Beijing doesn't want to let the market fall too quickly.

Takeaway from Nomura Global Memory The AI spending race has become extremely difficult for Big Tech to stop, even as it eats into free cash flow. 1. AI capex is becoming a game of chicken Stop spending → risk falling behind in AI and potentially hurting your existing business. Keep spending → accept negative FCF and increasingly rely on external funding 2. Why 2027 memory looks protected 2027 memory is basically already booked. 2028 is the bigger question. More supply could arrive, while power constraints and demand become increasingly important 3. What could actually stop the AI spending race? The AI race probably doesn't stop because Big Tech wants to spend less. It stops if Big Tech can no longer easily raise the money to keep spending.

$INTC earnings Beat and raise please !!!

🌈🐻Citi is saying the semiconductor-equipment story may still be strong, but the stocks could fall further because companies may not be able to build and ship enough equipment. Too much demand can actually become a short-term problem. AI demand ↑ ↓ Equipment demand ↑↑ ↓ Component shortages ↑ ↓ Equipment shipments constrained ↓ WFE growth limited ↓ $AMAT / $KLAC / $LRCX earnings upside limited ↓ Stocks de-rate Sound confusing? The original text is even more confusing

AI is now the top priority for CIOs. 85% want AI agents. 17% are seeing company-wide returns today. And there’s a second effect more AI → more security spending. For the chart 2 The three numbers measure different things: 84.5% = plan to deploy agents within 12 months, from BofA Exhibit 7. 37.3% = more than 25% of AI pilots progressed into production over the last 12 months, from Exhibit 13. 17.2% = already generating measurable enterprise-wide ROI, from Exhibit 11. Because the three bars sit together, some of you could interpret them as a funnel: 85% want agents → 37% reach production → 17% get ROI That's not what the survey measured. They are three separate questions, and the middle number refers to pilot conversion over the last twelve months.

First, CFROI is basically how much cash return a company generates from the money invested in the business. And right now, tech returns is verystrong. But the market is already expecting these returns to come back down. Memory is the clearest example Record returns are forecast, yet the market is pricing in a much weaker future. At the same time, funding the massive AI buildout is becoming another question. Strong today. But for how long, and at what cost?

How much are customers willing to pay upfront? $CRWV Customer prepayments change the financing equation dramatically. For the 4.2 GW scenario: No customer prepayment → $84.5B new debt 20% customer prepayment → $38.0B new debt So, customers effectively help fund the infrastructure needed to serve their own AI workloads.

All human wisdom is contained in these two words, Wait and hope 10 years from now, the only problem you're going to have is that you didn't buy more!

Why does HBM need hybrid bonding? HBM stacks are getting taller → hybrid bonding helps make them thinner. The bigger technological benefits are: Hybrid bonding → denser interconnects → faster communication + lower latency + lower power → better thermal management It's potentially about making HBM perform better and consume power more efficiently. Fun Fact: Intel remains positioned for a further hybrid bonding ramp. BESI stated that Intel's substantial HB tool orders from 2024/25 have now been delivered and installed. $INTC With Intel already utilizing hybrid bonding technology and seeing improved end-market demand, management suggested that a further round of capacity expansion could materialize in 2027.

Takeaway 1. Why is the market worried? CoreWeave's unsecured debt came under heavy pressure, with the report noting yields around 15%. That raised fears that financing for the broader AI buildout could be cracking. 2. Why does Barclays think that's the wrong conclusion? Because the Because the bonds getting crushed are only a small and particularly risky piece of CoreWeave's financing structure. CoreWeave has roughly: $35B total debt Less than $2B in unsecured bonds And importantly, those unsecured bonds sit far down the repayment hierarchy. Think about it like this: CoreWeave owes a lot of people money. If things go wrong, creditors don't all get paid equally. First in line: Secured lenders Then: Other creditors Finally: Unsecured bondholders So, these bonds offer a very high yield. 3. What if CoreWeave really does run into trouble? Even then, Barclays estimates CoreWeave accounts for only roughly 4-5% of total AI capex. If CoreWeave's spending were cut in half, Barclays estimates the impact on total AI capex would be around 2%, or roughly $20B on a $1T+ 2027 estimate. Not nothing. But not the cycle 4. So what should investors actually watch? Watch the price and demand for compute, not just CoreWeave's bond price. $CRWV If a company has a $35bn mortgage and a $1.5bn credit card balance, and the credit card rate spikes to 15%, that tells you something about the credit card. It doesn't tell you that all the mortgage borrowers in that city are in trouble.

AI isn't only eating compute. It's eating the capital markets too. AI infrastructure is now ~48% of the U.S. convertible market by value. U.S. convertible issuance has already reached $137B through Oct. 21, surpassing the $121B full-year record in 2025.

This is actually new to me. I thought tech was issuing debt because AI capex was getting so huge that cash flow couldn't cover it. Apparently, tech's aggregate cash flow still covers capex through 2030. The financing gap shows up after buybacks + dividends. So tech isn't necessarily borrowing because it can't afford the AI buildout. They want to fund the AI buildout AND keep shareholders happy.

More agents = more work = more CPUs. AI can now do things for us!!! Look at Grok, Muse and Dots That means more computing happening behind the scenes, and potentially a much bigger role for CPUs.

Each new AI chip generation uses more power, but the amount of work it can do is growing much faster. The chips are getting more efficient. Our demand for AI is just growing even faster. $NVDA

Wait, so Samsung and $SKHY are already pricing in a 2027 memory downcycle and negative memory pricing?

AI capex is not slowing down! This isn't one hyperscaler going all in. They're ALL going all in.

If semiconductor demand is about to peak, why are customers placing orders this far out? $ASML

Before anyone takes these charts as gospel They’re based on JPMorgan’s model, not my forecasts or necessarily SpaceX’s own plans. The further out you go, the more assumptions are involved. $SPCX Chart 1 — revenue mix Plots three segment revenues as 100% stacked columns; the printed labels are the SI share. The three shares always add to 1 by construction, which is why the stacked columns total 100%. Chain worth knowing: 2025A SI revenue is SI!H42 = SUM(H31,H37) = infrastructure H31 = SUM(D31:G31) (1,356) + advertising H37 = SUM(D37:G37) (1,844) = 3,200. From 2028E, infrastructure is SI!S31 = S27*AVERAGE(R20,S20) ($/Watt × average nameplate MW) and ads are SI!S37 = R37*(1+S38). Chart 2 — compute capacity Plots end-of-year nameplate compute in GW: total (row 19) as the column height, in-orbit (row 9) as the orange top, on-Earth (row 14) as the blue bottom. So the chart's stack is internally consistent by construction: from 2030E the orbital figure is defined as total minus terrestrial, and terrestrial is frozen (V14 = U14 = T14 = 20.877936). Chart 3 — capex vs free cash flow Both series are the same unit (Minthesheet,plottedinMinthesheet,plottedinB). Capex is stored negative, plotted positive. Printed labels: CF!S17 −389,118.8 → "389B";‘CF!R41‘−81,608.7→"−389B";‘CF!R41‘−81,608.7→"−82B"; CF!V17 −818,965.1 → "$819B". Note CF!A41 is labelled Free Cash Flow while DCF!A17 is Unlevered free cash flow — a different series (DCF!C17 = C10+SUM(C14:C16)); the chart uses only the CF row. Chart 4 — margins Three ratios from the IS tab, all divided by the same revenue cell. Row labels are formula-linked (IS!A91 = A12, A94 = A18, A97 = A39), so the margin block never drifts from the P&L. Printed labels: IS!H94 −13.86% → "−14%", IS!V94 65.56% → "66%", headline $500 from Rel Val!C7.