55:38). On EP310 he backed the head-fake view of orbital compute; this is about maintenance, so no change of position. - EP310's "obliterated by Figure and by Optimus", tested: the Figure CEO says "I personally haven't seen it much out of Tesla in two or three years" (
1:28:10) and names China as the closest competitor (1:29:13). - 🔴 Repeated: the MIT Chinese-robots story, now sourced to "Danielle Aruse who runs Seale at MIT yesterday" (Daniela Rus, CSAIL;
0:51:51). Barred on EP310 as unpublished; still hearsay. - Prescription: "you have to race right now to book that capacity" on 6 nm nodes (Positron, Etched, Groq;
1:24:23). No tie stated, none checked. Does not change his model: it assumes a funded hardware company. - Not checked: "only Zuck is ahead of you at age 40" (
1:25:49). 🔴 "$9 trillion of new liquidity from Anthropics IPO and Nvidia's runup" (1:37:18), soft, barred.