Replying to @herbertong @thejefflutz
No, we will build and run the fab. Let there be ZERO doubt about that.
Maybe TSMC subleases part of the Terafab if they want, but nothing more than that.
W. retweeted
Replying to @Restructuring__
You are totally wrong. I’ve had even worst slides. Thank god I change my mind frequently!
W. retweeted
Retail investors are trading less.
Over the last 20 trading days, retail investors have purchased just +$1 billion of single stocks, near the lowest 20-day total in at least 2 years.
In early August, this figure even briefly turned negative for the first time since at least October 2024.
By comparison, retail's 20-day purchases stood at +$20 billion in April 2025.
Furthermore, total retail equity purchases have declined to +$10 billion over the last 20 trading days, near their lowest reading since at least October 2024.
Retail equity purchases are now down -67% from the +$30 billion recorded over the 20-day period in February 2026.
Retail investors are becoming more cautious.
W. retweeted
This is why Trump doesn’t think a deal is urgent.
Saudi oil exports have recovered and are about to reach their 2025 monthly average, as the US military has carved out a new shipping lane along the Omani coast.
Oil is flowing and the market is strong, so don’t expect a deal.
W. retweeted
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
W. retweeted
To my right is Kelvin Chiu — the first Hong Kong trader to be featured in the Market Wizards series.
I had the pleasure of meeting him in Singapore today and having lunch together. (Photo posted with Kelvin’s permission. He also knows that I use the J Law avatar on social media.)
If you’ve read the newly released Market Wizards: The Next Generation, you’ll know that Kelvin grew up in Hong Kong, graduated from Cambridge, and went on to work at Goldman Sachs and Vitol, one of the world’s largest energy trading firms. During his proprietary trading years at Vitol, his best year generated $35 million in profits.
After leaving Vitol and trading his own capital, he compounded at 108.7% annually for 7 consecutive years, with a Sharpe Ratio of 2.0. Today, he focuses on his own family office.
So when you sit down with a trader like this, naturally I was ready to pull out my notebook and ask him everything I could about trading and asset management.
Instead, our conversation started with… our kids’ education. 😂
Long story short, here are some of the biggest takeaways I got from both his interview in the book and our lunch conversation today:
@KC_SilverCape @Clement_Ang17
SpaceXAI engineer (ex-Cursor):
"right now I'm running 10-20 GrokBot agents that automate 90% of my routine
i have a Chief of Staff agent. He knows about all my other bots and manages everything"
in a 50-minutes podcast, a SpaceXAI engineer showed how to build a team of agents that will work for you 24/7
worth more than a $500 course on agentic engineering
watch today, then read how to build a Grok agents team from scratch in the article below
This video is larger than Cloudflare's 512 MB cache, so it can't be played through. More donations are needed to cover a larger cache. Donate
W. retweeted
Hedge funds are ramping up bets against the US Dollar:
1-month risk reversals on the Bloomberg Dollar Spot Index are down to -0.25, their lowest since February.
This measures the difference in demand between bullish and bearish Dollar options.
By comparison, this metric was positive from March through July, marking a sharp reversal in positioning over the last few weeks.
Furthermore, demand for US Dollar put options versus the Euro was +47% higher than US Dollar calls on August 21st, according to Depository Trust and Clearing Corp data.
Meanwhile, the Bloomberg Dollar Spot Index has declined -2.8% over the last 2 months, to its lowest since May.
Bets against the US Dollar are surging.
W. retweeted
We've kicked off The Wallet Co, a mobile app that pairs the usability of modern fintech with self custody and blockchain native products. I'm excited to announce the initial team soon.
One role we still need to fill: operations and compliance. And this is not a typical ops/compliance hire.
First, why this is a big opportunity. The team behind The Wallet Co founded SoFi and Figure. We put the first consumer loans on blockchain, originated nearly $30 billion, launched the first SEC registered interest bearing stablecoin, and listed the first public equity on chain. The Wallet Co is the consumer layer on those rails: yielding cash you can spend, real world asset yield no bank can match, securities prediction markets, and an AI agent in every wallet. Money is moving on chain. The app that makes that usable for normal people, without giving up custody of your assets, is a generational product. That's what we're building.
Now the role. Ops and compliance scale linearly with volume. Double the customers, double the KYC reviews, the transaction alerts, the reconciliations, the support escalations, the headcount. Every fintech accepts this. I don't want to.
The mandate is to run both functions the traditional way on day one, and then systematically hand the repeatable work to AI. Alert triage, document review, reconciliation, regulatory change tracking, audit prep, vendor monitoring. Build the playbook by hand, teach agents to execute it, and reserve human judgment for the decisions that genuinely require it. Accountability always stays with a person. Leverage is the goal, not replacement.
What you need: real experience across both sides of the house. On compliance, KYC/AML, state licensing, and audit. On operations, payments and money movement, customer onboarding, reconciliation, and bank and vendor partner management. Enough scar tissue to know what can never be automated. And a builder's instinct, because the honest job description is to make your own job smaller every quarter. Most careers in this function reward growing a team. This one rewards shrinking the work.
If we get this right, ops and compliance cost scales like software while volume scales like a marketplace. That's a moat, and you'd own it from day one, with founding team equity.
DM me if this sounds like you, or tag someone I should reach out to.
W. retweeted
If this market is going to have a normal pre-midterm election correction, it should start today and continue through September (historically speaking)
h/t @ISABELNET_SA
W. retweeted
As a gauge of AI compute demand, we like looking across the full cross-section of GPUs and the term structure. But if we have to pick a single favorite metric, it’s not the H100 rental rate — it’s the A100.
Released in May 2020, this “ancient” chip is still going strong and rented out at near-full capacity (as CoreWeave’s recent earnings call confirmed). A100s are now used predominantly for inference.
The fact that A100 rental rates have held so steady — even as newer and far more powerful GPUs keep coming online — is a clear signal of the unrelenting strength and growth in inference demand.
W. retweeted
First impressions of the NVDA Deal. 500BN financing with 6 major credit managers
By far the most notable issue is that NVDA is shifting from direct off balance sheet vendor financing to slightly less direct off balance sheet vendor financing. The big takeaway is this changes the nature of the counterparty NVDA is taking for its credit backstop. Prior to today when NVDA helped a compute buyer they got equity or debt from those buyers in exchange for chip purchases.
With this vehicle the counterparty NVDA backstops owns the full stack of data center assets including but not limited to chips. In addition the project managers in the structure are not single data center companies but a portfolio of data center companies. NVDA is taking less concentrated risk than in the past regarding the specific collateral AND the counterparty.
Let's look at the structure it is very similar to a CDO with third party credit protection built in. Instead of monoline insurers or GSE's providing the credit backstop NVDA provides that protection. But remember this structure isn't in itself new financing it's a shift in financing from direct vendor financing to indirect vendor financing.
How does a trade occur. A data center company has cash funded from its equity investors and needs chips and the full stack of AI infrastructure stuff and additional leverage. In the past the data center would buy those assets and buy the chips from NVDA with various forms of deals where the data center faced NVDA directly. With the new structure the Data center will pay for the chips and all the infrastructure directly. But put all the assets in the new CDO vehicle and they will retain the "equity tranche". But they need leverage to buy the full stack of assets. Private equity and private debt funds would perhaps buy some of the equity tranche and some of the high yield debt tranche (mezz). In the past those PE and PC funds would lend and co-invest with the data center as counterparty. With this structure they do the identical thing but invest in tranches of the CDO. Like all CDO's the senior tranche is where the "power" of the structure lies. In the past high grade bonds or bank loans were needed to provide senior funding. Hyperscalers like goog and orcl etc would issue high grade corporate debt and provide the proceeds to help buy the assets of the data center. The market would accept these bonds because of the credit quality of the issuer. With this financing CDO they can cut out the middle man. The high grade asset backed buyer can buy the senior tranche of the new AI CDO having senior exposure to the asset stack and having NVDA asset guarantee and the mezz and equity tranche to absorb any losses before they take a loss. I'm sure the rating agencies will provide high quality ratings to any super senior or senior tranche that this structure spins off. The mezz and equity tranches and the assets themselves will benefit from any additional access to capital or lower cost that this structure realizes and reduce some of the growth in off balance sheet risk Orcl et al are taking.
So on the main this is mostly a swap from existing financing to this new version which is pretty irrelevant but on the margin may provide more attractive aggregate cost of financing (provided by all tranche owners preference but particularly senior tranche demand).
NVDA has probably run out of capacity off balance sheet for concentrated counterparty and asset (its own chips) exposure and so this allows them to extend additional vendor financing in a less direct way. The net is NVDA running out of capacity is probably as big a signal as them creatively coming up with new capacity but that's hard to judge.
Hope this helps.
Incredible. Jensen is completing the circle.
- Bankers don’t like GPUs as collateral because the depreciation is unpredictable
- It’s unpredictable because a new GPU can obsolete an old one
- Jensen knows his own roadmap
- so he’s offering depreciation insurance to the banks
- the depreciation insurance (up to 25%) helps the banks get marginal deals over the line
Speculation
- Nvidia will also advise the banks on “reference designs” for datacenters that will make them fungible
- Having them be fungible means that the debt can repackaged into Asset Backed Securities, Collateralized Loan Obligations and Collateralized Debt Obligation (ABS, CLOs and CDOs from 2008 haha)
- This allows tranching to get investment grade ratings on the debt so that it can be resold to pension funds and insurance firms
- It also allows the banks to trade idiosyncratic project specific credit risk for sector wide credit risk
So Jensen is trying to get his customers the same cost of financing as real estate rather than venture equity.
This is going to move the data center game out of the VCs and into the big leagues.
W. retweeted
They don't have enough inference capacity to fulfil demand, so they are trying to raise prices and destroy demand
One of the reasons why I'm not worried about closed-source lab revenue is that they or their investors (hyperscalers like Google, Amazon, Microsoft) control the majority of global inference capacity
DeepSeek, Kimi, Zhipu (have less than 400MW of combined compute capacity) can't challenge Ant/OAI (have over 6GW of compute)
The real threat for Ant/OAI is a company like Meta/xAI (have 3-4 GW of compute) starting a price war in the near term
The long-term threat is on-prem adoption of open models with enterprises setting up mini clusters (Jensen wants this to happen as it reduces customer concentration risk)
W. retweeted
$AXTI is holding $134.7M of customer money for wafers it has not made yet. Lumentum, Coherent and Casela each prepaid to lock indium phosphide substrate, one of them through 2031!
Napkin math:
• Current revenue: ~$190M/yr (Q2 annualized)
• FY2028 revenue: $705M (3.7x)
• Market cap today: ~$4B
• FY2028 target: ~$6B
• Upside: 1.5x from here
1/ AXT grows indium phosphide substrates. Every EML and CW laser in an AI optical link is epitaxially grown on one. Silicon cannot emit light, so there is no substitute material.
2/ Q2 2026, reported July 30:• Revenue $47.6M, the highest quarter in company history!
• Up 77% sequential, 165% year over year
• InP revenue $30.7M, a record, driven by datacenter
• Gross margin 44.9%, up from 8.0% a year ago
The Casela, Coherent and Lumentum agreements did not materially contribute to that quarter. All three are incremental from here.
3/ Three companies hold over 90% of global InP substrate supply. Sumitomo at roughly 42% share and 800k wafers. AXT at 36% and 300k. JX at 13% and 200k.
-=-=-=-=-=-
Sumitomo is saturated and consuming more of its own output internally. JX has announced no expansion. AXT is the only one of the three adding capacity, and the only one whose capacity is for sale.
-=-=-=-=-=-
4/ Management's own targets for InP revenue capacity:• $60M per quarter exiting 2026
• $130M per quarter exiting 2027
From $30.7M in Q2, that is a double in two quarters and another double in four more. Their words: this would make AXT "by far the largest indium phosphide producer in the world."
5/ On the demand side, from the same call
"Customer demand continues to outpace supply, no matter how fast we add capacity". This is what we like to hear, keep talking dirty to us managment.
• Backlog above $100M, and management says that figure no longer reflects all available demand
• China demand more than doubled
• Materials now in "multiple US hyperscalers"
• Targeting gross margin "that begins with a five"
6/ The margin math is more interesting than the headline number suggests. AXT sells five things and they earn very differently.Q2 2026, derived from the reported split
• Core InP substrates: ~$28M at ~54% gross margin
• 6-inch InP for CPO: ~$2.7M at ~58%
• GaAs: ~$6.3M at ~30%
• Germanium: ~$0.3M at 25%
• Raw-material JVs: ~$10.3M at ~26%
-=-=-=-=-=-
The core product already earns in the mid-fifties. The reported 44.9% is legacy businesses diluting it. "A number that begins with a five" needs no pricing improvement at all, only for InP to keep growing as a share of the mix.
-=-=-=-=-=-
7/ As InP goes from 64% to 86% of revenue on the two capacity doublings, blended margin reaches 55% by late 2027 on arithmetic alone. Shortage pricing sits on top of that.Revenue up 3.7x and margins expanding at the same time.
8/ Where $AXTI sits:AI optical link
→ transceiver / optical engine
→ EML and CW lasers from $LITE $COHR $AAOI $SIVE
→ epitaxy on InP wafer
→ AXT (substrate)Lumentum runs five of its own fabs and is converting a Greensboro site to InP. It still put $87M down to reserve AXT capacity through 2031. That tells you what its internal position is worth.
9/ Risks
Export permits. Manufacturing is in Beijing, and management names China export licensing as its single biggest challenge
• Coherent is building its own 6-inch InP capacity and has a CHIPS LOI for Sherman, Texas. That competes with AXT's highest-margin line
• The prepayments cut both ways. Coherent can terminate on a six-month capacity miss
• VCSEL and microLED could take short-reach optical volume from InP. Partly hedged by AXT's GaAs lineCounter-evidence on the last one: NPO and CPO both need more InP, not less, and the industry is moving toward them.
(image courtesy of @aleabitoreddit)
W. retweeted
Spending a week with my family and working most of it! Being a few hours ahead of US markets has given me some time to digest a lot of incredible markets discourse on X - there is too much to try and consume about AI but not enough about leverage. We are seeing a classic head hunt of the most levered players in the equity and convert market re AI globally. Hyperscaler and associated credit spreads in IG are wider as they should be (portfolio construction by notional and duration matter in credit because we don’t have the payout that equity does) and debt is being added to compute and power as another constraint on the AI theme. Govt regulation remains a massive wildcard but a longer cycle isn’t necessary a worse one. I would look for the forced sellers of assets trading at or below contract value with counterparties you feel good about that have positive optionality on growth opportunities. Think about the impact on spot and next 1-2 year curves for compute, power, and shell - those who are long and don’t need financing + can term out contracts now are materially advantaged. If this is the whole cycle being elongated and the curve flattened there are a lot of interesting securities to buy from forced sellers. More time for competition and technology to emerge in the intermediate term isn’t necessarily a bad thing for many infrastructure assets. I started my career in the middle of the early 2000s telecom cycle - Nortel, Lucent, Cisco, etc were financing their customers. There have been some very astute comments on this platform from people who understand the AI echosystem far better than I do about Nvidia and Broadcoms business model decision to become the working capital bank of the AI build - bridging the industry to revenue and cash flow. My sense is the focus in credit markets right now is too much on Meta Google Amazon etc and not enough on that business model change which liquifies the compute roll out in the near term and shifts the credit risk to those large semiconductor companies. It’s fun to seeing liquidity having a price again and god forbid IG companies cost of debt having to compete with their cost to equity.
“There is a contradiction in the AI picks-and-shovels trade between the multiples on memory and storage stocks trading at ~6x earnings (suggesting the cycle is nearly over) and the much higher multiples on chipmakers like Intel, AMD, and Cerebras—as well as other parts of the data-center value chain such as cooling and power (suggesting the cycle has much longer to run)? One of them has to be wrong.”
I’ve heard this argument repeated a few times lately, initially from @GavinSBaker. Is the stock market making a plain mistake in broad daylight?
I think the quoted post is a good illustration of why the market is not making an elementary internal inconsistency error. The demand elasticity of substitution between HBM and DRAM is well over 1, so total memory-maker revenue can slow even as HBM grows, because HBM is a more efficient form of memory.
What the market is clearly pricing is that memory and the higher-multiple parts of the value chain are fundamentally different businesses. Memory is assumed to be much more of a commodity with high elasticity of substitution. The market assumes it is far easier for supply response or efficiency gains to erode pricing power and earnings in memory than in the rest of the chain. That assumption could prove technologically false, but it is not a plain logical inconsistency.
In fact, much of what is being priced today may depend on the tacit assumption that meaningful efficiency gains in memory and storage will arrive soon, enough for the overall cycle to continue, thereby justifying the elevated multiples on the rest of the value chain through a lengthened runway.
An indefinitely long period of explosive demand for memory and storage at current efficiency levels, while still commanding 80–90% gross margins, is surely inconsistent with the cycle lasting much longer. So either efficiencies are found, or the capex cycle slows or dies under the gravity of its own growing weight at successive stages of hardware supply bottlenecks.
Of course the magnitudes could still be wrong: memory margins could stay elevated longer than the 5–6× implies, efficiency gains could arrive faster or slower than assumed, or power and cooling constraints could prove stickier than expected. Those are quantitative or technological forecasting disagreements, not a plain logical inconsistency in the cross-section of multiples. The market is treating memory as the more mean-reverting, high-elasticity layer of the stack and the rest as more structural.
In that sense, the cross-section of stock valuations can be quite rational and internally consistent, even if it turns out to be false from a technological point of view.
This quoted post is unavailable.
W. retweeted
Ce qu’on observe ces derniers jours sur les semiconducteurs, ce n’est pas un éclatement de bulle ou autre (certains aimeraient mais quand on voit les valos, on reste loin d’une bulle).
Non, ce qu’on voit, c’est l’éclatement d’un narratif. Les gens avaient une conviction dans le graphique et le momentum, pas dans les entreprises. Ce qu’on a vu, c’est de la spéculation sur de belles histoires.
Certains y sont allé comme des bourrins avec du levier sur des valeurs high beta dont ils n’avaient même pas entendu parler une semaine avant. Donc quand ça baisse et qu’il n’y a aucune conviction sur le business, le seul truc que ça fait, c’est que ça liquide pour éviter l’appel de marge, et surtout de se retrouver à poil.
Quasiment rien n’a changé depuis le début de l’année, et c’est pas en 3 semaines que les entreprises sont passées de magnifiques à complètement pourries.
Bref, comme d’habitude, DYOR et faites-vous votre propre opinion avant de suivre aveuglément quelqu’un sur un trade qu’il ne maîtrise pas forcément
W. retweeted
Not many people will get the reference, but this reads like the Damned United, about when Brian Clough took over at Leeds, and the players didnt respect him because he had spent years sniping at them through the media. They lost three of their first four games...
Replying to @NickTimiraos
Warsh had to persuade the most rate-cut-hungry president in modern memory to give him the job. Now he has a new challenge: convince those 18 to give up the intellectual habits of a profession he thinks led them astray.
The first test comes Wednesday wsj.com/economy/central-bank…
W. retweeted
Google also distilled Microsoft Word, Excel, PowerPoint by running testing files to build compatibility.
Microsoft distilled WordPerfect & Lotus. WordPerfect distilled WordStar.
Microsoft C++ and Borland C++ distilled each other.
BSD distilled Unix.
The list is infinite.
David Friedberg says Google basically used to ‘distill’ Microsoft and Yahoo
“At Google, in the early days, we would submit millions of search queries to Yahoo and Microsoft search engines to see what the result sets were. We would compare our results against theirs as a way of improving our search engine rankings and our algorithm.”
“It was a very common technique. It doesn’t mean we were stealing their algorithm. We didn’t go into their servers and steal their software. We looked at the output of their software and used that to improve our software.”