Analyst. Semiconductors, data centers & physical AI. Portfolio and investment theses on Substack Join 14,000+ readers: https://xitter.cf/t.co/poHmmwvC5M
Spain
Joined April 2024
- Tweets9.1K
- Following344
- Followers46.8K
- Likes3K
Pinned Tweet
YTD performance: +91%
After such a run, I’m now looking at what comes next
> 71 companies on the watchlist, mapped out over the coming months
I share it all here
hypertechinvest.com/p/my-ful…
$NVDA CEO Jensen Huang Rejects Memory Oversupply Fears Through 2029
Nvidia CEO Jensen Huang dismissed concerns about a potential memory glut by 2029, arguing that “you can always use more of it.”
Huang expects HBM demand to remain strong “for years to come” and confirmed Nvidia will continue using large quantities of LPDDR memory
Regarding potential cooperation with China's CXMT, Huang said “We're open-minded,” adding that “no country collectively is better at manufacturing than China.”
He also argued that US export restrictions have accelerated China's domestic semiconductor development, while maintaining that American companies are ready to compete
$TSLA and $NVDA Invest $2.8B to Secure Samsung Electro-Mechanics ABF Capacity
Nvidia will invest $2 billion and Tesla $830 million, supporting SEMCO's broader $3.2 billion expansion
Construction is expected to finish by late 2027, with equipment installation and trials in H1 2028 and mass production starting in H2 2028
Tesla has secured a five-year supply agreement (2029–2033) for AI chips powering autonomous vehicles and Optimus robots, while Nvidia's capacity will support AI computing and data center networking
Every time you doubt the AI supercycle, consider one simple thing:
> AI is, at this moment, the dumbest it’ll ever be
> The most expensive it’ll ever be
> The slowest it’ll ever be
> And it has the smallest context windows it’ll ever have
Now reconsider
$TSM Plans New CoPoS Advanced Packaging Facility in Taiwan
TSMC is reportedly planning a new CoPoS advanced packaging facility at its 104.65-hectare Longtan site in Taiwan, expanding beyond its existing AP7 and AP8 projects to meet surging AI demand
TSMC is also outsourcing interposer production to Vanguard and $GFS, including a $2 billion, five-year agreement with the latter, while ASE, SPIL, and Powertech expand alternative advanced packaging capacity
10 years from now, you'll look at the stock chart of any decent AI company and think:
Why was I stressing so much?
All I had to do was keep buying
We're in the early innings of the biggest revolution in human history
All this doubt surrounding AI will look so freaking dumb in just a few years
We'll look back and think: How was I not buying at any price, on any day, with every dollar I had?
$NFLX is fundamentally broken as a business
People's attention spans are fried
And it's not even just a Western phenomenon. In developing countries, it's even worse
People can't even sit through a movie or a single episode of a series in one go anymore
Sad reality
Samsung Electro-Mechanics' MLCC Lead Times Reach 61 Weeks
Semco is facing severe MLCC shortages, with lead times reaching 61 weeks, up from 39–49 weeks in September as AI servers require over 10 times more MLCCs than conventional servers
With capacity utilization at 91%, Semco is gaining pricing power. AI server MLCC prices are forecast to rise 10–20% in Q4, while average selling prices increased 13.9% YoY in H1
Semco has signed $3.2 billion in major supply agreements since May, securing demand through 2027. Its new Philippine factory is expected to begin operations in H1 2027
Samsung Orders Final Equipment for Its First Texas 2nm Fab
Samsung has reportedly ordered additional equipment to bring its first Taylor, Texas fab to 20,000 wafers per month by H2 2027, reaching its reported design capacity
The facility is already qualifying equipment, with initial operations expected by the end of 2026. Tesla is its first confirmed customer, under a $16.5 billion AI chip manufacturing agreement
Samsung also plans to begin construction of a second Taylor fab in November 2026, targeting 1.4nm production in 2030
$MU Taiwan strike ends, but labor dispute remains unresolved
Micron’s Taoyuan union concluded its 24-hour warning strike on October 23, marking Taiwan’s first semiconductor industry strike
Of the 1,637 union members who voted, 1,633 rejected Micron’s revised compensation proposal, representing 99% of voters and nearly 70% of total union membership
While Micron’s Taichung union accepted the new profit-sharing and bonus structure, the Taoyuan union continues to demand better terms. Negotiations will resume on October 27, after being postponed due to a typhoon
$LITE down 10% just doesn't match what we've been seeing lately
The market moves in strange ways
Voting machine, weighing machine, whatever
Neoclouds getting crushed despite payback periods of less than a year?
With B200 and B300 pricing at record highs?
With hyperscalers reaffirming their capex plans?
With AWS saying they're booked through 2028?
$NBIS at $160 now
Impressive.
$NBIS Secures $3.5 Billion Financing Facility With Citi
The facility has a 364-day maturity, with an initial interest rate of SOFR + 2.75%, potentially increasing to SOFR + 3.50%
One-year maturity
Nebius is being really aggressive with its buildout
Their payback expectations must be extremely strong.
Memory stocks could fall to a 0.01 P/E, and people would still say, "It's not cheap, it's cyclical."
$AMD Argues CPUs and FPGAs Are Just as Critical as GPUs for Physical AI
AMD argues that raw computing power is not enough for robotics and physical AI. Predictable latency, real-time responsiveness, and efficient data movement are more important
AMD advocates distributed computing architectures that combine GPUs and NPUs for AI inference, FPGAs for real-time control, and CPUs for coordination
Good lord
JPM with a massive revision on $SPCX
I understand the bullishness, but $500 is a bit absurd
It's quite obvious that these banks, which have profited so much from SpaceX's IPO, are pretty biased
JPM doubled $SPCX target price to $500
TLDR
JPM isn't raising SpaceX to $500 because rockets got better.
It's because its model effectively turns SpaceX into one of the world's largest AI infrastructure operators.
Takeaway:
How aggressive is the AI compute buildout?
Extremely aggressive.
Which GPUs are driving this?
Almost all of the 10 GW incremental compute planned for 2027 will be Nvidia Vera Rubin generation. SpaceX expects 440,000 Vera Rubin GPUs deployed by the end of 2Q27
Why does Vera Rubin matter so much?
It's not just more compute. Vera Rubin monetization could be roughly 40% higher than the strongest GB300 systems.
Better GPUs → more intelligence per watt → higher monetization → potentially faster payback → more compute investment
How much money could the AI business actually make?
By 2027, JPM expects SI to contribute 86% of total SpaceX revenue.
