$MU $SKHY $LITE $COHR $AXTI $AAOI Agreed, and the prospectus backs the money t… · @carrioresearch · TwKit

@carrioresearch

$MU $SKHY $LITE $COHR $AXTI $AAOI Agreed, and the prospectus backs the money trail with more than the FCF line. Part of October's selloff in memory and optics was the worry that the labs would never make money. Anthropic's S-1, filed Monday, shows $3.93B of operating cash in the first half of 2026 against $482M of capex. A year earlier that line was $1.51B going out. The bigger number is a few pages later. As of June 30, Anthropic had about $265.6B in fixed or minimum guaranteed compute and infrastructure commitments, with terms of one to 20 years. Revenue for the half was $16.3B, 14 times a year earlier, and $8.9B went to compute. That tells me where AI is right now. The buyer at the top of the chain pays its way out of operating cash and is signing for capacity a decade out. Spending that sits on contracts holds up a lot better than spending that waits on the next funding round. So where does $265.6B end up? Anthropic says it in its own risk factors. It names SK hynix, Samsung and Micron, and says there is "no guarantee" their expansions bring supply "fast enough to meet demand, particularly if large customers or competitors secure exclusive supply commitments or otherwise reserve substantially all of a manufacturer's available production capacity." A buyer holding a quarter trillion in commitments is worried about getting memory. The sellers say the same from their side. Samsung's preliminary Q3 operating profit was a record ₩107.4T, and Reuters reports that Samsung and Micron both expect supply to stay tight through 2028. TrendForce still has Q4 DRAM contract prices up 10-15% for conventional parts and 15-20% including HBM. The increases are smaller off a much higher base, and prices are still going up. Memory is turning into a long-contract business, and that deserves a different multiple from a cyclical one. The part that surprised me: "laser" and "optic" get zero hits in the S-1. The companies selling the lasers sound tighter than the buyer does. - Andy Bechtolsheim put this year's OSFP module demand closer to 150M, with shipments a bit over 120M because module makers can't get enough lasers. At OCP he said lead times went from about 3 months to 9-12. - $LITE CEO Michael Hurlston said on Oct 8 that some products are about 70% short next year, others about 30% short through 2028, and 2029 is nearly sold out. - NVIDIA put $2B each into Lumentum and $COHR in March, with purchase commitments and capacity rights attached. - In China, Innolight says overseas orders run into 2027 and some key customers have already given 2028 guidance. CW laser chipmaker Yuanjie guided Q3 net margin to roughly 62-65%. In that chain the profit sits at the laser. Go one more layer up and the scarcity has a home country. InP substrates, which these lasers are grown on, have been on China's export control list since February 2025, and $AXTI ships from its China plant under permits approved one application at a time. Washington and Beijing extended their truce to Jan 10, 2027 last month, which keeps the suspended rare earth curbs suspended. The InP licensing from February 2025 still applies. The rivalry pushes both sides to build. Anthropic's filing says future export controls could limit where chips can be installed. China's CXMT went from 7.6% to 9.5% of global DRAM revenue between Q1 and Q2, and industry media report it sampling 8-high HBM3 to domestic customers. NVIDIA is paying to expand laser fabs in the US. Two supply chains need two sets of fabs, substrates and lasers, and that keeps every scarce layer scarce for longer. A bounce can come from anything, and the $265.6B is already signed. I'm bullish on memory and optics.

曝光 513 · 评论 1 · 点赞 10 · 书签 3