Humanoid robots: 1. Forecasts: Goldman says $38bn by 2035. Barclays says $40bn base, $200bn if things go right. BofA models 10 million units that year, up from 20,000 in 2025. Morgan Stanley waves at $60tn. Near term beneficiaries are factories: 72% of 2027 installs in logistics, manufacturing and automotive. My view: technology is moving faster than expected. I think we are only three years away from seeing millions of humanoid robots deployed. 2. The bottleneck is not the AI: The brain is the cheap, solved part. Memory will still be the constraint but the hard part is the body. This is a machine-tool problem, not software. A humanoid is a car in miniature, thousands of fine-tolerance parts, and the edge sits with whoever already has automotive and precision-mechanics DNA. Guess who have ALREADY the capacity to produce millions of EV cars? CHINA. 3. The map, by layer: • Brain (semis + models): $NVDA, and the memory it needs in $MU , SK Hynix, Samsung. Cheap, crowded, commoditizing. • Brawn (the actual bottleneck): screws and reducers in Hiwin, THK, Harmonic Drive, Nabtesco, Jiangsu Hengli. Motors in Nidec, Estun, Inovance. • Sensors and structure: force/torque, vision and connectors in Sensata, TE Connectivity, Sony, Keyence, $OUST. • Rare earths (the input under all of it): $MP, Lynas, JL Mag. China controls over 90% of magnetic rare earths. • Integrators (the full robot): Tesla, plus a heavily Chinese field, UBTech, Xiaomi, BYD, XPeng, and Hyundai via Boston Dynamics. Morgan Stanley data point that gets buried: 73% of confirmed humanoid participants are Asian, 56% Chinese. The US has $TSLA , $NVDA , and not much else on the hardware side. As always I prefer to play the picks-and-shovels under it, the screws, the rare earths, the grinding-machine bottleneck, get paid whether or not Optimus or Figure is the one that scales.

Humanoid robots: 1. Forecasts: Goldman says $38bn by 2035. Barclays says $40bn base, $200bn if things go right. BofA models 10 million units that year, up from 20,000 in 2025. Morgan Stanley waves at $60tn. Near term beneficiaries are factories: 72% of 2027 installs in logistics, manufacturing and automotive. My view: technology is moving faster than expected. I think we are only three years away from seeing millions of humanoid robots deployed. 2. The bottleneck is not the AI: The brain is the cheap, solved part. Memory will still be the constraint but the hard part is the body. This is a machine-tool problem, not software. A humanoid is a car in miniature, thousands of fine-tolerance parts, and the edge sits with whoever already has automotive and precision-mechanics DNA. Guess who have ALREADY the capacity to produce millions of EV cars? CHINA. 3. The map, by layer: • Brain (semis + models): $NVDA, and the memory it needs in $MU , SK Hynix, Samsung. Cheap, crowded, commoditizing. • Brawn (the actual bottleneck): screws and reducers in Hiwin, THK, Harmonic Drive, Nabtesco, Jiangsu Hengli. Motors in Nidec, Estun, Inovance. • Sensors and structure: force/torque, vision and connectors in Sensata, TE Connectivity, Sony, Keyence, $OUST. • Rare earths (the input under all of it): $MP, Lynas, JL Mag. China controls over 90% of magnetic rare earths. • Integrators (the full robot): Tesla, plus a heavily Chinese field, UBTech, Xiaomi, BYD, XPeng, and Hyundai via Boston Dynamics. Morgan Stanley data point that gets buried: 73% of confirmed humanoid participants are Asian, 56% Chinese. The US has $TSLA , $NVDA , and not much else on the hardware side. As always I prefer to play the picks-and-shovels under it, the screws, the rare earths, the grinding-machine bottleneck, get paid whether or not Optimus or Figure is the one that scales.

Moving $IREN from my long term portfolio to my short term portfolio. The competitive advantage hasnt changed. I still believe the business multiplies its value several times over the long run BUT the shareholder will not capture that growth to the same degree, and that is the whole point. > The dilution stack: - $5bn in convertible notes - $NVDA right for 30M shares at $70 - The co-CEO grant: 18.2M RSUs - More equity and debt still coming to build the other 4.5 GW ($6bn ATM) > Facts: 5 GW secured, a $9.7bn $MSFT contract, a $3.4bn $NVDA contract, investment grade financing. > Assumptions: In a base case scenario, the share count climbs toward 650M. The market cap can multiply by roughly 20x as the 5 GW converts to revenue. But the price per share captures a 55% of that, closer to 12x, because dilution eats nearly half the growth. > What bothers me about the co CEOs grant: The RSUs will be vesting on continued employment rather than on execution, so what bothers me is not the dilution itself but the incentives (non execution milestones) > Im giving them until December 2026 to show real execution and new contracts. > What they need to achieve: 480 MW operational, $3.7bn ARR target. If they convert capacity into revenue at the margins they promise and the dilution slows, $IREN earns its way back into my long term book.

Space Infrastructure AI Neocloud / Power Infrastructure AI Biotech / Digital Health Photonics Energy Storage / Grid Resilience Physical AI / Robotics Sensing Memory Infrastructure / CXL Counter-UAS / Defense Tech SiC Power Semiconductors

I started a position in $WOLF a month ago because of the design shift in data center power architecture, moving to 800V DC to eliminate the inefficiencies of traditional silicon. Traditional silicon suffers from switching losses and major efficiency drops when operating continuously at voltages as high as 800V. The solution is to use silicon carbide, which can handle high power densities while dissipating the heat that traditional silicon would generate, therefore allowing power supplies inside solana:4aG547Y9PPAgpd8dQEMn8n87oVtFYPAP6vhqEYDHmoon new Vera Rubin racks to be miniaturized. The rise of solid-state transformers (SSTs): With the introduction of SSTs, extreme industrial-grade power semiconductors will be required. $WOLF is a pioneer in 3.3kV, 6.5kV and even 10kV SiC MOSFETs, designed for these SSTs that connect data centers directly to the medium voltage grid. With the move to 800V, BESS will become much more relevant in 2027-2028. The problem is that AI workloads create millisecond scale fluctuations that destabilize the power bus: BESS inverters must react and inject energy at extremely high speeds. SiC switches up to 10x faster than traditional silicon. Since these batteries sit on the same 800V bus, inverters built with $WOLF SiC technology eliminate the need to convert energy back into alternating current (AC), achieving conversion efficiencies above 98%.

Why I like $LIB.V / $VLTLF as a strategic lithium play: The constraint in lithium is the refining. China mines only about 18% of the world lithium but refines roughly 70% of it, and supplies 85% of the cathode and anode materials that go into every battery. Batteries are becoming a US strategic product, and not just for EVs. Military drones, and increasingly BESS for data centers, are going to pull serious lithium demand over the next few years, and the US will want that supply chain to run outside China. So the structural setup is: Rising strategic demand for domestic, non China, refined lithium, against a refining capacity that barely exists in America today. Thats why LibertyStream interests me. It doesnt just extract. It refines oilfield brine into lithium carbonate on site in Texas, on top of existing oil and gas infrastructure. The rare part is not the extraction, its doing the refining domestically. Important: First commercial tonne already delivered, with a 600 tpa offtake term sheet in hand. My thesis is structural: this is a multi-year play, with several factors converging into one single point: the US needs to refine a LOT of #lithium# at low cost and in a scalable way. $LIB.V is a early exposure to a domestic lithium supply chain the US increasingly has to build.

Thoughts about $CCXI Agility Robotics: > People see the tech as inferior to competitors because of the appearance of Digit, the company main robot. I think it is a mistake to confuse appearance with the real functionality of the robot. I prefer a robot that looks like an ugly fridge but whose qualities and tech are superior to the competition. > Following the previous point, its limbs are specifically designed for handling plastic containers (totes). > The main beneficiaries of robots are factories, and years later, homes, where the appearance of the robot will be adapted according to the day to day functionalities in the home. > They have RoboFab in Salem, Oregon, the first factory dedicated exclusively to humanoids, designed to scale production up to 10000 units per year. > One KPI I like is accumulated hours in customer facilities. These hours are data, and we already know the role of data in AI. > Customers: $AMZN, $GXO, $MELI, Toyota. > Business model: Robotics as a Service. You pay an hourly fee. I like it quite a lot because as production increases, the €/h will decrease and customers will first reduce personnel expenses and then opex, improving margins. > Investors: $NVDA, which provides AI infrastructure, and Foxconn, which provides scalability in production. Amazon and Schaeffler have also invested to secure supply.

Rocket Lab is acquiring Iridium Communications Inc – one of the most transformative deals in the space industry. By combining our launch capability and satellite manufacturing with @IridiumComm’s global satellite communications network and rare spectrum, Rocket Lab becomes a fully integrated, self-launching, tier-1 space power, delivering critical communications capability to millions of users worldwide. Full details and important information:

We just hit 600 subscribers on Substack! I have been a bit quiet lately because my day job has been very busy. I work in Due Diligence, and sometimes the workload gets crazy. Yesterday I finally published a new article on $MU results and their impact on the full supply chain. I break down the companies that could benefit and the ones that could be hurt. Thank you all so much for the support! Full article in the comments.

We just hit 600 subscribers on Substack! I have been a bit quiet lately because my day job has been very busy. I work in Due Diligence, and sometimes the workload gets crazy. Yesterday I finally published a new article on $MU results and their impact on the full supply chain. I break down the companies that could benefit and the ones that could be hurt. Thank you all so much for the support! Full article in the comments.

Robotics VC investment just hit an all time high: $16B in a single quarter: > Robotics is gaining traction, and I do believe the theme will be successful. But we are still probably 3 - 5 years away from seeing real breakthroughs translate into meaningful volumes. However, I'm long $OUST as a picks and shovels play here: > We already know that not all robotics companies will use LiDAR, some go pure vision. > $OUST bought StereoLabs, the ZED stereo-camera company. So the company now sells LiDAR, cameras, and the fusion of both, so I don't have to guess which robot or which sensing approach wins. > As the category scales, $OUST becomes one of the cleaner ways to get exposure to robotics. Risk: There's a window where price can run ahead of fundamentals and then correct when the market gets impatient with robotics theme.

Agree here with @ren_stocks: The more I study the HBM4 ramp, the more bullish I get on $PDFS: > HBM4 is not just a supply problem, it is also a yield problem. > Memory is getting more complex: more dies per stack, more bonding steps, more test steps, more packaging complexity, and more points of failure across the manufacturing flow. Implication? That creates a lot more data. > And if you are trying to ramp a complex product like HBM4, that data becomes extremely valuable. You need to know where defects are happening, why yield is breaking, and how to fix the process faster. That is my $PDFS thesis: > It helps semiconductor companies connect manufacturing and test data, understand yield issues, and improve the ramp from development to high-volume production. > If AI memory keeps getting harder to manufacture, the software layer that helps companies improve yield and control the process becomes more important.

I guess selling $SIVE and rotating into $IQE was not a bad decision after all. If you want to know why I made that move, and how I think about both companies, I wrote a full breakdown on the $SIVE vs $IQE thesis. Link below.