# AI capex — X 热门讨论 (2026-09-28 10:23 UTC)
## @MadarM49564 (Cobra_Web3) · 09-28 08:50 · ♥67 ↻1 💬84 Most AI projects are still training on the internet.
Vangrid is trying to solve a different problem how to collect real world ground truth data without the usual cost and complexity.
Their zero capex sensor swarm approach stands out because traditional physical data networks are expensive centralized and slow to scale Vangrid is building around continuous decentralized data collection with a focus on sovereignty and privacy.
What feels interesting:
📍 Edge computed privacy instead of just cloud dependence
📍 Cryptographic provenance for data trust
📍 Clear targeting of high value domains like embodied AI defense logistics and critical infrastructure
Physical AI won’t scale properly without better real world data This feels like an attempt to build that missing layer.
Still early but the thesis is strong.
@vangrid_io https://x.com/MadarM49564/status/2104493917618254244
## @cherryPayment (mark) · 09-27 14:02 · ♥31 ↻5 💬15 其实,能达成共识的是政府采购量可以忽略,但从黑名单到采购禁令,我们能看到方向越来越窄,可操作空间几乎为0。
当然,旭创的泰国工厂也绕不开的,需要说清楚子公司或者关联公司都不行。
然后我们来思考具体会有什么影响?
首先,就和我一开始说的那样就算国防和政府网络计划在在5年内换掉旭创、新易盛的模块,但这里面的量太小了,不足以支撑给COHR/LITE带来巨大增量,我认为象征意义大于实际用途,就是为了给"美国供应商"贴上合规溢价标签。
但不能说不重要哦!
毕竟以后如果股价上涨,但没有任何利好推动的情况下,市场可以给出 “美国供应商”的理由,自圆其说了也算。
这个弹性是很大的
对于云厂商?
近1-2年,我认为会很严重,毕竟西方产能补不上,800G/1.6T模块自然而然会供不应求,价格和毛利上行。COHR、AAOI、FN等美国及盟友产能议价能力大增,但可能出现"订单有、交付不了"。
那么如果光模块成为瓶颈,会如何?
那么数据中心上线会被延迟,反过来压制GPU和整条AI capex的交付节奏。
这个我们可以看 $nvda有多聪明
2-4年之后,产能实现大转移,云厂商会预付款、签长单,扶持美国、泰国以及马来西亚等产能地区
那么光模块都这样了,上游怎么办?
模块产能翻了几倍,上游的InP衬底、DFB/EML激光器、外延片需求会被放大。
但别忽视了地缘政治:
AXTI 的难点是它主要产能在中国,反而可能被要求迁移或受限,要特别留意。
SIVE、IQE这类盟友国家的上游节点,会同时享受需求增长和"可信供应链"溢价。
LITE、COHR 的激光器业务从"也卖给中国厂"变成"必须优先供应西方模块厂",定价权上升。
溢价会很恐怖 > 引用 @The_AI_Investor: COHR LITE. Washington just handed US optics a demand tailwind. Senators Cornyn, McCormick, Fetterman and Gallego introduced bipartisan legislation to bar optical transceivers from Eoptolink and Zhongji Innolight in sensitive US national security systems. Optical transceivers are the modules that convert electrical signals into light pulses to move data between servers, and large AI data centers need millions of them.
The bill gives the government five years to comply and lets the Commerce and Defense secretaries add more Chinese vendors to the list. The industry pushback is already on the record: the ITI wrote the FCC in August warning that supply disruption would slow US data center deployment, and Reuters reports Coherent and Lumentum sell competitive technology but lack the scale to replace Chinese vendors today.
The government slice of transceiver demand is small, but the signal matters more than the volume. If federal procurement goes domestic first, the same logic spreads to hyperscaler supply chains worried about firmware risk in reprogrammable transceivers. Coherent is the prime beneficiary by position, the risk is that Western capacity cannot scale fast enough and the buildout slows instead. https://x.com/cherryPayment/status/2104210159639633927
## @KrisPatel99 (Kris Patel 🇺🇸) · 09-28 07:32 · ♥31 ↻0 💬7 Long term treasuries are rising globally as there is little sign of AI Capex spending moderating or Oil prices coming down because of the Strait of Hormuz closure.
Alot of countries and their reserve levels are low and the end is nowhere in sight. https://t.co/7pTFdkHNyR https://x.com/KrisPatel99/status/2104474314728673322