1) $NVDA hit ATH yesterday. Other AI related names are at or near ATHs. Makes sense. AI demand & supply data keep improving almost daily within an exponential framework.
It's not impossible that CQ4 NVL racks may be ~ 12-14K vs ~ 6K in CQ2. Prior post in May was more a 20-80 guestimate. Q4 variance is still very large, but smaller than in May. Track data real time going forward. Subject to quick change.
That's with BW Ultra only starting to kick in. B300 production prob ~ 100K's in Q3, with nominal GB300 shipments (given lag to racks) but some GB300 volume CQ4, followed by bigger ramp in 2026.
More important LT. AI inference demand went into higher gear months ago, and that acceleration has continued. Those cute AI insider hints about next gen models & apps show also show no signs of slowing.
Also where did all the bear stories about NVDA B200 GPU inventories go? Remember to review NVDA rev rec, which is product dependent.
2) Jensen's high level TAM comments at yesterday's annual meeting was nice but they're hard to distinguish from previous comments. What are durations of these TAMs and shape of deployments?
- Robots and AI infrastructure to train them will be the next multi-trillion $ industry
- 100 NVIDIA AI-powered AI factories, buildouts around the world. That's double a year ago, and they're larger. Average number of GPUs/factory has also doubled and requires tens of GW of AI infrastructure in coming years.
- Robots and AI infrastructure to train them will be the next multi-trillion $ industry
- 100 NVIDIA AI-powered AI factories, buildouts around the world. That's double a year ago, and they're larger. Average number of GPUs/factory has also doubled and requires tens of GW of AI infrastructure in coming years.
3) If one were worried about BW transitions. Some were very worried about Hopper last year.
Wistron told clients "if you don't buy GB200 you won't get GB300" via Jefferies 10D ago.
These are fun but wouldn't keep coming up without strong NVDA leverage.
"If China doesn't speed up those approvals, companies have warned the White House, auto plants may have to idle pandemic-style stoppages." (WSJ)
Rare earths were more important for initial trade detente with China than some realized.
Chinese rare earth controls are similar in nature to US tech export controls. But why are mostly unknown elements so important?
2) Rare earths are necessary for huge ranges of industrial, tech, automotive, robotic, defense (e.g. drones, missiles, fighters) and consumer products.
Light (LREE) & heavy (HREE) rare earths elements aren't always rare, but even when mined elsewhere, Chinese companies refine > 90% for critical markets.
E.g. NdFeB magnets are very strong, high perf and necessary for key electromechanical products, such as EVs.
US has some mining but *zero* refining capacity of such magnets. MP is the only US company that mines rare earths in volume.
MP mined 12K metric tons (MT) of REO in Q1, which is ok but much smaller than China's capacity. Sales were lower. MP stopped shipping to Chinese refiners due to tariffs and is trying to drive midstream, stockpiling, diversification and longer term vertical integration thru Stage 3.
3) When will MP produce Stage 3 high perf magnets?
MP is ramping its downstream "Independence" magnet mfg facility. And in Q1, MP sold $5.2M in NdPr precursors to $GM and provided GM with first automotive-grade magnets for "them to validate."
There are many grades of magnets. Back to EV grade NdFeB magnets that require HREE for processing. HREEs are sourced globally, but MP also plans to separate HREEs such as DY & TB (for NdFeB) from its Mountain Pass ores.
Furnaces to support 1,000 MT are now installed at Independence. LT capacity targets are higher.
1) Storage stocks (components) reacted ~ as expected Fri.
$STX $WDC (mostly DC) up Fri despite EU & $AAPL threats. Both up big YTD.
$MU & $SNDK (large CE exposure) down Fri despite strong ST data points. Hard to believe but MU is +11% YTD. Again, fears of pull-in and net demand destruction from unknown future electronics tariffs, such as 232.
High level anecdote. I have an AI output folder. Folder size is growing exponentially. Text was tiny. Multimodal i/o such as document formats, audio e.g. NotebookLM, images, video = storage rose from tiny -> 100's KB -> MB's -> 10's MB -> 100 MB or more/file. More later in thread.
This also needs inference & storage on AI cloud side.
I've never ever never seen spot DRAM (NAND-related spot, card, SSD prices are also up but not as much) do this without some stocks rising sharply.
2) So stocks are saying "this isn't a real cycle." DDR4 will be legacy soon but still.
Is the cycle 100% non-organic? There is some pull-in, but it's unclear how dramatic. May & June PC & SP data will help. Net demand destruction still depends on "policy" i.e. a megalomaniac.
But are there organic elements too? Edge & DC AI both help (let's leave out HBM for now). PCs are being futureproofed for edge AI and Windows 10 expiration with higher configs. Rumors of higher DRAM content in future iPhones & SPs due to AI do make sense.
3) *If* spot pulls back but not fully, that could be bullish at these valuations depending on demand outlook post-'25.
And god forbid $AAPL Apple Intelligence makes real progress. $MSFT Copilot too. User base for these 2 companies alone is massive.
Is even SNDK a buy? Beat & raised but still losses due to what should've a mid-cycle inv correction plus underutilization charges & startup costs. Ex-$110M charges which go away if inv correction ends, SNDK guided to > 6% higher GM (ex charges again) and ~ $2.5 EPS run rate.
Given its leverage, peak $10+ EPS power still isn't impossible in a normal upcycle. Normalized EPS is lower but will increase if this downturn is short & moderate. What's discounted at 0.67 EV/C26 rev?
1) $MSFT DC scare 🧵. High level with details if have time. Some know all datapoints mentioned. Or more (if you have non-MNPI data pls DM me).
a) Conclusion 1st. Global AI capex next few years went UP recently. Even MSFT.
Is this just the latest AI scare? Again, Trump & co finally began scaring investors. Have pet theories about scenarios but a different thread.
The usual discussion about LT AI variability later in 🧵. Even Satya, Sam, Ilya, Jensen, Dario, Jeff etc don't know > 5-10 year consequences of AI. I have no idea. There's never been a tech cycle like this in human history.
But let's start with MSFT Dec Q & 10Q. Keep in mind, that as of 1/29, MSFT already embedded diffusion rule, DS, @OpenAI, Stargate, Cowen "news" and prob much more we don't know.
- Commercial bookings grew 75% Y/Y (!) but guided flat Y/Y Mar Q. Watch OAI commitments going forward
- Capex ~ flat Q/Q Mar & Jun but rises again F26, which is some info. Lower growth rate? No real info, since range is huge
- Have discussed leases being much larger % of MSFT capex last few years. It helps to analyze lease data in 10Q/K's
- "Expect capex to increase in coming years" in 10Q/K's have been stated forever and is a general statement. Has occurred before down capex years
- MSFT clearly said spend shifts back to "short-lived assets (i.e. infra) which are more correlated to sales growth. That's real info
2) Next elements of Cowen note. International first.
Channel checks "indicate MSFT is also re-allocating a considerable portion of intl spend to the US, which suggests to us a material slowdown in intl leasing."
b) Are we sure?
Mentioned "pre-diffusion" for APAC capacity plans on purpose. This is one estimate (Citi) of APAC ex-China DC capacity. Only months old but it's outdated. Possible new projected would've created GWs of spikes.
Japan, Australia, Korea are Tier 1. Consider Tier 2. India, Malaysia, Indonesia etc. Comb thru diffusion rules. E.g. Malaysia was already higher (but pre-diffusion). Well over 3 GW is nowhere near happening with low GPU limits.
Who's who of hyperscalers and DC developers are involved in some projects. Could this be why capex is being reallocated back to the US or other Tier 1's?
3) Here are locations of Malaysian DCs & projects from one source. Details around involvement of US hyperscalers and developers $EQIX $GDS etc can sometimes be found.
I've been posting "every self-respecting economic & military power" has no choice but to invest because AI = economic, political, military survival.
And "India has grand ambitions because India has no choice. Not sure how long diffusion rules lasts."
Outcomes are super variable if policy changes for India alone due to Trump. Or others. No clue but look for hints.
2) to frame an objective reality with his EPR paper (QM is correct but are there hidden variables?).
Bell's inequality was demonstrated by '22 Nobel Laureates. Almost a "what is reality" Nobel. So this sometimes shows up in pop science "what is AI & consciousness" discussions.
But that's besides the point. This is just 2M after NotebookLM. It's gonna be incredibly useful. Not even a smarter model yet but workflow & UI productivity is off the charts.