That’s the tension under this a16z conversation between Baker, CIO of Atreides Management, and general partner David George. Public AI stocks fell for two months while usage and revenue kept climbing. Their contrarian read: the market should be bracing for a shortage, not a glut.

• Zero deteriorating data points. Baker found none across the AI industry in July and August, even as AI stocks fell for two months Gavin @ 1:22.

• A $360B swing lever. Frontier labs can cut reported revenue from $480B to $120B by shifting compute from inference to training Gavin @ 6:20.

• $15B against $50B. NVIDIA’s equity check per data center, the rest debt-backed by its own guarantees, a moat as real as its chip lead Gavin @ 64:11.

• No spare capacity through 2028. Political delays push that further out, making the real risk higher AI prices, not a glut David @ 29:40.

This breakdown is for paid subscribers. What follows covers the revenue lever, the undersupply case, NVIDIA’s financing moat, and SpaceX’s orbital bet.

Read more

• Visser’s $700 trillion framework implies Bitcoin at $600,000 to $1 million within a decade, but he separately expects Bitcoin’s own crypto-market share to shrink. Jordi @ 51:21

• He now holds more Ethereum than he ever has, as new capital shifts away from Bitcoin. Jordi @ 54:45

• Nvidia crossed $1 billion of revenue per day and guided to 70% growth next year, yet its muted 25% one-year return reflects opportunity cost, not weak conviction. Anthony @ 27:53

• Visser’s Bitcoin-as-AI-hedge thesis independently converges with an AI-safety researcher’s warning of a 99.99% chance AI destroys humanity. Jordi @ 23:46

Paid subscribers only. Below: the assumption buried in Visser’s price-target math, the kill conditions on all three trades, and where Bitcoin’s real edge sits once you strip out the headline number. Join for full access.

Read more

• Nvidia’s 96.2B quarter and its 70% guide against a 45% Street estimate cut against the AI-capex-bubble thesis.

• Salesforce jumped over 20%, backing up Chamath’s May call that “systems of record” would survive the SaaS-apocalypse narrative.

• Nvidia put roughly 20B toward acquihiring Poolside and buying Hugging Face, moving to lock up open-source distribution before OpenAI’s AMD deal reshapes the board.

• Friedberg puts the government on the hook to refinance 10 trillion in debt within 12 months, and Chamath calls a 6% 30-year yield the opening of a fiscal death spiral.

This breakdown is for paid subscribers: what complicates Nvidia’s bull case, where the systems-of-record thesis breaks, and why Moderna’s price tag looks like a short. Join to get full access.

Read more

This breakdown is for paid subscribers. Below: how the AI trade feeds into Bitcoin, the $50,000 gold option, Hayes’s portfolio ratio, and why he’d take Ethena over Circle. Join to read on.

Read more

Neil co-founded Sail Research, a “token factory,” and spent an hour with Patrick O’Shaughnessy on Invest Like the Best explaining how. The habit traces to his NVIDIA years building GPU kernels (low-level chip code) from 2016: chase the hardware’s absolute limit, not the limit relative to whoever you’re competing with.

• Background beats chatbot, 90-10. Neil predicts background AI agents overtake chatbots, from roughly 50-50 today toward 90-10 within a few years Neil @ 6:02.

• The KV cache is off by 10x-100x. Neil names it, not MoE sparsity, as the industry’s biggest remaining efficiency gap Neil @ 63:40.

• NVIDIA’s perf-per-watt has plateaued. Hopper to Rubin barely moved the needle, and Intel is at worst 2x behind TSMC, so a Taiwan shock would be a supply shock, not a catastrophe Neil @ 76:31.

• Sail is building a custom KV-cache chip. It offloads the cache to flash memory, hedging the HBM shortage Neil calls the industry’s tightest bottleneck Neil @ 74:21.

The Bottleneck Nobody Is Pricing

Neil ranks where AI compute gets wasted, dismissing the usual suspect first. MoE (models that fire only a slice of parameters per token) is already sparse, near 1% of experts active per token, with little room left to squeeze Neil @ 63:40.

The KV cache, memory that stores a model’s running context, is a different story.

We’re storing many kilobytes of data in the KV cache per token, and that’s probably off by an order of magnitude or two. Neil @ 63:40

DeepSeek’s research has compressed the cache by roughly another order of magnitude, repeated year over year. Sail’s answer: a custom chip offloading the cache to flash memory, aimed at one to ten tokens per second Neil @ 74:21.

Whoever closes that gap captures more value than the next chip generation. That’s the line to watch, not the next flops number.

NVIDIA: The Risk Premium Neil Says Doesn’t Add Up

The Street prices AI infrastructure with a real Taiwan-TSMC premium baked in. Performance-per-watt on core AI math hasn’t moved much from Hopper to Rubin, and Intel’s best process is at worst 2x behind TSMC on that metric.

If we lost access to TSMC for any reason, it wouldn’t be that bad. The gap is just far smaller than you would make it out to be if you follow the chip discourse. Neil @ 76:31

He’s still bullish on NVIDIA short term. A performance plateau across three chip generations, if it holds, argues for less geopolitical premium in AI infra names than the market assigns. One operator’s claim, not an audited number.

AMD’s Discount Is Real, and Closing

The Street prices AMD as NVIDIA’s cheaper, harder-to-program alternative. Neil agrees the gap is real, calling it “music to my ears,” an arbitrage his kernel team captures cheaply.

There’s no bad chips, there’s only bad pricing. Neil @ 44:57

That window is narrowing. AMD is “somewhat popular amongst some large buyers,” and he names Meta and OpenAI as buyers of real AMD volume, tightening the supply he’s arbitraging Neil @ 44:57. Sail’s real edge is speed porting to a new chip Neil @ 47:52, not a permanent lock.

The AMD discount is a trade with a shrinking half-life. More buyers on the same arbitrage is bullish for AMD’s re-rating, and bad news for anyone entering late.

Cerebras Isn’t the NVIDIA Killer

Cerebras and Groq get framed as a direct threat to NVIDIA’s dominance. Neil’s data argues otherwise. Cerebras’ on-chip SRAM (memory built into the chip) runs at roughly 21 petabytes per second, versus about 10 terabytes per second of HBM (memory stacked on a GPU) on NVIDIA’s Blackwell Neil @ 23:43.

His read: the winning system pairs the two, fixed model weights on SRAM-heavy chips like Cerebras, the growing KV cache on GPUs.

You should think of them as accelerators, used in conjunction with a more traditional GPU-like device. Neil @ 30:57

That’s not NVIDIA versus Cerebras. It’s NVIDIA plus Cerebras. Size it as a complement, not a bet against NVIDIA’s share.

What Breaks the Thesis

Advising hardware founders, Neil names four chip-supply bottlenecks: TSMC wafer, HBM, packaging, power. HBM is the one he’d attack, since new memory fabs take years to come online Neil @ 79:04. “The boys in Boise,” his term for Micron, “don’t love huge capex for cyclical” demand, and have been burned before.

iPhones are going to go up in price and we’re just going to deal with it. Neil @ 80:18

If HBM supply doesn’t ease, every KV-cache-heavy strategy stays squeezed, Sail’s flash-offload bet included, and memory margins stay elevated. Watch iPhone component pricing as the leading tell.

One flag: no revenue, funding, headcount, or customer count beyond Parallel Web Systems is disclosed. Neil’s cost-per-token numbers are his own, and the 95% uptime bet is unproven.

Alpha Takeaway

• KV cache compression, not the next GPU, is the real prize. MoE is tapped out, and the cache runs 10x-100x oversized.

  • Implication: Watch DeepSeek’s compression work and Sail’s flash chip next.

• AMD’s discount is closing. Meta and OpenAI already absorb the supply Sail is arbitraging.

  • Implication: Model AMD’s mispricing as shrinking, not durable.

• NVIDIA’s perf-per-watt has plateaued for three chip generations. A Taiwan shock would be a supply hit, not a re-rating event.

  • Implication: Stress-test the Taiwan premium in AI infra names.

• HBM, not TSMC, is the bottleneck Neil would attack first. Memory makers are reluctant to fund cyclical capex.

  • Implication: Watch iPhone pricing as confirmation scarcity isn’t easing.

Bottom line: Sail’s argument: the industry solved the wrong problem, chatbot speed instead of background-agent cost, leaving the KV cache and memory supply as the two places a patient operator can still find edge.

This newsletter is for informational and educational purposes only. It is not investment advice, not a research report, and not an offer or solicitation to buy or sell any security.

Any valuations, price scenarios, projections, and forward-looking statements are hypothetical illustrations based on assumptions that can change without notice. They are not price targets, forecasts, or guarantees of future results. Model outputs depend heavily on their inputs, and small changes in assumptions can produce large changes in the results. Actual outcomes will differ, often materially. Do not rely on any figure here as a prediction or a basis for an investment decision.

Information is drawn from third-party sources we believe to be reliable, including public filings and podcast interviews, but we do not guarantee its accuracy or completeness. All views attributed to named speakers reflect their own opinions, not Podcast Alpha’s.

This content is general in nature and does not account for your financial situation, objectives, or risk tolerance. The author may hold positions in securities mentioned. Consult a licensed financial, legal, or tax advisor, and do your own research, before making any investment decision.

Lee is head of research at Fundstrat and chairman of Bitmine Immersion Technologies (BMNR), the largest corporate Ethereum treasury. He joined the Milk Road Show days after Ethereum’s largest short-covering rally in history to explain why he still calls ETH cheap.

• Bitmine’s $4 billion buyback retired nearly 20 million shares in five weeks, on top of 65 straight weeks of ETH buying Tom @ 20:53.

• Lee expects the ETH/BTC ratio to clear its 2021 high of 0.08 by year end, a scenario that puts Ethereum near $6,000 Tom @ 32:56.

• Four catalysts could drive that, led by the Clarity Act in September Tom @ 32:56.

• His “uncanny valley of wealth” thesis says AI agents need crypto rails to stay auditable, though he admits he can’t say for sure Tom @ 5:02.

This breakdown is for paid subscribers. Below: the buyback mechanics, the dated catalysts behind the ratio call, and what breaks this trade. Join to get full access.

Read more

Give twenty people a billion dollars now and they put all of it to good use. A decade ago that money would have gone to waste, and Martin Casado argues the market hasn’t caught up to why Martin @ 39:48.

Casado leads enterprise and infrastructure investing at a16z. Sinofsky once ran Windows at Microsoft. The two spend an hour in disagreement, so their points of agreement carry weight.

This breakdown is for paid subscribers. Below: why Google’s model gap may be a cultural problem money can’t fix, an unverified token-rationing rumor worth tracking against earnings, and a one-line filter for every AI math headline. Join to read on.

Read more

Wall Street prices Robinhood as a beta play on retail trading and crypto volume, betting AI-era enthusiasm keeps feeding both. Tenev spent two hours with Jack Selby and Graham Stephan on The Iced Coffee Hour making that pitch: $125 billion in new assets, agentic trading, a private-markets fund, a tokenized blockchain. Twice, he broke from the script to name speculative excess in the trade his own products depend on.

• Tenev names hedge funds buying AI chips purely to resell as a live bubble risk sitting next to genuine compute demand Vlad @ 61:13.

• Robinhood added $125B in assets under custody in a year, and over 100,000 customers already trade through AI agents via an MCP server Vlad @ 11:55.

• Robinhood Ventures gives retail a stake in OpenAI and a Y Combinator seed fund, betting the accredited-investor rule gets rewritten Vlad @ 40:37.

• Robinhood Chain is already top-five in DEX volume weeks after launch, trading tokenized Tesla and Nvidia shares in 120+ countries Vlad @ 51:21.

This breakdown is for paid subscribers. Below: the resale mechanism Tenev says can tip a sound trade into a bubble, why Robinhood Ventures only pays off if Washington rewrites a decades-old rule, and the prediction-markets criticism he never fully answers. Join to get full access.

Read more

This breakdown is for paid subscribers. Below: the part that caps AI centralization, who out-earns both labs per megawatt, and the debt bill one economist calls a second Volcker shock. Join to read on.

Read more

Mike Stonebraker won the 2014 Turing Award for building Ingres and Postgres, the systems every enterprise database still runs on. Now running DBOS, he told Imagination in Action hosts John Werner and Alex Wissner-Gross that agentic AI’s real bottleneck isn’t model capability, it’s that agent frameworks throw structure away the moment they touch it.

• Agentic AI converts table joins into text, Stonebraker says, calling it “a terrible idea” and naming Anthropic’s text-only design as a structural weakness Mike @ 3:57.

• Only about 20% of enterprise AI pilots reach production, and text-to-SQL is “probably never” solved for a warehouse like MIT’s Mike @ 17:11.

• DBOS puts durable, reversible state inside the database for write-capable agents, borrowing a 1985 pattern called sagas to undo a bad agent write Mike @ 23:47.

• Open source beats closed labs once accuracy converges into “answers per dollar,” Stonebraker predicts, and separately says China overtakes the US within 15 years, disputed live by Alex Wissner-Gross Mike @ 75:19.

The Unpriced Risk Inside Anthropic’s Enterprise Pitch

Wall Street’s live Anthropic debate is IPO pricing: is the reported $2 trillion valuation sandbagged or fair? Stonebraker’s critique sits outside that conversation, exactly why it matters.

He calls Anthropic’s text-only architecture “one of their big flaws.” His case study: Munich’s Department of Transportation needs regulations, a CAD diagram, a trolley schedule, a map, and a school calendar just to answer one crosswalk complaint. Traditional agentic AI failed completely: everything but the regulations is structured data, lossy the moment it becomes text Mike @ 3:57.

Having done a bunch of agentic AI, if you have two data sources and you need to join them together, they will get turned into text, and you will do the join as text, and that throws away structure. Mike @ 3:57

No analyst pricing Anthropic’s enterprise pipeline is underwriting this constraint. It leaves consumer revenue untouched and hits the enterprise-data-agent business the IPO story leans on.

One point cuts the other way: enterprises sink roughly 95% of IT budget into unreadable COBOL Mike @ 51:53, and Stonebraker credits Claude for making full legacy-code rewrites viable Mike @ 48:14.

The Production Rate Nobody Prices In

Every AI capex bull case assumes pilots convert. Stonebraker puts a number on how many do: roughly 20% of enterprise AI pilots reach production Mike @ 17:11, his own estimate, not a cited study.

Asked when text-to-SQL gets solved for MIT’s own warehouse, he does not hedge.

Probably never. Mike @ 15:37

The reasons repeat everywhere: unreadable column names (”schema rot”), joins far more complex than benchmarks test, and no model trained on data behind a corporate firewall Mike @ 11:57. His fix runs on-prem or open source, since most enterprises won’t hand data to a closed model Mike @ 21:26.

The 20% figure, not pilot-announcement volume, is the number to track.

Open Source vs Closed: A Trade With a Trigger

The Street prices Anthropic and OpenAI on capability leadership. Stonebraker’s prediction has a trigger, not a vibe: once models converge to within a few percent on accuracy, competition shifts to “answers per dollar,” and he names DeepSeek’s cost advantage as the edge that wins Mike @ 40:25.

He has lived this before: Ingres was outgrowing Oracle in 1984 until IBM set the SQL standard and Oracle adopted it faster, pulling ahead for good Mike @ 36:22. His read: ownership structure decides these fights, not code quality, the same reason Postgres outlasted Oracle-owned MySQL.

It’s sad that OpenAI and Anthropic don’t take that point of view. Mike @ 40:25

Watch the convergence point, not the sentiment - the trade only triggers once accuracy actually converges.

DBOS and the Write-Capable Agent Gap

DBOS is private, no ticker, but the category it bets on isn’t: what happens once agents start writing data, not just reading it. Stonebraker, who co-founded DBOS with Databricks’s Matei Zaharia, frames durability as the missing “D” in ACID: state lives in the database, recovery resumes from the last completed step Mike @ 23:47.

If bad things happen, you don’t want to back up to the beginning. You want to back up to the last completed set of steps. Mike @ 23:47

His sharper example: a workflow overpays an employee and must undo it after someone else touched the record, the 1985 “sagas” pattern applied to modern agents Mike @ 28:36.

If write-capable agents scale as Stonebraker expects, durable-workflow infrastructure is a separable category worth tracking regardless of which model lab wins.

What Breaks It: The China Wildcard

Stonebraker’s biggest kill condition is macro: he tells listeners to learn Chinese, since SIGMOD authorship flipped from 80% US to roughly two-thirds Asian in twenty years Mike @ 75:19. Alex compares it to the 1980s fear that Japan would overtake America, a trend line that also looked airtight and did not hold Alex @ 75:55.

If you just view conference proceedings as a leading indicator, I think we are being overtaken by the Chinese. Mike @ 75:19

Neither man concedes. The unresolved split is the real read for anyone holding US AI names: China-catch-up risk is live, but a trend line alone does not settle it.

Alpha Takeaway

• Anthropic’s text-only architecture is a named, unpriced risk - the Munich case shows multi-source joins collapsing into lossy text.

  • Implication: Discount any Anthropic enterprise pitch that skips structured, multi-source data.

• Only 20% of enterprise AI pilots reach production - Stonebraker’s own estimate, delivered without a citation.

  • Implication: Track pilot-to-production conversion, not pilot-announcement volume, in the AI capex trade.

• Open source wins only once model accuracy converges into “answers per dollar” - the pattern that let Postgres outlast Oracle-owned MySQL.

  • Implication: Model DeepSeek and open-weight cost curves against Anthropic and OpenAI’s pricing power.

Bottom line: Anthropic’s nearest real risk is a structural enterprise-data gap, not IPO pricing, and nobody has priced it in.

This newsletter is for informational and educational purposes only. It is not investment advice, not a research report, and not an offer or solicitation to buy or sell any security.

Any valuations, price scenarios, projections, and forward-looking statements are hypothetical illustrations based on assumptions that can change without notice. They are not price targets, forecasts, or guarantees of future results. Model outputs depend heavily on their inputs, and small changes in assumptions can produce large changes in the results. Actual outcomes will differ, often materially. Do not rely on any figure here as a prediction or a basis for an investment decision.

Information is drawn from third-party sources we believe to be reliable, including public filings and podcast interviews, but we do not guarantee its accuracy or completeness. All views attributed to named speakers reflect their own opinions, not Podcast Alpha’s.

This content is general in nature and does not account for your financial situation, objectives, or risk tolerance. The author may hold positions in securities mentioned. Consult a licensed financial, legal, or tax advisor, and do your own research, before making any investment decision.

That trade-off says more about how Altman runs OpenAI than anything he says about safety. He spoke with David Senra, of the history podcast Founders, for two hours spanning Toby Lutke’s AI letters to OpenAI’s earliest days.

• OpenAI killed Sora and the Atlas browser to protect Codex’s compute, confirming a platform bet over a product portfolio Sam @ 52:11.

• Altman names power concentration alongside loss of control as his two biggest AI fears, calling AI safety’s “trust us” pitch anti-human Sam @ 30:21.

• Peter Thiel, not anyone inside OpenAI, is why ChatGPT never got diluted. He told Altman the “empty Google text box” was the point Sam @ 57:15.

• Altman admits he still resists using his own product, Codex, in daily work, calling it “psychologically inconsistent” Sam @ 6:07.

This breakdown is for paid subscribers. Below: the Sora and Atlas compute math, the Thiel conversation that saved ChatGPT, and the founding contradiction Altman never resolves. Join to get full access.

Read more

Stanford cryptographer Dan Boneh laid out the real roadmap at a16z crypto’s research lab, hosted by Tim Roughgarden, against another week of “quantum computers will kill Bitcoin” headlines, a category Boneh says produces new entries “every single week.” His answer carries real costs, including a cryptography bet he lost.

• Bitcoin’s Shrinks scheme grows signatures ~10x (64 bytes to hundreds), with the BIP drafting this summer and the block-size debate unresolved Dan @ 64:28.

• Hash-based beat lattice-based for both chains, despite Boneh’s own advocacy, trading away hardened key derivation Dan @ 0:29.

• Ledger and Trezor are “quite opposed” to hash-based signatures today, a chip-speed gap Boneh expects future hardware to close Dan @ 26:47.

• A new threshold scheme lets Fireblocks-style custodians use lattices privately, with the chain only ever verifying a standard signature Dan @ 37:27.

The FUD Is Noise. Boneh Lost the Real Argument.

Boneh names the noise directly: Jim Cramer selling Bitcoin over quantum risk, a D-Wave press line claiming quantum computing will “destroy” proof-of-work, all from a single week Dan @ 0:29. D-Wave’s [QBTS] own marketing helped seed that fear trade, and Boneh files it next to Cramer’s call: noise.

Boneh spent years advocating for lattice-based signatures, since lattices preserve the algebraic structure ECDSA and Schnorr already give Bitcoin and Ethereum. He lost that argument.

“For many years, I’ve been advocating for lattice-based signatures for blockchains. But unfortunately, I don’t think that’s where the world is going.” Dan @ 0:29

The reason is “fear of assumptions,” sharpened by an earlier lattice proposal called Hawk running into problems. Core developers are “quite excited and quite happy” to rest the whole chain on one thing: SHA-256.

Hardened key derivation, deriving one public key from another without exposing the private key, is “fundamentally not possible” with hash-based signatures Dan @ 31:37. That door is now closed on both chains.

The Number That Matters: 10x

Boneh gives a falsifiable figure. Bitcoin’s proposed scheme, Shrinks, produces signatures running hundreds of bytes, up from 64 today.

“The signature cost is going to go up by a factor of 10.” Dan @ 64:28

NIST’s full standard runs 8-kilobyte signatures, too big for a blockchain, so Shrinks trims it with a stateful, shorter-lived key design. Even trimmed, it’s roughly 10x today’s footprint.

Block size is, as of this talk, unresolved. “Some of them are saying we’re going to have to go back to the block size wars,” Boneh says, but “how they actually deal with it is currently unknown” Dan @ 64:28. That fight split Bitcoin once, in 2017, and is about to reopen.

Ethereum picked a different fix: SNARK-ify every signature, one proof every 12 seconds. Bitcoin’s culture treats that as radioactive, and core devs “have a heart attack when you ask them about this” Dan @ 64:28. The Street usually calls Ethereum the more technically sophisticated chain, but Bitcoin’s fix is “literally getting codified” this summer while Ethereum’s LeanSig “is a bit under flux” Dan @ 19:32.

The Bottleneck Is Hardware, Not Math

A chain-level fix doesn’t help if the hardware can’t run it. Ledger and Trezor, when Boneh talks to them directly, aren’t enthusiastic.

“If you talk to Ledger and Trezor and all those, they are also quite opposed to hash based signatures.” Dan @ 26:47

Current chips take “many tens of seconds” per signature, a gap Boneh expects hardware acceleration to eventually close. This is his secondhand account, not an on-record vendor claim.

A second constraint: stateful schemes need a counter tracking every use, a “foot gun” if a key ever repeats. Multi-chain wallets could run out of on-device memory for that state, forcing it onto a laptop and reintroducing the same risk. Boneh’s advice: pair any stateful scheme with ECDSA until quantum computers arrive.

A separate unpublished result fixes custody for the same problem. Existing ways to split a signing key either leak the signer count or need the chain to verify SNARK proofs, a non-starter for Bitcoin. Boneh’s scheme keeps the on-chain signature ordinary and hash-based while the key-splitting party privately uses a lattice mechanism “the blockchain is completely unaware of” Dan @ 37:27. He names Fireblocks as a likely adopter. Not yet peer-reviewed.

What Breaks It: A Resurrected Satoshi

Boneh saves the hardest problem for last, calling it “the elephant in the room”: what happens to Bitcoin in old or abandoned addresses, including Satoshi’s own coins, once ECDSA is retired.

The leading proposal seeds a “bait” address and waits: the first person to break it with a real quantum computer triggers a network-wide cancellation of ECDSA validation, and dormant coins then move through a still-undesigned recovery process. Boneh’s read on who breaks it first: “most likely it’s going to be one of the good guys.”

“In 2026, when there were no quantum computers, I created this transaction that shows I knew the secret key back in 2026. Therefore, you should give me my tokens.” Dan @ 67:53

Boneh calls the question “extremely complicated” and says the community wants to “delay that as much as we can.” Whoever writes the recovery procedure decides what counts as circulating Bitcoin supply once quantum computers arrive.

Alpha Takeaway

• A security upgrade is about to reopen Bitcoin’s oldest fight. The ~10x signature-size increase from Shrinks has core devs floating a return to the block-size wars.

  • Implication: Watch the Shrinks BIP over the next 12-24 months for how the tradeoff resolves.

• The custody bottleneck is hardware, not cryptography. Ledger and Trezor’s resistance to hash-based signatures is chip speed, not math.

  • Implication: Model a hardware refresh cycle as a real rollout cost.

• Hash-based, not lattice-based, is now the default. The researcher who argued for lattices for years concedes the industry chose SHA-256-only signatures, closing off hardened key derivation.

  • Implication: Reset expectations for what post-quantum wallets can do.

Bottom line: Bitcoin and Ethereum both have real post-quantum plans, but the risk sits in execution cost, heavier signatures, unready hardware, an unsolved abandoned-coin problem, not whether a solution exists.

This newsletter is for informational and educational purposes only. It is not investment advice, not a research report, and not an offer or solicitation to buy or sell any security.

Any valuations, price scenarios, projections, and forward-looking statements are hypothetical illustrations based on assumptions that can change without notice. They are not price targets, forecasts, or guarantees of future results. Model outputs depend heavily on their inputs, and small changes in assumptions can produce large changes in the results. Actual outcomes will differ, often materially. Do not rely on any figure here as a prediction or a basis for an investment decision.

Information is drawn from third-party sources we believe to be reliable, including public filings and podcast interviews, but we do not guarantee its accuracy or completeness. All views attributed to named speakers reflect their own opinions, not Podcast Alpha’s.

This content is general in nature and does not account for your financial situation, objectives, or risk tolerance. The author may hold positions in securities mentioned. Consult a licensed financial, legal, or tax advisor, and do your own research, before making any investment decision.

Murdock, founder of Insight Partners, invested through the dot-com bust, 2008, and the cloud buildout. On 20VC, he tells Harry Stebbings: this correction is credit, not demand.

• October 2026 to March 2027 is Murdock’s correction window, a credit dislocation tied to the Iran war, not an AI demand failure Jerry @ 2:44.

• At least half of today’s neoclouds disappear within 36 months, while hyperscalers survive and buy the wreckage cheap Jerry @ 7:20.

• Fireworks beats Base10 roughly ten to one on capital efficiency, Murdock’s screen for the whole inference layer Jerry @ 8:30.

• Anthropic and OpenAI’s $100-150B mega-rounds may prove cheap, but neoclouds, app-layer AI, and legal AI don’t get the same pass Jerry @ 34:43.

• Nvidia is a buy to $10 trillion in five years, its flat chart just circular hyperscaler transactions hiding real demand Jerry @ 59:51.

This breakdown is for paid subscribers. Below: the capital-efficiency screen, the mechanism behind the Nvidia call, and two security names still unpriced. Join to get full access.

Read more

• Bitcoin’s move computes to seven sigma, the third such move in a decade, both priors preceding large rallies. Jordi @ 4:24

• The trigger was Bessent’s first yen intervention since 1998, paired with a debt-buyback language change, not a Bitcoin headline. Jordi @ 4:04

• Nvidia’s 80% gain since June 2024 landed in four of the last 26 months, Visser’s preview of compression spreading to Micron and the frontier labs. Jordi @ 14:53

• Stripe bought Open Router for a reported $7-8 billion, Ramp answered with router.com, both betting on the entrepreneurial economy. Anthony @ 29:52

Read more

For paid subscribers: the actuator supply chain that could gate Tesla’s humanoid bet, and why SpaceX’s own numbers argue against the story Wall Street prices it on. Join to get full access.

Read more

• The capital squeeze: doomerism, two governors’ data-center pushback, and Treasury doubling its bond buying hit labs that are not investment-grade, right as they need more compute Chamath @ 5:29.

• The ban mechanism: Sacks says Dario’s essay dodged the real charge, then maps a FINRA-style regulator’s path to a de facto open-source ban David @ 29:51.

• The RSI wildcard: Friedberg argues recursive self-improvement, if real, only needs chips, power, and a connection, making jurisdiction-based regulation pointless David F @ 47:33.

• The DSA math: the platform costs $71 to $212 trillion over a decade against $23 trillion in wealth above $50 million David F @ 66:47.

Below the fold: the dollar mechanism Chamath lays out, the three-step playbook that ends open-weight models without ever passing a ban, and the math that breaks the DSA platform.

Read more

This breakdown is for paid subscribers. Below: the actuator supply chain Kang says is the real trade, why Tesla is RoboStrategy’s own biggest risk, and the discount math BOT could sit through for years. Join to get full access.

Andrew Kang does not think chips are the bottleneck for humanoid robots. He thinks it’s a part smaller than a fist, machined to ASML-grade tolerances by a handful of companies most investors have never heard of.

Kang is CEO of RoboStrategy (NASDAQ: BOT), the closed-end fund we covered in July after it raised $100 million in institutional PIPE rounds at 3-4x NAV. He’s back on Raoul Pal’s Journeyman seven weeks later with sharper specifics on the supply chain, Washington, and the fund’s real numbers.

Read more

This breakdown is for paid subscribers. Below: the kill condition that ends this trade before the midterms finish, the currency-adjusted number that undercuts “weak dollar is bullish for stocks,” and why Moderna and Eli Lilly, not chased semiconductor names, are the AI trade left standing. Join to get full access.

Read more

This breakdown is for paid subscribers: what BVNK monetizes, the growth math needing no stablecoin bet, and the headcount question nobody asked. Join for full access.

Read more

That gap is Ryan Detrick’s own example of AI capex circularity (spending that becomes another company’s revenue, which funds more spending), raised unprompted while making the bull case. Detrick, Chief Market Strategist at $8 billion Carson Group, is defending a 15-18% S&P target and three sector picks, tech among them, with the same number a skeptic would use to call the top.

• Carson Group raised its S&P 500 target to 15-18% after staying overweight through a spring drawdown that took the index down 5% Ryan @ 0:15.

• Bank stocks are breaking out above their 2007 peak, which Detrick reads as bullish confirmation, not the pre-crisis warning it once was Ryan @ 7:53.

• The tech overweight leans on $1 trillion of projected 2027 hyperscaler capex, with Alphabet’s relayed $9-a-share beat from stake markups as the circularity risk underneath Ryan @ 22:30.

• Three contrarian mid-year calls, no Fed hike, a shallower correction, a stronger labor market, have all held Ryan @ 32:07.

The Bull Case That Survived a Drawdown

Carson came into the year with a 12-15% target. The index fell 5% in the spring, long enough that “a lot of long-term bulls were cutting their targets.” Carson didn’t, and raised its target to 15-18% by mid-year Ryan @ 0:15.

“We hung tough...and then we upgraded up, up, up to our target.” Ryan @ 0:15

One stat backs it: this year’s Q2 was the worst of 16 presidential-cycle quarters, then rallied over 15%, the best midterm Q2 on record Ryan @ 4:06.

Breadth Confirms It. Bank Stocks Are the Tell.

Detrick’s real technical argument is the advance/decline line (a running count of stocks rising versus falling each day), which is making new highs now, unlike its divergence before the narrow-leadership 2000 and 2007 tops Ryan @ 7:53.

He applies the same logic to bank stocks, which just broke out above their 2007 peak after 19 years of going nowhere, a warning sign then, bullish confirmation now across community, regional, and large banks, plus Europe. His reversal trigger: a close back below that level is a “major, major warning sign” Ryan @ 17:49.

One wrinkle: Phil Rosen’s last four guests flagged financials independently Phil @ 19:16. Crowded doesn’t mean wrong. It changes the sizing math.

Industrials and the Barbell Underneath It

Industrials is the second pick, chosen for the highest correlation to the broad market of any sector he tracks, a bet on being broadly bullish, not on one theme. “A lot of people talk about industrials. I don’t think so,” he says, pointing to a broad base of names, Uber and 3M among them, cheap valuations, and earnings upside few strategists discuss Ryan @ 16:37.

Underneath all three sits a barbell, momentum against low-volatility exposure since the 2024 election Ryan @ 12:07.

The Tech Overweight Is a Chip Trade in Costume

Tech is the third leg, where the central tension lives. Ryan’s case starts with history: tech has led most bull markets for a century. It continues with a number every guest now cites: the top five hyperscalers are projected to spend $1 trillion on 2027 capex, 2.8% of GDP, up from $600 billion this year Ryan @ 19:55. Phil’s pushback: is that sustainable, and is 2.8% of GDP too much in one trade Phil @ 22:10.

Ryan agrees, then argues one company’s capex is another’s revenue: hyperscaler cash flow is declining while semiconductor cash flow rises, and that divergence is the real trade. His example is Alphabet: over $9 a share against an expected $2 to $3, and Amazon’s hundreds of billions in profit, both from marking up private stakes in SpaceX and Anthropic Ryan @ 22:30, relayed rather than sourced. ⚠ Uncertain: treat the per-share number as unconfirmed.

Podcast Alpha found harder numbers on the same mechanism three weeks earlier: 85% of Amazon’s and 87% of Alphabet’s Q2 net income came from unrealized AI stake gains, not operations, per Aswath Damodaran (full breakdown). Detrick’s figure may be loose. The direction isn’t.

If Anthropic or SpaceX’s private valuation gets marked down, that gain reverses.

What Would Actually Change His Mind

The dollar, not the capex math, is the risk he actually watches, tracking the Dollar Index around 100, citing the 2022 bear market and the 2024 yen-carry unwind Ryan @ 27:37. He calls it the cleanest shirt in the dirty laundry: when something breaks beneath the surface, capital flees into dollar safety even as stocks, bonds, and gold sell off together. A sustained DXY break above 100, not the next capex headline, is his trigger to turn cautious.

“We think the Fed’s gonna run it hot. We don’t see any hikes.” Ryan @ 32:07

Three contrarian mid-year calls have each held:

• No Fed hike, even as core inflation broadens: components above 3% year-over-year rose from 41% to 54% last month.

• A shallower correction than typical: this year’s 9.1% drawdown came in well under the ~17.5% average midterm correction.

• A stronger labor market than headlines suggest: prime-age (25-54) participation holds above 80%, with job growth near 50,000 a month Ryan @ 32:07.

Alpha Takeaway

• The tech overweight is a chip trade wearing a hyperscaler costume. One company’s capex is another’s revenue, so the real exposure is chip cash flow.

  • Implication: Model the tech overweight as chip exposure, not hyperscaler earnings.

• 85-87% of Amazon’s and Alphabet’s Q2 net income came from stake markups, not operations, per Damodaran’s breakdown.

  • Implication: Strip stake gains before trusting either stock’s PE as cheap.

• The dollar, not the capex number, is what would change Detrick’s mind. A DXY breakout above 100 is his explicit trigger.

  • Implication: Track DXY next earnings season, not the next capex headline.

Bottom line: Carson’s target survived a drawdown and three contrarian calls, but its tech leg runs through the same markups a bear would call an earnings-quality problem.

This newsletter is for informational and educational purposes only. It is not investment advice, not a research report, and not an offer or solicitation to buy or sell any security.

Any valuations, price scenarios, projections, and forward-looking statements are hypothetical illustrations based on assumptions that can change without notice. They are not price targets, forecasts, or guarantees of future results. Model outputs depend heavily on their inputs, and small changes in assumptions can produce large changes in the results. Actual outcomes will differ, often materially. Do not rely on any figure here as a prediction or a basis for an investment decision.

Information is drawn from third-party sources we believe to be reliable, including public filings and podcast interviews, but we do not guarantee its accuracy or completeness. All views attributed to named speakers reflect their own opinions, not Podcast Alpha’s.

This content is general in nature and does not account for your financial situation, objectives, or risk tolerance. The author may hold positions in securities mentioned. Consult a licensed financial, legal, or tax advisor, and do your own research, before making any investment decision.