Trail

I remain bullish on AI and the global opportunity for $NVDA, $MSFT and $AMZN. But the threat to US data-center construction deserves a much better answer than last quarter’s sales. Texas voters elected Gina Hinojosa after she promised a ban on data-center approvals until the Legislature passes protections. I take her words seriously. A project waiting for permission can lose years. A developer paying interest while waiting can lose the business. I understand why investors are selling. But I believe AI can keep expanding through this, because construction is already becoming international. In July, Meta broke ground on a planned one-gigawatt AI campus in Alberta, Canada. Amazon has announced a €33.7 billion cloud and AI infrastructure commitment in Spain through 2035. OpenAI announced a one-gigawatt Stargate cluster in Abu Dhabi back in May 2025, with its first 200 megawatts expected in 2026. I am careful with that last word: expected. An announcement does not tell me the machines are switched on. Those overseas projects have their own power, financing and regulatory risks. But they show that building outside America was already part of the plan before this election. It is a route companies have been developing, with named partners and specific locations. And the US construction pipeline does not all face the same deadline. Nomura’s assessment reported today is that much of the capacity due in 2027 is already under construction, with the bigger permitting threat falling on 2028 and beyond. I see time to adapt. I also see a much tougher contest over who can actually deliver. Moving construction abroad takes years. Advanced chips face export controls. Some customer data must stay close to home. I would never assume another country can replace every blocked American project. Still, I believe companies will keep pursuing places where they can secure power and permission. The chips can serve customers from more than one country. The opportunity reaches beyond a single state’s construction calendar. My bullish view has room for failed projects. Some developers will suffer badly. Some shareholders will own the wrong company even as AI adoption grows. I want companies that can fund construction, secure electricity and deliver computing customers will pay for. I will be watching actual project progress and customer commitments much more closely than ambitious campus announcements. Today has made me more demanding about which businesses I trust. It has not convinced me that AI’s useful years are behind us. I believe we are still early. I expect the next phase to be harder to build, more international, and far less forgiving of weak businesses. I remain strongly bullish on the companies that can deliver it.

$SMOL plans to launch its first capacitors in 2027 in optical and high-frequency chips, the same two markets $SIVE sells into, and it ended September with SEK 45.5M of cash to get there. It sold nothing in the quarter. A year ago it held SEK 18.5M; the summer raises filled the gap. The first ultra-thin capacitors for customers to test should be ready toward the end of this year. On the hydrogen side, Heraeus is now running a 3,000-hour durability test on Smoltek's electrode. The quarter used SEK 11.0M of cash in operations, down from 19.5M a year earlier. At that pace the money covers about four more quarters, and Smoltek's own target for positive cash flow is the end of 2027. That leaves no room for the first paying deal to slip past next year.

Starcloud, a startup that wants to sell AI computing from orbit, ordered more than 50 laser terminals from $SPCX, the company racing it to do the same thing. Those lasers are the wiring of a space data center. A satellite can't plug into a cable, so it talks to the next one with a beam of light. Each Starcloud satellite will carry two Starlink Mini Lasers. Starcloud's CEO said that link is what "turns individual satellites into a functioning distributed data center." $GOOGL wants the same thing. Its next two Suncatcher satellites, planned for 2027, will test laser links between spacecraft. Mynaric, once Europe's listed laser-terminal maker, had delivered more than 100 terminals to the US Space Development Agency by June 2025. Its shareholders still got nothing. $RKLB took the company over on 14 April for $155.3M, paid in its own shares. Rocket Lab's CEO called laser communication "a supply chain pain point" for satellite constellations. The laser inside these terminals is usually a 1550 nm part borrowed from fiber networks. The slow, low-yield step is the aiming: keeping a narrow beam locked on a satellite thousands of kilometers away, with mirrors aligned to a few millionths of a radian. So if you own laser-chip makers for this theme, orbit adds little to the chip. The money goes to whoever builds the aimed terminal at volume. Today that's SpaceX, and it sells to its own rivals.

$NBIS, $CRWV, $IREN, $WULF, $RIOT and $APLD got dumped today for one reason: nobody knows yet what the election means for data centers. That's all it is. Not knowing. Goldman warned that new state and local leaders could pause new data center building, and Texas alone holds about a quarter of what's planned. Brad Gerstner said 10 to 15 GW of next year's build may slip. Scary headlines, zero new rules written. When nobody knows, people sell first and think later. That's how fear works, and that's where the wild swings come from. Now look at what happened on the very same day. Riot's quarterly filing showed Anthropic signed a 15-year lease for 910 MW in Texas, worth about $28B. Google agreed to stand behind $19B of Anthropic's rent at TeraWulf in Kentucky. Microsoft and Meta have already signed more than $44B with Nebius, which has no data center in Texas. The money is signed. The demand for AI computing didn't change overnight. Only the mood did. Markets hate a blank page. Once the new rules are written and people can read them, the fear gets a size, and a fear with a size is easy to price. I think this is the dip people will wish they had bought. I'm buying it.

On 1 November, $GOOGL said its AI chip in orbit, on a satellite built by $PL, had answered its first question. It wrote a haiku about life in space. That satellite carries four Google AI chips. Its solar panels make about 1 kW, enough to run one of them for roughly 20 minutes at a time. Starcloud got a data-center chip up first: an $NVDA H100, in November 2025, the same chip that fills AI data centers on the ground. It trained a small language model on the complete works of Shakespeare. The model answered in Shakespearean English. That mission ended around 2 October. The rideshare orbit it got put it in Earth's shadow for part of every lap, running on battery. Starcloud's next satellite was due in October 2026. It now flies in 2027!! On 14 October, a Chinese company called ADA Space put eight more computing satellites in orbit, which takes its constellation to 20. They're testing 10,000-watt class liquid cooling loops, big folding radiators and laser links to the ground. Starcloud has one in orbit. Google has one. Planet got paid to build Google's satellite and SpaceX got paid to launch it, whatever the haiku proves.

Swedish $MYCR grew sales 38% last quarter, and 2/3 of the extra came from one unit whose die bonders place tiny chips, like the lasers $SIVE and $COHR make, inside the optical links of AI data centers. Operating profit more than doubled, from SEK 255M to SEK 610M.

$AAOI signed for 900,000 indium phosphide wafers from a Chinese supplier, the same material $AXTI and $COHR are spending to make more of. Yunnan Germanium filed the deal today: 180,000 four-inch wafers a year from 2027 to 2031. The 2027 contract alone is $194M, about $1,080 a wafer. That one year is 122% of what Yunnan Germanium sold in all of 2025. Every shipment needs a Chinese export license for dual-use goods, and if it isn't granted, AAOI can cancel and get its money back. Five years of laser wafers now hang on a signature from China's Commerce Ministry.

$NBIS rents out $NVDA B300 AI chips by the hour, like a hotel for computing power. Full price is $9.50 an hour, and that chip is yours until you're done. It also sells spot chips, which are the leftovers. They're normally much cheaper because Nebius can take them back with 60 seconds' notice. Until today, spot could never cost more than full price. Yesterday a US B300 on spot cost $1.24. Nebius removed that cap today. Tonight buyers in Europe are paying $10.09 an hour for spot, more than full price, for a chip they can lose at any minute. That happened on the same day $NBIS, $IREN and $CRWV stock sold off.

I think the selloff in $NBIS, $IREN and $CRWV reads Musk backwards. On 25 Oct he said 200+ neoclouds face "guppy-level death". On 1 Nov he said the cost of serving Grok Bot was "unsustainably high relative to revenue". Neither line says demand for compute is falling.

$MRVL is paying $115B in stock for $LITE, a company whose AI lasers are sold out through early 2029. $NOK paid $2.3B for Infinera in 2025 and got its own indium phosphide laser fab in Sunnyvale with it. Marvell just told the market what owning lasers is worth.

Buy $SPCX and thank me later. This huge potenial of growth in the next few years.

The market sold $NBIS 5.7%, $IREN 9% and $CRWV 5% today after $MSFT and $GOOGL said they still can't get enough compute. What they said: Microsoft: Azure +48%, guiding 49-50% next quarter. Amy Hood: "we still have demand that exceeds supply." The third-party capacity that slipped now lands in the first half of 2027. Google: Cloud +103% to $30.7B, with a $561B backlog. It started using third-party compute at the end of Q3, will ramp it in Q4, and expects a margin hit for doing it. Then Amazon, after the close: AWS +39%, its fastest in nearly five years. 2026 capex held at $220B, and Andy Jassy says they'll "spend meaningfully more in 2027." Together that's about $590B of capex this year. Google and Amazon both say 2027 goes higher. What the market did: sold the companies that rent out extra capacity. Google is spending $195-205B on its own build this year and still pays someone else for compute, margin hit included. That's the customer neoclouds were built for.

$MSFT has $27.1B of contracts with $NBIS and $IREN running through 2031, and last night it said the third-party capacity it expected arrived late. Nebius holds $17.4B of it and says every tranche is already delivered. IREN holds $9.7B, with its first Childress hall accepted in August and its last halls targeted for this quarter. A delay on a signed contract is revenue that lands later, not revenue that disappears.

BREAKING: $NOK passed a new Drop Test $NOK closed at $7.88 the day after its Q3 report, a 15% drop in one session. Today it's trading around $9.20. Five sessions for the whole round trip. The quarter Nokia guided to next got lost in the selling: Q4 sales up 20% to 24% from Q3, and full-year operating profit above the midpoint of its €2.1B to €2.6B range. The Infinera integration also wraps up by the end of this year, a full year ahead of plan, with €200M of net synergies. Anyone who sold that drop now pays more to get back in, three months before Q4 lands on January 28.

$MSFT spent $45.2B on capex last quarter against a guide of over $50B, and $NBIS, $IREN and $CRWV holders woke up to this screen. Microsoft says the rest installs next quarter.

$ORCL's $225B revenue target for FY30 is the demand number $NBIS, $CRWV and $NVDA holders have been waiting for. Oracle laid it out at its investor day today: 35%+ growth a year from FY26, a 40%+ operating margin, and $30B+ of free cash flow in FY29. It also capped new capital raises at $20B through FY30, half equity and half debt. That answers the biggest worry hanging over the stock. When one AI cloud plans four more years at 35%, the compute shortage isn't going anywhere.

$AXTI says demand for its indium phosphide wafers "is outpacing supply," and it's raising up to $545M to catch up. Its China unit Tongmei signed for at least $285M of private equity, and up to $545M, to build more wafer and raw-material capacity. Closing is expected in December. Indium phosphide is the starting material for the lasers in AI optical links, the ones $LITE and $COHR sell into every new data center. A substrate maker doesn't raise half a billion unless laser demand is running ahead of supply.

NVIDIA paid $6.01 a share for its $1B stake in $NOK Since that deal, Nokia, T-Mobile and NVIDIA completed the world's first AI-RAN call on a live operator network. The radio upgrade NVIDIA bought into hasn't started paying yet. It will. $NVDA $TMUS