Marvell made a significant acquisition to address the network wall problem and become the leading option in this category, especially if we look at the best of the network connectivity, but it does have expensive hardware requirements (Disaggregation: Use massive optical bandwidth to separate memory from compute, eliminating the "memory wall"), which is not bad perse but it's harder to scale. My guess once the tech (Photonic Fabric Gen2 - Targeted at 28.8 Tbps (3.6 TB/s) per chiplet) is in production, meaning well tested, it's going to be bought or acquired by any of the hyperscalers, especially Microsoft or Nvidia, as Google and Amazon have gone with slower but more efficient and scalable routes. BTW, great article!
That Agentforce jump from 6k to 9.5k paid deals in one quarter is what matters. When you see 58% sequential growth in actual deployments (not just pilot credits), it means enterprises are moving past testing phase. The Informatica timing also matters because you need clean data pipelines before agents can do anything useful in production.
Excellnt framing on Meta's pivot. The Horizon Worlds budget cuts hit diffeerntly when you realize Zuck burned $70B racing against himself while Ray-Ban Meta shipped 2M units and is scaling to 10M. Your point about Marvell shifting from copper to photonics is understated becauase bandwidth bottlenecks at 100k+ GPU scale are the real infrastructure crisis no one talks about. If Celestial hits that $500M run-rate by Q4 FY28, they're not just selling chips but owning the physical layer that makes frontier models trainable.
Marvell made a significant acquisition to address the network wall problem and become the leading option in this category, especially if we look at the best of the network connectivity, but it does have expensive hardware requirements (Disaggregation: Use massive optical bandwidth to separate memory from compute, eliminating the "memory wall"), which is not bad perse but it's harder to scale. My guess once the tech (Photonic Fabric Gen2 - Targeted at 28.8 Tbps (3.6 TB/s) per chiplet) is in production, meaning well tested, it's going to be bought or acquired by any of the hyperscalers, especially Microsoft or Nvidia, as Google and Amazon have gone with slower but more efficient and scalable routes. BTW, great article!
That Agentforce jump from 6k to 9.5k paid deals in one quarter is what matters. When you see 58% sequential growth in actual deployments (not just pilot credits), it means enterprises are moving past testing phase. The Informatica timing also matters because you need clean data pipelines before agents can do anything useful in production.
Excellnt framing on Meta's pivot. The Horizon Worlds budget cuts hit diffeerntly when you realize Zuck burned $70B racing against himself while Ray-Ban Meta shipped 2M units and is scaling to 10M. Your point about Marvell shifting from copper to photonics is understated becauase bandwidth bottlenecks at 100k+ GPU scale are the real infrastructure crisis no one talks about. If Celestial hits that $500M run-rate by Q4 FY28, they're not just selling chips but owning the physical layer that makes frontier models trainable.