Research-focused. Photonics. AI. Defense. Tech. Space. Optic Supercycle. NFA DYDD
New York, NY
Joined October 2021
- Tweets2.8K
- Following623
- Followers119K
- Likes4.7K
WOW. Brazil’s ETF $EWZ just plunged nearly 20% in overnight trading.
Lula wins a fourth term after beating Flávio Bolsonaro.
An entire country ETF trading like a memecoin.
Wouldn't be surprised if Anthropic revenue is reaccelerating in October.
OpenRouter usage has recovered sharply from August, with several recent days around 450–600B tokens.
In the most recent breakdown, Opus 5.5 accounts for 197B of the 361B tokens recorded so far. Fable 5.1 adds another 32B.
Together, almost 64% of usage.
At standard uncached pricing, Opus 5.5 costs twice as much as Sonnet 5.5, while Fable 5.1 costs five times as much.
Assuming the same input/output mix, Fable’s 32B tokens would therefore generate the same bill as approximately 160B Sonnet tokens.
There’s been a lot of concern recently about slowing growth at the frontier labs and whether revenues can keep up with spending on compute.
If Anthropic’s growth is picking up again, that would challenge an important part of the bearish case.
AI agents are already accessing bank accounts and trying to move money, per Axios.
Banks are scrambling to restrict them. Could be a pretty interesting setup for some of the smaller cyber names.
$OSPN already serves more than 60% of the world's 100 largest banks, according to its investor deck.
They already sell those banks authentication and transaction security. Now add agents logging in, buying things and making payments for customers.
The bank has to know who gave the agent access, what it's allowed to do, and which transactions still need the customer's approval.
I think banks could end up spending a lot more on this as agent use grows. $OSPN already has relationships with many of the banks that would be buying.
$MITK, $YOU and a few others are also names I like here.
Disclosure: I'm long $OSPN.
$CRDO getting hit pretty hard today following $MXL’s plans to enter the 1.6T AEC market.
Competition isn’t anything new here. $ALAB and several others have already been making moves into the space, and given how attractive the market is, more competition was inevitable.
Credo still has an estimated 90%+ share of the AEC market and relationships with 5/6 major hyperscalers.
$MXL isn’t expecting initial revenues until 2027, with a more meaningful ramp in 2028.
Definitely something to monitor, but I think the market is overreacting here. $CRDO is starting to look ridiculously cheap.
The alt data on $META’s Muse is mixed heading into earnings.
Sensor Tower estimates 10.2M US downloads through Oct. 17, with daily active users up 15% week over week.
Early retention looks encouraging too. 23% of September installers were still coming back after two weeks. But search interest has cooled, and Meta is spending heavily to promote the app.
Meta’s $130–145B capex guide came before Muse launched. With ~$51B spent in the first half, I’m thinking $40–45B for Q3. If usage keeps growing, I wouldn’t be surprised to see them raise the annual guide or land near the high end.
Meta already expects a ~$10B Q3 legal charge, and today’s $25B New Mexico ruling adds more uncertainty.
The $10B charge alone is going to weigh heavily on reported earnings.
WSJ just reported Anthropic’s infrastructure commitments could reach $750B.
Revenue is currently growing much faster than compute costs though.
H1 2026 vs. H1 2025:
> Revenue: $16.3B vs. $1.1B
> Compute + infrastructure: nearly $9B vs. $2.3B
They went from spending over $2 on compute for every dollar of revenue to about 55 cents.
The $450B cloud portion runs over 7–10 years. Average that out and you get $45–64B a year. Keeping those payments at roughly 55% of revenue would take $82–116B in annual sales, assuming even payments and excluding the other infrastructure commitments.
Most of those cloud deals are noncancelable. Some are payable “regardless of actual usage.”
Dario put it pretty bluntly when discussing a hypothetical $1T annual compute commitment: “If I’m just off by a year … then you go bankrupt.”
$MSFT $AMZN $ORCL $GOOGL $META
How many hyped IPOs have burned retail in the past couple years?
$CRWV. $CBRS. $BLSH. $FIG. Even $SPCX got hammered after the initial hype.
Now Nscale reportedly wants ~$30B, but $MSFT has pulled out of a UK data center deal and some investors are pushing for $16–19B. I wouldn’t be surprised if this IPO gets delayed.
Last week, Reuters reported some Anthropic investors were discussing a valuation closer to ~$1.5T, versus earlier expectations of ~$2T.
Retail shouldn't have to bet on five years of perfect execution just to earn a decent return.
Leave something on the table.
RAMmageddon continues to show up in phone shipments.
IDC says global smartphone shipments fell 22% in Q3, the worst decline on record. Growth isn’t expected to recover until 2028 at the earliest.
IDC research director Nabila Popal says the smartphone market “has not yet reached the bottom.”
“Memory costs have surged approximately 350% since last year and are still rising.”
“The worst is yet to come.”
$MU $SNDK $AAPL $QCOM
Update on VARA: looks like the fund didn’t get liquidated after all.
The blowup rumor appears to have been wrong.
But judging by the tape, wouldn’t be surprised if someone else did today.
VARA AI Fund had 9 employees and about $26.4B under management at the end of June, 16 months after launching its hedge fund.
Rumor going around today that it blew up.
Their 3/31 13F book at today’s prices is $13.1B across 60 names.
$NBIS $1.28B
$BE $1.27B
$LITE $1.03B
$CRWV $889M
$STX $728M
$VRT $720M
$WDC $624M
$CLS $537M
$COHR $492M
$FIX $426M
In hindsight, the guy with a decade-long thesis buying calls with a two-week expiration probably wasn’t a great sign.
VARA AI Fund had 9 employees and about $26.4B under management at the end of June, 16 months after launching its hedge fund.
Rumor going around today that it blew up.
Their 3/31 13F book at today’s prices is $13.1B across 60 names.
$NBIS $1.28B
$BE $1.27B
$LITE $1.03B
$CRWV $889M
$STX $728M
$VRT $720M
$WDC $624M
$CLS $537M
$COHR $492M
$FIX $426M
KawzInvests retweeted
3 reasons:
1. AI slower, less corp bonds, less supply on the long end
2. AI slower, US GDP slower, real yield down.
3. AI slower, less memory/computer component supply shock, less inflation.
All points to lower US10Y
$CRWV isn’t paying 15% interest. Only 4% of their debt costs that.
I went through the Q2 10-Q line by line. Out of ~$35B in debt, exactly one loan sits at 15%. It’s $1.3B, due March 2028, and already being paid down.
The 15% chart going around is the market yield on their 2032 bond. A higher yield doesn’t change what CoreWeave pays. It works like a mortgage. If you locked in at 3% and rates go to 7%, you’re still paying 3%. CoreWeave locked that bond at 9.625% and still pays 9.625%.
What they actually pay
> $6.6B of convertibles at ~2%
> ~$10B of bonds at 9-10%
> Most secured loans around 9%
> ~8.4% blended across everything
They have roughly $20B of 2026 capex left to fund, and 2027 will need more. Their main funding tool has been loans backed by customer contracts. The investment grade one backed by Meta priced around 5.9% in March. If they keep borrowing that way, I think the math holds.
Zooming out, the AI trade still looks intact to me. $ASML reported record Q3 profit this morning and said lithography demand will outrun their supply in 2027 and 2028, even after adding ~30% capacity each year. Customers are already asking about 2029 and 2030 slots.
Demand is still outpacing supply everywhere I look. $CRWV $ASML
the real risk is if they stop getting contract-backed loans and have to borrow at that 15%.
say they need $20B more. at the ~6-9% they’ve been getting, that’s around $1.3-1.8B a year in interest. at 15% it’s $3B a year.
for context, they paid $640M in interest last quarter, so about $2.6B a year on everything they owe right now. one new $20B chunk at 15% would cost more than all of that. and they’re still not profitable.
the next financing deal is critical. if it’s contract-backed at single digits, fine. if they’re forced into 15% bonds, that’s when it gets really risky.