James Altucher
Investor & Author, Technology & Markets
AI's power problem is really a connectivity problem, and Marvell, Astera Labs, Fabric AI and Kopin are working on the fix. A 100,000-GPU data center needs a million or more copper interconnects; switching to MicroLED optical links would use about 1/30 of the power, under 1 pJ/bit versus up to 5 for copper and 15 for laser. x.com x.comÂ
Interconnects are the trade beneath open models: sparse mixture-of-experts models like Kimi K3 are slow mainly because GPU-to-GPU traffic runs over links at about 1% of in-GPU speed. Faster interconnects would let open models surpass closed ones. Credo's copper active electrical cables are hitting physical limits. Marvell is co-developing MicroLED optical links with Mojo Vision, and Fabric AI, the highest-reward play, is building MicroLED interconnects with Kopin, targeting a demo by end-2026. x.com x.comÂ
Marvell looks well positioned for scale-up compute domains, where interconnect is the wall that limits performance. Scale-up domains should be thought of differently from simply networking racks together, and Marvell's own commentary now focuses on exactly that opportunity. x.com x.comÂ
Inference's move toward edge devices and dense, memory-hungry racks should drive demand for specialized chips, links, power delivery and cooling. McKinsey projects inference at 60% of AI computing demand by 2030. Ambarella makes on-device AI chips; Astera Labs links GPUs, memory and networks; Credo supplies short copper connections; Rambus earns from memory interfaces; Navitas reduces power waste; Modine cools dense racks; and Super Micro builds inference-specific server racks at volume. x.com … Show more
Marvell risks paying a peak-scarcity price for Lumentum if optical supply catches up by 2028-30. The proposed ~$99B deal values Lumentum at roughly 20x guided annualized revenue and 58x estimated after-tax operating income, while Lumentum, Coherent and Chinese suppliers are adding laser capacity. The DSP-and-laser combination has strategic logic, but durable scarcity economics and Nvidia's response remain uncertain. x.comÂ
AI data-center spending flows beyond GPUs and operators into networking, optics, power and cooling. Nvidia, AMD, Broadcom and Marvell supply chips; Arista, Nokia and Credo connectivity; Applied Optoelectronics and Lumentum optics; Vistra, Constellation and Bloom power; Vertiv and Modine cooling. Meta's $140B spending this year illustrates the scale, per a commenter with whom the point is agreed. x.com x.comÂ
Rising 1.6T switching demand favors optical suppliers, though their execution risks vary. Bank of America says the market could grow ~101% annually through 2030. Marvell's leading DSPs and Lumentum's supply-short 200G EML lasers are lowest risk; Applied Optoelectronics' capacity expansion and Fabrinet's customer concentration make them medium risk. Aeluma remains pre-commercial, while Lightwave Logic's polymer modulators are a high-risk moonshot despite its TSM partnership. x.comÂ
Marvell's management is running the stock as a pump and dump to move shareholders' money into their own pockets. Wall Street prices Marvell as an ASIC design house without realizing that ASIC design is just IT outsourcing. It is not a real CXL play either: CXL IP is sold by Cadence, Synopsys and Rambus with no moat, and Astera Labs, which builds the switches and software around CXL, is the only good CXL play. x.comÂ
Marvell CEO Matt Murphy's vision of fully disaggregated data centers opens up future optical interconnect demand for Coherent and Lumentum. At Computex he said AI's next bottleneck is connectivity: copper reaches about 2.5m at 200Gbps per lane and barely spans a rack at 400Gbps. Optics would let CPUs, GPUs and memory sit in separate pools, moving optical engines into the compute layer, perhaps 10x per rack, a net-new TAM. Jensen Huang: copper where you can, optics where you must. x.com x.comÂ
A potential Lumentum acquisition could let Marvell pair custom AI chips with the optical infrastructure connecting them. Following Celestial AI, it could become one of the few companies offering both, moving toward an AI-focused version of Broadcom; financing a deal this large remains a major concern. x.comÂ
Astera Labs carries a huge premium for flawless execution against a CXL market that is small near term. Switch-based CXL memory pooling capturing $15-16B in 2026-2028 is low probability: Marvell guides its XConn switch business to only about $100M by fiscal 2028, PCIe included. Pooling software likely stays immature into 2028, per a Samsung white paper, and Nvidia is contesting the warm KV-cache tier. Astera does CXL best, but SK hynix and Micron offer better risk-adjusted returns. x.com x.comÂ
More lasers per module and new 400G DSPs could keep the pluggable-optics runway open for years, supporting Lumentum, Coherent, Marvell, Applied Optoelectronics and Sivers even as co-packaged optics approaches. NVIDIA and Cisco co-designed a module with four times the bandwidth in the same slot; fewer modules per GPU look bearish, but the added optical content points the other way. x.comÂ
