Barbell

Convert podcasts into alpha. Barbell mines podcasts and qualitative data for bullish/bearish takes.

[Barbell @barbell_fi]

@Restructuring__ In fairness, we’ve roundtripped to Gavin-on-ILTB levels

https://t.co/sOnaIkskes

[Barbell @barbell_fi]

Sam outlines two possible scenarios of compute oversupply:

1) Models become so efficient that human needs and attention are easily satisfied

2) A scaling wall prevents their ability to drive the cost curve down

Earlier in the conversation, Sam emphasized that "demand for AI at a sufficiently high level and a sufficiently low price was basically uncapped." The key phrase is "at a sufficiently low price."

If scaling laws stop working and costs can't continue dropping, then AI remains expensive, which caps demand at lower levels than the industry is building capacity for. In other words: the massive compute buildout happening now is predicated on costs continuing to fall dramatically through scaling, which would unlock exponentially more use cases and demand. If that cost curve flattens due to a scaling wall, you'd have built enormous capacity for a market that can't afford to use it all at the higher price point, creating oversupply.

However, his skepticism of this scenario was evident, calling scaling laws "the most hated prediction of all time."

"Everybody always wants to say they're going run out. They can't be like this. And yet it keeps going."

[Barbell @barbell_fi]

https://t.co/JZ93BLjpU4

[Barbell @barbell_fi]

https://t.co/JZ93BLjpU4

[Barbell @barbell_fi]

Sam Altman @sama on whether we have AGI:

“It's very hard for me to say what I want from this model that it can't do, but there are clearly some things. You can't yet go say like ‘cure cancer’ and get cancer cured. You can't yet say ‘go do this complicated physical thing’ in the robot. The model also, although brilliant, is still not learning continuously as it goes. And that feels to me like maybe not a hard requirement for AGI, but certainly something that I'd like.”

[Barbell @barbell_fi]

@GavinSBaker time to revisit his Dwarkesh ep
https://t.co/yxFrj9cwuF

[Barbell @barbell_fi]

Our notes on the conversation:
https://t.co/QYB1EK4jXf

[Barbell @barbell_fi]

Nvidia’s CUDA moat may be shrinking much faster than the market understands.

$CBRS CEO Andrew Feldman @andrewdfeldman told @mattturck that the change is already visible at the frontier.

Two years ago, virtually every state-of-the-art model was trained in a CUDA flow. Today, Gemini is trained on Google TPUs, Claude is trained on AWS Trainium, while OpenAI remains on NVIDIA.

By Feldman's rough count, losing 2 of the 3 major frontier labs is effectively losing 70% market share for state-of-the-art model training.

For inference, Feldman goes further: he says there is no CUDA moat whatsoever.

His argument is that inference can increasingly be switched between providers with relatively little friction. (Feldman emphasizes the "eight keystrokes" it takes to switch to Cerebras.)

As training fragments across GPUs, TPUs, and Trainium, and inference becomes an API-level decision, the walled garden of CUDA becomes less and less insurmountable.

His position isn't that NVIDIA will fail, but rather that the AI chip market has evolved into a healthy, multi-vendor ecosystem where "bigger chips are undoubtedly the best way to go" for AI workloads, and where architectural innovation matters more than software lock-in.

[Barbell @barbell_fi]

@tristanwalker congrats!

[Barbell @barbell_fi]

@edzitron @edels0n banger ep - we tried to organize the ideas:
https://t.co/sdKBr5z6gT

[Barbell @barbell_fi]

Chat with transcript: https://t.co/kv9HammbJI

Full interview: https://t.co/cacMYSspw2

[Barbell @barbell_fi]

Goldman’s base case: no Fed hikes this year.

Goldman Sachs chief U.S. economist David Mericle puts the odds of further Fed hikes at roughly 25%—well below the market’s implied probability.

Mericle interprets the market pricing for ~1.5 hikes as something like a 50/50 chance of two or three hikes.

Goldman sees current inflation as supply-driven (war, tariffs, measurement issues), not demand-driven overheating. The Fed has signaled they're done "litigating" inflation causes and will respond if it stays high, regardless of source.

Mericle expects inflation to soften in H2 2026 as tariff effects fade, oil stabilizes below its April–May peaks, and AI-related price distortions ease. June’s weak CPI may mark the start of that trend, though it probably overstates the improvement.

His core PCE forecast: roughly 0.2% per month. That should be soft enough to keep the Fed comfortable—but, as Mericle puts it, there’s “not a lot of margin for error.”

[Barbell @barbell_fi]

Mark Cuban @mcuban: "If you need to have forward-deployed engineers, that tells you all you need to know about AI."

The fact that companies like Microsoft, Anthropic, and OpenAI are borrowing from the $PLTR playbook and hiring thousands of implementation engineers reveals AI isn't ready for autonomous deployment.

Cuban's bottom line: AI is simultaneously the most impactful technology we may ever see AND nowhere near ready to replace human workers. As it diffuses, that means a massive market for experts and entrepreneurs who can bridge the gap between its promise and current brittle reality.

[Barbell @barbell_fi]

Scott Galloway @ProfGMarkets urges diversification away from AI: "If the AI trade sneezes, we're not catching a cold. We're getting pneumonia."

His hedge? Long GLP-1s.

"Of the big tech companies we talk about, who's the least concentrated or dependent upon AI? $AAPL. And what's Apple stock's one-year performance? It's up 60%. Year-to-date, it's up 23%. All the other guys, my pick, Amazon, still related to AI, but mostly diversified, it's up 13% this year."

"As a general theme, if I could go long a basket of stocks, it would be GLP-1. And if I could go short a basket, it's AI."

[Barbell @barbell_fi]

@moritzseibert @opmpod Our notes:
https://t.co/Y5GKckS4nH

[Barbell @barbell_fi]

Moritz Seibert @MoritzSeibert on @OPMpod says it's been "a great year" for trend following. His current portfolio:

"We're still long most of the precious metal markets, definitely we're long gold and silver. We're also long platinum."

"We're short Bitcoin and Ethereum. The crypto markets aren't trading that well, or at least these two."

"We are long the equity markets. No surprise there... We're long the petroleum markets, no surprise there either."

"We're short cocoa since a long time. Again, it's it's a great, great trade."

"Long bean oil is a great one. So, bean oil has been going higher and higher and higher for quite some time now. And that's a market that has become larger in our portfolio, for sure."

"We're still long cotton... It's also a position that now has a footprint in our portfolio."

[Barbell @barbell_fi]

@signulll You're not gonna believe this

[Barbell @barbell_fi]

"It's uncontroversial to say we're gonna have $10T companies.

Because margins are going to drop, all the returns are going to accrue to scale—because, obviously, low-margin, low-scale is not a very good business.

It's crudely a bit of a Walmart effect in software, which is that the future looks like low gross margins, razor-thin net margins, huge scale. And this is probably a problem if the SaaS provider is the mom and pop shop like Walmart's coming to town." -@jeremygiffon

[Barbell Insights @BarbellHQ]

❚█═[ BARBELL ]═█❚

INVESTOR JEREMY GIFFON: "WE ARE AT PEAK GUY. THE BILLIONAIRE CLASS HAS BECOME SUBSERVIENT TO THE POSTING CLASS" - SAYS POSTERS NOW WIELD MORE CULTURAL POWER THAN BILLIONAIRES [ILTB]

[Barbell @barbell_fi]

@BoringBiz_ instant fintwit legend if they made an X account

[Barbell @barbell_fi]

@hamptonism Big Karp Mode

[Barbell @barbell_fi]

Our notes from the conversation:
https://t.co/Gm8PNOYLt8

[Barbell @barbell_fi]

Lenovo's CFO Winston Cheng positioned $LNVGY as an AI beneficiary on Odd Lots:

"The solutions and services group at Lenovo can also build data centers. So, different from our OEM-like competitors, they can actually build data centers. People who are not even hyperscalers, people who are landowners and they have power in specific areas, they come to Lenovo, and they say, 'help me build a data center' because we also have a modular solution that can build data centers within 9 months. And depending on what you already have available in terms of infrastructure, we can do it as fast as six months...

So that is very, very fast, and that gets the time to market and enable revenues immediately for our partners.

And then in terms of that total infrastructure, if they can have GPU compute in the local market, then, of course, Lenovo can also provide that, particularly with our 11,000 rack liquid cooling capability to be able to service a GPU compute today. Right? And we're expanding on that capability. So today, we're the most end-to-end and relevant partner really from that domain."