#我的七夕交易分享 AI demand is reshaping the value proposition of every layer of the supply chain, but the pace and certainty vary greatly across layers.
In the short term, consumer NAND is moving sideways while enterprise NAND is still rising, with Kioxia allocating 60% to 70% of its capacity to enterprise products. Eight 64GB memory modules now cost close to the price of an entire server, while enterprise rack-server shipments are shrinking. Supernodes will reach 7,000 racks in 2026, with mass production beginning in October; a full rack with 64 cards will cost 10 million to 11 million. Indium phosphide will face a clear shortfall in 2026, red phosphorus prices have tripled, and MOCVD lead times are 15 to 20 months. Meta will have 7GW of computing capacity by year-end, and the Personal Super Intelligence Agent may be launched in September. Toyoda weaving machines have a 1.5-year lead time, the electronic-fabric shortfall is one-third, and new kilns coming online in September will release monthly capacity of 4 million meters.
In the medium term, SanDisk's NBM has locked in $93 billion in revenue over three years; HBF samples will be delivered in 2027 and ramp up in 2028, with a maximum option value of $330 billion. Meta will reach 16GW by the end of 2027, split evenly between training and inference. Supernodes will reach 22,000 racks in 2027, with penetration of 35% to 38%, while 224G backplane connectors will begin ramping up. Domestic AI chip demand will reach 4 million to 5 million units in 2027, but 7nm process capacity is the bottleneck. The risk of a consumer NAND shortage will rise in 2027; materials bought for 1 billion in the past now cost 16 billion. Indium phosphide supply will exceed demand in the second half of 2027, and prices may fall. The electronic-fabric shortage will initially ease by the end of June 2027, while the trajectory of low-end fabric prices will depend on whether domestic weaving machines can achieve a breakthrough. Demand for high-end electronic fabric will rise to several times its current level in Q1 to Q2 2027, and the shortfall will persist.
In the long term, NAND will be revalued from a cyclical product into AI-inference storage. HBF places Flash inside the GPU package, offering 8 to 16 times the capacity of HBM at one-fifth to one-eighth the price. The baseline HBF TAM in 2030 is $47 billion. The inference market will be highly diversified, diluting Nvidia's share. Domestic chips are entering the elimination phase. SanDisk's $100 billion base market capitalization is the floor, while successful HBF validation will open a second growth curve. The next stage will focus on four things: whether HBF samples will be validated by cloud providers in 2027; whether Meta will launch its superintelligent Agent at Connect in September; whether indium phosphide supply will exceed demand as scheduled in Q4 2027; and whether domestic weaving machines can achieve a breakthrough in 2026, which will determine the trajectory of electronic fabric in 2027. $SNDK
In the short term, consumer NAND is moving sideways while enterprise NAND is still rising, with Kioxia allocating 60% to 70% of its capacity to enterprise products. Eight 64GB memory modules now cost close to the price of an entire server, while enterprise rack-server shipments are shrinking. Supernodes will reach 7,000 racks in 2026, with mass production beginning in October; a full rack with 64 cards will cost 10 million to 11 million. Indium phosphide will face a clear shortfall in 2026, red phosphorus prices have tripled, and MOCVD lead times are 15 to 20 months. Meta will have 7GW of computing capacity by year-end, and the Personal Super Intelligence Agent may be launched in September. Toyoda weaving machines have a 1.5-year lead time, the electronic-fabric shortfall is one-third, and new kilns coming online in September will release monthly capacity of 4 million meters.
In the medium term, SanDisk's NBM has locked in $93 billion in revenue over three years; HBF samples will be delivered in 2027 and ramp up in 2028, with a maximum option value of $330 billion. Meta will reach 16GW by the end of 2027, split evenly between training and inference. Supernodes will reach 22,000 racks in 2027, with penetration of 35% to 38%, while 224G backplane connectors will begin ramping up. Domestic AI chip demand will reach 4 million to 5 million units in 2027, but 7nm process capacity is the bottleneck. The risk of a consumer NAND shortage will rise in 2027; materials bought for 1 billion in the past now cost 16 billion. Indium phosphide supply will exceed demand in the second half of 2027, and prices may fall. The electronic-fabric shortage will initially ease by the end of June 2027, while the trajectory of low-end fabric prices will depend on whether domestic weaving machines can achieve a breakthrough. Demand for high-end electronic fabric will rise to several times its current level in Q1 to Q2 2027, and the shortfall will persist.
In the long term, NAND will be revalued from a cyclical product into AI-inference storage. HBF places Flash inside the GPU package, offering 8 to 16 times the capacity of HBM at one-fifth to one-eighth the price. The baseline HBF TAM in 2030 is $47 billion. The inference market will be highly diversified, diluting Nvidia's share. Domestic chips are entering the elimination phase. SanDisk's $100 billion base market capitalization is the floor, while successful HBF validation will open a second growth curve. The next stage will focus on four things: whether HBF samples will be validated by cloud providers in 2027; whether Meta will launch its superintelligent Agent at Connect in September; whether indium phosphide supply will exceed demand as scheduled in Q4 2027; and whether domestic weaving machines can achieve a breakthrough in 2026, which will determine the trajectory of electronic fabric in 2027. $SNDK
